11cf9b237SStefano Zampini #include <../src/mat/impls/aij/seq/aij.h> 2ab5c6b0cSJed Brown #include <../src/ksp/pc/impls/bddc/bddc.h> 3ab5c6b0cSJed Brown #include <../src/ksp/pc/impls/bddc/bddcprivate.h> 4674ae819SStefano Zampini #include <petscblaslapack.h> 5674ae819SStefano Zampini 6906d46d4SStefano Zampini static PetscErrorCode PCBDDCMatMultTranspose_Private(Mat A, Vec x, Vec y); 7906d46d4SStefano Zampini static PetscErrorCode PCBDDCMatMult_Private(Mat A, Vec x, Vec y); 8906d46d4SStefano Zampini 9a3df083aSStefano Zampini #undef __FUNCT__ 10a3df083aSStefano Zampini #define __FUNCT__ "PCBDDCBenignMatMult_Private_Private" 11a3df083aSStefano Zampini PetscErrorCode PCBDDCBenignMatMult_Private_Private(Mat A, Vec x, Vec y, PetscBool transpose) 12a3df083aSStefano Zampini { 13a3df083aSStefano Zampini PCBDDCBenignMatMult_ctx ctx; 14a3df083aSStefano Zampini PetscErrorCode ierr; 15a3df083aSStefano Zampini PetscBool apply_right,apply_left,reset_x; 16a3df083aSStefano Zampini 17a3df083aSStefano Zampini PetscFunctionBegin; 18a3df083aSStefano Zampini ierr = MatShellGetContext(A,&ctx);CHKERRQ(ierr); 19a3df083aSStefano Zampini if (transpose) { 20a3df083aSStefano Zampini apply_right = ctx->apply_left; 21a3df083aSStefano Zampini apply_left = ctx->apply_right; 22a3df083aSStefano Zampini } else { 23a3df083aSStefano Zampini apply_right = ctx->apply_right; 24a3df083aSStefano Zampini apply_left = ctx->apply_left; 25a3df083aSStefano Zampini } 26a3df083aSStefano Zampini reset_x = PETSC_FALSE; 27a3df083aSStefano Zampini if (apply_right) { 28a3df083aSStefano Zampini const PetscScalar *ax; 29a3df083aSStefano Zampini PetscInt nl,i; 30a3df083aSStefano Zampini 31a3df083aSStefano Zampini ierr = VecGetLocalSize(x,&nl);CHKERRQ(ierr); 32a3df083aSStefano Zampini ierr = VecGetArrayRead(x,&ax);CHKERRQ(ierr); 33a3df083aSStefano Zampini ierr = PetscMemcpy(ctx->work,ax,nl*sizeof(PetscScalar));CHKERRQ(ierr); 34a3df083aSStefano Zampini ierr = VecRestoreArrayRead(x,&ax);CHKERRQ(ierr); 35a3df083aSStefano Zampini for (i=0;i<ctx->benign_n;i++) { 36a3df083aSStefano Zampini PetscScalar sum,val; 37a3df083aSStefano Zampini const PetscInt *idxs; 38a3df083aSStefano Zampini PetscInt nz,j; 39a3df083aSStefano Zampini ierr = ISGetLocalSize(ctx->benign_zerodiag_subs[i],&nz);CHKERRQ(ierr); 40a3df083aSStefano Zampini ierr = ISGetIndices(ctx->benign_zerodiag_subs[i],&idxs);CHKERRQ(ierr); 41a3df083aSStefano Zampini sum = 0.; 42a3df083aSStefano Zampini if (ctx->apply_p0) { 43a3df083aSStefano Zampini val = ctx->work[idxs[nz-1]]; 44a3df083aSStefano Zampini for (j=0;j<nz-1;j++) { 45a3df083aSStefano Zampini sum += ctx->work[idxs[j]]; 46a3df083aSStefano Zampini ctx->work[idxs[j]] += val; 47a3df083aSStefano Zampini } 48a3df083aSStefano Zampini } else { 49a3df083aSStefano Zampini for (j=0;j<nz-1;j++) { 50a3df083aSStefano Zampini sum += ctx->work[idxs[j]]; 51a3df083aSStefano Zampini } 52a3df083aSStefano Zampini } 53a3df083aSStefano Zampini ctx->work[idxs[nz-1]] -= sum; 54a3df083aSStefano Zampini ierr = ISRestoreIndices(ctx->benign_zerodiag_subs[i],&idxs);CHKERRQ(ierr); 55a3df083aSStefano Zampini } 56a3df083aSStefano Zampini ierr = VecPlaceArray(x,ctx->work);CHKERRQ(ierr); 57a3df083aSStefano Zampini reset_x = PETSC_TRUE; 58a3df083aSStefano Zampini } 59a3df083aSStefano Zampini if (transpose) { 60a3df083aSStefano Zampini ierr = MatMultTranspose(ctx->A,x,y);CHKERRQ(ierr); 61a3df083aSStefano Zampini } else { 62a3df083aSStefano Zampini ierr = MatMult(ctx->A,x,y);CHKERRQ(ierr); 63a3df083aSStefano Zampini } 64a3df083aSStefano Zampini if (reset_x) { 65a3df083aSStefano Zampini ierr = VecResetArray(x);CHKERRQ(ierr); 66a3df083aSStefano Zampini } 67a3df083aSStefano Zampini if (apply_left) { 68a3df083aSStefano Zampini PetscScalar *ay; 69a3df083aSStefano Zampini PetscInt i; 70a3df083aSStefano Zampini 71a3df083aSStefano Zampini ierr = VecGetArray(y,&ay);CHKERRQ(ierr); 72a3df083aSStefano Zampini for (i=0;i<ctx->benign_n;i++) { 73a3df083aSStefano Zampini PetscScalar sum,val; 74a3df083aSStefano Zampini const PetscInt *idxs; 75a3df083aSStefano Zampini PetscInt nz,j; 76a3df083aSStefano Zampini ierr = ISGetLocalSize(ctx->benign_zerodiag_subs[i],&nz);CHKERRQ(ierr); 77a3df083aSStefano Zampini ierr = ISGetIndices(ctx->benign_zerodiag_subs[i],&idxs);CHKERRQ(ierr); 78a3df083aSStefano Zampini val = -ay[idxs[nz-1]]; 79a3df083aSStefano Zampini if (ctx->apply_p0) { 80a3df083aSStefano Zampini sum = 0.; 81a3df083aSStefano Zampini for (j=0;j<nz-1;j++) { 82a3df083aSStefano Zampini sum += ay[idxs[j]]; 83a3df083aSStefano Zampini ay[idxs[j]] += val; 84a3df083aSStefano Zampini } 85a3df083aSStefano Zampini ay[idxs[nz-1]] += sum; 86a3df083aSStefano Zampini } else { 87a3df083aSStefano Zampini for (j=0;j<nz-1;j++) { 88a3df083aSStefano Zampini ay[idxs[j]] += val; 89a3df083aSStefano Zampini } 90a3df083aSStefano Zampini ay[idxs[nz-1]] = 0.; 91a3df083aSStefano Zampini } 92a3df083aSStefano Zampini ierr = ISRestoreIndices(ctx->benign_zerodiag_subs[i],&idxs);CHKERRQ(ierr); 93a3df083aSStefano Zampini } 94a3df083aSStefano Zampini ierr = VecRestoreArray(y,&ay);CHKERRQ(ierr); 95a3df083aSStefano Zampini } 96a3df083aSStefano Zampini PetscFunctionReturn(0); 97a3df083aSStefano Zampini } 98a3df083aSStefano Zampini 99a3df083aSStefano Zampini #undef __FUNCT__ 100a3df083aSStefano Zampini #define __FUNCT__ "PCBDDCBenignMatMultTranspose_Private" 101a3df083aSStefano Zampini PetscErrorCode PCBDDCBenignMatMultTranspose_Private(Mat A, Vec x, Vec y) 102a3df083aSStefano Zampini { 103a3df083aSStefano Zampini PetscErrorCode ierr; 104a3df083aSStefano Zampini 105a3df083aSStefano Zampini PetscFunctionBegin; 106a3df083aSStefano Zampini ierr = PCBDDCBenignMatMult_Private_Private(A,x,y,PETSC_TRUE);CHKERRQ(ierr); 107a3df083aSStefano Zampini PetscFunctionReturn(0); 108a3df083aSStefano Zampini } 109a3df083aSStefano Zampini 110a3df083aSStefano Zampini #undef __FUNCT__ 111a3df083aSStefano Zampini #define __FUNCT__ "PCBDDCBenignMatMult_Private" 112a3df083aSStefano Zampini PetscErrorCode PCBDDCBenignMatMult_Private(Mat A, Vec x, Vec y) 113a3df083aSStefano Zampini { 114a3df083aSStefano Zampini PetscErrorCode ierr; 115a3df083aSStefano Zampini 116a3df083aSStefano Zampini PetscFunctionBegin; 117a3df083aSStefano Zampini ierr = PCBDDCBenignMatMult_Private_Private(A,x,y,PETSC_FALSE);CHKERRQ(ierr); 118a3df083aSStefano Zampini PetscFunctionReturn(0); 119a3df083aSStefano Zampini } 120a3df083aSStefano Zampini 121a3df083aSStefano Zampini #undef __FUNCT__ 122a3df083aSStefano Zampini #define __FUNCT__ "PCBDDCBenignShellMat" 123a3df083aSStefano Zampini PetscErrorCode PCBDDCBenignShellMat(PC pc, PetscBool restore) 124a3df083aSStefano Zampini { 125a3df083aSStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 126a3df083aSStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 127a3df083aSStefano Zampini Mat_IS *matis = (Mat_IS*)pc->pmat->data; 128a3df083aSStefano Zampini PCBDDCBenignMatMult_ctx ctx; 129a3df083aSStefano Zampini PetscErrorCode ierr; 130a3df083aSStefano Zampini 131a3df083aSStefano Zampini PetscFunctionBegin; 132a3df083aSStefano Zampini if (!restore) { 133a3df083aSStefano Zampini Mat A,A_IB,A_BI; 134a3df083aSStefano Zampini PetscScalar *work; 135df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse = pcbddc->sub_schurs->reuse_solver; 136a3df083aSStefano Zampini 137a3df083aSStefano Zampini /* local mat */ 138a3df083aSStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&A);CHKERRQ(ierr); 139a3df083aSStefano Zampini ierr = MatSetSizes(A,pcis->n,pcis->n,PETSC_DECIDE,PETSC_DECIDE);CHKERRQ(ierr); 140a3df083aSStefano Zampini ierr = MatSetType(A,MATSHELL);CHKERRQ(ierr); 141a3df083aSStefano Zampini ierr = MatShellSetOperation(A,MATOP_MULT,(void (*)(void))PCBDDCBenignMatMult_Private);CHKERRQ(ierr); 142a3df083aSStefano Zampini ierr = MatShellSetOperation(A,MATOP_MULT_TRANSPOSE,(void (*)(void))PCBDDCBenignMatMultTranspose_Private);CHKERRQ(ierr); 143a3df083aSStefano Zampini ierr = PetscNew(&ctx);CHKERRQ(ierr); 144a3df083aSStefano Zampini ierr = MatShellSetContext(A,ctx);CHKERRQ(ierr); 145a3df083aSStefano Zampini ctx->apply_left = PETSC_TRUE; 146a3df083aSStefano Zampini ctx->apply_right = PETSC_TRUE; 147a3df083aSStefano Zampini ctx->apply_p0 = PETSC_TRUE; 148a3df083aSStefano Zampini ctx->benign_n = pcbddc->benign_n; 149a3df083aSStefano Zampini ctx->benign_zerodiag_subs = pcbddc->benign_zerodiag_subs; 150a3df083aSStefano Zampini ctx->A = matis->A; 151a3df083aSStefano Zampini ierr = PetscMalloc1(pcis->n,&work);CHKERRQ(ierr); 152a3df083aSStefano Zampini ctx->work = work; 153a3df083aSStefano Zampini ierr = MatSetUp(A);CHKERRQ(ierr); 154a3df083aSStefano Zampini ierr = MatAssemblyBegin(A,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 155a3df083aSStefano Zampini ierr = MatAssemblyEnd(A,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 156a3df083aSStefano Zampini matis->A = A; 157a3df083aSStefano Zampini 158a3df083aSStefano Zampini /* A_IB */ 159a3df083aSStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&A_IB);CHKERRQ(ierr); 160a3df083aSStefano Zampini ierr = MatSetSizes(A_IB,pcis->n-pcis->n_B,pcis->n_B,PETSC_DECIDE,PETSC_DECIDE);CHKERRQ(ierr); 161a3df083aSStefano Zampini ierr = MatSetType(A_IB,MATSHELL);CHKERRQ(ierr); 162a3df083aSStefano Zampini ierr = MatShellSetOperation(A_IB,MATOP_MULT,(void (*)(void))PCBDDCBenignMatMult_Private);CHKERRQ(ierr); 163a3df083aSStefano Zampini ierr = MatShellSetOperation(A_IB,MATOP_MULT_TRANSPOSE,(void (*)(void))PCBDDCBenignMatMultTranspose_Private);CHKERRQ(ierr); 164a3df083aSStefano Zampini ierr = PetscNew(&ctx);CHKERRQ(ierr); 165a3df083aSStefano Zampini ierr = MatShellSetContext(A_IB,ctx);CHKERRQ(ierr); 166a3df083aSStefano Zampini ctx->apply_left = PETSC_TRUE; 167a3df083aSStefano Zampini ctx->apply_right = PETSC_FALSE; 168a3df083aSStefano Zampini ctx->apply_p0 = PETSC_FALSE; 169a3df083aSStefano Zampini ctx->benign_n = pcbddc->benign_n; 170059032f7SStefano Zampini if (reuse) { 171a3df083aSStefano Zampini ctx->benign_zerodiag_subs = reuse->benign_zerodiag_subs; 172059032f7SStefano Zampini } else { /* TODO: could be optimized for successive solves */ 173059032f7SStefano Zampini ISLocalToGlobalMapping N_to_D; 174059032f7SStefano Zampini PetscInt i; 175059032f7SStefano Zampini 176059032f7SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(pcis->is_I_local,&N_to_D);CHKERRQ(ierr); 177059032f7SStefano Zampini ierr = PetscMalloc1(pcbddc->benign_n,&ctx->benign_zerodiag_subs);CHKERRQ(ierr); 178059032f7SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 179059032f7SStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(N_to_D,IS_GTOLM_DROP,pcbddc->benign_zerodiag_subs[i],&ctx->benign_zerodiag_subs[i]);CHKERRQ(ierr); 180059032f7SStefano Zampini } 181059032f7SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&N_to_D);CHKERRQ(ierr); 182059032f7SStefano Zampini } 183a3df083aSStefano Zampini ctx->A = pcis->A_IB; 184a3df083aSStefano Zampini ctx->work = work; 185a3df083aSStefano Zampini ierr = MatSetUp(A_IB);CHKERRQ(ierr); 186a3df083aSStefano Zampini ierr = MatAssemblyBegin(A_IB,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 187a3df083aSStefano Zampini ierr = MatAssemblyEnd(A_IB,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 188a3df083aSStefano Zampini pcis->A_IB = A_IB; 189a3df083aSStefano Zampini 190a3df083aSStefano Zampini /* A_BI as A_IB^T */ 191a3df083aSStefano Zampini ierr = MatCreateTranspose(A_IB,&A_BI);CHKERRQ(ierr); 192a3df083aSStefano Zampini pcbddc->benign_original_mat = pcis->A_BI; 193a3df083aSStefano Zampini pcis->A_BI = A_BI; 194a3df083aSStefano Zampini } else { 195059032f7SStefano Zampini PCBDDCReuseSolvers reuse = pcbddc->sub_schurs->reuse_solver; 196059032f7SStefano Zampini 197a3df083aSStefano Zampini ierr = MatShellGetContext(matis->A,&ctx);CHKERRQ(ierr); 198a3df083aSStefano Zampini ierr = MatDestroy(&matis->A);CHKERRQ(ierr); 199a3df083aSStefano Zampini matis->A = ctx->A; 200a3df083aSStefano Zampini ierr = PetscFree(ctx);CHKERRQ(ierr); 201a3df083aSStefano Zampini ierr = MatShellGetContext(pcis->A_IB,&ctx);CHKERRQ(ierr); 202a3df083aSStefano Zampini ierr = MatDestroy(&pcis->A_IB);CHKERRQ(ierr); 203a3df083aSStefano Zampini pcis->A_IB = ctx->A; 204059032f7SStefano Zampini if (!reuse) { 205059032f7SStefano Zampini PetscInt i; 206059032f7SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 207059032f7SStefano Zampini ierr = ISDestroy(&ctx->benign_zerodiag_subs[i]);CHKERRQ(ierr); 208059032f7SStefano Zampini } 209059032f7SStefano Zampini ierr = PetscFree(ctx->benign_zerodiag_subs);CHKERRQ(ierr); 210059032f7SStefano Zampini } 211a3df083aSStefano Zampini ierr = PetscFree(ctx->work);CHKERRQ(ierr); 212a3df083aSStefano Zampini ierr = PetscFree(ctx);CHKERRQ(ierr); 213a3df083aSStefano Zampini ierr = MatDestroy(&pcis->A_BI);CHKERRQ(ierr); 214a3df083aSStefano Zampini pcis->A_BI = pcbddc->benign_original_mat; 215a3df083aSStefano Zampini pcbddc->benign_original_mat = NULL; 216a3df083aSStefano Zampini } 217a3df083aSStefano Zampini PetscFunctionReturn(0); 218a3df083aSStefano Zampini } 219a3df083aSStefano Zampini 220a3df083aSStefano Zampini /* used just in bddc debug mode */ 221a3df083aSStefano Zampini #undef __FUNCT__ 222a3df083aSStefano Zampini #define __FUNCT__ "PCBDDCBenignProject" 223a3df083aSStefano Zampini PetscErrorCode PCBDDCBenignProject(PC pc, IS is1, IS is2, Mat *B) 224a3df083aSStefano Zampini { 225a3df083aSStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 226a3df083aSStefano Zampini Mat_IS *matis = (Mat_IS*)pc->pmat->data; 227a3df083aSStefano Zampini Mat An; 228a3df083aSStefano Zampini PetscErrorCode ierr; 229a3df083aSStefano Zampini 230a3df083aSStefano Zampini PetscFunctionBegin; 231a3df083aSStefano Zampini ierr = MatPtAP(matis->A,pcbddc->benign_change,MAT_INITIAL_MATRIX,2.0,&An);CHKERRQ(ierr); 232a3df083aSStefano Zampini ierr = MatZeroRowsColumns(An,pcbddc->benign_n,pcbddc->benign_p0_lidx,1.0,NULL,NULL);CHKERRQ(ierr); 233a3df083aSStefano Zampini if (is1) { 234a3df083aSStefano Zampini ierr = MatGetSubMatrix(An,is1,is2,MAT_INITIAL_MATRIX,B);CHKERRQ(ierr); 235a3df083aSStefano Zampini ierr = MatDestroy(&An);CHKERRQ(ierr); 236a3df083aSStefano Zampini } else { 237a3df083aSStefano Zampini *B = An; 238a3df083aSStefano Zampini } 239a3df083aSStefano Zampini PetscFunctionReturn(0); 240a3df083aSStefano Zampini } 241a3df083aSStefano Zampini 2421cf9b237SStefano Zampini /* TODO: add reuse flag */ 2431cf9b237SStefano Zampini #undef __FUNCT__ 2441cf9b237SStefano Zampini #define __FUNCT__ "MatSeqAIJCompress" 2451cf9b237SStefano Zampini PetscErrorCode MatSeqAIJCompress(Mat A, Mat *B) 2461cf9b237SStefano Zampini { 2471cf9b237SStefano Zampini Mat Bt; 2481cf9b237SStefano Zampini PetscScalar *a,*bdata; 2491cf9b237SStefano Zampini const PetscInt *ii,*ij; 2501cf9b237SStefano Zampini PetscInt m,n,i,nnz,*bii,*bij; 2511cf9b237SStefano Zampini PetscBool flg_row; 2521cf9b237SStefano Zampini PetscErrorCode ierr; 2531cf9b237SStefano Zampini 2541cf9b237SStefano Zampini PetscFunctionBegin; 2551cf9b237SStefano Zampini ierr = MatGetSize(A,&n,&m);CHKERRQ(ierr); 2561cf9b237SStefano Zampini ierr = MatGetRowIJ(A,0,PETSC_FALSE,PETSC_FALSE,&n,&ii,&ij,&flg_row);CHKERRQ(ierr); 2571cf9b237SStefano Zampini ierr = MatSeqAIJGetArray(A,&a);CHKERRQ(ierr); 2581cf9b237SStefano Zampini nnz = n; 2591cf9b237SStefano Zampini for (i=0;i<ii[n];i++) { 2601cf9b237SStefano Zampini if (PetscLikely(PetscAbsScalar(a[i]) > PETSC_SMALL)) nnz++; 2611cf9b237SStefano Zampini } 2621cf9b237SStefano Zampini ierr = PetscMalloc1(n+1,&bii);CHKERRQ(ierr); 2631cf9b237SStefano Zampini ierr = PetscMalloc1(nnz,&bij);CHKERRQ(ierr); 2641cf9b237SStefano Zampini ierr = PetscMalloc1(nnz,&bdata);CHKERRQ(ierr); 2651cf9b237SStefano Zampini nnz = 0; 2661cf9b237SStefano Zampini bii[0] = 0; 2671cf9b237SStefano Zampini for (i=0;i<n;i++) { 2681cf9b237SStefano Zampini PetscInt j; 2691cf9b237SStefano Zampini for (j=ii[i];j<ii[i+1];j++) { 2701cf9b237SStefano Zampini PetscScalar entry = a[j]; 2711cf9b237SStefano Zampini if (PetscLikely(PetscAbsScalar(entry) > PETSC_SMALL) || ij[j] == i) { 2721cf9b237SStefano Zampini bij[nnz] = ij[j]; 2731cf9b237SStefano Zampini bdata[nnz] = entry; 2741cf9b237SStefano Zampini nnz++; 2751cf9b237SStefano Zampini } 2761cf9b237SStefano Zampini } 2771cf9b237SStefano Zampini bii[i+1] = nnz; 2781cf9b237SStefano Zampini } 2791cf9b237SStefano Zampini ierr = MatSeqAIJRestoreArray(A,&a);CHKERRQ(ierr); 2801cf9b237SStefano Zampini ierr = MatCreateSeqAIJWithArrays(PetscObjectComm((PetscObject)A),n,m,bii,bij,bdata,&Bt);CHKERRQ(ierr); 2811cf9b237SStefano Zampini ierr = MatRestoreRowIJ(A,0,PETSC_FALSE,PETSC_FALSE,&n,&ii,&ij,&flg_row);CHKERRQ(ierr); 2821cf9b237SStefano Zampini { 2831cf9b237SStefano Zampini Mat_SeqAIJ *b = (Mat_SeqAIJ*)(Bt->data); 2841cf9b237SStefano Zampini b->free_a = PETSC_TRUE; 2851cf9b237SStefano Zampini b->free_ij = PETSC_TRUE; 2861cf9b237SStefano Zampini } 2871cf9b237SStefano Zampini *B = Bt; 2881cf9b237SStefano Zampini PetscFunctionReturn(0); 2891cf9b237SStefano Zampini } 2901cf9b237SStefano Zampini 291674ae819SStefano Zampini #undef __FUNCT__ 2924f1b2e48SStefano Zampini #define __FUNCT__ "MatDetectDisconnectedComponents" 2934f1b2e48SStefano Zampini PetscErrorCode MatDetectDisconnectedComponents(Mat A, PetscBool filter, PetscInt *ncc, IS* cc[]) 2944f1b2e48SStefano Zampini { 2954f1b2e48SStefano Zampini Mat B; 2964f1b2e48SStefano Zampini IS is_dummy,*cc_n; 2974f1b2e48SStefano Zampini ISLocalToGlobalMapping l2gmap_dummy; 2984f1b2e48SStefano Zampini PCBDDCGraph graph; 2994f1b2e48SStefano Zampini PetscInt i,n; 3004f1b2e48SStefano Zampini PetscInt *xadj,*adjncy; 3014f1b2e48SStefano Zampini PetscInt *xadj_filtered,*adjncy_filtered; 3024f1b2e48SStefano Zampini PetscBool flg_row,isseqaij; 3034f1b2e48SStefano Zampini PetscErrorCode ierr; 3044f1b2e48SStefano Zampini 3054f1b2e48SStefano Zampini PetscFunctionBegin; 30663c961adSStefano Zampini if (!A->rmap->N || !A->cmap->N) { 30763c961adSStefano Zampini *ncc = 0; 30863c961adSStefano Zampini *cc = NULL; 30963c961adSStefano Zampini PetscFunctionReturn(0); 31063c961adSStefano Zampini } 3114f1b2e48SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)A,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 3124f1b2e48SStefano Zampini if (!isseqaij && filter) { 3131cf9b237SStefano Zampini PetscBool isseqdense; 3141cf9b237SStefano Zampini 3151cf9b237SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)A,MATSEQDENSE,&isseqdense);CHKERRQ(ierr); 3161cf9b237SStefano Zampini if (!isseqdense) { 3174f1b2e48SStefano Zampini ierr = MatConvert(A,MATSEQAIJ,MAT_INITIAL_MATRIX,&B);CHKERRQ(ierr); 3181cf9b237SStefano Zampini } else { /* TODO: rectangular case and LDA */ 3191cf9b237SStefano Zampini PetscScalar *array; 3201cf9b237SStefano Zampini PetscReal chop=1.e-6; 3211cf9b237SStefano Zampini 3221cf9b237SStefano Zampini ierr = MatDuplicate(A,MAT_COPY_VALUES,&B);CHKERRQ(ierr); 3231cf9b237SStefano Zampini ierr = MatDenseGetArray(B,&array);CHKERRQ(ierr); 3241cf9b237SStefano Zampini ierr = MatGetSize(B,&n,NULL);CHKERRQ(ierr); 3251cf9b237SStefano Zampini for (i=0;i<n;i++) { 3261cf9b237SStefano Zampini PetscInt j; 3271cf9b237SStefano Zampini for (j=i+1;j<n;j++) { 3281cf9b237SStefano Zampini PetscReal thresh = chop*(PetscAbsScalar(array[i*(n+1)])+PetscAbsScalar(array[j*(n+1)])); 3291cf9b237SStefano Zampini if (PetscAbsScalar(array[i*n+j]) < thresh) array[i*n+j] = 0.; 3301cf9b237SStefano Zampini if (PetscAbsScalar(array[j*n+i]) < thresh) array[j*n+i] = 0.; 3311cf9b237SStefano Zampini } 3321cf9b237SStefano Zampini } 3331cf9b237SStefano Zampini ierr = MatDenseRestoreArray(B,&array);CHKERRQ(ierr); 3341cf9b237SStefano Zampini ierr = MatConvert(B,MATSEQAIJ,MAT_REUSE_MATRIX,&B);CHKERRQ(ierr); 3351cf9b237SStefano Zampini } 3364f1b2e48SStefano Zampini } else { 3374f1b2e48SStefano Zampini B = A; 3384f1b2e48SStefano Zampini } 3394f1b2e48SStefano Zampini ierr = MatGetRowIJ(B,0,PETSC_TRUE,PETSC_FALSE,&n,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 3404f1b2e48SStefano Zampini 3414f1b2e48SStefano Zampini /* if filter is true, then removes entries lower than PETSC_SMALL in magnitude */ 3424f1b2e48SStefano Zampini if (filter) { 3434f1b2e48SStefano Zampini PetscScalar *data; 3444f1b2e48SStefano Zampini PetscInt j,cum; 3454f1b2e48SStefano Zampini 3464f1b2e48SStefano Zampini ierr = PetscCalloc2(n+1,&xadj_filtered,xadj[n],&adjncy_filtered);CHKERRQ(ierr); 3474f1b2e48SStefano Zampini ierr = MatSeqAIJGetArray(B,&data);CHKERRQ(ierr); 3484f1b2e48SStefano Zampini cum = 0; 3494f1b2e48SStefano Zampini for (i=0;i<n;i++) { 3504f1b2e48SStefano Zampini PetscInt t; 3514f1b2e48SStefano Zampini 3524f1b2e48SStefano Zampini for (j=xadj[i];j<xadj[i+1];j++) { 3534f1b2e48SStefano Zampini if (PetscUnlikely(PetscAbsScalar(data[j]) < PETSC_SMALL)) { 3544f1b2e48SStefano Zampini continue; 3554f1b2e48SStefano Zampini } 3564f1b2e48SStefano Zampini adjncy_filtered[cum+xadj_filtered[i]++] = adjncy[j]; 3574f1b2e48SStefano Zampini } 3584f1b2e48SStefano Zampini t = xadj_filtered[i]; 3594f1b2e48SStefano Zampini xadj_filtered[i] = cum; 3604f1b2e48SStefano Zampini cum += t; 3614f1b2e48SStefano Zampini } 3624f1b2e48SStefano Zampini ierr = MatSeqAIJRestoreArray(B,&data);CHKERRQ(ierr); 3634f1b2e48SStefano Zampini } else { 3644f1b2e48SStefano Zampini xadj_filtered = NULL; 3654f1b2e48SStefano Zampini adjncy_filtered = NULL; 3664f1b2e48SStefano Zampini } 3674f1b2e48SStefano Zampini 3684f1b2e48SStefano Zampini /* compute local connected components using PCBDDCGraph */ 3694f1b2e48SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,n,0,1,&is_dummy);CHKERRQ(ierr); 3704f1b2e48SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(is_dummy,&l2gmap_dummy);CHKERRQ(ierr); 3714f1b2e48SStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 3724f1b2e48SStefano Zampini ierr = PCBDDCGraphCreate(&graph);CHKERRQ(ierr); 3734f1b2e48SStefano Zampini ierr = PCBDDCGraphInit(graph,l2gmap_dummy,n);CHKERRQ(ierr); 3744f1b2e48SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&l2gmap_dummy);CHKERRQ(ierr); 3754f1b2e48SStefano Zampini if (xadj_filtered) { 3764f1b2e48SStefano Zampini graph->xadj = xadj_filtered; 3774f1b2e48SStefano Zampini graph->adjncy = adjncy_filtered; 3784f1b2e48SStefano Zampini } else { 3794f1b2e48SStefano Zampini graph->xadj = xadj; 3804f1b2e48SStefano Zampini graph->adjncy = adjncy; 3814f1b2e48SStefano Zampini } 3824f1b2e48SStefano Zampini ierr = PCBDDCGraphSetUp(graph,1,NULL,NULL,0,NULL,NULL);CHKERRQ(ierr); 3834f1b2e48SStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(graph);CHKERRQ(ierr); 3844f1b2e48SStefano Zampini /* partial clean up */ 3854f1b2e48SStefano Zampini ierr = PetscFree2(xadj_filtered,adjncy_filtered);CHKERRQ(ierr); 3864f1b2e48SStefano Zampini ierr = MatRestoreRowIJ(B,0,PETSC_TRUE,PETSC_FALSE,&n,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 3871cf9b237SStefano Zampini if (A != B) { 3884f1b2e48SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 3894f1b2e48SStefano Zampini } 3904f1b2e48SStefano Zampini 3914f1b2e48SStefano Zampini /* get back data */ 3921cf9b237SStefano Zampini if (ncc) *ncc = graph->ncc; 3931cf9b237SStefano Zampini if (cc) { 3944f1b2e48SStefano Zampini ierr = PetscMalloc1(graph->ncc,&cc_n);CHKERRQ(ierr); 3954f1b2e48SStefano Zampini for (i=0;i<graph->ncc;i++) { 3964f1b2e48SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,graph->cptr[i+1]-graph->cptr[i],graph->queue+graph->cptr[i],PETSC_COPY_VALUES,&cc_n[i]);CHKERRQ(ierr); 3974f1b2e48SStefano Zampini } 3984f1b2e48SStefano Zampini *cc = cc_n; 3991cf9b237SStefano Zampini } 4004f1b2e48SStefano Zampini /* clean up graph */ 4014f1b2e48SStefano Zampini graph->xadj = 0; 4024f1b2e48SStefano Zampini graph->adjncy = 0; 4034f1b2e48SStefano Zampini ierr = PCBDDCGraphDestroy(&graph);CHKERRQ(ierr); 4044f1b2e48SStefano Zampini PetscFunctionReturn(0); 4054f1b2e48SStefano Zampini } 4064f1b2e48SStefano Zampini 4074f1b2e48SStefano Zampini #undef __FUNCT__ 4085408967cSStefano Zampini #define __FUNCT__ "PCBDDCBenignCheck" 4095408967cSStefano Zampini PetscErrorCode PCBDDCBenignCheck(PC pc, IS zerodiag) 4105408967cSStefano Zampini { 4115408967cSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 4125408967cSStefano Zampini PC_IS* pcis = (PC_IS*)(pc->data); 413dee84bffSStefano Zampini IS dirIS = NULL; 4144f1b2e48SStefano Zampini PetscInt i; 4155408967cSStefano Zampini PetscErrorCode ierr; 4165408967cSStefano Zampini 4175408967cSStefano Zampini PetscFunctionBegin; 418dee84bffSStefano Zampini ierr = PCBDDCGraphGetDirichletDofs(pcbddc->mat_graph,&dirIS);CHKERRQ(ierr); 4195408967cSStefano Zampini if (zerodiag) { 4205408967cSStefano Zampini Mat A; 4215408967cSStefano Zampini Vec vec3_N; 4225408967cSStefano Zampini PetscScalar *vals; 4235408967cSStefano Zampini const PetscInt *idxs; 424d12d3064SStefano Zampini PetscInt nz,*count; 4255408967cSStefano Zampini 4265408967cSStefano Zampini /* p0 */ 4275408967cSStefano Zampini ierr = VecSet(pcis->vec1_N,0.);CHKERRQ(ierr); 4285408967cSStefano Zampini ierr = PetscMalloc1(pcis->n,&vals);CHKERRQ(ierr); 4295408967cSStefano Zampini ierr = ISGetLocalSize(zerodiag,&nz);CHKERRQ(ierr); 4305408967cSStefano Zampini ierr = ISGetIndices(zerodiag,&idxs);CHKERRQ(ierr); 4314f1b2e48SStefano Zampini for (i=0;i<nz;i++) vals[i] = 1.; 4325408967cSStefano Zampini ierr = VecSetValues(pcis->vec1_N,nz,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 4335408967cSStefano Zampini ierr = VecAssemblyBegin(pcis->vec1_N);CHKERRQ(ierr); 4345408967cSStefano Zampini ierr = VecAssemblyEnd(pcis->vec1_N);CHKERRQ(ierr); 4355408967cSStefano Zampini /* v_I */ 4365408967cSStefano Zampini ierr = VecSetRandom(pcis->vec2_N,NULL);CHKERRQ(ierr); 4375408967cSStefano Zampini for (i=0;i<nz;i++) vals[i] = 0.; 4385408967cSStefano Zampini ierr = VecSetValues(pcis->vec2_N,nz,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 4395408967cSStefano Zampini ierr = ISRestoreIndices(zerodiag,&idxs);CHKERRQ(ierr); 4405408967cSStefano Zampini ierr = ISGetIndices(pcis->is_B_local,&idxs);CHKERRQ(ierr); 4415408967cSStefano Zampini for (i=0;i<pcis->n_B;i++) vals[i] = 0.; 4425408967cSStefano Zampini ierr = VecSetValues(pcis->vec2_N,pcis->n_B,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 4435408967cSStefano Zampini ierr = ISRestoreIndices(pcis->is_B_local,&idxs);CHKERRQ(ierr); 4445408967cSStefano Zampini if (dirIS) { 4455408967cSStefano Zampini PetscInt n; 4465408967cSStefano Zampini 4475408967cSStefano Zampini ierr = ISGetLocalSize(dirIS,&n);CHKERRQ(ierr); 4485408967cSStefano Zampini ierr = ISGetIndices(dirIS,&idxs);CHKERRQ(ierr); 4495408967cSStefano Zampini for (i=0;i<n;i++) vals[i] = 0.; 4505408967cSStefano Zampini ierr = VecSetValues(pcis->vec2_N,n,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 4515408967cSStefano Zampini ierr = ISRestoreIndices(dirIS,&idxs);CHKERRQ(ierr); 4525408967cSStefano Zampini } 4535408967cSStefano Zampini ierr = VecAssemblyBegin(pcis->vec2_N);CHKERRQ(ierr); 4545408967cSStefano Zampini ierr = VecAssemblyEnd(pcis->vec2_N);CHKERRQ(ierr); 4555408967cSStefano Zampini ierr = VecDuplicate(pcis->vec1_N,&vec3_N);CHKERRQ(ierr); 4565408967cSStefano Zampini ierr = VecSet(vec3_N,0.);CHKERRQ(ierr); 4575408967cSStefano Zampini ierr = MatISGetLocalMat(pc->mat,&A);CHKERRQ(ierr); 4585408967cSStefano Zampini ierr = MatMult(A,pcis->vec1_N,vec3_N);CHKERRQ(ierr); 4595408967cSStefano Zampini ierr = VecDot(vec3_N,pcis->vec2_N,&vals[0]);CHKERRQ(ierr); 4604f1b2e48SStefano Zampini if (PetscAbsScalar(vals[0]) > PETSC_SMALL) { 4615408967cSStefano Zampini SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_SUP,"Benign trick can not be applied! b(v_I,p_0) = %f (should be numerically 0.)",PetscAbsScalar(vals[0])); 4625408967cSStefano Zampini } 4635408967cSStefano Zampini ierr = PetscFree(vals);CHKERRQ(ierr); 4645408967cSStefano Zampini ierr = VecDestroy(&vec3_N);CHKERRQ(ierr); 465d12d3064SStefano Zampini 466d12d3064SStefano Zampini /* there should not be any pressure dofs lying on the interface */ 467d12d3064SStefano Zampini ierr = PetscCalloc1(pcis->n,&count);CHKERRQ(ierr); 468d12d3064SStefano Zampini ierr = ISGetIndices(pcis->is_B_local,&idxs);CHKERRQ(ierr); 469d12d3064SStefano Zampini for (i=0;i<pcis->n_B;i++) count[idxs[i]]++; 470d12d3064SStefano Zampini ierr = ISRestoreIndices(pcis->is_B_local,&idxs);CHKERRQ(ierr); 471d12d3064SStefano Zampini ierr = ISGetIndices(zerodiag,&idxs);CHKERRQ(ierr); 472d4d8cf7bSStefano Zampini for (i=0;i<nz;i++) { 473d12d3064SStefano Zampini if (count[idxs[i]]) { 474d12d3064SStefano Zampini SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_SUP,"Benign trick can not be applied! pressure dof %d is an interface dof",idxs[i]); 475d12d3064SStefano Zampini } 476d4d8cf7bSStefano Zampini } 477d12d3064SStefano Zampini ierr = ISRestoreIndices(zerodiag,&idxs);CHKERRQ(ierr); 478d12d3064SStefano Zampini ierr = PetscFree(count);CHKERRQ(ierr); 4795408967cSStefano Zampini } 480dee84bffSStefano Zampini ierr = ISDestroy(&dirIS);CHKERRQ(ierr); 4815408967cSStefano Zampini 4825408967cSStefano Zampini /* check PCBDDCBenignGetOrSetP0 */ 4835408967cSStefano Zampini ierr = VecSetRandom(pcis->vec1_global,NULL);CHKERRQ(ierr); 4844f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) pcbddc->benign_p0[i] = -PetscGlobalRank-i; 4855408967cSStefano Zampini ierr = PCBDDCBenignGetOrSetP0(pc,pcis->vec1_global,PETSC_FALSE);CHKERRQ(ierr); 4864f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) pcbddc->benign_p0[i] = 1; 4875408967cSStefano Zampini ierr = PCBDDCBenignGetOrSetP0(pc,pcis->vec1_global,PETSC_TRUE);CHKERRQ(ierr); 4884f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 4894f1b2e48SStefano Zampini if ((PetscInt)PetscRealPart(pcbddc->benign_p0[i]) != -PetscGlobalRank-i) { 4904f1b2e48SStefano Zampini SETERRQ3(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Error testing PCBDDCBenignGetOrSetP0! Found %1.4e at %d instead of %1.4e\n",pcbddc->benign_p0[i],i,-PetscGlobalRank-i);CHKERRQ(ierr); 4914f1b2e48SStefano Zampini } 4925408967cSStefano Zampini } 4935408967cSStefano Zampini PetscFunctionReturn(0); 4945408967cSStefano Zampini } 4955408967cSStefano Zampini 4965408967cSStefano Zampini #undef __FUNCT__ 497339f8db1SStefano Zampini #define __FUNCT__ "PCBDDCBenignDetectSaddlePoint" 498339f8db1SStefano Zampini PetscErrorCode PCBDDCBenignDetectSaddlePoint(PC pc, IS *zerodiaglocal) 499339f8db1SStefano Zampini { 500339f8db1SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 5014f1b2e48SStefano Zampini IS pressures,zerodiag,*zerodiag_subs; 502b0f5fe93SStefano Zampini PetscInt nz,n; 5034f1b2e48SStefano Zampini PetscBool sorted,have_null,has_null_pressures,recompute_zerodiag; 504339f8db1SStefano Zampini PetscErrorCode ierr; 505339f8db1SStefano Zampini 506339f8db1SStefano Zampini PetscFunctionBegin; 5079f47a83aSStefano Zampini ierr = PetscSFDestroy(&pcbddc->benign_sf);CHKERRQ(ierr); 5089f47a83aSStefano Zampini ierr = MatDestroy(&pcbddc->benign_B0);CHKERRQ(ierr); 509339f8db1SStefano Zampini ierr = MatDestroy(&pcbddc->benign_original_mat);CHKERRQ(ierr); 510a3df083aSStefano Zampini for (n=0;n<pcbddc->benign_n;n++) { 511a3df083aSStefano Zampini ierr = ISDestroy(&pcbddc->benign_zerodiag_subs[n]);CHKERRQ(ierr); 512a3df083aSStefano Zampini } 513a3df083aSStefano Zampini ierr = PetscFree(pcbddc->benign_zerodiag_subs);CHKERRQ(ierr); 514a3df083aSStefano Zampini pcbddc->benign_n = 0; 5154f1b2e48SStefano Zampini /* if a local info on dofs is present, assumes the last field is represented by "pressures" 5164f1b2e48SStefano Zampini otherwise, it uses only zerodiagonal dofs (ok if the pressure block is all zero; it could fail if it is not) 5174f1b2e48SStefano Zampini Checks if all the pressure dofs in each subdomain have a zero diagonal 5184f1b2e48SStefano Zampini If not, a change of basis on pressures is not needed 5194f1b2e48SStefano Zampini since the local Schur complements are SPD 5204f1b2e48SStefano Zampini */ 5214f1b2e48SStefano Zampini has_null_pressures = PETSC_TRUE; 5224f1b2e48SStefano Zampini have_null = PETSC_TRUE; 52340fa8d13SStefano Zampini if (pcbddc->n_ISForDofsLocal) { 5244f1b2e48SStefano Zampini PetscInt npl,*idxs,p = pcbddc->n_ISForDofsLocal-1; 5254f1b2e48SStefano Zampini 5264f1b2e48SStefano Zampini /* Dofs splitting for BDDC cannot have PETSC_COMM_SELF, so create a sequential IS */ 5274f1b2e48SStefano Zampini ierr = ISGetLocalSize(pcbddc->ISForDofsLocal[p],&npl);CHKERRQ(ierr); 5284f1b2e48SStefano Zampini ierr = ISGetIndices(pcbddc->ISForDofsLocal[p],(const PetscInt**)&idxs);CHKERRQ(ierr); 5294f1b2e48SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,npl,idxs,PETSC_COPY_VALUES,&pressures);CHKERRQ(ierr); 530ba14f8e3SStefano Zampini ierr = ISRestoreIndices(pcbddc->ISForDofsLocal[p],(const PetscInt**)&idxs);CHKERRQ(ierr); 53140fa8d13SStefano Zampini ierr = ISSorted(pressures,&sorted);CHKERRQ(ierr); 53240fa8d13SStefano Zampini if (!sorted) { 53340fa8d13SStefano Zampini ierr = ISSort(pressures);CHKERRQ(ierr); 53440fa8d13SStefano Zampini } 53540fa8d13SStefano Zampini } else { 53640fa8d13SStefano Zampini pressures = NULL; 53740fa8d13SStefano Zampini } 53897d764eeSStefano Zampini /* pcis has not been setup yet, so get the local size from the subdomain matrix */ 53997d764eeSStefano Zampini ierr = MatGetLocalSize(pcbddc->local_mat,&n,NULL);CHKERRQ(ierr); 54097d764eeSStefano Zampini ierr = MatFindZeroDiagonals(pcbddc->local_mat,&zerodiag);CHKERRQ(ierr); 541339f8db1SStefano Zampini ierr = ISSorted(zerodiag,&sorted);CHKERRQ(ierr); 542339f8db1SStefano Zampini if (!sorted) { 543339f8db1SStefano Zampini ierr = ISSort(zerodiag);CHKERRQ(ierr); 544339f8db1SStefano Zampini } 545339f8db1SStefano Zampini ierr = ISGetLocalSize(zerodiag,&nz);CHKERRQ(ierr); 5464f1b2e48SStefano Zampini if (!nz) { 5474f1b2e48SStefano Zampini if (n) have_null = PETSC_FALSE; 5484f1b2e48SStefano Zampini has_null_pressures = PETSC_FALSE; 54940fa8d13SStefano Zampini ierr = ISDestroy(&zerodiag);CHKERRQ(ierr); 55040fa8d13SStefano Zampini } 5514f1b2e48SStefano Zampini recompute_zerodiag = PETSC_FALSE; 5524f1b2e48SStefano Zampini /* in case disconnected subdomains info is present, split the pressures accordingly (otherwise the benign trick could fail) */ 5534f1b2e48SStefano Zampini zerodiag_subs = NULL; 5544f1b2e48SStefano Zampini pcbddc->benign_n = 0; 5554f1b2e48SStefano Zampini if (has_null_pressures) { 5564f1b2e48SStefano Zampini IS *subs; 5574f1b2e48SStefano Zampini PetscInt nsubs,i; 5584f1b2e48SStefano Zampini 5594f1b2e48SStefano Zampini subs = pcbddc->local_subs; 5604f1b2e48SStefano Zampini nsubs = pcbddc->n_local_subs; 5614f1b2e48SStefano Zampini if (nsubs > 1) { 5624f1b2e48SStefano Zampini ierr = PetscCalloc1(nsubs,&zerodiag_subs);CHKERRQ(ierr); 5634f1b2e48SStefano Zampini for (i=0;i<nsubs;i++) { 5644f1b2e48SStefano Zampini ISLocalToGlobalMapping l2g; 5654f1b2e48SStefano Zampini IS t_zerodiag_subs; 5664f1b2e48SStefano Zampini PetscInt nl; 5674f1b2e48SStefano Zampini 5684f1b2e48SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(subs[i],&l2g);CHKERRQ(ierr); 5694f1b2e48SStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(l2g,IS_GTOLM_DROP,zerodiag,&t_zerodiag_subs);CHKERRQ(ierr); 5704f1b2e48SStefano Zampini ierr = ISGetLocalSize(t_zerodiag_subs,&nl);CHKERRQ(ierr); 5714f1b2e48SStefano Zampini if (nl) { 5724f1b2e48SStefano Zampini PetscBool valid = PETSC_TRUE; 5734f1b2e48SStefano Zampini 5744f1b2e48SStefano Zampini if (pressures) { 5754f1b2e48SStefano Zampini IS t_pressure_subs; 5764f1b2e48SStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(l2g,IS_GTOLM_DROP,pressures,&t_pressure_subs);CHKERRQ(ierr); 5774f1b2e48SStefano Zampini ierr = ISEqual(t_pressure_subs,t_zerodiag_subs,&valid);CHKERRQ(ierr); 5784f1b2e48SStefano Zampini ierr = ISDestroy(&t_pressure_subs);CHKERRQ(ierr); 5794f1b2e48SStefano Zampini } 5804f1b2e48SStefano Zampini if (valid) { 5814f1b2e48SStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(l2g,t_zerodiag_subs,&zerodiag_subs[pcbddc->benign_n]);CHKERRQ(ierr); 5824f1b2e48SStefano Zampini pcbddc->benign_n++; 5834f1b2e48SStefano Zampini } else { 5844f1b2e48SStefano Zampini recompute_zerodiag = PETSC_TRUE; 5854f1b2e48SStefano Zampini } 5864f1b2e48SStefano Zampini } 5874f1b2e48SStefano Zampini ierr = ISDestroy(&t_zerodiag_subs);CHKERRQ(ierr); 5884f1b2e48SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&l2g);CHKERRQ(ierr); 5894f1b2e48SStefano Zampini } 5904f1b2e48SStefano Zampini } else { /* there's just one subdomain (or zero if they have not been detected */ 5914f1b2e48SStefano Zampini PetscBool valid = PETSC_TRUE; 5924f1b2e48SStefano Zampini if (pressures) { 5934f1b2e48SStefano Zampini ierr = ISEqual(pressures,zerodiag,&valid);CHKERRQ(ierr); 5944f1b2e48SStefano Zampini } 5954f1b2e48SStefano Zampini if (valid) { 5964f1b2e48SStefano Zampini pcbddc->benign_n = 1; 597ca92afb2SStefano Zampini ierr = PetscMalloc1(pcbddc->benign_n,&zerodiag_subs);CHKERRQ(ierr); 5984f1b2e48SStefano Zampini ierr = PetscObjectReference((PetscObject)zerodiag);CHKERRQ(ierr); 5994f1b2e48SStefano Zampini zerodiag_subs[0] = zerodiag; 6004f1b2e48SStefano Zampini } 6014f1b2e48SStefano Zampini } 6024f1b2e48SStefano Zampini } 6034f1b2e48SStefano Zampini 6044f1b2e48SStefano Zampini if (!pcbddc->benign_n) { 6054f1b2e48SStefano Zampini ierr = ISDestroy(&zerodiag);CHKERRQ(ierr); 6064f1b2e48SStefano Zampini recompute_zerodiag = PETSC_FALSE; 6074f1b2e48SStefano Zampini has_null_pressures = PETSC_FALSE; 6084f1b2e48SStefano Zampini have_null = PETSC_FALSE; 6094f1b2e48SStefano Zampini } 6104f1b2e48SStefano Zampini 6114f1b2e48SStefano Zampini /* final check for null pressures */ 6124f1b2e48SStefano Zampini if (zerodiag && pressures) { 6134f1b2e48SStefano Zampini PetscInt nz,np; 6144f1b2e48SStefano Zampini ierr = ISGetLocalSize(zerodiag,&nz);CHKERRQ(ierr); 6154f1b2e48SStefano Zampini ierr = ISGetLocalSize(pressures,&np);CHKERRQ(ierr); 6164f1b2e48SStefano Zampini if (nz != np) have_null = PETSC_FALSE; 6174f1b2e48SStefano Zampini } 6184f1b2e48SStefano Zampini 6194f1b2e48SStefano Zampini if (recompute_zerodiag) { 6204f1b2e48SStefano Zampini ierr = ISDestroy(&zerodiag);CHKERRQ(ierr); 6214f1b2e48SStefano Zampini if (pcbddc->benign_n == 1) { 6224f1b2e48SStefano Zampini ierr = PetscObjectReference((PetscObject)zerodiag_subs[0]);CHKERRQ(ierr); 6234f1b2e48SStefano Zampini zerodiag = zerodiag_subs[0]; 6244f1b2e48SStefano Zampini } else { 6254f1b2e48SStefano Zampini PetscInt i,nzn,*new_idxs; 6264f1b2e48SStefano Zampini 6274f1b2e48SStefano Zampini nzn = 0; 6284f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 6294f1b2e48SStefano Zampini PetscInt ns; 6304f1b2e48SStefano Zampini ierr = ISGetLocalSize(zerodiag_subs[i],&ns);CHKERRQ(ierr); 6314f1b2e48SStefano Zampini nzn += ns; 6324f1b2e48SStefano Zampini } 6334f1b2e48SStefano Zampini ierr = PetscMalloc1(nzn,&new_idxs);CHKERRQ(ierr); 6344f1b2e48SStefano Zampini nzn = 0; 6354f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 6364f1b2e48SStefano Zampini PetscInt ns,*idxs; 6374f1b2e48SStefano Zampini ierr = ISGetLocalSize(zerodiag_subs[i],&ns);CHKERRQ(ierr); 6384f1b2e48SStefano Zampini ierr = ISGetIndices(zerodiag_subs[i],(const PetscInt**)&idxs);CHKERRQ(ierr); 6394f1b2e48SStefano Zampini ierr = PetscMemcpy(new_idxs+nzn,idxs,ns*sizeof(PetscInt));CHKERRQ(ierr); 6404f1b2e48SStefano Zampini ierr = ISRestoreIndices(zerodiag_subs[i],(const PetscInt**)&idxs);CHKERRQ(ierr); 6414f1b2e48SStefano Zampini nzn += ns; 6424f1b2e48SStefano Zampini } 6434f1b2e48SStefano Zampini ierr = PetscSortInt(nzn,new_idxs);CHKERRQ(ierr); 6444f1b2e48SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,nzn,new_idxs,PETSC_OWN_POINTER,&zerodiag);CHKERRQ(ierr); 6454f1b2e48SStefano Zampini } 6464f1b2e48SStefano Zampini have_null = PETSC_FALSE; 6474f1b2e48SStefano Zampini } 6484f1b2e48SStefano Zampini 649b3afcdbeSStefano Zampini /* no-net-flux */ 650b3afcdbeSStefano Zampini if (pcbddc->benign_compute_nonetflux) { 651b3afcdbeSStefano Zampini Mat_IS* matis = (Mat_IS*)(pc->pmat->data); 652b3afcdbeSStefano Zampini MatNullSpace near_null_space; 653b3afcdbeSStefano Zampini ISLocalToGlobalMapping rmap; 654b3afcdbeSStefano Zampini Vec quad_vec; 655b3afcdbeSStefano Zampini PetscScalar *pvals; 656b3afcdbeSStefano Zampini PetscInt i,np,*dummyins; 657b3afcdbeSStefano Zampini 658b3afcdbeSStefano Zampini /* create vector to hold quadrature weights */ 659b3afcdbeSStefano Zampini ierr = MatCreateVecs(pc->pmat,&quad_vec,NULL);CHKERRQ(ierr); 660b3afcdbeSStefano Zampini ierr = VecSet(quad_vec,0.0);CHKERRQ(ierr); 661b3afcdbeSStefano Zampini ierr = MatGetLocalToGlobalMapping(pc->pmat,&rmap,NULL);CHKERRQ(ierr); 662b3afcdbeSStefano Zampini ierr = VecSetLocalToGlobalMapping(quad_vec,rmap);CHKERRQ(ierr); 663b3afcdbeSStefano Zampini 664b3afcdbeSStefano Zampini /* compute B^{(i)T} * 1_p */ 665b3afcdbeSStefano Zampini np = 0; 666b3afcdbeSStefano Zampini if (zerodiag) { 667b3afcdbeSStefano Zampini ierr = ISGetLocalSize(zerodiag,&np);CHKERRQ(ierr); 668b3afcdbeSStefano Zampini } else if (pressures) { 669b3afcdbeSStefano Zampini ierr = ISGetLocalSize(pressures,&np);CHKERRQ(ierr); 670b3afcdbeSStefano Zampini } 671b3afcdbeSStefano Zampini ierr = PetscMalloc1(np,&pvals);CHKERRQ(ierr); 672b3afcdbeSStefano Zampini for (i=0;i<np;i++) pvals[i] = 1.; 673b3afcdbeSStefano Zampini ierr = VecSet(matis->x,0.);CHKERRQ(ierr); 674b3afcdbeSStefano Zampini if (np) { 675b3afcdbeSStefano Zampini const PetscInt *pidxs; 676b3afcdbeSStefano Zampini 677b3afcdbeSStefano Zampini if (zerodiag) { 678b3afcdbeSStefano Zampini ierr = ISGetIndices(zerodiag,&pidxs);CHKERRQ(ierr); 679b3afcdbeSStefano Zampini } else if (pressures) { 680b3afcdbeSStefano Zampini ierr = ISGetIndices(pressures,&pidxs);CHKERRQ(ierr); 681b3afcdbeSStefano Zampini } 682b3afcdbeSStefano Zampini ierr = VecSetValues(matis->x,np,pidxs,pvals,INSERT_VALUES);CHKERRQ(ierr); 683b3afcdbeSStefano Zampini if (zerodiag) { 684b3afcdbeSStefano Zampini ierr = ISRestoreIndices(zerodiag,&pidxs);CHKERRQ(ierr); 685b3afcdbeSStefano Zampini } else if (pressures) { 686b3afcdbeSStefano Zampini ierr = ISRestoreIndices(pressures,&pidxs);CHKERRQ(ierr); 687b3afcdbeSStefano Zampini } 688b3afcdbeSStefano Zampini } 689b3afcdbeSStefano Zampini ierr = VecAssemblyBegin(matis->x);CHKERRQ(ierr); 690b3afcdbeSStefano Zampini ierr = VecAssemblyEnd(matis->x);CHKERRQ(ierr); 691b3afcdbeSStefano Zampini ierr = MatMult(matis->A,matis->x,matis->y);CHKERRQ(ierr); 692b3afcdbeSStefano Zampini ierr = PetscFree(pvals);CHKERRQ(ierr); 693b3afcdbeSStefano Zampini 694b3afcdbeSStefano Zampini /* decide which of the sharing ranks (per dof) has to insert the values (should just be a matter of having a different orientation) */ 695b3afcdbeSStefano Zampini ierr = MatISSetUpSF(pc->pmat);CHKERRQ(ierr); 696b3afcdbeSStefano Zampini for (i=0;i<matis->sf_nroots;i++) matis->sf_rootdata[i] = -1; 697b3afcdbeSStefano Zampini for (i=0;i<matis->sf_nleaves;i++) matis->sf_leafdata[i] = PetscGlobalRank; 698b3afcdbeSStefano Zampini ierr = PetscSFReduceBegin(matis->sf,MPIU_INT,matis->sf_leafdata,matis->sf_rootdata,MPI_MAX);CHKERRQ(ierr); 699b3afcdbeSStefano Zampini ierr = PetscSFReduceEnd(matis->sf,MPIU_INT,matis->sf_leafdata,matis->sf_rootdata,MPI_MAX);CHKERRQ(ierr); 700b3afcdbeSStefano Zampini ierr = PetscSFBcastBegin(matis->sf,MPIU_INT,matis->sf_rootdata,matis->sf_leafdata);CHKERRQ(ierr); 701b3afcdbeSStefano Zampini ierr = PetscSFBcastEnd(matis->sf,MPIU_INT,matis->sf_rootdata,matis->sf_leafdata);CHKERRQ(ierr); 702b3afcdbeSStefano Zampini ierr = VecGetArray(matis->y,&pvals);CHKERRQ(ierr); 703b3afcdbeSStefano Zampini for (i=0;i<matis->sf_nleaves;i++) { 704b3afcdbeSStefano Zampini if (PetscGlobalRank != matis->sf_leafdata[i]) { 705b3afcdbeSStefano Zampini pvals[i] = 0.; 706b3afcdbeSStefano Zampini } 707b3afcdbeSStefano Zampini } 708b3afcdbeSStefano Zampini ierr = PetscMalloc1(matis->sf_nleaves,&dummyins);CHKERRQ(ierr); 709b3afcdbeSStefano Zampini for (i=0;i<matis->sf_nleaves;i++) dummyins[i] = i; 710b3afcdbeSStefano Zampini ierr = VecSetValuesLocal(quad_vec,matis->sf_nleaves,dummyins,pvals,ADD_VALUES);CHKERRQ(ierr); 711b3afcdbeSStefano Zampini ierr = VecRestoreArray(matis->y,&pvals);CHKERRQ(ierr); 712b3afcdbeSStefano Zampini ierr = PetscFree(dummyins);CHKERRQ(ierr); 713b3afcdbeSStefano Zampini 714b3afcdbeSStefano Zampini /* assembly quadrature vec and attach near null space to pmat */ 715b3afcdbeSStefano Zampini ierr = VecAssemblyBegin(quad_vec);CHKERRQ(ierr); 716b3afcdbeSStefano Zampini ierr = VecAssemblyEnd(quad_vec);CHKERRQ(ierr); 717b3afcdbeSStefano Zampini ierr = VecNormalize(quad_vec,NULL);CHKERRQ(ierr); 718b3afcdbeSStefano Zampini ierr = MatNullSpaceCreate(PetscObjectComm((PetscObject)pc),PETSC_FALSE,1,&quad_vec,&near_null_space);CHKERRQ(ierr); 719b3afcdbeSStefano Zampini ierr = VecDestroy(&quad_vec);CHKERRQ(ierr); 720b3afcdbeSStefano Zampini ierr = MatSetNearNullSpace(pc->pmat,near_null_space);CHKERRQ(ierr); 721b3afcdbeSStefano Zampini ierr = MatNullSpaceDestroy(&near_null_space);CHKERRQ(ierr); 722b3afcdbeSStefano Zampini } 723b3afcdbeSStefano Zampini 724b3afcdbeSStefano Zampini /* change of basis and p0 dofs */ 7254f1b2e48SStefano Zampini if (has_null_pressures) { 7264f1b2e48SStefano Zampini IS zerodiagc; 7274f1b2e48SStefano Zampini const PetscInt *idxs,*idxsc; 7284f1b2e48SStefano Zampini PetscInt i,s,*nnz; 7294f1b2e48SStefano Zampini 7304f1b2e48SStefano Zampini ierr = ISGetLocalSize(zerodiag,&nz);CHKERRQ(ierr); 731339f8db1SStefano Zampini ierr = ISComplement(zerodiag,0,n,&zerodiagc);CHKERRQ(ierr); 732339f8db1SStefano Zampini ierr = ISGetIndices(zerodiagc,&idxsc);CHKERRQ(ierr); 733339f8db1SStefano Zampini /* local change of basis for pressures */ 734339f8db1SStefano Zampini ierr = MatDestroy(&pcbddc->benign_change);CHKERRQ(ierr); 73597d764eeSStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)pcbddc->local_mat),&pcbddc->benign_change);CHKERRQ(ierr); 736339f8db1SStefano Zampini ierr = MatSetType(pcbddc->benign_change,MATAIJ);CHKERRQ(ierr); 737339f8db1SStefano Zampini ierr = MatSetSizes(pcbddc->benign_change,n,n,PETSC_DECIDE,PETSC_DECIDE);CHKERRQ(ierr); 738339f8db1SStefano Zampini ierr = PetscMalloc1(n,&nnz);CHKERRQ(ierr); 7394f1b2e48SStefano Zampini for (i=0;i<n-nz;i++) nnz[idxsc[i]] = 1; /* identity on velocities plus pressure dofs for non-singular subdomains */ 7404f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 7414f1b2e48SStefano Zampini PetscInt nzs,j; 7424f1b2e48SStefano Zampini 7434f1b2e48SStefano Zampini ierr = ISGetLocalSize(zerodiag_subs[i],&nzs);CHKERRQ(ierr); 7444f1b2e48SStefano Zampini ierr = ISGetIndices(zerodiag_subs[i],&idxs);CHKERRQ(ierr); 7454f1b2e48SStefano Zampini for (j=0;j<nzs-1;j++) nnz[idxs[j]] = 2; /* change on pressures */ 7464f1b2e48SStefano Zampini nnz[idxs[nzs-1]] = nzs; /* last local pressure dof in subdomain */ 7474f1b2e48SStefano Zampini ierr = ISRestoreIndices(zerodiag_subs[i],&idxs);CHKERRQ(ierr); 7484f1b2e48SStefano Zampini } 749339f8db1SStefano Zampini ierr = MatSeqAIJSetPreallocation(pcbddc->benign_change,0,nnz);CHKERRQ(ierr); 750339f8db1SStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 751339f8db1SStefano Zampini /* set identity on velocities */ 752339f8db1SStefano Zampini for (i=0;i<n-nz;i++) { 753339f8db1SStefano Zampini ierr = MatSetValue(pcbddc->benign_change,idxsc[i],idxsc[i],1.,INSERT_VALUES);CHKERRQ(ierr); 754339f8db1SStefano Zampini } 7554f1b2e48SStefano Zampini ierr = ISRestoreIndices(zerodiagc,&idxsc);CHKERRQ(ierr); 7564f1b2e48SStefano Zampini ierr = ISDestroy(&zerodiagc);CHKERRQ(ierr); 7579f47a83aSStefano Zampini ierr = PetscFree3(pcbddc->benign_p0_lidx,pcbddc->benign_p0_gidx,pcbddc->benign_p0);CHKERRQ(ierr); 7584f1b2e48SStefano Zampini ierr = PetscMalloc3(pcbddc->benign_n,&pcbddc->benign_p0_lidx,pcbddc->benign_n,&pcbddc->benign_p0_gidx,pcbddc->benign_n,&pcbddc->benign_p0);CHKERRQ(ierr); 759339f8db1SStefano Zampini /* set change on pressures */ 7604f1b2e48SStefano Zampini for (s=0;s<pcbddc->benign_n;s++) { 7614f1b2e48SStefano Zampini PetscScalar *array; 7624f1b2e48SStefano Zampini PetscInt nzs; 7634f1b2e48SStefano Zampini 7644f1b2e48SStefano Zampini ierr = ISGetLocalSize(zerodiag_subs[s],&nzs);CHKERRQ(ierr); 7654f1b2e48SStefano Zampini ierr = ISGetIndices(zerodiag_subs[s],&idxs);CHKERRQ(ierr); 7664f1b2e48SStefano Zampini for (i=0;i<nzs-1;i++) { 767339f8db1SStefano Zampini PetscScalar vals[2]; 768339f8db1SStefano Zampini PetscInt cols[2]; 769339f8db1SStefano Zampini 770339f8db1SStefano Zampini cols[0] = idxs[i]; 7714f1b2e48SStefano Zampini cols[1] = idxs[nzs-1]; 772339f8db1SStefano Zampini vals[0] = 1.; 773b0f5fe93SStefano Zampini vals[1] = 1.; 7744f1b2e48SStefano Zampini ierr = MatSetValues(pcbddc->benign_change,1,cols,2,cols,vals,INSERT_VALUES);CHKERRQ(ierr); 775339f8db1SStefano Zampini } 7764f1b2e48SStefano Zampini ierr = PetscMalloc1(nzs,&array);CHKERRQ(ierr); 7774f1b2e48SStefano Zampini for (i=0;i<nzs-1;i++) array[i] = -1.; 7784f1b2e48SStefano Zampini array[nzs-1] = 1.; 7794f1b2e48SStefano Zampini ierr = MatSetValues(pcbddc->benign_change,1,idxs+nzs-1,nzs,idxs,array,INSERT_VALUES);CHKERRQ(ierr); 7804f1b2e48SStefano Zampini /* store local idxs for p0 */ 7814f1b2e48SStefano Zampini pcbddc->benign_p0_lidx[s] = idxs[nzs-1]; 7824f1b2e48SStefano Zampini ierr = ISRestoreIndices(zerodiag_subs[s],&idxs);CHKERRQ(ierr); 783339f8db1SStefano Zampini ierr = PetscFree(array);CHKERRQ(ierr); 7844f1b2e48SStefano Zampini } 785339f8db1SStefano Zampini ierr = MatAssemblyBegin(pcbddc->benign_change,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 786339f8db1SStefano Zampini ierr = MatAssemblyEnd(pcbddc->benign_change,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 787a3df083aSStefano Zampini /* project if needed */ 788a3df083aSStefano Zampini if (pcbddc->benign_change_explicit) { 78997d764eeSStefano Zampini ierr = MatPtAP(pcbddc->local_mat,pcbddc->benign_change,MAT_INITIAL_MATRIX,2.0,&pcbddc->benign_original_mat);CHKERRQ(ierr); 790339f8db1SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 79197d764eeSStefano Zampini ierr = MatSeqAIJCompress(pcbddc->benign_original_mat,&pcbddc->local_mat);CHKERRQ(ierr); 79297d764eeSStefano Zampini ierr = MatDestroy(&pcbddc->benign_original_mat);CHKERRQ(ierr); 793a3df083aSStefano Zampini } 7944f1b2e48SStefano Zampini /* store global idxs for p0 */ 7954f1b2e48SStefano Zampini ierr = ISLocalToGlobalMappingApply(pc->pmat->rmap->mapping,pcbddc->benign_n,pcbddc->benign_p0_lidx,pcbddc->benign_p0_gidx);CHKERRQ(ierr); 796339f8db1SStefano Zampini } 797ca92afb2SStefano Zampini pcbddc->benign_zerodiag_subs = zerodiag_subs; 7984f1b2e48SStefano Zampini ierr = ISDestroy(&pressures);CHKERRQ(ierr); 799b0f5fe93SStefano Zampini 800b0f5fe93SStefano Zampini /* determines if the coarse solver will be singular or not */ 801b0f5fe93SStefano Zampini ierr = MPI_Allreduce(&have_null,&pcbddc->benign_null,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 802339f8db1SStefano Zampini *zerodiaglocal = zerodiag; 803339f8db1SStefano Zampini PetscFunctionReturn(0); 804339f8db1SStefano Zampini } 805339f8db1SStefano Zampini 806339f8db1SStefano Zampini #undef __FUNCT__ 807015636ebSStefano Zampini #define __FUNCT__ "PCBDDCBenignGetOrSetP0" 808015636ebSStefano Zampini PetscErrorCode PCBDDCBenignGetOrSetP0(PC pc, Vec v, PetscBool get) 809efc2fbd9SStefano Zampini { 810efc2fbd9SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 811efc2fbd9SStefano Zampini PetscErrorCode ierr; 812efc2fbd9SStefano Zampini 813efc2fbd9SStefano Zampini PetscFunctionBegin; 814efc2fbd9SStefano Zampini if (!pcbddc->benign_sf) { 815efc2fbd9SStefano Zampini ierr = PetscSFCreate(PetscObjectComm((PetscObject)pc),&pcbddc->benign_sf);CHKERRQ(ierr); 8164f1b2e48SStefano Zampini ierr = PetscSFSetGraphLayout(pcbddc->benign_sf,pc->pmat->rmap,pcbddc->benign_n,NULL,PETSC_OWN_POINTER,pcbddc->benign_p0_gidx);CHKERRQ(ierr); 817efc2fbd9SStefano Zampini } 818015636ebSStefano Zampini if (get) { /* use SF to get values */ 819efc2fbd9SStefano Zampini PetscScalar *array; 820efc2fbd9SStefano Zampini 821efc2fbd9SStefano Zampini ierr = VecGetArrayRead(v,(const PetscScalar**)&array);CHKERRQ(ierr); 8224f1b2e48SStefano Zampini ierr = PetscSFBcastBegin(pcbddc->benign_sf,MPIU_SCALAR,array,pcbddc->benign_p0);CHKERRQ(ierr); 8234f1b2e48SStefano Zampini ierr = PetscSFBcastEnd(pcbddc->benign_sf,MPIU_SCALAR,array,pcbddc->benign_p0);CHKERRQ(ierr); 824efc2fbd9SStefano Zampini ierr = VecRestoreArrayRead(v,(const PetscScalar**)&array);CHKERRQ(ierr); 8254f1b2e48SStefano Zampini } else { /* use VecSetValues (not scalable, I should try to find a better solution (defining a new MPI_OP for reduction) */ 8264f1b2e48SStefano Zampini ierr = VecSetValues(v,pcbddc->benign_n,pcbddc->benign_p0_gidx,pcbddc->benign_p0,INSERT_VALUES);CHKERRQ(ierr); 827efc2fbd9SStefano Zampini ierr = VecAssemblyBegin(v);CHKERRQ(ierr); 828efc2fbd9SStefano Zampini ierr = VecAssemblyEnd(v);CHKERRQ(ierr); 829efc2fbd9SStefano Zampini } 830efc2fbd9SStefano Zampini PetscFunctionReturn(0); 831efc2fbd9SStefano Zampini } 832efc2fbd9SStefano Zampini 833efc2fbd9SStefano Zampini #undef __FUNCT__ 834c263805aSStefano Zampini #define __FUNCT__ "PCBDDCBenignPopOrPushB0" 835c263805aSStefano Zampini PetscErrorCode PCBDDCBenignPopOrPushB0(PC pc, PetscBool pop) 836c263805aSStefano Zampini { 837c263805aSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 838c263805aSStefano Zampini PetscErrorCode ierr; 839c263805aSStefano Zampini 840c263805aSStefano Zampini PetscFunctionBegin; 841c263805aSStefano Zampini /* TODO: add error checking 842c263805aSStefano Zampini - avoid nested pop (or push) calls. 843c263805aSStefano Zampini - cannot push before pop. 8441c604dc7SStefano Zampini - cannot call this if pcbddc->local_mat is NULL 845c263805aSStefano Zampini */ 8464f1b2e48SStefano Zampini if (!pcbddc->benign_n) { 847efc2fbd9SStefano Zampini PetscFunctionReturn(0); 848efc2fbd9SStefano Zampini } 849c263805aSStefano Zampini if (pop) { 850a3df083aSStefano Zampini if (pcbddc->benign_change_explicit) { 8514f1b2e48SStefano Zampini IS is_p0; 8524f1b2e48SStefano Zampini MatReuse reuse; 853c263805aSStefano Zampini 854c263805aSStefano Zampini /* extract B_0 */ 8554f1b2e48SStefano Zampini reuse = MAT_INITIAL_MATRIX; 8564f1b2e48SStefano Zampini if (pcbddc->benign_B0) { 8574f1b2e48SStefano Zampini reuse = MAT_REUSE_MATRIX; 8584f1b2e48SStefano Zampini } 8594f1b2e48SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,pcbddc->benign_n,pcbddc->benign_p0_lidx,PETSC_COPY_VALUES,&is_p0);CHKERRQ(ierr); 8604f1b2e48SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,is_p0,NULL,reuse,&pcbddc->benign_B0);CHKERRQ(ierr); 861c263805aSStefano Zampini /* remove rows and cols from local problem */ 862c263805aSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_KEEP_NONZERO_PATTERN,PETSC_TRUE);CHKERRQ(ierr); 86397d764eeSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_FALSE);CHKERRQ(ierr); 8644f1b2e48SStefano Zampini ierr = MatZeroRowsColumnsIS(pcbddc->local_mat,is_p0,1.0,NULL,NULL);CHKERRQ(ierr); 8654f1b2e48SStefano Zampini ierr = ISDestroy(&is_p0);CHKERRQ(ierr); 866a3df083aSStefano Zampini } else { 867a3df083aSStefano Zampini Mat_IS *matis = (Mat_IS*)pc->pmat->data; 868a3df083aSStefano Zampini PetscScalar *vals; 869a3df083aSStefano Zampini PetscInt i,n,*idxs_ins; 870a3df083aSStefano Zampini 871a3df083aSStefano Zampini ierr = VecGetLocalSize(matis->y,&n);CHKERRQ(ierr); 872a3df083aSStefano Zampini ierr = PetscMalloc2(n,&idxs_ins,n,&vals);CHKERRQ(ierr); 873a3df083aSStefano Zampini if (!pcbddc->benign_B0) { 8740b5adadeSStefano Zampini PetscInt *nnz; 875a3df083aSStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)pcbddc->local_mat),&pcbddc->benign_B0);CHKERRQ(ierr); 876a3df083aSStefano Zampini ierr = MatSetType(pcbddc->benign_B0,MATAIJ);CHKERRQ(ierr); 877a3df083aSStefano Zampini ierr = MatSetSizes(pcbddc->benign_B0,pcbddc->benign_n,n,PETSC_DECIDE,PETSC_DECIDE);CHKERRQ(ierr); 878331e053bSStefano Zampini ierr = PetscMalloc1(pcbddc->benign_n,&nnz);CHKERRQ(ierr); 879331e053bSStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 880331e053bSStefano Zampini ierr = ISGetLocalSize(pcbddc->benign_zerodiag_subs[i],&nnz[i]);CHKERRQ(ierr); 881331e053bSStefano Zampini nnz[i] = n - nnz[i]; 882331e053bSStefano Zampini } 883331e053bSStefano Zampini ierr = MatSeqAIJSetPreallocation(pcbddc->benign_B0,0,nnz);CHKERRQ(ierr); 884331e053bSStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 885331e053bSStefano Zampini } 886a3df083aSStefano Zampini 887a3df083aSStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 888a3df083aSStefano Zampini PetscScalar *array; 889a3df083aSStefano Zampini PetscInt *idxs,j,nz,cum; 890a3df083aSStefano Zampini 891a3df083aSStefano Zampini ierr = VecSet(matis->x,0.);CHKERRQ(ierr); 892a3df083aSStefano Zampini ierr = ISGetLocalSize(pcbddc->benign_zerodiag_subs[i],&nz);CHKERRQ(ierr); 893a3df083aSStefano Zampini ierr = ISGetIndices(pcbddc->benign_zerodiag_subs[i],(const PetscInt**)&idxs);CHKERRQ(ierr); 894a3df083aSStefano Zampini for (j=0;j<nz;j++) vals[j] = 1.; 895a3df083aSStefano Zampini ierr = VecSetValues(matis->x,nz,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 896a3df083aSStefano Zampini ierr = VecAssemblyBegin(matis->x);CHKERRQ(ierr); 897a3df083aSStefano Zampini ierr = VecAssemblyEnd(matis->x);CHKERRQ(ierr); 898a3df083aSStefano Zampini ierr = VecSet(matis->y,0.);CHKERRQ(ierr); 899a3df083aSStefano Zampini ierr = MatMult(matis->A,matis->x,matis->y);CHKERRQ(ierr); 900a3df083aSStefano Zampini ierr = VecGetArray(matis->y,&array);CHKERRQ(ierr); 901a3df083aSStefano Zampini cum = 0; 902a3df083aSStefano Zampini for (j=0;j<n;j++) { 903a3df083aSStefano Zampini if (PetscUnlikely(PetscAbsReal(array[j]) > PETSC_SMALL)) { 904a3df083aSStefano Zampini vals[cum] = array[j]; 905a3df083aSStefano Zampini idxs_ins[cum] = j; 906a3df083aSStefano Zampini cum++; 907a3df083aSStefano Zampini } 908a3df083aSStefano Zampini } 909a3df083aSStefano Zampini ierr = MatSetValues(pcbddc->benign_B0,1,&i,cum,idxs_ins,vals,INSERT_VALUES);CHKERRQ(ierr); 910a3df083aSStefano Zampini ierr = VecRestoreArray(matis->y,&array);CHKERRQ(ierr); 911a3df083aSStefano Zampini ierr = ISRestoreIndices(pcbddc->benign_zerodiag_subs[i],(const PetscInt**)&idxs);CHKERRQ(ierr); 912a3df083aSStefano Zampini } 913a3df083aSStefano Zampini ierr = MatAssemblyBegin(pcbddc->benign_B0,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 914a3df083aSStefano Zampini ierr = MatAssemblyEnd(pcbddc->benign_B0,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 915a3df083aSStefano Zampini ierr = PetscFree2(idxs_ins,vals);CHKERRQ(ierr); 916a3df083aSStefano Zampini } 917c263805aSStefano Zampini } else { /* push */ 918a3df083aSStefano Zampini if (pcbddc->benign_change_explicit) { 9194f1b2e48SStefano Zampini PetscInt i; 9204f1b2e48SStefano Zampini 9214f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 9224f1b2e48SStefano Zampini PetscScalar *B0_vals; 9234f1b2e48SStefano Zampini PetscInt *B0_cols,B0_ncol; 9244f1b2e48SStefano Zampini 9254f1b2e48SStefano Zampini ierr = MatGetRow(pcbddc->benign_B0,i,&B0_ncol,(const PetscInt**)&B0_cols,(const PetscScalar**)&B0_vals);CHKERRQ(ierr); 9264f1b2e48SStefano Zampini ierr = MatSetValues(pcbddc->local_mat,1,pcbddc->benign_p0_lidx+i,B0_ncol,B0_cols,B0_vals,INSERT_VALUES);CHKERRQ(ierr); 9277b034428SStefano Zampini ierr = MatSetValues(pcbddc->local_mat,B0_ncol,B0_cols,1,pcbddc->benign_p0_lidx+i,B0_vals,INSERT_VALUES);CHKERRQ(ierr); 9284f1b2e48SStefano Zampini ierr = MatSetValue(pcbddc->local_mat,pcbddc->benign_p0_lidx[i],pcbddc->benign_p0_lidx[i],0.0,INSERT_VALUES);CHKERRQ(ierr); 9294f1b2e48SStefano Zampini ierr = MatRestoreRow(pcbddc->benign_B0,i,&B0_ncol,(const PetscInt**)&B0_cols,(const PetscScalar**)&B0_vals);CHKERRQ(ierr); 9304f1b2e48SStefano Zampini } 931c263805aSStefano Zampini ierr = MatAssemblyBegin(pcbddc->local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 932c263805aSStefano Zampini ierr = MatAssemblyEnd(pcbddc->local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 933a3df083aSStefano Zampini } else { 934a3df083aSStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Cannot push B0!\n"); 935a3df083aSStefano Zampini } 936c263805aSStefano Zampini } 937c263805aSStefano Zampini PetscFunctionReturn(0); 938c263805aSStefano Zampini } 939c263805aSStefano Zampini 940c263805aSStefano Zampini #undef __FUNCT__ 941b1b3d7a2SStefano Zampini #define __FUNCT__ "PCBDDCAdaptiveSelection" 94208122e43SStefano Zampini PetscErrorCode PCBDDCAdaptiveSelection(PC pc) 943b1b3d7a2SStefano Zampini { 944b1b3d7a2SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 94508122e43SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 94608122e43SStefano Zampini PetscBLASInt B_dummyint,B_neigs,B_ierr,B_lwork; 94708122e43SStefano Zampini PetscBLASInt *B_iwork,*B_ifail; 94808122e43SStefano Zampini PetscScalar *work,lwork; 94908122e43SStefano Zampini PetscScalar *St,*S,*eigv; 95008122e43SStefano Zampini PetscScalar *Sarray,*Starray; 95108122e43SStefano Zampini PetscReal *eigs,thresh; 9521b968477SStefano Zampini PetscInt i,nmax,nmin,nv,cum,mss,cum2,cumarray,maxneigs; 953f6f667cfSStefano Zampini PetscBool allocated_S_St; 95408122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 95508122e43SStefano Zampini PetscReal *rwork; 95608122e43SStefano Zampini #endif 957b1b3d7a2SStefano Zampini PetscErrorCode ierr; 958b1b3d7a2SStefano Zampini 959b1b3d7a2SStefano Zampini PetscFunctionBegin; 960df4d28bfSStefano Zampini if (!sub_schurs->schur_explicit) { 961df4d28bfSStefano Zampini SETERRQ(PetscObjectComm((PetscObject)pc),PETSC_ERR_SUP,"Adaptive selection of constraints requires MUMPS and/or MKL_CPARDISO"); 96208122e43SStefano Zampini } 96308122e43SStefano Zampini 96406a4e24aSStefano Zampini if (sub_schurs->n_subs && (!sub_schurs->is_hermitian || !sub_schurs->is_posdef)) { 96506a4e24aSStefano Zampini SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_SUP,"Adaptive selection not yet implemented for general matrix pencils (herm %d, posdef %d)\n",sub_schurs->is_hermitian,sub_schurs->is_posdef); 96606a4e24aSStefano Zampini } 96706a4e24aSStefano Zampini 968fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 969fd14bc51SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 970fd14bc51SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 971fd14bc51SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Check adaptive selection of constraints\n");CHKERRQ(ierr); 9721575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 973fd14bc51SStefano Zampini } 974fd14bc51SStefano Zampini 975e496cd5dSStefano Zampini if (pcbddc->dbg_flag) { 976e496cd5dSStefano Zampini PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d cc %d (%d,%d).\n",PetscGlobalRank,sub_schurs->n_subs,sub_schurs->is_hermitian,sub_schurs->is_posdef); 977e496cd5dSStefano Zampini } 978e496cd5dSStefano Zampini 97908122e43SStefano Zampini /* max size of subsets */ 98008122e43SStefano Zampini mss = 0; 98108122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 98208122e43SStefano Zampini PetscInt subset_size; 983862806e4SStefano Zampini 98408122e43SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 98508122e43SStefano Zampini mss = PetscMax(mss,subset_size); 98608122e43SStefano Zampini } 98708122e43SStefano Zampini 98808122e43SStefano Zampini /* min/max and threshold */ 98908122e43SStefano Zampini nmax = pcbddc->adaptive_nmax > 0 ? pcbddc->adaptive_nmax : mss; 990f6f667cfSStefano Zampini nmin = pcbddc->adaptive_nmin > 0 ? pcbddc->adaptive_nmin : 0; 99108122e43SStefano Zampini nmax = PetscMax(nmin,nmax); 992f6f667cfSStefano Zampini allocated_S_St = PETSC_FALSE; 993f6f667cfSStefano Zampini if (nmin) { 994f6f667cfSStefano Zampini allocated_S_St = PETSC_TRUE; 995f6f667cfSStefano Zampini } 99608122e43SStefano Zampini 99708122e43SStefano Zampini /* allocate lapack workspace */ 99808122e43SStefano Zampini cum = cum2 = 0; 99908122e43SStefano Zampini maxneigs = 0; 100008122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 100108122e43SStefano Zampini PetscInt n,subset_size; 1002f6f667cfSStefano Zampini 100308122e43SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 100408122e43SStefano Zampini n = PetscMin(subset_size,nmax); 10059162d606SStefano Zampini cum += subset_size; 10069162d606SStefano Zampini cum2 += subset_size*n; 100708122e43SStefano Zampini maxneigs = PetscMax(maxneigs,n); 100808122e43SStefano Zampini } 100908122e43SStefano Zampini if (mss) { 10109ab7bb16SStefano Zampini if (sub_schurs->is_hermitian && sub_schurs->is_posdef) { 101108122e43SStefano Zampini PetscBLASInt B_itype = 1; 101208122e43SStefano Zampini PetscBLASInt B_N = mss; 10134c6709b3SStefano Zampini PetscReal zero = 0.0; 10144c6709b3SStefano Zampini PetscReal eps = 0.0; /* dlamch? */ 101508122e43SStefano Zampini 101608122e43SStefano Zampini B_lwork = -1; 101708122e43SStefano Zampini S = NULL; 101808122e43SStefano Zampini St = NULL; 1019a58a30b4SStefano Zampini eigs = NULL; 1020a58a30b4SStefano Zampini eigv = NULL; 1021a58a30b4SStefano Zampini B_iwork = NULL; 1022a58a30b4SStefano Zampini B_ifail = NULL; 1023d1710679SStefano Zampini #if defined(PETSC_USE_COMPLEX) 1024d1710679SStefano Zampini rwork = NULL; 1025d1710679SStefano Zampini #endif 10268bec7fa6SStefano Zampini thresh = 1.0; 102708122e43SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 102808122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 102908122e43SStefano Zampini PetscStackCallBLAS("LAPACKsygvx",LAPACKsygvx_(&B_itype,"V","V","L",&B_N,St,&B_N,S,&B_N,&zero,&thresh,&B_dummyint,&B_dummyint,&eps,&B_neigs,eigs,eigv,&B_N,&lwork,&B_lwork,rwork,B_iwork,B_ifail,&B_ierr)); 103008122e43SStefano Zampini #else 103108122e43SStefano Zampini PetscStackCallBLAS("LAPACKsygvx",LAPACKsygvx_(&B_itype,"V","V","L",&B_N,St,&B_N,S,&B_N,&zero,&thresh,&B_dummyint,&B_dummyint,&eps,&B_neigs,eigs,eigv,&B_N,&lwork,&B_lwork,B_iwork,B_ifail,&B_ierr)); 103208122e43SStefano Zampini #endif 103308122e43SStefano Zampini if (B_ierr != 0) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to SYGVX Lapack routine %d",(int)B_ierr); 103408122e43SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 103508122e43SStefano Zampini } else { 103608122e43SStefano Zampini /* TODO */ 103708122e43SStefano Zampini } 103808122e43SStefano Zampini } else { 103908122e43SStefano Zampini lwork = 0; 104008122e43SStefano Zampini } 104108122e43SStefano Zampini 104208122e43SStefano Zampini nv = 0; 1043d62866d3SStefano Zampini if (sub_schurs->is_vertices && pcbddc->use_vertices) { /* complement set of active subsets, each entry is a vertex (boundary made by active subsets, vertices and dirichlet dofs) */ 1044d62866d3SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_vertices,&nv);CHKERRQ(ierr); 104508122e43SStefano Zampini } 10464c6709b3SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lwork),&B_lwork);CHKERRQ(ierr); 1047f6f667cfSStefano Zampini if (allocated_S_St) { 1048f6f667cfSStefano Zampini ierr = PetscMalloc2(mss*mss,&S,mss*mss,&St);CHKERRQ(ierr); 1049f6f667cfSStefano Zampini } 1050f6f667cfSStefano Zampini ierr = PetscMalloc5(mss*mss,&eigv,mss,&eigs,B_lwork,&work,5*mss,&B_iwork,mss,&B_ifail);CHKERRQ(ierr); 105108122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 105208122e43SStefano Zampini ierr = PetscMalloc1(7*mss,&rwork);CHKERRQ(ierr); 105308122e43SStefano Zampini #endif 10549162d606SStefano Zampini ierr = PetscMalloc5(nv+sub_schurs->n_subs,&pcbddc->adaptive_constraints_n, 10559162d606SStefano Zampini nv+sub_schurs->n_subs+1,&pcbddc->adaptive_constraints_idxs_ptr, 10569162d606SStefano Zampini nv+sub_schurs->n_subs+1,&pcbddc->adaptive_constraints_data_ptr, 105708122e43SStefano Zampini nv+cum,&pcbddc->adaptive_constraints_idxs, 10589162d606SStefano Zampini nv+cum2,&pcbddc->adaptive_constraints_data);CHKERRQ(ierr); 105908122e43SStefano Zampini ierr = PetscMemzero(pcbddc->adaptive_constraints_n,(nv+sub_schurs->n_subs)*sizeof(PetscInt));CHKERRQ(ierr); 106008122e43SStefano Zampini 106108122e43SStefano Zampini maxneigs = 0; 106272b8c272SStefano Zampini cum = cumarray = 0; 10639162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr[0] = 0; 10649162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr[0] = 0; 1065d62866d3SStefano Zampini if (sub_schurs->is_vertices && pcbddc->use_vertices) { 106608122e43SStefano Zampini const PetscInt *idxs; 106708122e43SStefano Zampini 1068d62866d3SStefano Zampini ierr = ISGetIndices(sub_schurs->is_vertices,&idxs);CHKERRQ(ierr); 106908122e43SStefano Zampini for (cum=0;cum<nv;cum++) { 107008122e43SStefano Zampini pcbddc->adaptive_constraints_n[cum] = 1; 107108122e43SStefano Zampini pcbddc->adaptive_constraints_idxs[cum] = idxs[cum]; 107208122e43SStefano Zampini pcbddc->adaptive_constraints_data[cum] = 1.0; 10739162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr[cum+1] = pcbddc->adaptive_constraints_idxs_ptr[cum]+1; 10749162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr[cum+1] = pcbddc->adaptive_constraints_data_ptr[cum]+1; 107508122e43SStefano Zampini } 1076d62866d3SStefano Zampini ierr = ISRestoreIndices(sub_schurs->is_vertices,&idxs);CHKERRQ(ierr); 107708122e43SStefano Zampini } 107808122e43SStefano Zampini 107908122e43SStefano Zampini if (mss) { /* multilevel */ 108008122e43SStefano Zampini ierr = MatSeqAIJGetArray(sub_schurs->sum_S_Ej_inv_all,&Sarray);CHKERRQ(ierr); 108108122e43SStefano Zampini ierr = MatSeqAIJGetArray(sub_schurs->sum_S_Ej_tilda_all,&Starray);CHKERRQ(ierr); 108208122e43SStefano Zampini } 108308122e43SStefano Zampini 1084ffd830a3SStefano Zampini thresh = pcbddc->adaptive_threshold; 108508122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 108608122e43SStefano Zampini const PetscInt *idxs; 1087f6f667cfSStefano Zampini PetscReal infty = PETSC_MAX_REAL; 1088862806e4SStefano Zampini PetscInt j,subset_size,eigs_start = 0; 108908122e43SStefano Zampini PetscBLASInt B_N; 1090aff50787SStefano Zampini PetscBool same_data = PETSC_FALSE; 109108122e43SStefano Zampini 1092862806e4SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 1093ffd830a3SStefano Zampini ierr = ISGetIndices(sub_schurs->is_subs[i],&idxs);CHKERRQ(ierr); 1094f6f667cfSStefano Zampini ierr = PetscBLASIntCast(subset_size,&B_N);CHKERRQ(ierr); 1095f6f667cfSStefano Zampini if (allocated_S_St) { /* S and S_t should be copied since we could need them later */ 10969ab7bb16SStefano Zampini if (sub_schurs->is_hermitian) { 1097aff50787SStefano Zampini PetscInt j,k; 1098aff50787SStefano Zampini if (sub_schurs->n_subs == 1) { /* zeroing memory to use PetscMemcmp later */ 1099aff50787SStefano Zampini ierr = PetscMemzero(S,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 1100aff50787SStefano Zampini ierr = PetscMemzero(St,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 110108122e43SStefano Zampini } 110208122e43SStefano Zampini for (j=0;j<subset_size;j++) { 1103aff50787SStefano Zampini for (k=j;k<subset_size;k++) { 1104aff50787SStefano Zampini S [j*subset_size+k] = Sarray [cumarray+j*subset_size+k]; 1105aff50787SStefano Zampini St[j*subset_size+k] = Starray[cumarray+j*subset_size+k]; 1106aff50787SStefano Zampini } 110708122e43SStefano Zampini } 110808122e43SStefano Zampini } else { 110908122e43SStefano Zampini ierr = PetscMemcpy(S,Sarray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 111008122e43SStefano Zampini ierr = PetscMemcpy(St,Starray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 111108122e43SStefano Zampini } 11128bec7fa6SStefano Zampini } else { 1113f6f667cfSStefano Zampini S = Sarray + cumarray; 1114f6f667cfSStefano Zampini St = Starray + cumarray; 11158bec7fa6SStefano Zampini } 1116aff50787SStefano Zampini /* see if we can save some work */ 1117aff50787SStefano Zampini if (sub_schurs->n_subs == 1) { 1118aff50787SStefano Zampini ierr = PetscMemcmp(S,St,subset_size*subset_size*sizeof(PetscScalar),&same_data);CHKERRQ(ierr); 1119aff50787SStefano Zampini } 1120aff50787SStefano Zampini 1121aff50787SStefano Zampini if (same_data) { /* there's no need of constraints here, deluxe scaling is enough */ 1122aff50787SStefano Zampini B_neigs = 0; 1123aff50787SStefano Zampini } else { 11249ab7bb16SStefano Zampini if (sub_schurs->is_hermitian && sub_schurs->is_posdef) { 112508122e43SStefano Zampini PetscBLASInt B_itype = 1; 1126f6f667cfSStefano Zampini PetscBLASInt B_IL, B_IU; 11274c6709b3SStefano Zampini PetscReal eps = -1.0; /* dlamch? */ 11289552c7c7SStefano Zampini PetscInt nmin_s; 112908122e43SStefano Zampini 1130fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 11318bec7fa6SStefano Zampini PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Computing for sub %d/%d %d %d.\n",i,sub_schurs->n_subs,subset_size,pcbddc->mat_graph->count[idxs[0]]); 1132fd14bc51SStefano Zampini } 1133d16cbb6bSStefano Zampini 113408122e43SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 1135d16cbb6bSStefano Zampini if (thresh > 1.+PETSC_SMALL) { 1136d16cbb6bSStefano Zampini 1137d16cbb6bSStefano Zampini /* ask for eigenvalues larger than thresh */ 113808122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 1139f6f667cfSStefano Zampini PetscStackCallBLAS("LAPACKsygvx",LAPACKsygvx_(&B_itype,"V","V","L",&B_N,St,&B_N,S,&B_N,&thresh,&infty,&B_IL,&B_IU,&eps,&B_neigs,eigs,eigv,&B_N,work,&B_lwork,rwork,B_iwork,B_ifail,&B_ierr)); 114008122e43SStefano Zampini #else 1141f6f667cfSStefano Zampini PetscStackCallBLAS("LAPACKsygvx",LAPACKsygvx_(&B_itype,"V","V","L",&B_N,St,&B_N,S,&B_N,&thresh,&infty,&B_IL,&B_IU,&eps,&B_neigs,eigs,eigv,&B_N,work,&B_lwork,B_iwork,B_ifail,&B_ierr)); 114208122e43SStefano Zampini #endif 1143d16cbb6bSStefano Zampini } else { 1144d16cbb6bSStefano Zampini B_IU = PetscMax(1,PetscMin(B_N,nmax)); 1145d16cbb6bSStefano Zampini B_IL = 1; 1146d16cbb6bSStefano Zampini #if defined(PETSC_USE_COMPLEX) 1147d16cbb6bSStefano Zampini PetscStackCallBLAS("LAPACKsygvx",LAPACKsygvx_(&B_itype,"V","I","L",&B_N,St,&B_N,S,&B_N,&thresh,&infty,&B_IL,&B_IU,&eps,&B_neigs,eigs,eigv,&B_N,work,&B_lwork,rwork,B_iwork,B_ifail,&B_ierr)); 1148d16cbb6bSStefano Zampini #else 1149d16cbb6bSStefano Zampini PetscStackCallBLAS("LAPACKsygvx",LAPACKsygvx_(&B_itype,"V","I","L",&B_N,St,&B_N,S,&B_N,&thresh,&infty,&B_IL,&B_IU,&eps,&B_neigs,eigs,eigv,&B_N,work,&B_lwork,B_iwork,B_ifail,&B_ierr)); 1150d16cbb6bSStefano Zampini #endif 1151d16cbb6bSStefano Zampini } 115208122e43SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 115308122e43SStefano Zampini if (B_ierr) { 115408122e43SStefano Zampini if (B_ierr < 0 ) { 115508122e43SStefano Zampini SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYGVX Lapack routine: illegal value for argument %d",-(int)B_ierr); 115608122e43SStefano Zampini } else if (B_ierr <= B_N) { 115708122e43SStefano Zampini SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYGVX Lapack routine: %d eigenvalues failed to converge",(int)B_ierr); 115808122e43SStefano Zampini } else { 11599552c7c7SStefano Zampini SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYGVX Lapack routine: leading minor of order %d is not positive definite",(int)B_ierr-B_N-1); 116008122e43SStefano Zampini } 116108122e43SStefano Zampini } 116208122e43SStefano Zampini 116308122e43SStefano Zampini if (B_neigs > nmax) { 1164fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 1165fd14bc51SStefano Zampini PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," found %d eigs, more than maximum required %d.\n",B_neigs,nmax); 1166fd14bc51SStefano Zampini } 1167f6f667cfSStefano Zampini eigs_start = B_neigs -nmax; 116808122e43SStefano Zampini B_neigs = nmax; 116908122e43SStefano Zampini } 117008122e43SStefano Zampini 11719552c7c7SStefano Zampini nmin_s = PetscMin(nmin,B_N); 11729552c7c7SStefano Zampini if (B_neigs < nmin_s) { 117308122e43SStefano Zampini PetscBLASInt B_neigs2; 117408122e43SStefano Zampini 1175f6f667cfSStefano Zampini B_IU = B_N - B_neigs; 1176f6f667cfSStefano Zampini B_IL = B_N - nmin_s + 1; 1177fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 1178fd14bc51SStefano Zampini PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," found %d eigs, less than minimum required %d. Asking for %d to %d incl (fortran like)\n",B_neigs,nmin,B_IL,B_IU); 1179fd14bc51SStefano Zampini } 11809ab7bb16SStefano Zampini if (sub_schurs->is_hermitian) { 118108122e43SStefano Zampini PetscInt j; 118208122e43SStefano Zampini for (j=0;j<subset_size;j++) { 118308122e43SStefano Zampini ierr = PetscMemcpy(S+j*(subset_size+1),Sarray+cumarray+j*(subset_size+1),(subset_size-j)*sizeof(PetscScalar));CHKERRQ(ierr); 118408122e43SStefano Zampini } 118508122e43SStefano Zampini for (j=0;j<subset_size;j++) { 118608122e43SStefano Zampini ierr = PetscMemcpy(St+j*(subset_size+1),Starray+cumarray+j*(subset_size+1),(subset_size-j)*sizeof(PetscScalar));CHKERRQ(ierr); 118708122e43SStefano Zampini } 118808122e43SStefano Zampini } else { 118908122e43SStefano Zampini ierr = PetscMemcpy(S,Sarray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 119008122e43SStefano Zampini ierr = PetscMemcpy(St,Starray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 119108122e43SStefano Zampini } 119208122e43SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 119308122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 1194f6f667cfSStefano Zampini PetscStackCallBLAS("LAPACKsygvx",LAPACKsygvx_(&B_itype,"V","I","L",&B_N,St,&B_N,S,&B_N,&thresh,&infty,&B_IL,&B_IU,&eps,&B_neigs2,eigs+B_neigs,eigv+B_neigs*subset_size,&B_N,work,&B_lwork,rwork,B_iwork,B_ifail,&B_ierr)); 119508122e43SStefano Zampini #else 1196f6f667cfSStefano Zampini PetscStackCallBLAS("LAPACKsygvx",LAPACKsygvx_(&B_itype,"V","I","L",&B_N,St,&B_N,S,&B_N,&thresh,&infty,&B_IL,&B_IU,&eps,&B_neigs2,eigs+B_neigs,eigv+B_neigs*subset_size,&B_N,work,&B_lwork,B_iwork,B_ifail,&B_ierr)); 119708122e43SStefano Zampini #endif 119808122e43SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 119908122e43SStefano Zampini B_neigs += B_neigs2; 120008122e43SStefano Zampini } 120108122e43SStefano Zampini if (B_ierr) { 120208122e43SStefano Zampini if (B_ierr < 0 ) { 120308122e43SStefano Zampini SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYGVX Lapack routine: illegal value for argument %d",-(int)B_ierr); 120408122e43SStefano Zampini } else if (B_ierr <= B_N) { 120508122e43SStefano Zampini SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYGVX Lapack routine: %d eigenvalues failed to converge",(int)B_ierr); 120608122e43SStefano Zampini } else { 12079552c7c7SStefano Zampini SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYGVX Lapack routine: leading minor of order %d is not positive definite",(int)B_ierr-B_N-1); 120808122e43SStefano Zampini } 120908122e43SStefano Zampini } 1210fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 1211ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," -> Got %d eigs\n",B_neigs);CHKERRQ(ierr); 121208122e43SStefano Zampini for (j=0;j<B_neigs;j++) { 121308122e43SStefano Zampini if (eigs[j] == 0.0) { 1214ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," Inf\n");CHKERRQ(ierr); 121508122e43SStefano Zampini } else { 1216ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," %1.6e\n",eigs[j+eigs_start]);CHKERRQ(ierr); 1217fd14bc51SStefano Zampini } 121808122e43SStefano Zampini } 121908122e43SStefano Zampini } 122008122e43SStefano Zampini } else { 122108122e43SStefano Zampini /* TODO */ 122208122e43SStefano Zampini } 1223aff50787SStefano Zampini } 12246c3e6151SStefano Zampini /* change the basis back to the original one */ 12256c3e6151SStefano Zampini if (sub_schurs->change) { 122672b8c272SStefano Zampini Mat change,phi,phit; 12276c3e6151SStefano Zampini 12286c3e6151SStefano Zampini if (pcbddc->dbg_flag > 1) { 12296c3e6151SStefano Zampini PetscInt ii; 12306c3e6151SStefano Zampini for (ii=0;ii<B_neigs;ii++) { 12316c3e6151SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," -> Eigenvector (old basis) %d/%d (%d)\n",ii,B_neigs,B_N);CHKERRQ(ierr); 12326c3e6151SStefano Zampini for (j=0;j<B_N;j++) { 12336c3e6151SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," %1.4e\n",eigv[(ii+eigs_start)*subset_size+j]);CHKERRQ(ierr); 12346c3e6151SStefano Zampini } 12356c3e6151SStefano Zampini } 12366c3e6151SStefano Zampini } 123772b8c272SStefano Zampini ierr = KSPGetOperators(sub_schurs->change[i],&change,NULL);CHKERRQ(ierr); 12386c3e6151SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,subset_size,B_neigs,eigv+eigs_start*subset_size,&phit);CHKERRQ(ierr); 123972b8c272SStefano Zampini ierr = MatMatMult(change,phit,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&phi);CHKERRQ(ierr); 12406c3e6151SStefano Zampini ierr = MatCopy(phi,phit,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 12416c3e6151SStefano Zampini ierr = MatDestroy(&phit);CHKERRQ(ierr); 12426c3e6151SStefano Zampini ierr = MatDestroy(&phi);CHKERRQ(ierr); 12436c3e6151SStefano Zampini } 12448bec7fa6SStefano Zampini maxneigs = PetscMax(B_neigs,maxneigs); 12458bec7fa6SStefano Zampini pcbddc->adaptive_constraints_n[i+nv] = B_neigs; 12469162d606SStefano Zampini if (B_neigs) { 12479162d606SStefano Zampini ierr = PetscMemcpy(pcbddc->adaptive_constraints_data+pcbddc->adaptive_constraints_data_ptr[cum],eigv+eigs_start*subset_size,B_neigs*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 1248fd14bc51SStefano Zampini 1249fd14bc51SStefano Zampini if (pcbddc->dbg_flag > 1) { 12509552c7c7SStefano Zampini PetscInt ii; 12519552c7c7SStefano Zampini for (ii=0;ii<B_neigs;ii++) { 1252ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," -> Eigenvector %d/%d (%d)\n",ii,B_neigs,B_N);CHKERRQ(ierr); 12539552c7c7SStefano Zampini for (j=0;j<B_N;j++) { 1254ac47001eSStefano Zampini #if defined(PETSC_USE_COMPLEX) 1255ac47001eSStefano Zampini PetscReal r = PetscRealPart(pcbddc->adaptive_constraints_data[ii*subset_size+j+pcbddc->adaptive_constraints_data_ptr[cum]]); 1256ac47001eSStefano Zampini PetscReal c = PetscImaginaryPart(pcbddc->adaptive_constraints_data[ii*subset_size+j+pcbddc->adaptive_constraints_data_ptr[cum]]); 1257ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," %1.4e + %1.4e i\n",r,c);CHKERRQ(ierr); 1258ac47001eSStefano Zampini #else 1259ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," %1.4e\n",pcbddc->adaptive_constraints_data[ii*subset_size+j+pcbddc->adaptive_constraints_data_ptr[cum]]);CHKERRQ(ierr); 1260ac47001eSStefano Zampini #endif 12619552c7c7SStefano Zampini } 12629552c7c7SStefano Zampini } 1263fd14bc51SStefano Zampini } 12649162d606SStefano Zampini ierr = PetscMemcpy(pcbddc->adaptive_constraints_idxs+pcbddc->adaptive_constraints_idxs_ptr[cum],idxs,subset_size*sizeof(PetscInt));CHKERRQ(ierr); 12659162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr[cum+1] = pcbddc->adaptive_constraints_idxs_ptr[cum] + subset_size; 12669162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr[cum+1] = pcbddc->adaptive_constraints_data_ptr[cum] + subset_size*B_neigs; 12679162d606SStefano Zampini cum++; 126808122e43SStefano Zampini } 126908122e43SStefano Zampini ierr = ISRestoreIndices(sub_schurs->is_subs[i],&idxs);CHKERRQ(ierr); 127008122e43SStefano Zampini /* shift for next computation */ 127108122e43SStefano Zampini cumarray += subset_size*subset_size; 127208122e43SStefano Zampini } 1273fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 1274fd14bc51SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1275fd14bc51SStefano Zampini } 127608122e43SStefano Zampini 127708122e43SStefano Zampini if (mss) { 127808122e43SStefano Zampini ierr = MatSeqAIJRestoreArray(sub_schurs->sum_S_Ej_inv_all,&Sarray);CHKERRQ(ierr); 127908122e43SStefano Zampini ierr = MatSeqAIJRestoreArray(sub_schurs->sum_S_Ej_tilda_all,&Starray);CHKERRQ(ierr); 1280f6f667cfSStefano Zampini /* destroy matrices (junk) */ 1281f6f667cfSStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_inv_all);CHKERRQ(ierr); 1282f6f667cfSStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_tilda_all);CHKERRQ(ierr); 128308122e43SStefano Zampini } 1284f6f667cfSStefano Zampini if (allocated_S_St) { 1285f6f667cfSStefano Zampini ierr = PetscFree2(S,St);CHKERRQ(ierr); 1286f6f667cfSStefano Zampini } 1287f6f667cfSStefano Zampini ierr = PetscFree5(eigv,eigs,work,B_iwork,B_ifail);CHKERRQ(ierr); 128808122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 128908122e43SStefano Zampini ierr = PetscFree(rwork);CHKERRQ(ierr); 129008122e43SStefano Zampini #endif 129108122e43SStefano Zampini if (pcbddc->dbg_flag) { 12921b968477SStefano Zampini PetscInt maxneigs_r; 129308122e43SStefano Zampini ierr = MPI_Allreduce(&maxneigs,&maxneigs_r,1,MPIU_INT,MPI_MAX,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 12949b28b941SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Maximum number of constraints per cc %d\n",maxneigs_r);CHKERRQ(ierr); 129508122e43SStefano Zampini } 129608122e43SStefano Zampini PetscFunctionReturn(0); 129708122e43SStefano Zampini } 1298b1b3d7a2SStefano Zampini 1299674ae819SStefano Zampini #undef __FUNCT__ 1300c8587f34SStefano Zampini #define __FUNCT__ "PCBDDCSetUpSolvers" 1301c8587f34SStefano Zampini PetscErrorCode PCBDDCSetUpSolvers(PC pc) 1302c8587f34SStefano Zampini { 1303c8587f34SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 13048629588bSStefano Zampini PetscScalar *coarse_submat_vals; 1305c8587f34SStefano Zampini PetscErrorCode ierr; 1306c8587f34SStefano Zampini 1307c8587f34SStefano Zampini PetscFunctionBegin; 1308f4ddd8eeSStefano Zampini /* Setup local scatters R_to_B and (optionally) R_to_D */ 13095e8657edSStefano Zampini /* PCBDDCSetUpLocalWorkVectors should be called first! */ 1310c8587f34SStefano Zampini ierr = PCBDDCSetUpLocalScatters(pc);CHKERRQ(ierr); 1311c8587f34SStefano Zampini 1312684f6988SStefano Zampini /* Setup local neumann solver ksp_R */ 13130fccc4e9SStefano Zampini /* PCBDDCSetUpLocalScatters should be called first! */ 1314684f6988SStefano Zampini ierr = PCBDDCSetUpLocalSolvers(pc,PETSC_FALSE,PETSC_TRUE);CHKERRQ(ierr); 1315c8587f34SStefano Zampini 1316c8587f34SStefano Zampini /* Change global null space passed in by the user if change of basis has been requested */ 1317b9b85e73SStefano Zampini if (pcbddc->NullSpace && pcbddc->ChangeOfBasisMatrix) { 1318c8587f34SStefano Zampini ierr = PCBDDCNullSpaceAdaptGlobal(pc);CHKERRQ(ierr); 1319c8587f34SStefano Zampini } 1320c8587f34SStefano Zampini 13218629588bSStefano Zampini /* 13228629588bSStefano Zampini Setup local correction and local part of coarse basis. 13238629588bSStefano Zampini Gives back the dense local part of the coarse matrix in column major ordering 13248629588bSStefano Zampini */ 132547f4ddc3SStefano Zampini ierr = PCBDDCSetUpCorrection(pc,&coarse_submat_vals);CHKERRQ(ierr); 13268629588bSStefano Zampini 13278629588bSStefano Zampini /* Compute total number of coarse nodes and setup coarse solver */ 13288629588bSStefano Zampini ierr = PCBDDCSetUpCoarseSolver(pc,coarse_submat_vals);CHKERRQ(ierr); 13298629588bSStefano Zampini 13308629588bSStefano Zampini /* free */ 13318629588bSStefano Zampini ierr = PetscFree(coarse_submat_vals);CHKERRQ(ierr); 1332c8587f34SStefano Zampini PetscFunctionReturn(0); 1333c8587f34SStefano Zampini } 1334c8587f34SStefano Zampini 1335c8587f34SStefano Zampini #undef __FUNCT__ 1336674ae819SStefano Zampini #define __FUNCT__ "PCBDDCResetCustomization" 1337674ae819SStefano Zampini PetscErrorCode PCBDDCResetCustomization(PC pc) 1338674ae819SStefano Zampini { 1339674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 1340674ae819SStefano Zampini PetscErrorCode ierr; 1341674ae819SStefano Zampini 1342674ae819SStefano Zampini PetscFunctionBegin; 1343674ae819SStefano Zampini ierr = PCBDDCGraphResetCSR(pcbddc->mat_graph);CHKERRQ(ierr); 1344674ae819SStefano Zampini ierr = ISDestroy(&pcbddc->user_primal_vertices);CHKERRQ(ierr); 134530368db7SStefano Zampini ierr = ISDestroy(&pcbddc->user_primal_vertices_local);CHKERRQ(ierr); 1346674ae819SStefano Zampini ierr = MatNullSpaceDestroy(&pcbddc->NullSpace);CHKERRQ(ierr); 1347674ae819SStefano Zampini ierr = ISDestroy(&pcbddc->NeumannBoundaries);CHKERRQ(ierr); 1348785d1243SStefano Zampini ierr = ISDestroy(&pcbddc->NeumannBoundariesLocal);CHKERRQ(ierr); 1349674ae819SStefano Zampini ierr = ISDestroy(&pcbddc->DirichletBoundaries);CHKERRQ(ierr); 1350f4ddd8eeSStefano Zampini ierr = MatNullSpaceDestroy(&pcbddc->onearnullspace);CHKERRQ(ierr); 1351f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->onearnullvecs_state);CHKERRQ(ierr); 1352785d1243SStefano Zampini ierr = ISDestroy(&pcbddc->DirichletBoundariesLocal);CHKERRQ(ierr); 135363602bcaSStefano Zampini ierr = PCBDDCSetDofsSplitting(pc,0,NULL);CHKERRQ(ierr); 135463602bcaSStefano Zampini ierr = PCBDDCSetDofsSplittingLocal(pc,0,NULL);CHKERRQ(ierr); 1355674ae819SStefano Zampini PetscFunctionReturn(0); 1356674ae819SStefano Zampini } 1357674ae819SStefano Zampini 1358674ae819SStefano Zampini #undef __FUNCT__ 1359674ae819SStefano Zampini #define __FUNCT__ "PCBDDCResetTopography" 1360674ae819SStefano Zampini PetscErrorCode PCBDDCResetTopography(PC pc) 1361674ae819SStefano Zampini { 1362674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 13634f1b2e48SStefano Zampini PetscInt i; 1364674ae819SStefano Zampini PetscErrorCode ierr; 1365674ae819SStefano Zampini 1366674ae819SStefano Zampini PetscFunctionBegin; 1367b9b85e73SStefano Zampini ierr = MatDestroy(&pcbddc->user_ChangeOfBasisMatrix);CHKERRQ(ierr); 1368674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 1369674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 1370674ae819SStefano Zampini ierr = PCBDDCGraphReset(pcbddc->mat_graph);CHKERRQ(ierr); 13714f1b2e48SStefano Zampini for (i=0;i<pcbddc->n_local_subs;i++) { 13724f1b2e48SStefano Zampini ierr = ISDestroy(&pcbddc->local_subs[i]);CHKERRQ(ierr); 13734f1b2e48SStefano Zampini } 13744f1b2e48SStefano Zampini ierr = PetscFree(pcbddc->local_subs);CHKERRQ(ierr); 1375b96c3477SStefano Zampini ierr = PCBDDCSubSchursReset(pcbddc->sub_schurs);CHKERRQ(ierr); 1376674ae819SStefano Zampini PetscFunctionReturn(0); 1377674ae819SStefano Zampini } 1378674ae819SStefano Zampini 1379674ae819SStefano Zampini #undef __FUNCT__ 1380674ae819SStefano Zampini #define __FUNCT__ "PCBDDCResetSolvers" 1381674ae819SStefano Zampini PetscErrorCode PCBDDCResetSolvers(PC pc) 1382674ae819SStefano Zampini { 1383674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 1384674ae819SStefano Zampini PetscErrorCode ierr; 1385674ae819SStefano Zampini 1386674ae819SStefano Zampini PetscFunctionBegin; 1387674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->coarse_vec);CHKERRQ(ierr); 138858da7f69SStefano Zampini if (pcbddc->coarse_phi_B) { 1389ca92afb2SStefano Zampini PetscScalar *array; 139006656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&array);CHKERRQ(ierr); 139106656605SStefano Zampini ierr = PetscFree(array);CHKERRQ(ierr); 139258da7f69SStefano Zampini } 1393674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_B);CHKERRQ(ierr); 1394674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_D);CHKERRQ(ierr); 139515aaf578SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_B);CHKERRQ(ierr); 139615aaf578SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_D);CHKERRQ(ierr); 1397674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec1_P);CHKERRQ(ierr); 1398674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec1_C);CHKERRQ(ierr); 1399674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat2);CHKERRQ(ierr); 140006656605SStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat1);CHKERRQ(ierr); 1401674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec1_R);CHKERRQ(ierr); 1402674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec2_R);CHKERRQ(ierr); 14038ce42a96SStefano Zampini ierr = ISDestroy(&pcbddc->is_R_local);CHKERRQ(ierr); 1404674ae819SStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_B);CHKERRQ(ierr); 1405674ae819SStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_D);CHKERRQ(ierr); 1406674ae819SStefano Zampini ierr = VecScatterDestroy(&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 1407f4ddd8eeSStefano Zampini ierr = KSPDestroy(&pcbddc->ksp_D);CHKERRQ(ierr); 1408f4ddd8eeSStefano Zampini ierr = KSPDestroy(&pcbddc->ksp_R);CHKERRQ(ierr); 1409f4ddd8eeSStefano Zampini ierr = KSPDestroy(&pcbddc->coarse_ksp);CHKERRQ(ierr); 1410f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 1411727cdba6SStefano Zampini ierr = PetscFree(pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 14120e6343abSStefano Zampini ierr = PetscFree2(pcbddc->local_primal_ref_node,pcbddc->local_primal_ref_mult);CHKERRQ(ierr); 1413f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->global_primal_indices);CHKERRQ(ierr); 141470cf5478SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling);CHKERRQ(ierr); 14156e683305SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 141681d14e9dSStefano Zampini ierr = MatDestroy(&pcbddc->benign_change);CHKERRQ(ierr); 14170369aaf7SStefano Zampini ierr = VecDestroy(&pcbddc->benign_vec);CHKERRQ(ierr); 14188b9f24d4SStefano Zampini ierr = MatDestroy(&pcbddc->benign_original_mat);CHKERRQ(ierr); 14194f1b2e48SStefano Zampini ierr = MatDestroy(&pcbddc->benign_B0);CHKERRQ(ierr); 14208b9f24d4SStefano Zampini ierr = PetscSFDestroy(&pcbddc->benign_sf);CHKERRQ(ierr); 1421ca92afb2SStefano Zampini if (pcbddc->benign_zerodiag_subs) { 1422ca92afb2SStefano Zampini PetscInt i; 1423ca92afb2SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 1424ca92afb2SStefano Zampini ierr = ISDestroy(&pcbddc->benign_zerodiag_subs[i]);CHKERRQ(ierr); 1425ca92afb2SStefano Zampini } 1426ca92afb2SStefano Zampini ierr = PetscFree(pcbddc->benign_zerodiag_subs);CHKERRQ(ierr); 1427ca92afb2SStefano Zampini } 14284f1b2e48SStefano Zampini ierr = PetscFree3(pcbddc->benign_p0_lidx,pcbddc->benign_p0_gidx,pcbddc->benign_p0);CHKERRQ(ierr); 1429674ae819SStefano Zampini PetscFunctionReturn(0); 1430674ae819SStefano Zampini } 1431674ae819SStefano Zampini 1432674ae819SStefano Zampini #undef __FUNCT__ 1433f4ddd8eeSStefano Zampini #define __FUNCT__ "PCBDDCSetUpLocalWorkVectors" 1434f4ddd8eeSStefano Zampini PetscErrorCode PCBDDCSetUpLocalWorkVectors(PC pc) 14356bfb1811SStefano Zampini { 14366bfb1811SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 14376bfb1811SStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 14386bfb1811SStefano Zampini VecType impVecType; 14394f1b2e48SStefano Zampini PetscInt n_constraints,n_R,old_size; 14406bfb1811SStefano Zampini PetscErrorCode ierr; 14416bfb1811SStefano Zampini 14426bfb1811SStefano Zampini PetscFunctionBegin; 1443f4ddd8eeSStefano Zampini if (!pcbddc->ConstraintMatrix) { 1444019a44ceSStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"BDDC Constraint matrix has not been created"); 1445f4ddd8eeSStefano Zampini } 1446e7b262bdSStefano Zampini /* get sizes */ 14474f1b2e48SStefano Zampini n_constraints = pcbddc->local_primal_size - pcbddc->benign_n - pcbddc->n_vertices; 1448b371cd4fSStefano Zampini n_R = pcis->n - pcbddc->n_vertices; 14496bfb1811SStefano Zampini ierr = VecGetType(pcis->vec1_N,&impVecType);CHKERRQ(ierr); 1450e7b262bdSStefano Zampini /* local work vectors (try to avoid unneeded work)*/ 1451e7b262bdSStefano Zampini /* R nodes */ 1452e7b262bdSStefano Zampini old_size = -1; 1453e7b262bdSStefano Zampini if (pcbddc->vec1_R) { 1454e7b262bdSStefano Zampini ierr = VecGetSize(pcbddc->vec1_R,&old_size);CHKERRQ(ierr); 1455e7b262bdSStefano Zampini } 1456e7b262bdSStefano Zampini if (n_R != old_size) { 1457e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec1_R);CHKERRQ(ierr); 1458e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec2_R);CHKERRQ(ierr); 14596bfb1811SStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_R);CHKERRQ(ierr); 14606bfb1811SStefano Zampini ierr = VecSetSizes(pcbddc->vec1_R,PETSC_DECIDE,n_R);CHKERRQ(ierr); 14616bfb1811SStefano Zampini ierr = VecSetType(pcbddc->vec1_R,impVecType);CHKERRQ(ierr); 14626bfb1811SStefano Zampini ierr = VecDuplicate(pcbddc->vec1_R,&pcbddc->vec2_R);CHKERRQ(ierr); 1463e7b262bdSStefano Zampini } 1464e7b262bdSStefano Zampini /* local primal dofs */ 1465e7b262bdSStefano Zampini old_size = -1; 1466e7b262bdSStefano Zampini if (pcbddc->vec1_P) { 1467e7b262bdSStefano Zampini ierr = VecGetSize(pcbddc->vec1_P,&old_size);CHKERRQ(ierr); 1468e7b262bdSStefano Zampini } 1469e9189074SStefano Zampini if (pcbddc->local_primal_size != old_size) { 1470e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec1_P);CHKERRQ(ierr); 147183b7ccabSStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_P);CHKERRQ(ierr); 1472e9189074SStefano Zampini ierr = VecSetSizes(pcbddc->vec1_P,PETSC_DECIDE,pcbddc->local_primal_size);CHKERRQ(ierr); 14736bfb1811SStefano Zampini ierr = VecSetType(pcbddc->vec1_P,impVecType);CHKERRQ(ierr); 1474e7b262bdSStefano Zampini } 1475e7b262bdSStefano Zampini /* local explicit constraints */ 1476e7b262bdSStefano Zampini old_size = -1; 1477e7b262bdSStefano Zampini if (pcbddc->vec1_C) { 1478e7b262bdSStefano Zampini ierr = VecGetSize(pcbddc->vec1_C,&old_size);CHKERRQ(ierr); 1479e7b262bdSStefano Zampini } 1480e7b262bdSStefano Zampini if (n_constraints && n_constraints != old_size) { 1481e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec1_C);CHKERRQ(ierr); 148283b7ccabSStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_C);CHKERRQ(ierr); 148383b7ccabSStefano Zampini ierr = VecSetSizes(pcbddc->vec1_C,PETSC_DECIDE,n_constraints);CHKERRQ(ierr); 148483b7ccabSStefano Zampini ierr = VecSetType(pcbddc->vec1_C,impVecType);CHKERRQ(ierr); 148583b7ccabSStefano Zampini } 14866bfb1811SStefano Zampini PetscFunctionReturn(0); 14876bfb1811SStefano Zampini } 14886bfb1811SStefano Zampini 14896bfb1811SStefano Zampini #undef __FUNCT__ 149047f4ddc3SStefano Zampini #define __FUNCT__ "PCBDDCSetUpCorrection" 149147f4ddc3SStefano Zampini PetscErrorCode PCBDDCSetUpCorrection(PC pc, PetscScalar **coarse_submat_vals_n) 149288ebb749SStefano Zampini { 149325084f0cSStefano Zampini PetscErrorCode ierr; 149425084f0cSStefano Zampini /* pointers to pcis and pcbddc */ 149588ebb749SStefano Zampini PC_IS* pcis = (PC_IS*)pc->data; 149688ebb749SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 1497d62866d3SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 149825084f0cSStefano Zampini /* submatrices of local problem */ 149980677318SStefano Zampini Mat A_RV,A_VR,A_VV,local_auxmat2_R; 150006656605SStefano Zampini /* submatrices of local coarse problem */ 150106656605SStefano Zampini Mat S_VV,S_CV,S_VC,S_CC; 150225084f0cSStefano Zampini /* working matrices */ 150306656605SStefano Zampini Mat C_CR; 150425084f0cSStefano Zampini /* additional working stuff */ 150506656605SStefano Zampini PC pc_R; 15064f1b2e48SStefano Zampini Mat F; 15075cbda25cSStefano Zampini Vec dummy_vec; 1508a3df083aSStefano Zampini PetscBool isLU,isCHOL,isILU,need_benign_correction; 150925084f0cSStefano Zampini PetscScalar *coarse_submat_vals; /* TODO: use a PETSc matrix */ 151006656605SStefano Zampini PetscScalar *work; 151106656605SStefano Zampini PetscInt *idx_V_B; 1512ffd830a3SStefano Zampini PetscInt lda_rhs,n,n_vertices,n_constraints,*p0_lidx_I; 151306656605SStefano Zampini PetscInt i,n_R,n_D,n_B; 1514ffd830a3SStefano Zampini 151525084f0cSStefano Zampini /* some shortcuts to scalars */ 151606656605SStefano Zampini PetscScalar one=1.0,m_one=-1.0; 151788ebb749SStefano Zampini 151888ebb749SStefano Zampini PetscFunctionBegin; 1519ffd830a3SStefano Zampini if (!pcbddc->symmetric_primal && pcbddc->benign_n) { 1520ffd830a3SStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Non-symmetric primal basis computation with benign trick not yet implemented"); 1521ffd830a3SStefano Zampini } 1522ffd830a3SStefano Zampini 1523ffd830a3SStefano Zampini /* Set Non-overlapping dimensions */ 1524b371cd4fSStefano Zampini n_vertices = pcbddc->n_vertices; 15254f1b2e48SStefano Zampini n_constraints = pcbddc->local_primal_size - pcbddc->benign_n - n_vertices; 1526b371cd4fSStefano Zampini n_B = pcis->n_B; 1527b371cd4fSStefano Zampini n_D = pcis->n - n_B; 152888ebb749SStefano Zampini n_R = pcis->n - n_vertices; 152988ebb749SStefano Zampini 153088ebb749SStefano Zampini /* vertices in boundary numbering */ 1531785e854fSJed Brown ierr = PetscMalloc1(n_vertices,&idx_V_B);CHKERRQ(ierr); 15320e6343abSStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,n_vertices,pcbddc->local_primal_ref_node,&i,idx_V_B);CHKERRQ(ierr); 153388ebb749SStefano Zampini if (i != n_vertices) { 153422d5777bSStefano Zampini SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Error in boundary numbering for BDDC vertices! %d != %d\n",n_vertices,i); 153588ebb749SStefano Zampini } 153688ebb749SStefano Zampini 153706656605SStefano Zampini /* Subdomain contribution (Non-overlapping) to coarse matrix */ 1538019a44ceSStefano Zampini ierr = PetscCalloc1(pcbddc->local_primal_size*pcbddc->local_primal_size,&coarse_submat_vals);CHKERRQ(ierr); 153906656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_vertices,n_vertices,coarse_submat_vals,&S_VV);CHKERRQ(ierr); 154006656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_VV,pcbddc->local_primal_size);CHKERRQ(ierr); 154106656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_constraints,n_vertices,coarse_submat_vals+n_vertices,&S_CV);CHKERRQ(ierr); 154206656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_CV,pcbddc->local_primal_size);CHKERRQ(ierr); 154306656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_vertices,n_constraints,coarse_submat_vals+pcbddc->local_primal_size*n_vertices,&S_VC);CHKERRQ(ierr); 154406656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_VC,pcbddc->local_primal_size);CHKERRQ(ierr); 154506656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_constraints,n_constraints,coarse_submat_vals+(pcbddc->local_primal_size+1)*n_vertices,&S_CC);CHKERRQ(ierr); 154606656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_CC,pcbddc->local_primal_size);CHKERRQ(ierr); 154706656605SStefano Zampini 154806656605SStefano Zampini /* determine if can use MatSolve routines instead of calling KSPSolve on ksp_R */ 154906656605SStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_R);CHKERRQ(ierr); 155006656605SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_R,PCLU,&isLU);CHKERRQ(ierr); 155106656605SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_R,PCILU,&isILU);CHKERRQ(ierr); 155206656605SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_R,PCCHOLESKY,&isCHOL);CHKERRQ(ierr); 1553ffd830a3SStefano Zampini lda_rhs = n_R; 1554a3df083aSStefano Zampini need_benign_correction = PETSC_FALSE; 155506656605SStefano Zampini if (isLU || isILU || isCHOL) { 155606656605SStefano Zampini ierr = PCFactorGetMatrix(pc_R,&F);CHKERRQ(ierr); 1557df4d28bfSStefano Zampini } else if (sub_schurs->reuse_solver) { 1558df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 1559d62866d3SStefano Zampini MatFactorType type; 1560d62866d3SStefano Zampini 1561df4d28bfSStefano Zampini F = reuse_solver->F; 15626816873aSStefano Zampini ierr = MatGetFactorType(F,&type);CHKERRQ(ierr); 1563d62866d3SStefano Zampini if (type == MAT_FACTOR_CHOLESKY) isCHOL = PETSC_TRUE; 1564ffd830a3SStefano Zampini ierr = MatGetSize(F,&lda_rhs,NULL);CHKERRQ(ierr); 1565df4d28bfSStefano Zampini need_benign_correction = !!reuse_solver->benign_n; 156606656605SStefano Zampini } else { 156706656605SStefano Zampini F = NULL; 156806656605SStefano Zampini } 156906656605SStefano Zampini 1570ffd830a3SStefano Zampini /* allocate workspace */ 1571ffd830a3SStefano Zampini n = 0; 1572ffd830a3SStefano Zampini if (n_constraints) { 1573ffd830a3SStefano Zampini n += lda_rhs*n_constraints; 1574ffd830a3SStefano Zampini } 1575ffd830a3SStefano Zampini if (n_vertices) { 1576ffd830a3SStefano Zampini n = PetscMax(2*lda_rhs*n_vertices,n); 1577ffd830a3SStefano Zampini n = PetscMax((lda_rhs+n_B)*n_vertices,n); 1578ffd830a3SStefano Zampini } 1579ffd830a3SStefano Zampini ierr = PetscMalloc1(n,&work);CHKERRQ(ierr); 1580ffd830a3SStefano Zampini 15815cbda25cSStefano Zampini /* create dummy vector to modify rhs and sol of MatMatSolve (work array will never be used) */ 15825cbda25cSStefano Zampini dummy_vec = NULL; 15835cbda25cSStefano Zampini if (need_benign_correction && lda_rhs != n_R && F) { 15845cbda25cSStefano Zampini ierr = VecCreateSeqWithArray(PETSC_COMM_SELF,1,lda_rhs,work,&dummy_vec);CHKERRQ(ierr); 15855cbda25cSStefano Zampini } 15865cbda25cSStefano Zampini 158788ebb749SStefano Zampini /* Precompute stuffs needed for preprocessing and application of BDDC*/ 158888ebb749SStefano Zampini if (n_constraints) { 158972b8c272SStefano Zampini Mat M1,M2,M3,C_B; 159006656605SStefano Zampini IS is_aux; 159180677318SStefano Zampini PetscScalar *array,*array2; 159206656605SStefano Zampini 1593f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat1);CHKERRQ(ierr); 159480677318SStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat2);CHKERRQ(ierr); 159588ebb749SStefano Zampini 159625084f0cSStefano Zampini /* Extract constraints on R nodes: C_{CR} */ 159725084f0cSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,n_constraints,n_vertices,1,&is_aux);CHKERRQ(ierr); 15988ce42a96SStefano Zampini ierr = MatGetSubMatrix(pcbddc->ConstraintMatrix,is_aux,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&C_CR);CHKERRQ(ierr); 159972b8c272SStefano Zampini ierr = MatGetSubMatrix(pcbddc->ConstraintMatrix,is_aux,pcis->is_B_local,MAT_INITIAL_MATRIX,&C_B);CHKERRQ(ierr); 160088ebb749SStefano Zampini 160180677318SStefano Zampini /* Assemble local_auxmat2_R = (- A_{RR}^{-1} C^T_{CR}) needed by BDDC setup */ 160280677318SStefano Zampini /* Assemble pcbddc->local_auxmat2 = R_to_B (- A_{RR}^{-1} C^T_{CR}) needed by BDDC application */ 1603ffd830a3SStefano Zampini ierr = PetscMemzero(work,lda_rhs*n_constraints*sizeof(PetscScalar));CHKERRQ(ierr); 160488ebb749SStefano Zampini for (i=0;i<n_constraints;i++) { 160506656605SStefano Zampini const PetscScalar *row_cmat_values; 160606656605SStefano Zampini const PetscInt *row_cmat_indices; 160706656605SStefano Zampini PetscInt size_of_constraint,j; 160888ebb749SStefano Zampini 160906656605SStefano Zampini ierr = MatGetRow(C_CR,i,&size_of_constraint,&row_cmat_indices,&row_cmat_values);CHKERRQ(ierr); 161006656605SStefano Zampini for (j=0;j<size_of_constraint;j++) { 1611ffd830a3SStefano Zampini work[row_cmat_indices[j]+i*lda_rhs] = -row_cmat_values[j]; 161206656605SStefano Zampini } 161306656605SStefano Zampini ierr = MatRestoreRow(C_CR,i,&size_of_constraint,&row_cmat_indices,&row_cmat_values);CHKERRQ(ierr); 161406656605SStefano Zampini } 1615ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_constraints,NULL,&local_auxmat2_R);CHKERRQ(ierr); 161606656605SStefano Zampini if (F) { 161706656605SStefano Zampini Mat B; 161806656605SStefano Zampini 1619ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_constraints,work,&B);CHKERRQ(ierr); 1620a3df083aSStefano Zampini if (need_benign_correction) { 1621df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 1622a3df083aSStefano Zampini 162372b8c272SStefano Zampini /* rhs is already zero on interior dofs, no need to change the rhs */ 162472b8c272SStefano Zampini ierr = PetscMemzero(reuse_solver->benign_save_vals,pcbddc->benign_n*sizeof(PetscScalar));CHKERRQ(ierr); 1625a3df083aSStefano Zampini } 162680677318SStefano Zampini ierr = MatMatSolve(F,B,local_auxmat2_R);CHKERRQ(ierr); 1627a3df083aSStefano Zampini if (need_benign_correction) { 1628a3df083aSStefano Zampini PetscScalar *marr; 1629df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 1630a3df083aSStefano Zampini 1631a3df083aSStefano Zampini ierr = MatDenseGetArray(local_auxmat2_R,&marr);CHKERRQ(ierr); 16325cbda25cSStefano Zampini if (lda_rhs != n_R) { 16335cbda25cSStefano Zampini for (i=0;i<n_constraints;i++) { 16345cbda25cSStefano Zampini ierr = VecPlaceArray(dummy_vec,marr+i*lda_rhs);CHKERRQ(ierr); 16355cbda25cSStefano Zampini ierr = PCBDDCReuseSolversBenignAdapt(reuse_solver,dummy_vec,NULL,PETSC_TRUE,PETSC_TRUE);CHKERRQ(ierr); 16365cbda25cSStefano Zampini ierr = VecResetArray(dummy_vec);CHKERRQ(ierr); 16375cbda25cSStefano Zampini } 16385cbda25cSStefano Zampini } else { 1639a3df083aSStefano Zampini for (i=0;i<n_constraints;i++) { 1640a3df083aSStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,marr+i*lda_rhs);CHKERRQ(ierr); 16415cbda25cSStefano Zampini ierr = PCBDDCReuseSolversBenignAdapt(reuse_solver,pcbddc->vec1_R,NULL,PETSC_TRUE,PETSC_TRUE);CHKERRQ(ierr); 1642a3df083aSStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 1643a3df083aSStefano Zampini } 16445cbda25cSStefano Zampini } 1645a3df083aSStefano Zampini ierr = MatDenseRestoreArray(local_auxmat2_R,&marr);CHKERRQ(ierr); 1646a3df083aSStefano Zampini } 164706656605SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 164806656605SStefano Zampini } else { 164980677318SStefano Zampini PetscScalar *marr; 165080677318SStefano Zampini 165180677318SStefano Zampini ierr = MatDenseGetArray(local_auxmat2_R,&marr);CHKERRQ(ierr); 165206656605SStefano Zampini for (i=0;i<n_constraints;i++) { 1653ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*lda_rhs);CHKERRQ(ierr); 1654ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,marr+i*lda_rhs);CHKERRQ(ierr); 165506656605SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 165606656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 165706656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 165806656605SStefano Zampini } 165980677318SStefano Zampini ierr = MatDenseRestoreArray(local_auxmat2_R,&marr);CHKERRQ(ierr); 166006656605SStefano Zampini } 166180677318SStefano Zampini if (!pcbddc->switch_static) { 166280677318SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,n_constraints,NULL,&pcbddc->local_auxmat2);CHKERRQ(ierr); 166380677318SStefano Zampini ierr = MatDenseGetArray(pcbddc->local_auxmat2,&array);CHKERRQ(ierr); 166480677318SStefano Zampini ierr = MatDenseGetArray(local_auxmat2_R,&array2);CHKERRQ(ierr); 166580677318SStefano Zampini for (i=0;i<n_constraints;i++) { 1666ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,array2+i*lda_rhs);CHKERRQ(ierr); 166780677318SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,array+i*n_B);CHKERRQ(ierr); 166880677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 166980677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 167080677318SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 167180677318SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 167280677318SStefano Zampini } 167380677318SStefano Zampini ierr = MatDenseRestoreArray(local_auxmat2_R,&array2);CHKERRQ(ierr); 167480677318SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->local_auxmat2,&array);CHKERRQ(ierr); 167572b8c272SStefano Zampini ierr = MatMatMult(C_B,pcbddc->local_auxmat2,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&M3);CHKERRQ(ierr); 167680677318SStefano Zampini } else { 1677ffd830a3SStefano Zampini if (lda_rhs != n_R) { 1678ffd830a3SStefano Zampini IS dummy; 1679ffd830a3SStefano Zampini 1680ffd830a3SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,n_R,0,1,&dummy);CHKERRQ(ierr); 168172b8c272SStefano Zampini ierr = MatGetSubMatrix(local_auxmat2_R,dummy,NULL,MAT_INITIAL_MATRIX,&pcbddc->local_auxmat2);CHKERRQ(ierr); 1682ffd830a3SStefano Zampini ierr = ISDestroy(&dummy);CHKERRQ(ierr); 1683ffd830a3SStefano Zampini } else { 168480677318SStefano Zampini ierr = PetscObjectReference((PetscObject)local_auxmat2_R);CHKERRQ(ierr); 168580677318SStefano Zampini pcbddc->local_auxmat2 = local_auxmat2_R; 1686ffd830a3SStefano Zampini } 168725084f0cSStefano Zampini ierr = MatMatMult(C_CR,pcbddc->local_auxmat2,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&M3);CHKERRQ(ierr); 168880677318SStefano Zampini } 168980677318SStefano Zampini ierr = ISDestroy(&is_aux);CHKERRQ(ierr); 169080677318SStefano Zampini /* Assemble explicitly S_CC = ( C_{CR} A_{RR}^{-1} C^T_{CR} )^{-1} */ 169180677318SStefano Zampini ierr = MatScale(M3,m_one);CHKERRQ(ierr); 169206656605SStefano Zampini ierr = MatDuplicate(M3,MAT_DO_NOT_COPY_VALUES,&M1);CHKERRQ(ierr); 169306656605SStefano Zampini ierr = MatDuplicate(M3,MAT_DO_NOT_COPY_VALUES,&M2);CHKERRQ(ierr); 169480677318SStefano Zampini if (isCHOL) { 169580677318SStefano Zampini ierr = MatCholeskyFactor(M3,NULL,NULL);CHKERRQ(ierr); 169680677318SStefano Zampini } else { 169725084f0cSStefano Zampini ierr = MatLUFactor(M3,NULL,NULL,NULL);CHKERRQ(ierr); 169880677318SStefano Zampini } 169980677318SStefano Zampini ierr = VecSet(pcbddc->vec1_C,one);CHKERRQ(ierr); 170006656605SStefano Zampini ierr = MatDiagonalSet(M2,pcbddc->vec1_C,INSERT_VALUES);CHKERRQ(ierr); 170125084f0cSStefano Zampini ierr = MatMatSolve(M3,M2,M1);CHKERRQ(ierr); 170225084f0cSStefano Zampini ierr = MatDestroy(&M2);CHKERRQ(ierr); 170325084f0cSStefano Zampini ierr = MatDestroy(&M3);CHKERRQ(ierr); 170480677318SStefano Zampini /* Assemble local_auxmat1 = S_CC*C_{CB} needed by BDDC application in KSP and in preproc */ 170572b8c272SStefano Zampini ierr = MatMatMult(M1,C_B,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&pcbddc->local_auxmat1);CHKERRQ(ierr); 170672b8c272SStefano Zampini ierr = MatDestroy(&C_B);CHKERRQ(ierr); 170706656605SStefano Zampini ierr = MatCopy(M1,S_CC,SAME_NONZERO_PATTERN);CHKERRQ(ierr); /* S_CC can have a different LDA, MatMatSolve doesn't support it */ 170806656605SStefano Zampini ierr = MatDestroy(&M1);CHKERRQ(ierr); 1709f4ddd8eeSStefano Zampini } 1710fc227af8SStefano Zampini 1711fc227af8SStefano Zampini /* Get submatrices from subdomain matrix */ 171288ebb749SStefano Zampini if (n_vertices) { 171306656605SStefano Zampini IS is_aux; 17143a50541eSStefano Zampini 1715df4d28bfSStefano Zampini if (sub_schurs->reuse_solver) { /* is_R_local is not sorted, ISComplement doesn't like it */ 17166816873aSStefano Zampini IS tis; 17176816873aSStefano Zampini 17186816873aSStefano Zampini ierr = ISDuplicate(pcbddc->is_R_local,&tis);CHKERRQ(ierr); 17196816873aSStefano Zampini ierr = ISSort(tis);CHKERRQ(ierr); 17206816873aSStefano Zampini ierr = ISComplement(tis,0,pcis->n,&is_aux);CHKERRQ(ierr); 17216816873aSStefano Zampini ierr = ISDestroy(&tis);CHKERRQ(ierr); 17226816873aSStefano Zampini } else { 17233a50541eSStefano Zampini ierr = ISComplement(pcbddc->is_R_local,0,pcis->n,&is_aux);CHKERRQ(ierr); 17246816873aSStefano Zampini } 17259577ea80SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,pcbddc->is_R_local,is_aux,MAT_INITIAL_MATRIX,&A_RV);CHKERRQ(ierr); 17269577ea80SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,is_aux,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&A_VR);CHKERRQ(ierr); 172704708bb6SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,is_aux,is_aux,MAT_INITIAL_MATRIX,&A_VV);CHKERRQ(ierr); 172825084f0cSStefano Zampini ierr = ISDestroy(&is_aux);CHKERRQ(ierr); 172988ebb749SStefano Zampini } 173088ebb749SStefano Zampini 173188ebb749SStefano Zampini /* Matrix of coarse basis functions (local) */ 1732f4ddd8eeSStefano Zampini if (pcbddc->coarse_phi_B) { 173306656605SStefano Zampini PetscInt on_B,on_primal,on_D=n_D; 173406656605SStefano Zampini if (pcbddc->coarse_phi_D) { 173506656605SStefano Zampini ierr = MatGetSize(pcbddc->coarse_phi_D,&on_D,NULL);CHKERRQ(ierr); 173606656605SStefano Zampini } 1737f4ddd8eeSStefano Zampini ierr = MatGetSize(pcbddc->coarse_phi_B,&on_B,&on_primal);CHKERRQ(ierr); 173806656605SStefano Zampini if (on_B != n_B || on_primal != pcbddc->local_primal_size || on_D != n_D) { 173906656605SStefano Zampini PetscScalar *marray; 174006656605SStefano Zampini 174106656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&marray);CHKERRQ(ierr); 174206656605SStefano Zampini ierr = PetscFree(marray);CHKERRQ(ierr); 1743f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_B);CHKERRQ(ierr); 1744f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_B);CHKERRQ(ierr); 1745f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_D);CHKERRQ(ierr); 1746f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_D);CHKERRQ(ierr); 1747f4ddd8eeSStefano Zampini } 1748f4ddd8eeSStefano Zampini } 174906656605SStefano Zampini 1750f4ddd8eeSStefano Zampini if (!pcbddc->coarse_phi_B) { 175106656605SStefano Zampini PetscScalar *marray; 175288ebb749SStefano Zampini 175306656605SStefano Zampini n = n_B*pcbddc->local_primal_size; 17548eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 175506656605SStefano Zampini n += n_D*pcbddc->local_primal_size; 175688ebb749SStefano Zampini } 17573301b35fSStefano Zampini if (!pcbddc->symmetric_primal) { 175806656605SStefano Zampini n *= 2; 175988ebb749SStefano Zampini } 176006656605SStefano Zampini ierr = PetscCalloc1(n,&marray);CHKERRQ(ierr); 176106656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,pcbddc->local_primal_size,marray,&pcbddc->coarse_phi_B);CHKERRQ(ierr); 176206656605SStefano Zampini n = n_B*pcbddc->local_primal_size; 17638eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 176406656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_D,pcbddc->local_primal_size,marray+n,&pcbddc->coarse_phi_D);CHKERRQ(ierr); 176506656605SStefano Zampini n += n_D*pcbddc->local_primal_size; 176688ebb749SStefano Zampini } 17673301b35fSStefano Zampini if (!pcbddc->symmetric_primal) { 176806656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,pcbddc->local_primal_size,marray+n,&pcbddc->coarse_psi_B);CHKERRQ(ierr); 17698eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 177006656605SStefano Zampini n = n_B*pcbddc->local_primal_size; 177106656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_D,pcbddc->local_primal_size,marray+n,&pcbddc->coarse_psi_D);CHKERRQ(ierr); 177288ebb749SStefano Zampini } 177388ebb749SStefano Zampini } else { 1774c0553b1fSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->coarse_phi_B);CHKERRQ(ierr); 1775c0553b1fSStefano Zampini pcbddc->coarse_psi_B = pcbddc->coarse_phi_B; 17761b968477SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 1777c0553b1fSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->coarse_phi_D);CHKERRQ(ierr); 1778c0553b1fSStefano Zampini pcbddc->coarse_psi_D = pcbddc->coarse_phi_D; 1779c0553b1fSStefano Zampini } 178088ebb749SStefano Zampini } 178106656605SStefano Zampini } 1782019a44ceSStefano Zampini 178306656605SStefano Zampini /* We are now ready to evaluate coarse basis functions and subdomain contribution to coarse problem */ 17844f1b2e48SStefano Zampini p0_lidx_I = NULL; 17854f1b2e48SStefano Zampini if (pcbddc->benign_n && (pcbddc->switch_static || pcbddc->dbg_flag)) { 1786d12edf2fSStefano Zampini const PetscInt *idxs; 1787d12edf2fSStefano Zampini 1788d12edf2fSStefano Zampini ierr = ISGetIndices(pcis->is_I_local,&idxs);CHKERRQ(ierr); 17894f1b2e48SStefano Zampini ierr = PetscMalloc1(pcbddc->benign_n,&p0_lidx_I);CHKERRQ(ierr); 17904f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 17914f1b2e48SStefano Zampini ierr = PetscFindInt(pcbddc->benign_p0_lidx[i],pcis->n-pcis->n_B,idxs,&p0_lidx_I[i]);CHKERRQ(ierr); 17924f1b2e48SStefano Zampini } 1793d12edf2fSStefano Zampini ierr = ISRestoreIndices(pcis->is_I_local,&idxs);CHKERRQ(ierr); 1794d12edf2fSStefano Zampini } 1795d16cbb6bSStefano Zampini 179606656605SStefano Zampini /* vertices */ 179706656605SStefano Zampini if (n_vertices) { 179816f15bc4SStefano Zampini 1799ffd830a3SStefano Zampini ierr = MatConvert(A_VV,MATDENSE,MAT_REUSE_MATRIX,&A_VV);CHKERRQ(ierr); 180004708bb6SStefano Zampini 180116f15bc4SStefano Zampini if (n_R) { 180214393ed6SStefano Zampini Mat A_RRmA_RV,A_RV_bcorr=NULL,S_VVt; /* S_VVt with LDA=N */ 180306656605SStefano Zampini PetscBLASInt B_N,B_one = 1; 180416f15bc4SStefano Zampini PetscScalar *x,*y; 180504708bb6SStefano Zampini PetscBool isseqaij; 180606656605SStefano Zampini 180721eccb56SStefano Zampini ierr = MatScale(A_RV,m_one);CHKERRQ(ierr); 180814393ed6SStefano Zampini if (need_benign_correction) { 180914393ed6SStefano Zampini ISLocalToGlobalMapping RtoN; 181014393ed6SStefano Zampini IS is_p0; 181114393ed6SStefano Zampini PetscInt *idxs_p0,n; 181214393ed6SStefano Zampini 181314393ed6SStefano Zampini ierr = PetscMalloc1(pcbddc->benign_n,&idxs_p0);CHKERRQ(ierr); 181414393ed6SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(pcbddc->is_R_local,&RtoN);CHKERRQ(ierr); 181514393ed6SStefano Zampini ierr = ISGlobalToLocalMappingApply(RtoN,IS_GTOLM_DROP,pcbddc->benign_n,pcbddc->benign_p0_lidx,&n,idxs_p0);CHKERRQ(ierr); 181614393ed6SStefano Zampini if (n != pcbddc->benign_n) { 181714393ed6SStefano Zampini SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Error in R numbering for benign p0! %d != %d\n",n,pcbddc->benign_n); 181814393ed6SStefano Zampini } 181914393ed6SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&RtoN);CHKERRQ(ierr); 182014393ed6SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,n,idxs_p0,PETSC_OWN_POINTER,&is_p0);CHKERRQ(ierr); 182114393ed6SStefano Zampini ierr = MatGetSubMatrix(A_RV,is_p0,NULL,MAT_INITIAL_MATRIX,&A_RV_bcorr);CHKERRQ(ierr); 182214393ed6SStefano Zampini ierr = ISDestroy(&is_p0);CHKERRQ(ierr); 182314393ed6SStefano Zampini } 182414393ed6SStefano Zampini 1825ffd830a3SStefano Zampini if (lda_rhs == n_R) { 1826ffd830a3SStefano Zampini ierr = MatConvert(A_RV,MATDENSE,MAT_REUSE_MATRIX,&A_RV);CHKERRQ(ierr); 1827ffd830a3SStefano Zampini } else { 1828ca92afb2SStefano Zampini PetscScalar *av,*array; 1829ca92afb2SStefano Zampini const PetscInt *xadj,*adjncy; 1830ca92afb2SStefano Zampini PetscInt n; 1831ca92afb2SStefano Zampini PetscBool flg_row; 1832ffd830a3SStefano Zampini 1833ca92afb2SStefano Zampini array = work+lda_rhs*n_vertices; 1834ca92afb2SStefano Zampini ierr = PetscMemzero(array,lda_rhs*n_vertices*sizeof(PetscScalar));CHKERRQ(ierr); 1835ca92afb2SStefano Zampini ierr = MatConvert(A_RV,MATSEQAIJ,MAT_REUSE_MATRIX,&A_RV);CHKERRQ(ierr); 1836ca92afb2SStefano Zampini ierr = MatGetRowIJ(A_RV,0,PETSC_FALSE,PETSC_FALSE,&n,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 1837ca92afb2SStefano Zampini ierr = MatSeqAIJGetArray(A_RV,&av);CHKERRQ(ierr); 1838ca92afb2SStefano Zampini for (i=0;i<n;i++) { 1839ca92afb2SStefano Zampini PetscInt j; 1840ca92afb2SStefano Zampini for (j=xadj[i];j<xadj[i+1];j++) array[lda_rhs*adjncy[j]+i] = av[j]; 1841ffd830a3SStefano Zampini } 1842ca92afb2SStefano Zampini ierr = MatRestoreRowIJ(A_RV,0,PETSC_FALSE,PETSC_FALSE,&n,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 1843ca92afb2SStefano Zampini ierr = MatDestroy(&A_RV);CHKERRQ(ierr); 1844ca92afb2SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_vertices,array,&A_RV);CHKERRQ(ierr); 1845ffd830a3SStefano Zampini } 1846ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_vertices,work,&A_RRmA_RV);CHKERRQ(ierr); 1847a3df083aSStefano Zampini if (need_benign_correction) { 1848df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 1849a3df083aSStefano Zampini PetscScalar *marr; 1850a3df083aSStefano Zampini 1851a3df083aSStefano Zampini ierr = MatDenseGetArray(A_RV,&marr);CHKERRQ(ierr); 185214393ed6SStefano Zampini /* need \Phi^T A_RV = (I+L)A_RV, L given by 185314393ed6SStefano Zampini 185414393ed6SStefano Zampini | 0 0 0 | (V) 185514393ed6SStefano Zampini L = | 0 0 -1 | (P-p0) 185614393ed6SStefano Zampini | 0 0 -1 | (p0) 185714393ed6SStefano Zampini 185814393ed6SStefano Zampini */ 1859df4d28bfSStefano Zampini for (i=0;i<reuse_solver->benign_n;i++) { 186014393ed6SStefano Zampini const PetscScalar *vals; 186114393ed6SStefano Zampini const PetscInt *idxs,*idxs_zero; 186214393ed6SStefano Zampini PetscInt n,j,nz; 186314393ed6SStefano Zampini 1864df4d28bfSStefano Zampini ierr = ISGetLocalSize(reuse_solver->benign_zerodiag_subs[i],&nz);CHKERRQ(ierr); 1865df4d28bfSStefano Zampini ierr = ISGetIndices(reuse_solver->benign_zerodiag_subs[i],&idxs_zero);CHKERRQ(ierr); 186614393ed6SStefano Zampini ierr = MatGetRow(A_RV_bcorr,i,&n,&idxs,&vals);CHKERRQ(ierr); 186714393ed6SStefano Zampini for (j=0;j<n;j++) { 186814393ed6SStefano Zampini PetscScalar val = vals[j]; 186914393ed6SStefano Zampini PetscInt k,col = idxs[j]; 187014393ed6SStefano Zampini for (k=0;k<nz;k++) marr[idxs_zero[k]+lda_rhs*col] -= val; 187114393ed6SStefano Zampini } 187214393ed6SStefano Zampini ierr = MatRestoreRow(A_RV_bcorr,i,&n,&idxs,&vals);CHKERRQ(ierr); 1873df4d28bfSStefano Zampini ierr = ISRestoreIndices(reuse_solver->benign_zerodiag_subs[i],&idxs_zero);CHKERRQ(ierr); 187414393ed6SStefano Zampini } 187572b8c272SStefano Zampini ierr = MatDenseRestoreArray(A_RV,&marr);CHKERRQ(ierr); 187672b8c272SStefano Zampini } 187772b8c272SStefano Zampini if (F) { 187814393ed6SStefano Zampini /* need to correct the rhs */ 187972b8c272SStefano Zampini if (need_benign_correction) { 188072b8c272SStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 188172b8c272SStefano Zampini PetscScalar *marr; 188272b8c272SStefano Zampini 188372b8c272SStefano Zampini ierr = MatDenseGetArray(A_RV,&marr);CHKERRQ(ierr); 18845cbda25cSStefano Zampini if (lda_rhs != n_R) { 18855cbda25cSStefano Zampini for (i=0;i<n_vertices;i++) { 18865cbda25cSStefano Zampini ierr = VecPlaceArray(dummy_vec,marr+i*lda_rhs);CHKERRQ(ierr); 18875cbda25cSStefano Zampini ierr = PCBDDCReuseSolversBenignAdapt(reuse_solver,dummy_vec,NULL,PETSC_FALSE,PETSC_TRUE);CHKERRQ(ierr); 18885cbda25cSStefano Zampini ierr = VecResetArray(dummy_vec);CHKERRQ(ierr); 18895cbda25cSStefano Zampini } 18905cbda25cSStefano Zampini } else { 1891a3df083aSStefano Zampini for (i=0;i<n_vertices;i++) { 1892a3df083aSStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,marr+i*lda_rhs);CHKERRQ(ierr); 18935cbda25cSStefano Zampini ierr = PCBDDCReuseSolversBenignAdapt(reuse_solver,pcbddc->vec1_R,NULL,PETSC_FALSE,PETSC_TRUE);CHKERRQ(ierr); 1894a3df083aSStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 1895a3df083aSStefano Zampini } 18965cbda25cSStefano Zampini } 1897a3df083aSStefano Zampini ierr = MatDenseRestoreArray(A_RV,&marr);CHKERRQ(ierr); 1898a3df083aSStefano Zampini } 189906656605SStefano Zampini ierr = MatMatSolve(F,A_RV,A_RRmA_RV);CHKERRQ(ierr); 190014393ed6SStefano Zampini /* need to correct the solution */ 1901a3df083aSStefano Zampini if (need_benign_correction) { 1902df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 1903a3df083aSStefano Zampini PetscScalar *marr; 1904a3df083aSStefano Zampini 1905a3df083aSStefano Zampini ierr = MatDenseGetArray(A_RRmA_RV,&marr);CHKERRQ(ierr); 19065cbda25cSStefano Zampini if (lda_rhs != n_R) { 19075cbda25cSStefano Zampini for (i=0;i<n_vertices;i++) { 19085cbda25cSStefano Zampini ierr = VecPlaceArray(dummy_vec,marr+i*lda_rhs);CHKERRQ(ierr); 19095cbda25cSStefano Zampini ierr = PCBDDCReuseSolversBenignAdapt(reuse_solver,dummy_vec,NULL,PETSC_TRUE,PETSC_TRUE);CHKERRQ(ierr); 19105cbda25cSStefano Zampini ierr = VecResetArray(dummy_vec);CHKERRQ(ierr); 19115cbda25cSStefano Zampini } 19125cbda25cSStefano Zampini } else { 1913a3df083aSStefano Zampini for (i=0;i<n_vertices;i++) { 1914a3df083aSStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,marr+i*lda_rhs);CHKERRQ(ierr); 19155cbda25cSStefano Zampini ierr = PCBDDCReuseSolversBenignAdapt(reuse_solver,pcbddc->vec1_R,NULL,PETSC_TRUE,PETSC_TRUE);CHKERRQ(ierr); 1916a3df083aSStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 1917a3df083aSStefano Zampini } 19185cbda25cSStefano Zampini } 1919a3df083aSStefano Zampini ierr = MatDenseRestoreArray(A_RRmA_RV,&marr);CHKERRQ(ierr); 1920a3df083aSStefano Zampini } 192106656605SStefano Zampini } else { 192206656605SStefano Zampini ierr = MatDenseGetArray(A_RV,&y);CHKERRQ(ierr); 192306656605SStefano Zampini for (i=0;i<n_vertices;i++) { 1924ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,y+i*lda_rhs);CHKERRQ(ierr); 1925ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,work+i*lda_rhs);CHKERRQ(ierr); 192606656605SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 192706656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 192806656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 192906656605SStefano Zampini } 193006656605SStefano Zampini ierr = MatDenseRestoreArray(A_RV,&y);CHKERRQ(ierr); 193106656605SStefano Zampini } 193280677318SStefano Zampini ierr = MatDestroy(&A_RV);CHKERRQ(ierr); 1933ffd830a3SStefano Zampini /* S_VV and S_CV */ 193406656605SStefano Zampini if (n_constraints) { 193506656605SStefano Zampini Mat B; 193680677318SStefano Zampini 1937ffd830a3SStefano Zampini ierr = PetscMemzero(work+lda_rhs*n_vertices,n_B*n_vertices*sizeof(PetscScalar));CHKERRQ(ierr); 193880677318SStefano Zampini for (i=0;i<n_vertices;i++) { 1939ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*lda_rhs);CHKERRQ(ierr); 1940ffd830a3SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,work+lda_rhs*n_vertices+i*n_B);CHKERRQ(ierr); 194180677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 194280677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 194380677318SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 194480677318SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 194580677318SStefano Zampini } 1946ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,n_vertices,work+lda_rhs*n_vertices,&B);CHKERRQ(ierr); 194780677318SStefano Zampini ierr = MatMatMult(pcbddc->local_auxmat1,B,MAT_REUSE_MATRIX,PETSC_DEFAULT,&S_CV);CHKERRQ(ierr); 194880677318SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 1949ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_vertices,work+lda_rhs*n_vertices,&B);CHKERRQ(ierr); 195080677318SStefano Zampini ierr = MatMatMult(local_auxmat2_R,S_CV,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B);CHKERRQ(ierr); 195106656605SStefano Zampini ierr = MatScale(S_CV,m_one);CHKERRQ(ierr); 1952ffd830a3SStefano Zampini ierr = PetscBLASIntCast(lda_rhs*n_vertices,&B_N);CHKERRQ(ierr); 1953ffd830a3SStefano Zampini PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&B_N,&one,work+lda_rhs*n_vertices,&B_one,work,&B_one)); 195406656605SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 195506656605SStefano Zampini } 195604708bb6SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)A_VR,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 195704708bb6SStefano Zampini if (!isseqaij) { /* MatMatMult with SEQ(S)BAIJ below will raise an error */ 195804708bb6SStefano Zampini ierr = MatConvert(A_VR,MATSEQAIJ,MAT_REUSE_MATRIX,&A_VR);CHKERRQ(ierr); 195904708bb6SStefano Zampini } 1960ffd830a3SStefano Zampini if (lda_rhs != n_R) { 1961ffd830a3SStefano Zampini ierr = MatDestroy(&A_RRmA_RV);CHKERRQ(ierr); 1962ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_vertices,work,&A_RRmA_RV);CHKERRQ(ierr); 1963ffd830a3SStefano Zampini ierr = MatSeqDenseSetLDA(A_RRmA_RV,lda_rhs);CHKERRQ(ierr); 1964ffd830a3SStefano Zampini } 196506656605SStefano Zampini ierr = MatMatMult(A_VR,A_RRmA_RV,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&S_VVt);CHKERRQ(ierr); 196614393ed6SStefano Zampini /* need A_VR * \Phi * A_RRmA_RV = A_VR * (I+L)^T * A_RRmA_RV, L given as before */ 196714393ed6SStefano Zampini if (need_benign_correction) { 1968df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 196914393ed6SStefano Zampini PetscScalar *marr,*sums; 197014393ed6SStefano Zampini 197114393ed6SStefano Zampini ierr = PetscMalloc1(n_vertices,&sums);CHKERRQ(ierr); 197214393ed6SStefano Zampini ierr = MatDenseGetArray(S_VVt,&marr); 1973df4d28bfSStefano Zampini for (i=0;i<reuse_solver->benign_n;i++) { 197414393ed6SStefano Zampini const PetscScalar *vals; 197514393ed6SStefano Zampini const PetscInt *idxs,*idxs_zero; 197614393ed6SStefano Zampini PetscInt n,j,nz; 197714393ed6SStefano Zampini 1978df4d28bfSStefano Zampini ierr = ISGetLocalSize(reuse_solver->benign_zerodiag_subs[i],&nz);CHKERRQ(ierr); 1979df4d28bfSStefano Zampini ierr = ISGetIndices(reuse_solver->benign_zerodiag_subs[i],&idxs_zero);CHKERRQ(ierr); 198014393ed6SStefano Zampini for (j=0;j<n_vertices;j++) { 198114393ed6SStefano Zampini PetscInt k; 198214393ed6SStefano Zampini sums[j] = 0.; 198314393ed6SStefano Zampini for (k=0;k<nz;k++) sums[j] += work[idxs_zero[k]+j*lda_rhs]; 198414393ed6SStefano Zampini } 198514393ed6SStefano Zampini ierr = MatGetRow(A_RV_bcorr,i,&n,&idxs,&vals);CHKERRQ(ierr); 198614393ed6SStefano Zampini for (j=0;j<n;j++) { 198714393ed6SStefano Zampini PetscScalar val = vals[j]; 198814393ed6SStefano Zampini PetscInt k; 198914393ed6SStefano Zampini for (k=0;k<n_vertices;k++) { 199014393ed6SStefano Zampini marr[idxs[j]+k*n_vertices] += val*sums[k]; 199114393ed6SStefano Zampini } 199214393ed6SStefano Zampini } 199314393ed6SStefano Zampini ierr = MatRestoreRow(A_RV_bcorr,i,&n,&idxs,&vals);CHKERRQ(ierr); 1994df4d28bfSStefano Zampini ierr = ISRestoreIndices(reuse_solver->benign_zerodiag_subs[i],&idxs_zero);CHKERRQ(ierr); 199514393ed6SStefano Zampini } 199614393ed6SStefano Zampini ierr = PetscFree(sums);CHKERRQ(ierr); 199714393ed6SStefano Zampini ierr = MatDenseRestoreArray(S_VVt,&marr); 199814393ed6SStefano Zampini ierr = MatDestroy(&A_RV_bcorr);CHKERRQ(ierr); 199914393ed6SStefano Zampini } 200080677318SStefano Zampini ierr = MatDestroy(&A_RRmA_RV);CHKERRQ(ierr); 200106656605SStefano Zampini ierr = PetscBLASIntCast(n_vertices*n_vertices,&B_N);CHKERRQ(ierr); 200206656605SStefano Zampini ierr = MatDenseGetArray(A_VV,&x);CHKERRQ(ierr); 200306656605SStefano Zampini ierr = MatDenseGetArray(S_VVt,&y);CHKERRQ(ierr); 200406656605SStefano Zampini PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&B_N,&one,x,&B_one,y,&B_one)); 200506656605SStefano Zampini ierr = MatDenseRestoreArray(A_VV,&x);CHKERRQ(ierr); 200606656605SStefano Zampini ierr = MatDenseRestoreArray(S_VVt,&y);CHKERRQ(ierr); 200706656605SStefano Zampini ierr = MatCopy(S_VVt,S_VV,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 2008d16cbb6bSStefano Zampini ierr = MatDestroy(&S_VVt);CHKERRQ(ierr); 2009019a44ceSStefano Zampini } else { 2010d16cbb6bSStefano Zampini ierr = MatCopy(A_VV,S_VV,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 2011d16cbb6bSStefano Zampini } 201221eccb56SStefano Zampini ierr = MatDestroy(&A_VV);CHKERRQ(ierr); 2013d16cbb6bSStefano Zampini 201406656605SStefano Zampini /* coarse basis functions */ 201506656605SStefano Zampini for (i=0;i<n_vertices;i++) { 201616f15bc4SStefano Zampini PetscScalar *y; 201716f15bc4SStefano Zampini 2018ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+lda_rhs*i);CHKERRQ(ierr); 201906656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 202006656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*i);CHKERRQ(ierr); 202106656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 202206656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 202306656605SStefano Zampini y[n_B*i+idx_V_B[i]] = 1.0; 202406656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 202506656605SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 202606656605SStefano Zampini 202706656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 20284f1b2e48SStefano Zampini PetscInt j; 20294f1b2e48SStefano Zampini 203006656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 203106656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*i);CHKERRQ(ierr); 203206656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 203306656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 203406656605SStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 20354f1b2e48SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) y[n_D*i+p0_lidx_I[j]] = 0.0; 203606656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 203706656605SStefano Zampini } 203806656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 203906656605SStefano Zampini } 204004708bb6SStefano Zampini /* if n_R == 0 the object is not destroyed */ 204104708bb6SStefano Zampini ierr = MatDestroy(&A_RV);CHKERRQ(ierr); 204206656605SStefano Zampini } 20435cbda25cSStefano Zampini ierr = VecDestroy(&dummy_vec);CHKERRQ(ierr); 204406656605SStefano Zampini 204506656605SStefano Zampini if (n_constraints) { 204606656605SStefano Zampini Mat B; 204706656605SStefano Zampini 2048ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_constraints,work,&B);CHKERRQ(ierr); 204906656605SStefano Zampini ierr = MatScale(S_CC,m_one);CHKERRQ(ierr); 205080677318SStefano Zampini ierr = MatMatMult(local_auxmat2_R,S_CC,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B);CHKERRQ(ierr); 205106656605SStefano Zampini ierr = MatScale(S_CC,m_one);CHKERRQ(ierr); 205206656605SStefano Zampini if (n_vertices) { 205380677318SStefano Zampini if (isCHOL) { /* if we can solve the interior problem with cholesky, we should also be fine with transposing here */ 205480677318SStefano Zampini ierr = MatTranspose(S_CV,MAT_REUSE_MATRIX,&S_VC);CHKERRQ(ierr); 205580677318SStefano Zampini } else { 205680677318SStefano Zampini Mat S_VCt; 205780677318SStefano Zampini 2058ffd830a3SStefano Zampini if (lda_rhs != n_R) { 2059ffd830a3SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 206072b8c272SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_constraints,work,&B);CHKERRQ(ierr); 2061ffd830a3SStefano Zampini ierr = MatSeqDenseSetLDA(B,lda_rhs);CHKERRQ(ierr); 2062ffd830a3SStefano Zampini } 206380677318SStefano Zampini ierr = MatMatMult(A_VR,B,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&S_VCt);CHKERRQ(ierr); 206480677318SStefano Zampini ierr = MatCopy(S_VCt,S_VC,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 206580677318SStefano Zampini ierr = MatDestroy(&S_VCt);CHKERRQ(ierr); 206680677318SStefano Zampini } 206706656605SStefano Zampini } 206806656605SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 206906656605SStefano Zampini /* coarse basis functions */ 207006656605SStefano Zampini for (i=0;i<n_constraints;i++) { 207106656605SStefano Zampini PetscScalar *y; 207206656605SStefano Zampini 2073ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+lda_rhs*i);CHKERRQ(ierr); 207406656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 207506656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*(i+n_vertices));CHKERRQ(ierr); 207606656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 207706656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 207806656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 207906656605SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 208006656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 20814f1b2e48SStefano Zampini PetscInt j; 20824f1b2e48SStefano Zampini 208306656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 208406656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*(i+n_vertices));CHKERRQ(ierr); 208506656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 208606656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 208706656605SStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 20884f1b2e48SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) y[n_D*i+p0_lidx_I[j]] = 0.0; 208906656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 209006656605SStefano Zampini } 209106656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 209206656605SStefano Zampini } 209306656605SStefano Zampini } 209480677318SStefano Zampini if (n_constraints) { 209580677318SStefano Zampini ierr = MatDestroy(&local_auxmat2_R);CHKERRQ(ierr); 209680677318SStefano Zampini } 20974f1b2e48SStefano Zampini ierr = PetscFree(p0_lidx_I);CHKERRQ(ierr); 209872b8c272SStefano Zampini 209972b8c272SStefano Zampini /* coarse matrix entries relative to B_0 */ 210072b8c272SStefano Zampini if (pcbddc->benign_n) { 210172b8c272SStefano Zampini Mat B0_B,B0_BPHI; 210272b8c272SStefano Zampini IS is_dummy; 210372b8c272SStefano Zampini PetscScalar *data; 210472b8c272SStefano Zampini PetscInt j; 210572b8c272SStefano Zampini 210672b8c272SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcbddc->benign_n,0,1,&is_dummy);CHKERRQ(ierr); 210772b8c272SStefano Zampini ierr = MatGetSubMatrix(pcbddc->benign_B0,is_dummy,pcis->is_B_local,MAT_INITIAL_MATRIX,&B0_B);CHKERRQ(ierr); 210872b8c272SStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 210972b8c272SStefano Zampini ierr = MatMatMult(B0_B,pcbddc->coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&B0_BPHI);CHKERRQ(ierr); 211072b8c272SStefano Zampini ierr = MatConvert(B0_BPHI,MATSEQDENSE,MAT_REUSE_MATRIX,&B0_BPHI);CHKERRQ(ierr); 211172b8c272SStefano Zampini ierr = MatDenseGetArray(B0_BPHI,&data);CHKERRQ(ierr); 211272b8c272SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) { 211372b8c272SStefano Zampini PetscInt primal_idx = pcbddc->local_primal_size - pcbddc->benign_n + j; 211472b8c272SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 211572b8c272SStefano Zampini coarse_submat_vals[primal_idx*pcbddc->local_primal_size+i] = data[i*pcbddc->benign_n+j]; 211672b8c272SStefano Zampini coarse_submat_vals[i*pcbddc->local_primal_size+primal_idx] = data[i*pcbddc->benign_n+j]; 211772b8c272SStefano Zampini } 211872b8c272SStefano Zampini } 211972b8c272SStefano Zampini ierr = MatDenseRestoreArray(B0_BPHI,&data);CHKERRQ(ierr); 212072b8c272SStefano Zampini ierr = MatDestroy(&B0_B);CHKERRQ(ierr); 212172b8c272SStefano Zampini ierr = MatDestroy(&B0_BPHI);CHKERRQ(ierr); 212272b8c272SStefano Zampini } 2123019a44ceSStefano Zampini 212406656605SStefano Zampini /* compute other basis functions for non-symmetric problems */ 21253301b35fSStefano Zampini if (!pcbddc->symmetric_primal) { 2126ffd830a3SStefano Zampini Mat B_V=NULL,B_C=NULL; 2127ffd830a3SStefano Zampini PetscScalar *marray; 212806656605SStefano Zampini 212906656605SStefano Zampini if (n_constraints) { 2130ffd830a3SStefano Zampini Mat S_CCT,C_CRT; 213106656605SStefano Zampini 2132ffd830a3SStefano Zampini ierr = MatTranspose(C_CR,MAT_INITIAL_MATRIX,&C_CRT);CHKERRQ(ierr); 213306656605SStefano Zampini ierr = MatTranspose(S_CC,MAT_INITIAL_MATRIX,&S_CCT);CHKERRQ(ierr); 2134ffd830a3SStefano Zampini ierr = MatMatMult(C_CRT,S_CCT,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&B_C);CHKERRQ(ierr); 213516f15bc4SStefano Zampini ierr = MatDestroy(&S_CCT);CHKERRQ(ierr); 213606656605SStefano Zampini if (n_vertices) { 2137ffd830a3SStefano Zampini Mat S_VCT; 213806656605SStefano Zampini 213906656605SStefano Zampini ierr = MatTranspose(S_VC,MAT_INITIAL_MATRIX,&S_VCT);CHKERRQ(ierr); 2140ffd830a3SStefano Zampini ierr = MatMatMult(C_CRT,S_VCT,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&B_V);CHKERRQ(ierr); 214116f15bc4SStefano Zampini ierr = MatDestroy(&S_VCT);CHKERRQ(ierr); 214206656605SStefano Zampini } 2143ffd830a3SStefano Zampini ierr = MatDestroy(&C_CRT);CHKERRQ(ierr); 214406656605SStefano Zampini } 214516f15bc4SStefano Zampini if (n_vertices && n_R) { 2146ffd830a3SStefano Zampini PetscScalar *av,*marray; 2147ffd830a3SStefano Zampini const PetscInt *xadj,*adjncy; 2148ffd830a3SStefano Zampini PetscInt n; 2149ffd830a3SStefano Zampini PetscBool flg_row; 215006656605SStefano Zampini 2151ffd830a3SStefano Zampini /* B_V = B_V - A_VR^T */ 2152ffd830a3SStefano Zampini ierr = MatConvert(A_VR,MATSEQAIJ,MAT_REUSE_MATRIX,&A_VR);CHKERRQ(ierr); 2153ffd830a3SStefano Zampini ierr = MatGetRowIJ(A_VR,0,PETSC_FALSE,PETSC_FALSE,&n,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 2154ffd830a3SStefano Zampini ierr = MatSeqAIJGetArray(A_VR,&av);CHKERRQ(ierr); 2155ffd830a3SStefano Zampini ierr = MatDenseGetArray(B_V,&marray);CHKERRQ(ierr); 2156ffd830a3SStefano Zampini for (i=0;i<n;i++) { 2157ffd830a3SStefano Zampini PetscInt j; 2158ffd830a3SStefano Zampini for (j=xadj[i];j<xadj[i+1];j++) marray[i*n_R + adjncy[j]] -= av[j]; 2159ffd830a3SStefano Zampini } 2160ffd830a3SStefano Zampini ierr = MatDenseRestoreArray(B_V,&marray);CHKERRQ(ierr); 2161ffd830a3SStefano Zampini ierr = MatRestoreRowIJ(A_VR,0,PETSC_FALSE,PETSC_FALSE,&n,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 2162ffd830a3SStefano Zampini ierr = MatDestroy(&A_VR);CHKERRQ(ierr); 216306656605SStefano Zampini } 216406656605SStefano Zampini 2165ffd830a3SStefano Zampini /* currently there's no support for MatTransposeMatSolve(F,B,X) */ 2166ffd830a3SStefano Zampini ierr = MatDenseGetArray(B_V,&marray);CHKERRQ(ierr); 2167ffd830a3SStefano Zampini for (i=0;i<n_vertices;i++) { 2168ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,marray+i*n_R);CHKERRQ(ierr); 2169ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,work+i*n_R);CHKERRQ(ierr); 217006656605SStefano Zampini ierr = KSPSolveTranspose(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 217106656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 217206656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 217306656605SStefano Zampini } 2174ffd830a3SStefano Zampini ierr = MatDenseRestoreArray(B_V,&marray);CHKERRQ(ierr); 2175ffd830a3SStefano Zampini ierr = MatDenseGetArray(B_C,&marray);CHKERRQ(ierr); 2176ffd830a3SStefano Zampini for (i=n_vertices;i<n_constraints+n_vertices;i++) { 2177ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,marray+(i-n_vertices)*n_R);CHKERRQ(ierr); 2178ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,work+i*n_R);CHKERRQ(ierr); 2179ffd830a3SStefano Zampini ierr = KSPSolveTranspose(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 2180ffd830a3SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 2181ffd830a3SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 218206656605SStefano Zampini } 2183ffd830a3SStefano Zampini ierr = MatDenseRestoreArray(B_C,&marray);CHKERRQ(ierr); 218406656605SStefano Zampini /* coarse basis functions */ 218506656605SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 218606656605SStefano Zampini PetscScalar *y; 218706656605SStefano Zampini 2188ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*n_R);CHKERRQ(ierr); 218906656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_psi_B,&y);CHKERRQ(ierr); 219006656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*i);CHKERRQ(ierr); 219106656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 219206656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 219306656605SStefano Zampini if (i<n_vertices) { 219406656605SStefano Zampini y[n_B*i+idx_V_B[i]] = 1.0; 219506656605SStefano Zampini } 219606656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_psi_B,&y);CHKERRQ(ierr); 219706656605SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 219806656605SStefano Zampini 219906656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 220006656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_psi_D,&y);CHKERRQ(ierr); 220106656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*i);CHKERRQ(ierr); 220206656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 220306656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 220406656605SStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 220506656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_psi_D,&y);CHKERRQ(ierr); 220606656605SStefano Zampini } 220706656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 220806656605SStefano Zampini } 2209ffd830a3SStefano Zampini ierr = MatDestroy(&B_V);CHKERRQ(ierr); 2210ffd830a3SStefano Zampini ierr = MatDestroy(&B_C);CHKERRQ(ierr); 221106656605SStefano Zampini } 2212d62866d3SStefano Zampini /* free memory */ 221388ebb749SStefano Zampini ierr = PetscFree(idx_V_B);CHKERRQ(ierr); 221406656605SStefano Zampini ierr = MatDestroy(&S_VV);CHKERRQ(ierr); 221506656605SStefano Zampini ierr = MatDestroy(&S_CV);CHKERRQ(ierr); 221606656605SStefano Zampini ierr = MatDestroy(&S_VC);CHKERRQ(ierr); 221706656605SStefano Zampini ierr = MatDestroy(&S_CC);CHKERRQ(ierr); 2218d62866d3SStefano Zampini ierr = PetscFree(work);CHKERRQ(ierr); 2219d62866d3SStefano Zampini if (n_vertices) { 2220d62866d3SStefano Zampini ierr = MatDestroy(&A_VR);CHKERRQ(ierr); 2221d62866d3SStefano Zampini } 2222d62866d3SStefano Zampini if (n_constraints) { 2223d62866d3SStefano Zampini ierr = MatDestroy(&C_CR);CHKERRQ(ierr); 2224d62866d3SStefano Zampini } 222588ebb749SStefano Zampini /* Checking coarse_sub_mat and coarse basis functios */ 222688ebb749SStefano Zampini /* Symmetric case : It should be \Phi^{(j)^T} A^{(j)} \Phi^{(j)}=coarse_sub_mat */ 222788ebb749SStefano Zampini /* Non-symmetric case : It should be \Psi^{(j)^T} A^{(j)} \Phi^{(j)}=coarse_sub_mat */ 2228d12edf2fSStefano Zampini if (pcbddc->dbg_flag) { 222988ebb749SStefano Zampini Mat coarse_sub_mat; 223025084f0cSStefano Zampini Mat AUXMAT,TM1,TM2,TM3,TM4; 223188ebb749SStefano Zampini Mat coarse_phi_D,coarse_phi_B; 223288ebb749SStefano Zampini Mat coarse_psi_D,coarse_psi_B; 223388ebb749SStefano Zampini Mat A_II,A_BB,A_IB,A_BI; 22348bec7fa6SStefano Zampini Mat C_B,CPHI; 22358bec7fa6SStefano Zampini IS is_dummy; 22368bec7fa6SStefano Zampini Vec mones; 223788ebb749SStefano Zampini MatType checkmattype=MATSEQAIJ; 223888ebb749SStefano Zampini PetscReal real_value; 223988ebb749SStefano Zampini 2240a3df083aSStefano Zampini if (pcbddc->benign_n && !pcbddc->benign_change_explicit) { 2241a3df083aSStefano Zampini Mat A; 2242a3df083aSStefano Zampini ierr = PCBDDCBenignProject(pc,NULL,NULL,&A);CHKERRQ(ierr); 2243a3df083aSStefano Zampini ierr = MatGetSubMatrix(A,pcis->is_I_local,pcis->is_I_local,MAT_INITIAL_MATRIX,&A_II);CHKERRQ(ierr); 2244a3df083aSStefano Zampini ierr = MatGetSubMatrix(A,pcis->is_I_local,pcis->is_B_local,MAT_INITIAL_MATRIX,&A_IB);CHKERRQ(ierr); 2245a3df083aSStefano Zampini ierr = MatGetSubMatrix(A,pcis->is_B_local,pcis->is_I_local,MAT_INITIAL_MATRIX,&A_BI);CHKERRQ(ierr); 2246a3df083aSStefano Zampini ierr = MatGetSubMatrix(A,pcis->is_B_local,pcis->is_B_local,MAT_INITIAL_MATRIX,&A_BB);CHKERRQ(ierr); 2247a3df083aSStefano Zampini ierr = MatDestroy(&A);CHKERRQ(ierr); 2248a3df083aSStefano Zampini } else { 224988ebb749SStefano Zampini ierr = MatConvert(pcis->A_II,checkmattype,MAT_INITIAL_MATRIX,&A_II);CHKERRQ(ierr); 225088ebb749SStefano Zampini ierr = MatConvert(pcis->A_IB,checkmattype,MAT_INITIAL_MATRIX,&A_IB);CHKERRQ(ierr); 225188ebb749SStefano Zampini ierr = MatConvert(pcis->A_BI,checkmattype,MAT_INITIAL_MATRIX,&A_BI);CHKERRQ(ierr); 225288ebb749SStefano Zampini ierr = MatConvert(pcis->A_BB,checkmattype,MAT_INITIAL_MATRIX,&A_BB);CHKERRQ(ierr); 2253a3df083aSStefano Zampini } 225488ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_phi_D,checkmattype,MAT_INITIAL_MATRIX,&coarse_phi_D);CHKERRQ(ierr); 225588ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_phi_B,checkmattype,MAT_INITIAL_MATRIX,&coarse_phi_B);CHKERRQ(ierr); 2256ffd830a3SStefano Zampini if (!pcbddc->symmetric_primal) { 225788ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_psi_D,checkmattype,MAT_INITIAL_MATRIX,&coarse_psi_D);CHKERRQ(ierr); 225888ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_psi_B,checkmattype,MAT_INITIAL_MATRIX,&coarse_psi_B);CHKERRQ(ierr); 225988ebb749SStefano Zampini } 226088ebb749SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,pcbddc->local_primal_size,pcbddc->local_primal_size,coarse_submat_vals,&coarse_sub_mat);CHKERRQ(ierr); 226188ebb749SStefano Zampini 226225084f0cSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 22633301b35fSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Check coarse sub mat computation (symmetric %d)\n",pcbddc->symmetric_primal);CHKERRQ(ierr); 226425084f0cSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 2265ffd830a3SStefano Zampini if (!pcbddc->symmetric_primal) { 226688ebb749SStefano Zampini ierr = MatMatMult(A_II,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 226788ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM1);CHKERRQ(ierr); 226888ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 226988ebb749SStefano Zampini ierr = MatMatMult(A_BB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 227088ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM2);CHKERRQ(ierr); 227188ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 227288ebb749SStefano Zampini ierr = MatMatMult(A_IB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 227388ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM3);CHKERRQ(ierr); 227488ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 227588ebb749SStefano Zampini ierr = MatMatMult(A_BI,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 227688ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM4);CHKERRQ(ierr); 227788ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 227888ebb749SStefano Zampini } else { 227988ebb749SStefano Zampini ierr = MatPtAP(A_II,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&TM1);CHKERRQ(ierr); 228088ebb749SStefano Zampini ierr = MatPtAP(A_BB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&TM2);CHKERRQ(ierr); 228188ebb749SStefano Zampini ierr = MatMatMult(A_IB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 228288ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_phi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM3);CHKERRQ(ierr); 228388ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 228488ebb749SStefano Zampini ierr = MatMatMult(A_BI,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 228588ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_phi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM4);CHKERRQ(ierr); 228688ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 228788ebb749SStefano Zampini } 228888ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM2,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 228988ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM3,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 229088ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM4,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 229188ebb749SStefano Zampini ierr = MatConvert(TM1,MATSEQDENSE,MAT_REUSE_MATRIX,&TM1);CHKERRQ(ierr); 22924f1b2e48SStefano Zampini if (pcbddc->benign_n) { 2293fc227af8SStefano Zampini Mat B0_B,B0_BPHI; 2294d12edf2fSStefano Zampini PetscScalar *data,*data2; 22954f1b2e48SStefano Zampini PetscInt j; 2296d12edf2fSStefano Zampini 22974f1b2e48SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcbddc->benign_n,0,1,&is_dummy);CHKERRQ(ierr); 2298fc227af8SStefano Zampini ierr = MatGetSubMatrix(pcbddc->benign_B0,is_dummy,pcis->is_B_local,MAT_INITIAL_MATRIX,&B0_B);CHKERRQ(ierr); 2299d12edf2fSStefano Zampini ierr = MatMatMult(B0_B,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&B0_BPHI);CHKERRQ(ierr); 2300d12edf2fSStefano Zampini ierr = MatConvert(B0_BPHI,MATSEQDENSE,MAT_REUSE_MATRIX,&B0_BPHI);CHKERRQ(ierr); 2301d12edf2fSStefano Zampini ierr = MatDenseGetArray(TM1,&data);CHKERRQ(ierr); 2302d12edf2fSStefano Zampini ierr = MatDenseGetArray(B0_BPHI,&data2);CHKERRQ(ierr); 23034f1b2e48SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) { 23044f1b2e48SStefano Zampini PetscInt primal_idx = pcbddc->local_primal_size - pcbddc->benign_n + j; 2305d12edf2fSStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 23064f1b2e48SStefano Zampini data[primal_idx*pcbddc->local_primal_size+i] += data2[i*pcbddc->benign_n+j]; 23074f1b2e48SStefano Zampini data[i*pcbddc->local_primal_size+primal_idx] += data2[i*pcbddc->benign_n+j]; 23084f1b2e48SStefano Zampini } 2309d12edf2fSStefano Zampini } 2310d12edf2fSStefano Zampini ierr = MatDenseRestoreArray(TM1,&data);CHKERRQ(ierr); 2311d12edf2fSStefano Zampini ierr = MatDenseRestoreArray(B0_BPHI,&data2);CHKERRQ(ierr); 2312d12edf2fSStefano Zampini ierr = MatDestroy(&B0_B);CHKERRQ(ierr); 2313d12edf2fSStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 2314d12edf2fSStefano Zampini ierr = MatDestroy(&B0_BPHI);CHKERRQ(ierr); 2315d12edf2fSStefano Zampini } 2316d12edf2fSStefano Zampini #if 0 2317d12edf2fSStefano Zampini { 2318d12edf2fSStefano Zampini PetscViewer viewer; 2319d12edf2fSStefano Zampini char filename[256]; 2320ffd830a3SStefano Zampini sprintf(filename,"details_local_coarse_mat%d_level%d.m",PetscGlobalRank,pcbddc->current_level); 2321d12edf2fSStefano Zampini ierr = PetscViewerASCIIOpen(PETSC_COMM_SELF,filename,&viewer);CHKERRQ(ierr); 2322d12edf2fSStefano Zampini ierr = PetscViewerSetFormat(viewer,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 2323ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)coarse_sub_mat,"computed");CHKERRQ(ierr); 2324ffd830a3SStefano Zampini ierr = MatView(coarse_sub_mat,viewer);CHKERRQ(ierr); 2325ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)TM1,"projected");CHKERRQ(ierr); 2326d12edf2fSStefano Zampini ierr = MatView(TM1,viewer);CHKERRQ(ierr); 232772b8c272SStefano Zampini if (save_change) { 232872b8c272SStefano Zampini Mat phi_B; 232972b8c272SStefano Zampini ierr = MatMatMult(save_change,pcbddc->coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&phi_B);CHKERRQ(ierr); 233072b8c272SStefano Zampini ierr = PetscObjectSetName((PetscObject)phi_B,"phi_B");CHKERRQ(ierr); 233172b8c272SStefano Zampini ierr = MatView(phi_B,viewer);CHKERRQ(ierr); 233272b8c272SStefano Zampini ierr = MatDestroy(&phi_B);CHKERRQ(ierr); 233372b8c272SStefano Zampini } else { 2334ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)pcbddc->coarse_phi_B,"phi_B");CHKERRQ(ierr); 2335ffd830a3SStefano Zampini ierr = MatView(pcbddc->coarse_phi_B,viewer);CHKERRQ(ierr); 233672b8c272SStefano Zampini } 2337ffd830a3SStefano Zampini if (pcbddc->coarse_phi_D) { 2338ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)pcbddc->coarse_phi_D,"phi_D");CHKERRQ(ierr); 2339ffd830a3SStefano Zampini ierr = MatView(pcbddc->coarse_phi_D,viewer);CHKERRQ(ierr); 2340ffd830a3SStefano Zampini } 2341ffd830a3SStefano Zampini if (pcbddc->coarse_psi_B) { 2342ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)pcbddc->coarse_psi_B,"psi_B");CHKERRQ(ierr); 2343ffd830a3SStefano Zampini ierr = MatView(pcbddc->coarse_psi_B,viewer);CHKERRQ(ierr); 2344ffd830a3SStefano Zampini } 234572b8c272SStefano Zampini if (pcbddc->coarse_psi_D) { 2346ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)pcbddc->coarse_psi_D,"psi_D");CHKERRQ(ierr); 2347ffd830a3SStefano Zampini ierr = MatView(pcbddc->coarse_psi_D,viewer);CHKERRQ(ierr); 2348ffd830a3SStefano Zampini } 2349d12edf2fSStefano Zampini ierr = PetscViewerDestroy(&viewer);CHKERRQ(ierr); 2350d12edf2fSStefano Zampini } 2351d12edf2fSStefano Zampini #endif 235281d9aea3SBarry Smith ierr = MatAXPY(TM1,m_one,coarse_sub_mat,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 23538bec7fa6SStefano Zampini ierr = MatNorm(TM1,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 23541575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 235506656605SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d matrix error % 1.14e\n",PetscGlobalRank,real_value);CHKERRQ(ierr); 23568bec7fa6SStefano Zampini 23578bec7fa6SStefano Zampini /* check constraints */ 2358*a00504b5SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcbddc->local_primal_size-pcbddc->benign_n,0,1,&is_dummy);CHKERRQ(ierr); 2359*a00504b5SStefano Zampini ierr = MatGetSubMatrix(pcbddc->ConstraintMatrix,is_dummy,pcis->is_B_local,MAT_INITIAL_MATRIX,&C_B);CHKERRQ(ierr); 23604f1b2e48SStefano Zampini if (!pcbddc->benign_n) { /* TODO: add benign case */ 23618bec7fa6SStefano Zampini ierr = MatMatMult(C_B,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&CPHI);CHKERRQ(ierr); 2362*a00504b5SStefano Zampini } else { 2363*a00504b5SStefano Zampini PetscScalar *data; 2364*a00504b5SStefano Zampini Mat tmat; 2365*a00504b5SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&data);CHKERRQ(ierr); 2366*a00504b5SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,pcis->n_B,pcbddc->local_primal_size-pcbddc->benign_n,data,&tmat);CHKERRQ(ierr); 2367*a00504b5SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_B,&data);CHKERRQ(ierr); 2368*a00504b5SStefano Zampini ierr = MatMatMult(C_B,tmat,MAT_INITIAL_MATRIX,1.0,&CPHI);CHKERRQ(ierr); 2369*a00504b5SStefano Zampini ierr = MatDestroy(&tmat);CHKERRQ(ierr); 2370*a00504b5SStefano Zampini } 23718bec7fa6SStefano Zampini ierr = MatCreateVecs(CPHI,&mones,NULL);CHKERRQ(ierr); 23728bec7fa6SStefano Zampini ierr = VecSet(mones,-1.0);CHKERRQ(ierr); 23738bec7fa6SStefano Zampini ierr = MatDiagonalSet(CPHI,mones,ADD_VALUES);CHKERRQ(ierr); 23748bec7fa6SStefano Zampini ierr = MatNorm(CPHI,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 2375bdae7319SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d phi constraints error % 1.14e\n",PetscGlobalRank,real_value);CHKERRQ(ierr); 2376ffd830a3SStefano Zampini if (!pcbddc->symmetric_primal) { 2377bdae7319SStefano Zampini ierr = MatMatMult(C_B,coarse_psi_B,MAT_REUSE_MATRIX,1.0,&CPHI);CHKERRQ(ierr); 2378bdae7319SStefano Zampini ierr = VecSet(mones,-1.0);CHKERRQ(ierr); 2379bdae7319SStefano Zampini ierr = MatDiagonalSet(CPHI,mones,ADD_VALUES);CHKERRQ(ierr); 2380bdae7319SStefano Zampini ierr = MatNorm(CPHI,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 2381bdae7319SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d psi constraints error % 1.14e\n",PetscGlobalRank,real_value);CHKERRQ(ierr); 238288ebb749SStefano Zampini } 23838bec7fa6SStefano Zampini ierr = MatDestroy(&C_B);CHKERRQ(ierr); 23848bec7fa6SStefano Zampini ierr = MatDestroy(&CPHI);CHKERRQ(ierr); 23858bec7fa6SStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 23868bec7fa6SStefano Zampini ierr = VecDestroy(&mones);CHKERRQ(ierr); 238725084f0cSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 238888ebb749SStefano Zampini ierr = MatDestroy(&A_II);CHKERRQ(ierr); 238988ebb749SStefano Zampini ierr = MatDestroy(&A_BB);CHKERRQ(ierr); 239088ebb749SStefano Zampini ierr = MatDestroy(&A_IB);CHKERRQ(ierr); 239188ebb749SStefano Zampini ierr = MatDestroy(&A_BI);CHKERRQ(ierr); 239288ebb749SStefano Zampini ierr = MatDestroy(&TM1);CHKERRQ(ierr); 239388ebb749SStefano Zampini ierr = MatDestroy(&TM2);CHKERRQ(ierr); 239488ebb749SStefano Zampini ierr = MatDestroy(&TM3);CHKERRQ(ierr); 239588ebb749SStefano Zampini ierr = MatDestroy(&TM4);CHKERRQ(ierr); 239688ebb749SStefano Zampini ierr = MatDestroy(&coarse_phi_D);CHKERRQ(ierr); 239788ebb749SStefano Zampini ierr = MatDestroy(&coarse_phi_B);CHKERRQ(ierr); 2398ffd830a3SStefano Zampini if (!pcbddc->symmetric_primal) { 239988ebb749SStefano Zampini ierr = MatDestroy(&coarse_psi_D);CHKERRQ(ierr); 240088ebb749SStefano Zampini ierr = MatDestroy(&coarse_psi_B);CHKERRQ(ierr); 240188ebb749SStefano Zampini } 240288ebb749SStefano Zampini ierr = MatDestroy(&coarse_sub_mat);CHKERRQ(ierr); 240388ebb749SStefano Zampini } 24048629588bSStefano Zampini /* get back data */ 24058629588bSStefano Zampini *coarse_submat_vals_n = coarse_submat_vals; 240688ebb749SStefano Zampini PetscFunctionReturn(0); 240788ebb749SStefano Zampini } 240888ebb749SStefano Zampini 240988ebb749SStefano Zampini #undef __FUNCT__ 2410d65f70fdSStefano Zampini #define __FUNCT__ "MatGetSubMatrixUnsorted" 2411d648f858SStefano Zampini PetscErrorCode MatGetSubMatrixUnsorted(Mat A, IS isrow, IS iscol, Mat* B) 2412aa0d41d4SStefano Zampini { 2413d65f70fdSStefano Zampini Mat *work_mat; 2414d65f70fdSStefano Zampini IS isrow_s,iscol_s; 2415d65f70fdSStefano Zampini PetscBool rsorted,csorted; 2416d65f70fdSStefano Zampini PetscInt rsize,*idxs_perm_r,csize,*idxs_perm_c; 2417aa0d41d4SStefano Zampini PetscErrorCode ierr; 2418aa0d41d4SStefano Zampini 2419aa0d41d4SStefano Zampini PetscFunctionBegin; 2420d65f70fdSStefano Zampini ierr = ISSorted(isrow,&rsorted);CHKERRQ(ierr); 2421d65f70fdSStefano Zampini ierr = ISSorted(iscol,&csorted);CHKERRQ(ierr); 2422d65f70fdSStefano Zampini ierr = ISGetLocalSize(isrow,&rsize);CHKERRQ(ierr); 2423d65f70fdSStefano Zampini ierr = ISGetLocalSize(iscol,&csize);CHKERRQ(ierr); 2424aa0d41d4SStefano Zampini 2425d65f70fdSStefano Zampini if (!rsorted) { 2426906d46d4SStefano Zampini const PetscInt *idxs; 2427906d46d4SStefano Zampini PetscInt *idxs_sorted,i; 2428aa0d41d4SStefano Zampini 2429d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_perm_r);CHKERRQ(ierr); 2430d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_sorted);CHKERRQ(ierr); 2431d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 2432d65f70fdSStefano Zampini idxs_perm_r[i] = i; 2433aa0d41d4SStefano Zampini } 2434d65f70fdSStefano Zampini ierr = ISGetIndices(isrow,&idxs);CHKERRQ(ierr); 2435d65f70fdSStefano Zampini ierr = PetscSortIntWithPermutation(rsize,idxs,idxs_perm_r);CHKERRQ(ierr); 2436d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 2437d65f70fdSStefano Zampini idxs_sorted[i] = idxs[idxs_perm_r[i]]; 2438aa0d41d4SStefano Zampini } 2439d65f70fdSStefano Zampini ierr = ISRestoreIndices(isrow,&idxs);CHKERRQ(ierr); 2440d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,rsize,idxs_sorted,PETSC_OWN_POINTER,&isrow_s);CHKERRQ(ierr); 2441d65f70fdSStefano Zampini } else { 2442d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)isrow);CHKERRQ(ierr); 2443d65f70fdSStefano Zampini isrow_s = isrow; 2444aa0d41d4SStefano Zampini } 2445906d46d4SStefano Zampini 2446d65f70fdSStefano Zampini if (!csorted) { 2447d65f70fdSStefano Zampini if (isrow == iscol) { 2448d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)isrow_s);CHKERRQ(ierr); 2449d65f70fdSStefano Zampini iscol_s = isrow_s; 2450d65f70fdSStefano Zampini } else { 2451d65f70fdSStefano Zampini const PetscInt *idxs; 2452d65f70fdSStefano Zampini PetscInt *idxs_sorted,i; 2453906d46d4SStefano Zampini 2454d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_perm_c);CHKERRQ(ierr); 2455d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_sorted);CHKERRQ(ierr); 2456d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 2457d65f70fdSStefano Zampini idxs_perm_c[i] = i; 2458d65f70fdSStefano Zampini } 2459d65f70fdSStefano Zampini ierr = ISGetIndices(iscol,&idxs);CHKERRQ(ierr); 2460d65f70fdSStefano Zampini ierr = PetscSortIntWithPermutation(csize,idxs,idxs_perm_c);CHKERRQ(ierr); 2461d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 2462d65f70fdSStefano Zampini idxs_sorted[i] = idxs[idxs_perm_c[i]]; 2463d65f70fdSStefano Zampini } 2464d65f70fdSStefano Zampini ierr = ISRestoreIndices(iscol,&idxs);CHKERRQ(ierr); 2465d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,csize,idxs_sorted,PETSC_OWN_POINTER,&iscol_s);CHKERRQ(ierr); 2466d65f70fdSStefano Zampini } 2467d65f70fdSStefano Zampini } else { 2468d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)iscol);CHKERRQ(ierr); 2469d65f70fdSStefano Zampini iscol_s = iscol; 2470d65f70fdSStefano Zampini } 2471d65f70fdSStefano Zampini 2472d648f858SStefano Zampini ierr = MatGetSubMatrices(A,1,&isrow_s,&iscol_s,MAT_INITIAL_MATRIX,&work_mat);CHKERRQ(ierr); 2473d65f70fdSStefano Zampini 2474d65f70fdSStefano Zampini if (!rsorted || !csorted) { 2475906d46d4SStefano Zampini Mat new_mat; 2476d65f70fdSStefano Zampini IS is_perm_r,is_perm_c; 2477906d46d4SStefano Zampini 2478d65f70fdSStefano Zampini if (!rsorted) { 2479d65f70fdSStefano Zampini PetscInt *idxs_r,i; 2480d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_r);CHKERRQ(ierr); 2481d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 2482d65f70fdSStefano Zampini idxs_r[idxs_perm_r[i]] = i; 2483906d46d4SStefano Zampini } 2484d65f70fdSStefano Zampini ierr = PetscFree(idxs_perm_r);CHKERRQ(ierr); 2485d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,rsize,idxs_r,PETSC_OWN_POINTER,&is_perm_r);CHKERRQ(ierr); 2486d65f70fdSStefano Zampini } else { 2487d65f70fdSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,rsize,0,1,&is_perm_r);CHKERRQ(ierr); 2488906d46d4SStefano Zampini } 2489d65f70fdSStefano Zampini ierr = ISSetPermutation(is_perm_r);CHKERRQ(ierr); 2490d65f70fdSStefano Zampini 2491d65f70fdSStefano Zampini if (!csorted) { 2492d65f70fdSStefano Zampini if (isrow_s == iscol_s) { 2493d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)is_perm_r);CHKERRQ(ierr); 2494d65f70fdSStefano Zampini is_perm_c = is_perm_r; 2495d65f70fdSStefano Zampini } else { 2496d65f70fdSStefano Zampini PetscInt *idxs_c,i; 2497d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_c);CHKERRQ(ierr); 2498d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 2499d65f70fdSStefano Zampini idxs_c[idxs_perm_c[i]] = i; 2500d65f70fdSStefano Zampini } 2501d65f70fdSStefano Zampini ierr = PetscFree(idxs_perm_c);CHKERRQ(ierr); 2502d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,csize,idxs_c,PETSC_OWN_POINTER,&is_perm_c);CHKERRQ(ierr); 2503d65f70fdSStefano Zampini } 2504d65f70fdSStefano Zampini } else { 2505d65f70fdSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,csize,0,1,&is_perm_c);CHKERRQ(ierr); 2506d65f70fdSStefano Zampini } 2507d65f70fdSStefano Zampini ierr = ISSetPermutation(is_perm_c);CHKERRQ(ierr); 2508d65f70fdSStefano Zampini 2509d65f70fdSStefano Zampini ierr = MatPermute(work_mat[0],is_perm_r,is_perm_c,&new_mat);CHKERRQ(ierr); 2510d65f70fdSStefano Zampini ierr = MatDestroy(&work_mat[0]);CHKERRQ(ierr); 2511d65f70fdSStefano Zampini work_mat[0] = new_mat; 2512d65f70fdSStefano Zampini ierr = ISDestroy(&is_perm_r);CHKERRQ(ierr); 2513d65f70fdSStefano Zampini ierr = ISDestroy(&is_perm_c);CHKERRQ(ierr); 2514d65f70fdSStefano Zampini } 2515d65f70fdSStefano Zampini 2516d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)work_mat[0]);CHKERRQ(ierr); 2517d65f70fdSStefano Zampini *B = work_mat[0]; 2518d65f70fdSStefano Zampini ierr = MatDestroyMatrices(1,&work_mat);CHKERRQ(ierr); 2519d65f70fdSStefano Zampini ierr = ISDestroy(&isrow_s);CHKERRQ(ierr); 2520d65f70fdSStefano Zampini ierr = ISDestroy(&iscol_s);CHKERRQ(ierr); 2521d65f70fdSStefano Zampini PetscFunctionReturn(0); 2522d65f70fdSStefano Zampini } 2523d65f70fdSStefano Zampini 2524d65f70fdSStefano Zampini #undef __FUNCT__ 25255e8657edSStefano Zampini #define __FUNCT__ "PCBDDCComputeLocalMatrix" 25265e8657edSStefano Zampini PetscErrorCode PCBDDCComputeLocalMatrix(PC pc, Mat ChangeOfBasisMatrix) 2527aa0d41d4SStefano Zampini { 2528aa0d41d4SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 25295e8657edSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 2530d65f70fdSStefano Zampini Mat new_mat; 25315e8657edSStefano Zampini IS is_local,is_global; 2532d65f70fdSStefano Zampini PetscInt local_size; 2533d65f70fdSStefano Zampini PetscBool isseqaij; 2534aa0d41d4SStefano Zampini PetscErrorCode ierr; 2535aa0d41d4SStefano Zampini 2536aa0d41d4SStefano Zampini PetscFunctionBegin; 2537aa0d41d4SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 25385e8657edSStefano Zampini ierr = MatGetSize(matis->A,&local_size,NULL);CHKERRQ(ierr); 25395e8657edSStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)matis->A),local_size,0,1,&is_local);CHKERRQ(ierr); 2540b087196eSStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(pc->pmat->rmap->mapping,is_local,&is_global);CHKERRQ(ierr); 2541aa0d41d4SStefano Zampini ierr = ISDestroy(&is_local);CHKERRQ(ierr); 2542d648f858SStefano Zampini ierr = MatGetSubMatrixUnsorted(ChangeOfBasisMatrix,is_global,is_global,&new_mat);CHKERRQ(ierr); 2543aa0d41d4SStefano Zampini ierr = ISDestroy(&is_global);CHKERRQ(ierr); 2544906d46d4SStefano Zampini 2545906d46d4SStefano Zampini /* check */ 2546906d46d4SStefano Zampini if (pcbddc->dbg_flag) { 2547906d46d4SStefano Zampini Vec x,x_change; 2548906d46d4SStefano Zampini PetscReal error; 2549906d46d4SStefano Zampini 25505e8657edSStefano Zampini ierr = MatCreateVecs(ChangeOfBasisMatrix,&x,&x_change);CHKERRQ(ierr); 2551906d46d4SStefano Zampini ierr = VecSetRandom(x,NULL);CHKERRQ(ierr); 25525e8657edSStefano Zampini ierr = MatMult(ChangeOfBasisMatrix,x,x_change);CHKERRQ(ierr); 2553e176bc59SStefano Zampini ierr = VecScatterBegin(matis->cctx,x,matis->x,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2554e176bc59SStefano Zampini ierr = VecScatterEnd(matis->cctx,x,matis->x,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2555d65f70fdSStefano Zampini ierr = MatMult(new_mat,matis->x,matis->y);CHKERRQ(ierr); 2556e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,matis->y,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2557e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,matis->y,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2558906d46d4SStefano Zampini ierr = VecAXPY(x,-1.0,x_change);CHKERRQ(ierr); 2559906d46d4SStefano Zampini ierr = VecNorm(x,NORM_INFINITY,&error);CHKERRQ(ierr); 2560906d46d4SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 2561906d46d4SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Error global vs local change on N: %1.6e\n",error);CHKERRQ(ierr); 2562906d46d4SStefano Zampini ierr = VecDestroy(&x);CHKERRQ(ierr); 2563906d46d4SStefano Zampini ierr = VecDestroy(&x_change);CHKERRQ(ierr); 2564906d46d4SStefano Zampini } 2565906d46d4SStefano Zampini 256622d5777bSStefano Zampini /* TODO: HOW TO WORK WITH BAIJ and SBAIJ and SEQDENSE? */ 25679b28b3ffSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)matis->A,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 256822d5777bSStefano Zampini if (isseqaij) { 2569*a00504b5SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 2570*a00504b5SStefano Zampini ierr = MatPtAP(matis->A,new_mat,MAT_INITIAL_MATRIX,2.0,&pcbddc->local_mat);CHKERRQ(ierr); 2571aa0d41d4SStefano Zampini } else { 2572*a00504b5SStefano Zampini Mat work_mat; 25731cf9b237SStefano Zampini 2574*a00504b5SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 2575aa0d41d4SStefano Zampini ierr = MatConvert(matis->A,MATSEQAIJ,MAT_INITIAL_MATRIX,&work_mat);CHKERRQ(ierr); 2576*a00504b5SStefano Zampini ierr = MatPtAP(work_mat,new_mat,MAT_INITIAL_MATRIX,2.0,&pcbddc->local_mat);CHKERRQ(ierr); 2577aa0d41d4SStefano Zampini } 25783301b35fSStefano Zampini if (matis->A->symmetric_set) { 25793301b35fSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_SYMMETRIC,matis->A->symmetric);CHKERRQ(ierr); 2580e496cd5dSStefano Zampini #if !defined(PETSC_USE_COMPLEX) 25813301b35fSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_HERMITIAN,matis->A->symmetric);CHKERRQ(ierr); 2582e496cd5dSStefano Zampini #endif 25833301b35fSStefano Zampini } 258445a1bb75SStefano Zampini /* 258545a1bb75SStefano Zampini ierr = PetscViewerSetFormat(PETSC_VIEWER_STDOUT_SELF,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 2586d65f70fdSStefano Zampini ierr = MatView(new_mat,(PetscViewer)0);CHKERRQ(ierr); 258745a1bb75SStefano Zampini */ 2588d65f70fdSStefano Zampini ierr = MatDestroy(&new_mat);CHKERRQ(ierr); 2589aa0d41d4SStefano Zampini PetscFunctionReturn(0); 2590aa0d41d4SStefano Zampini } 2591aa0d41d4SStefano Zampini 2592aa0d41d4SStefano Zampini #undef __FUNCT__ 2593a64d13efSStefano Zampini #define __FUNCT__ "PCBDDCSetUpLocalScatters" 25948ce42a96SStefano Zampini PetscErrorCode PCBDDCSetUpLocalScatters(PC pc) 2595a64d13efSStefano Zampini { 2596a64d13efSStefano Zampini PC_IS* pcis = (PC_IS*)(pc->data); 2597a64d13efSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 2598d62866d3SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 259953892102SStefano Zampini PetscInt *idx_R_local=NULL; 26003a50541eSStefano Zampini PetscInt n_vertices,i,j,n_R,n_D,n_B; 26013a50541eSStefano Zampini PetscInt vbs,bs; 26026816873aSStefano Zampini PetscBT bitmask=NULL; 2603a64d13efSStefano Zampini PetscErrorCode ierr; 2604a64d13efSStefano Zampini 2605a64d13efSStefano Zampini PetscFunctionBegin; 2606b23d619eSStefano Zampini /* 2607b23d619eSStefano Zampini No need to setup local scatters if 2608b23d619eSStefano Zampini - primal space is unchanged 2609b23d619eSStefano Zampini AND 2610b23d619eSStefano Zampini - we actually have locally some primal dofs (could not be true in multilevel or for isolated subdomains) 2611b23d619eSStefano Zampini AND 2612b23d619eSStefano Zampini - we are not in debugging mode (this is needed since there are Synchronized prints at the end of the subroutine 2613b23d619eSStefano Zampini */ 2614b23d619eSStefano Zampini if (!pcbddc->new_primal_space_local && pcbddc->local_primal_size && !pcbddc->dbg_flag) { 2615f4ddd8eeSStefano Zampini PetscFunctionReturn(0); 2616f4ddd8eeSStefano Zampini } 2617f4ddd8eeSStefano Zampini /* destroy old objects */ 2618f4ddd8eeSStefano Zampini ierr = ISDestroy(&pcbddc->is_R_local);CHKERRQ(ierr); 2619f4ddd8eeSStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_B);CHKERRQ(ierr); 2620f4ddd8eeSStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_D);CHKERRQ(ierr); 2621a64d13efSStefano Zampini /* Set Non-overlapping dimensions */ 2622b371cd4fSStefano Zampini n_B = pcis->n_B; 2623b371cd4fSStefano Zampini n_D = pcis->n - n_B; 2624b371cd4fSStefano Zampini n_vertices = pcbddc->n_vertices; 26253a50541eSStefano Zampini 2626a64d13efSStefano Zampini /* Dohrmann's notation: dofs splitted in R (Remaining: all dofs but the vertices) and V (Vertices) */ 26276816873aSStefano Zampini 262853892102SStefano Zampini /* create auxiliary bitmask and allocate workspace */ 2629df4d28bfSStefano Zampini if (!sub_schurs->reuse_solver) { 2630854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n-n_vertices,&idx_R_local);CHKERRQ(ierr); 2631a64d13efSStefano Zampini ierr = PetscBTCreate(pcis->n,&bitmask);CHKERRQ(ierr); 2632a64d13efSStefano Zampini for (i=0;i<n_vertices;i++) { 26330e6343abSStefano Zampini ierr = PetscBTSet(bitmask,pcbddc->local_primal_ref_node[i]);CHKERRQ(ierr); 2634a64d13efSStefano Zampini } 2635a64d13efSStefano Zampini 2636a64d13efSStefano Zampini for (i=0, n_R=0; i<pcis->n; i++) { 26374641a718SStefano Zampini if (!PetscBTLookup(bitmask,i)) { 26386816873aSStefano Zampini idx_R_local[n_R++] = i; 2639a64d13efSStefano Zampini } 2640a64d13efSStefano Zampini } 2641df4d28bfSStefano Zampini } else { /* A different ordering (already computed) is present if we are reusing the Schur solver */ 2642df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 26436816873aSStefano Zampini 2644df4d28bfSStefano Zampini ierr = ISGetIndices(reuse_solver->is_R,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 2645df4d28bfSStefano Zampini ierr = ISGetLocalSize(reuse_solver->is_R,&n_R);CHKERRQ(ierr); 26466816873aSStefano Zampini } 26473a50541eSStefano Zampini 26483a50541eSStefano Zampini /* Block code */ 26493a50541eSStefano Zampini vbs = 1; 26503a50541eSStefano Zampini ierr = MatGetBlockSize(pcbddc->local_mat,&bs);CHKERRQ(ierr); 26513a50541eSStefano Zampini if (bs>1 && !(n_vertices%bs)) { 26523a50541eSStefano Zampini PetscBool is_blocked = PETSC_TRUE; 26533a50541eSStefano Zampini PetscInt *vary; 2654df4d28bfSStefano Zampini if (!sub_schurs->reuse_solver) { 2655785e854fSJed Brown ierr = PetscMalloc1(pcis->n/bs,&vary);CHKERRQ(ierr); 26563a50541eSStefano Zampini ierr = PetscMemzero(vary,pcis->n/bs*sizeof(PetscInt));CHKERRQ(ierr); 2657d3df7717SStefano Zampini /* Verify that the vertex indices correspond to each element in a block (code taken from sbaij2.c) */ 2658d3df7717SStefano 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 */ 26590e6343abSStefano Zampini for (i=0; i<n_vertices; i++) vary[pcbddc->local_primal_ref_node[i]/bs]++; 2660d3df7717SStefano Zampini for (i=0; i<pcis->n/bs; i++) { 26613a50541eSStefano Zampini if (vary[i]!=0 && vary[i]!=bs) { 26623a50541eSStefano Zampini is_blocked = PETSC_FALSE; 26633a50541eSStefano Zampini break; 26643a50541eSStefano Zampini } 26653a50541eSStefano Zampini } 2666d3df7717SStefano Zampini ierr = PetscFree(vary);CHKERRQ(ierr); 2667d3df7717SStefano Zampini } else { 2668d3df7717SStefano Zampini /* Verify directly the R set */ 2669d3df7717SStefano Zampini for (i=0; i<n_R/bs; i++) { 2670d3df7717SStefano Zampini PetscInt j,node=idx_R_local[bs*i]; 2671d3df7717SStefano Zampini for (j=1; j<bs; j++) { 2672d3df7717SStefano Zampini if (node != idx_R_local[bs*i+j]-j) { 2673d3df7717SStefano Zampini is_blocked = PETSC_FALSE; 2674d3df7717SStefano Zampini break; 2675d3df7717SStefano Zampini } 2676d3df7717SStefano Zampini } 2677d3df7717SStefano Zampini } 2678d3df7717SStefano Zampini } 26793a50541eSStefano Zampini if (is_blocked) { /* build compressed IS for R nodes (complement of vertices) */ 26803a50541eSStefano Zampini vbs = bs; 26813a50541eSStefano Zampini for (i=0;i<n_R/vbs;i++) { 26823a50541eSStefano Zampini idx_R_local[i] = idx_R_local[vbs*i]/vbs; 26833a50541eSStefano Zampini } 26843a50541eSStefano Zampini } 26853a50541eSStefano Zampini } 26863a50541eSStefano Zampini ierr = ISCreateBlock(PETSC_COMM_SELF,vbs,n_R/vbs,idx_R_local,PETSC_COPY_VALUES,&pcbddc->is_R_local);CHKERRQ(ierr); 2687df4d28bfSStefano Zampini if (sub_schurs->reuse_solver) { 2688df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 268953892102SStefano Zampini 2690df4d28bfSStefano Zampini ierr = ISRestoreIndices(reuse_solver->is_R,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 2691df4d28bfSStefano Zampini ierr = ISDestroy(&reuse_solver->is_R);CHKERRQ(ierr); 269253892102SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->is_R_local);CHKERRQ(ierr); 2693df4d28bfSStefano Zampini reuse_solver->is_R = pcbddc->is_R_local; 269453892102SStefano Zampini } else { 26953a50541eSStefano Zampini ierr = PetscFree(idx_R_local);CHKERRQ(ierr); 269653892102SStefano Zampini } 2697a64d13efSStefano Zampini 2698a64d13efSStefano Zampini /* print some info if requested */ 2699a64d13efSStefano Zampini if (pcbddc->dbg_flag) { 2700a64d13efSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 2701a64d13efSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 27021575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 2703a64d13efSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d local dimensions\n",PetscGlobalRank);CHKERRQ(ierr); 2704a64d13efSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"local_size = %d, dirichlet_size = %d, boundary_size = %d\n",pcis->n,n_D,n_B);CHKERRQ(ierr); 27054f1b2e48SStefano 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); 2706a64d13efSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 2707a64d13efSStefano Zampini } 2708a64d13efSStefano Zampini 2709a64d13efSStefano Zampini /* VecScatters pcbddc->R_to_B and (optionally) pcbddc->R_to_D */ 2710df4d28bfSStefano Zampini if (!sub_schurs->reuse_solver) { 27116816873aSStefano Zampini IS is_aux1,is_aux2; 27126816873aSStefano Zampini PetscInt *aux_array1,*aux_array2,*is_indices,*idx_R_local; 27136816873aSStefano Zampini 27143a50541eSStefano Zampini ierr = ISGetIndices(pcbddc->is_R_local,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 2715854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n_B-n_vertices,&aux_array1);CHKERRQ(ierr); 2716854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n_B-n_vertices,&aux_array2);CHKERRQ(ierr); 2717a64d13efSStefano Zampini ierr = ISGetIndices(pcis->is_I_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 27184641a718SStefano Zampini for (i=0; i<n_D; i++) { 27194641a718SStefano Zampini ierr = PetscBTSet(bitmask,is_indices[i]);CHKERRQ(ierr); 27204641a718SStefano Zampini } 2721a64d13efSStefano Zampini ierr = ISRestoreIndices(pcis->is_I_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 2722a64d13efSStefano Zampini for (i=0, j=0; i<n_R; i++) { 27234641a718SStefano Zampini if (!PetscBTLookup(bitmask,idx_R_local[i])) { 27244641a718SStefano Zampini aux_array1[j++] = i; 2725a64d13efSStefano Zampini } 2726a64d13efSStefano Zampini } 2727a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array1,PETSC_OWN_POINTER,&is_aux1);CHKERRQ(ierr); 2728a64d13efSStefano Zampini ierr = ISGetIndices(pcis->is_B_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 2729a64d13efSStefano Zampini for (i=0, j=0; i<n_B; i++) { 27304641a718SStefano Zampini if (!PetscBTLookup(bitmask,is_indices[i])) { 27314641a718SStefano Zampini aux_array2[j++] = i; 2732a64d13efSStefano Zampini } 2733a64d13efSStefano Zampini } 2734a64d13efSStefano Zampini ierr = ISRestoreIndices(pcis->is_B_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 2735a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array2,PETSC_OWN_POINTER,&is_aux2);CHKERRQ(ierr); 2736a64d13efSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,is_aux1,pcis->vec1_B,is_aux2,&pcbddc->R_to_B);CHKERRQ(ierr); 2737a64d13efSStefano Zampini ierr = ISDestroy(&is_aux1);CHKERRQ(ierr); 2738a64d13efSStefano Zampini ierr = ISDestroy(&is_aux2);CHKERRQ(ierr); 2739a64d13efSStefano Zampini 27408eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 2741785e854fSJed Brown ierr = PetscMalloc1(n_D,&aux_array1);CHKERRQ(ierr); 2742a64d13efSStefano Zampini for (i=0, j=0; i<n_R; i++) { 27434641a718SStefano Zampini if (PetscBTLookup(bitmask,idx_R_local[i])) { 27444641a718SStefano Zampini aux_array1[j++] = i; 2745a64d13efSStefano Zampini } 2746a64d13efSStefano Zampini } 2747a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array1,PETSC_OWN_POINTER,&is_aux1);CHKERRQ(ierr); 2748a64d13efSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,is_aux1,pcis->vec1_D,(IS)0,&pcbddc->R_to_D);CHKERRQ(ierr); 2749a64d13efSStefano Zampini ierr = ISDestroy(&is_aux1);CHKERRQ(ierr); 2750a64d13efSStefano Zampini } 27514641a718SStefano Zampini ierr = PetscBTDestroy(&bitmask);CHKERRQ(ierr); 27523a50541eSStefano Zampini ierr = ISRestoreIndices(pcbddc->is_R_local,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 2753d62866d3SStefano Zampini } else { 2754df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 27556816873aSStefano Zampini IS tis; 27566816873aSStefano Zampini PetscInt schur_size; 27576816873aSStefano Zampini 2758df4d28bfSStefano Zampini ierr = ISGetLocalSize(reuse_solver->is_B,&schur_size);CHKERRQ(ierr); 27596816873aSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,schur_size,n_D,1,&tis);CHKERRQ(ierr); 2760df4d28bfSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,tis,pcis->vec1_B,reuse_solver->is_B,&pcbddc->R_to_B);CHKERRQ(ierr); 27616816873aSStefano Zampini ierr = ISDestroy(&tis);CHKERRQ(ierr); 27626816873aSStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 27636816873aSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,n_D,0,1,&tis);CHKERRQ(ierr); 27646816873aSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,tis,pcis->vec1_D,(IS)0,&pcbddc->R_to_D);CHKERRQ(ierr); 27656816873aSStefano Zampini ierr = ISDestroy(&tis);CHKERRQ(ierr); 2766d62866d3SStefano Zampini } 2767d62866d3SStefano Zampini } 2768a64d13efSStefano Zampini PetscFunctionReturn(0); 2769a64d13efSStefano Zampini } 2770a64d13efSStefano Zampini 2771304d26faSStefano Zampini 2772304d26faSStefano Zampini #undef __FUNCT__ 2773304d26faSStefano Zampini #define __FUNCT__ "PCBDDCSetUpLocalSolvers" 2774684f6988SStefano Zampini PetscErrorCode PCBDDCSetUpLocalSolvers(PC pc, PetscBool dirichlet, PetscBool neumann) 2775304d26faSStefano Zampini { 2776304d26faSStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 2777304d26faSStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 2778304d26faSStefano Zampini PC pc_temp; 2779304d26faSStefano Zampini Mat A_RR; 2780f4ddd8eeSStefano Zampini MatReuse reuse; 2781304d26faSStefano Zampini PetscScalar m_one = -1.0; 2782304d26faSStefano Zampini PetscReal value; 278304708bb6SStefano Zampini PetscInt n_D,n_R; 27849577ea80SStefano Zampini PetscBool use_exact,use_exact_reduced,issbaij; 2785304d26faSStefano Zampini PetscErrorCode ierr; 2786e604994aSStefano Zampini /* prefixes stuff */ 2787312be037SStefano Zampini char dir_prefix[256],neu_prefix[256],str_level[16]; 2788e604994aSStefano Zampini size_t len; 2789304d26faSStefano Zampini 2790304d26faSStefano Zampini PetscFunctionBegin; 2791304d26faSStefano Zampini 2792e604994aSStefano Zampini /* compute prefixes */ 2793e604994aSStefano Zampini ierr = PetscStrcpy(dir_prefix,"");CHKERRQ(ierr); 2794e604994aSStefano Zampini ierr = PetscStrcpy(neu_prefix,"");CHKERRQ(ierr); 2795e604994aSStefano Zampini if (!pcbddc->current_level) { 2796e604994aSStefano Zampini ierr = PetscStrcpy(dir_prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 2797e604994aSStefano Zampini ierr = PetscStrcpy(neu_prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 2798e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,"pc_bddc_dirichlet_");CHKERRQ(ierr); 2799e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,"pc_bddc_neumann_");CHKERRQ(ierr); 2800e604994aSStefano Zampini } else { 2801e604994aSStefano Zampini ierr = PetscStrcpy(str_level,"");CHKERRQ(ierr); 2802312be037SStefano Zampini sprintf(str_level,"l%d_",(int)(pcbddc->current_level)); 2803e604994aSStefano Zampini ierr = PetscStrlen(((PetscObject)pc)->prefix,&len);CHKERRQ(ierr); 2804e604994aSStefano Zampini len -= 15; /* remove "pc_bddc_coarse_" */ 2805312be037SStefano Zampini if (pcbddc->current_level>1) len -= 3; /* remove "lX_" with X level number */ 2806312be037SStefano Zampini if (pcbddc->current_level>10) len -= 1; /* remove another char from level number */ 280734d6797cSStefano Zampini ierr = PetscStrncpy(dir_prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 280834d6797cSStefano Zampini ierr = PetscStrncpy(neu_prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 2809e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,"pc_bddc_dirichlet_");CHKERRQ(ierr); 2810e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,"pc_bddc_neumann_");CHKERRQ(ierr); 2811e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,str_level);CHKERRQ(ierr); 2812e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,str_level);CHKERRQ(ierr); 2813e604994aSStefano Zampini } 2814e604994aSStefano Zampini 2815304d26faSStefano Zampini /* DIRICHLET PROBLEM */ 2816684f6988SStefano Zampini if (dirichlet) { 2817d5574798SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 2818450f8f5eSStefano Zampini if (pcbddc->benign_n && !pcbddc->benign_change_explicit) { 2819450f8f5eSStefano Zampini if (!sub_schurs->reuse_solver) { 2820450f8f5eSStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Not yet implemented\n"); 2821450f8f5eSStefano Zampini } 2822450f8f5eSStefano Zampini if (pcbddc->dbg_flag) { 2823a3df083aSStefano Zampini Mat A_IIn; 2824a3df083aSStefano Zampini 2825a3df083aSStefano Zampini ierr = PCBDDCBenignProject(pc,pcis->is_I_local,pcis->is_I_local,&A_IIn);CHKERRQ(ierr); 2826a3df083aSStefano Zampini ierr = MatDestroy(&pcis->A_II);CHKERRQ(ierr); 2827a3df083aSStefano Zampini pcis->A_II = A_IIn; 2828a3df083aSStefano Zampini } 2829450f8f5eSStefano Zampini } 28303301b35fSStefano Zampini if (pcbddc->local_mat->symmetric_set) { 28313301b35fSStefano Zampini ierr = MatSetOption(pcis->A_II,MAT_SYMMETRIC,pcbddc->local_mat->symmetric_set);CHKERRQ(ierr); 2832964fefecSStefano Zampini } 2833ac78edfcSStefano Zampini /* Matrix for Dirichlet problem is pcis->A_II */ 2834964fefecSStefano Zampini n_D = pcis->n - pcis->n_B; 2835304d26faSStefano Zampini if (!pcbddc->ksp_D) { /* create object if not yet build */ 2836304d26faSStefano Zampini ierr = KSPCreate(PETSC_COMM_SELF,&pcbddc->ksp_D);CHKERRQ(ierr); 2837304d26faSStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->ksp_D,(PetscObject)pc,1);CHKERRQ(ierr); 2838304d26faSStefano Zampini /* default */ 2839304d26faSStefano Zampini ierr = KSPSetType(pcbddc->ksp_D,KSPPREONLY);CHKERRQ(ierr); 2840e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->ksp_D,dir_prefix);CHKERRQ(ierr); 28419577ea80SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcis->A_II,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 2842304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_D,&pc_temp);CHKERRQ(ierr); 28439577ea80SStefano Zampini if (issbaij) { 28449577ea80SStefano Zampini ierr = PCSetType(pc_temp,PCCHOLESKY);CHKERRQ(ierr); 28459577ea80SStefano Zampini } else { 2846304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCLU);CHKERRQ(ierr); 28479577ea80SStefano Zampini } 2848304d26faSStefano Zampini /* Allow user's customization */ 2849304d26faSStefano Zampini ierr = KSPSetFromOptions(pcbddc->ksp_D);CHKERRQ(ierr); 2850304d26faSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 2851304d26faSStefano Zampini } 2852d1e9a80fSBarry Smith ierr = KSPSetOperators(pcbddc->ksp_D,pcis->A_II,pcis->A_II);CHKERRQ(ierr); 2853df4d28bfSStefano Zampini if (sub_schurs->reuse_solver) { 2854df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 2855d62866d3SStefano Zampini 2856df4d28bfSStefano Zampini ierr = KSPSetPC(pcbddc->ksp_D,reuse_solver->interior_solver);CHKERRQ(ierr); 2857d5574798SStefano Zampini } 2858304d26faSStefano Zampini /* umfpack interface has a bug when matrix dimension is zero. TODO solve from umfpack interface */ 2859304d26faSStefano Zampini if (!n_D) { 2860304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_D,&pc_temp);CHKERRQ(ierr); 2861304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCNONE);CHKERRQ(ierr); 2862304d26faSStefano Zampini } 2863304d26faSStefano Zampini /* Set Up KSP for Dirichlet problem of BDDC */ 2864304d26faSStefano Zampini ierr = KSPSetUp(pcbddc->ksp_D);CHKERRQ(ierr); 2865304d26faSStefano Zampini /* set ksp_D into pcis data */ 2866304d26faSStefano Zampini ierr = KSPDestroy(&pcis->ksp_D);CHKERRQ(ierr); 2867304d26faSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->ksp_D);CHKERRQ(ierr); 2868304d26faSStefano Zampini pcis->ksp_D = pcbddc->ksp_D; 2869684f6988SStefano Zampini } 2870304d26faSStefano Zampini 2871304d26faSStefano Zampini /* NEUMANN PROBLEM */ 2872684f6988SStefano Zampini A_RR = 0; 2873684f6988SStefano Zampini if (neumann) { 2874d62866d3SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 287504708bb6SStefano Zampini PetscInt ibs,mbs; 287604708bb6SStefano Zampini PetscBool issbaij; 287704708bb6SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 2878f4ddd8eeSStefano Zampini /* Matrix for Neumann problem is A_RR -> we need to create/reuse it at this point */ 28798ce42a96SStefano Zampini ierr = ISGetSize(pcbddc->is_R_local,&n_R);CHKERRQ(ierr); 2880f4ddd8eeSStefano Zampini if (pcbddc->ksp_R) { /* already created ksp */ 2881f4ddd8eeSStefano Zampini PetscInt nn_R; 288281d9aea3SBarry Smith ierr = KSPGetOperators(pcbddc->ksp_R,NULL,&A_RR);CHKERRQ(ierr); 2883f4ddd8eeSStefano Zampini ierr = PetscObjectReference((PetscObject)A_RR);CHKERRQ(ierr); 2884f4ddd8eeSStefano Zampini ierr = MatGetSize(A_RR,&nn_R,NULL);CHKERRQ(ierr); 2885f4ddd8eeSStefano Zampini if (nn_R != n_R) { /* old ksp is not reusable, so reset it */ 2886f4ddd8eeSStefano Zampini ierr = KSPReset(pcbddc->ksp_R);CHKERRQ(ierr); 2887f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 2888f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 2889f4ddd8eeSStefano Zampini } else { /* same sizes, but nonzero pattern depend on primal vertices so it can be changed */ 2890727cdba6SStefano Zampini if (pcbddc->new_primal_space_local) { /* we are not sure the matrix will have the same nonzero pattern */ 2891f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 2892f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 2893f4ddd8eeSStefano Zampini } else { /* safe to reuse the matrix */ 2894f4ddd8eeSStefano Zampini reuse = MAT_REUSE_MATRIX; 2895f4ddd8eeSStefano Zampini } 2896f4ddd8eeSStefano Zampini } 2897f4ddd8eeSStefano Zampini /* last check */ 2898d1e9a80fSBarry Smith if (pc->flag == DIFFERENT_NONZERO_PATTERN) { 2899f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 2900f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 2901f4ddd8eeSStefano Zampini } 2902f4ddd8eeSStefano Zampini } else { /* first time, so we need to create the matrix */ 2903f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 2904f4ddd8eeSStefano Zampini } 2905*a00504b5SStefano Zampini /* convert pcbddc->local_mat if needed later in PCBDDCSetUpCorrection */ 2906af732b37SStefano Zampini ierr = MatGetBlockSize(pcbddc->local_mat,&mbs);CHKERRQ(ierr); 2907af732b37SStefano Zampini ierr = ISGetBlockSize(pcbddc->is_R_local,&ibs);CHKERRQ(ierr); 290804708bb6SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcbddc->local_mat,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 290904708bb6SStefano Zampini if (ibs != mbs) { /* need to convert to SEQAIJ to extract any submatrix with is_R_local */ 291004708bb6SStefano Zampini if (matis->A == pcbddc->local_mat) { 291104708bb6SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 291204708bb6SStefano Zampini ierr = MatConvert(matis->A,MATSEQAIJ,MAT_INITIAL_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 2913af732b37SStefano Zampini } else { 291404708bb6SStefano Zampini ierr = MatConvert(pcbddc->local_mat,MATSEQAIJ,MAT_REUSE_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 29156816873aSStefano Zampini } 291604708bb6SStefano Zampini } else if (issbaij) { /* need to convert to BAIJ to get offdiagonal blocks */ 291704708bb6SStefano Zampini if (matis->A == pcbddc->local_mat) { 291804708bb6SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 291904708bb6SStefano Zampini ierr = MatConvert(matis->A,MATSEQBAIJ,MAT_INITIAL_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 292004708bb6SStefano Zampini } else { 292104708bb6SStefano Zampini ierr = MatConvert(pcbddc->local_mat,MATSEQBAIJ,MAT_REUSE_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 292204708bb6SStefano Zampini } 292304708bb6SStefano Zampini } 2924*a00504b5SStefano Zampini /* extract A_RR */ 2925*a00504b5SStefano Zampini if (sub_schurs->reuse_solver) { 2926*a00504b5SStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 2927*a00504b5SStefano Zampini 2928*a00504b5SStefano Zampini if (pcbddc->dbg_flag) { /* we need A_RR to test the solver later */ 292916e386b8SStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 2930*a00504b5SStefano Zampini if (reuse_solver->benign_n) { /* we are not using the explicit change of basis on the pressures */ 293116e386b8SStefano Zampini ierr = PCBDDCBenignProject(pc,pcbddc->is_R_local,pcbddc->is_R_local,&A_RR);CHKERRQ(ierr); 293216e386b8SStefano Zampini } else { 2933*a00504b5SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,pcbddc->is_R_local,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&A_RR);CHKERRQ(ierr); 2934*a00504b5SStefano Zampini } 2935*a00504b5SStefano Zampini } else { 2936*a00504b5SStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 2937*a00504b5SStefano Zampini ierr = PCGetOperators(reuse_solver->correction_solver,&A_RR,NULL);CHKERRQ(ierr); 2938*a00504b5SStefano Zampini ierr = PetscObjectReference((PetscObject)A_RR);CHKERRQ(ierr); 2939*a00504b5SStefano Zampini } 2940*a00504b5SStefano Zampini } else { /* we have to build the neumann solver, so we need to extract the relevant matrix */ 2941f4ddd8eeSStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,pcbddc->is_R_local,pcbddc->is_R_local,reuse,&A_RR);CHKERRQ(ierr); 294216e386b8SStefano Zampini } 29433301b35fSStefano Zampini if (pcbddc->local_mat->symmetric_set) { 29443301b35fSStefano Zampini ierr = MatSetOption(A_RR,MAT_SYMMETRIC,pcbddc->local_mat->symmetric_set);CHKERRQ(ierr); 29456816873aSStefano Zampini } 2946f4ddd8eeSStefano Zampini if (!pcbddc->ksp_R) { /* create object if not present */ 2947304d26faSStefano Zampini ierr = KSPCreate(PETSC_COMM_SELF,&pcbddc->ksp_R);CHKERRQ(ierr); 2948304d26faSStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->ksp_R,(PetscObject)pc,1);CHKERRQ(ierr); 2949304d26faSStefano Zampini /* default */ 2950304d26faSStefano Zampini ierr = KSPSetType(pcbddc->ksp_R,KSPPREONLY);CHKERRQ(ierr); 2951e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->ksp_R,neu_prefix);CHKERRQ(ierr); 2952304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_temp);CHKERRQ(ierr); 29539577ea80SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)A_RR,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 29549577ea80SStefano Zampini if (issbaij) { 29559577ea80SStefano Zampini ierr = PCSetType(pc_temp,PCCHOLESKY);CHKERRQ(ierr); 29569577ea80SStefano Zampini } else { 2957304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCLU);CHKERRQ(ierr); 29589577ea80SStefano Zampini } 2959304d26faSStefano Zampini /* Allow user's customization */ 2960304d26faSStefano Zampini ierr = KSPSetFromOptions(pcbddc->ksp_R);CHKERRQ(ierr); 2961304d26faSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 2962304d26faSStefano Zampini } 2963304d26faSStefano Zampini /* umfpack interface has a bug when matrix dimension is zero. TODO solve from umfpack interface */ 2964304d26faSStefano Zampini if (!n_R) { 2965304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_temp);CHKERRQ(ierr); 2966304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCNONE);CHKERRQ(ierr); 2967304d26faSStefano Zampini } 29685cbda25cSStefano Zampini ierr = KSPSetOperators(pcbddc->ksp_R,A_RR,A_RR);CHKERRQ(ierr); 2969df4d28bfSStefano Zampini /* Reuse solver if it is present */ 2970df4d28bfSStefano Zampini if (sub_schurs->reuse_solver) { 2971df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 2972d62866d3SStefano Zampini 2973df4d28bfSStefano Zampini ierr = KSPSetPC(pcbddc->ksp_R,reuse_solver->correction_solver);CHKERRQ(ierr); 2974d62866d3SStefano Zampini } 2975304d26faSStefano Zampini /* Set Up KSP for Neumann problem of BDDC */ 2976304d26faSStefano Zampini ierr = KSPSetUp(pcbddc->ksp_R);CHKERRQ(ierr); 2977684f6988SStefano Zampini } 2978304d26faSStefano Zampini 2979304d26faSStefano Zampini /* check Dirichlet and Neumann solvers and adapt them if a nullspace correction is needed */ 29800fccc4e9SStefano Zampini if (pcbddc->NullSpace || pcbddc->dbg_flag) { 2981684f6988SStefano Zampini if (pcbddc->dbg_flag) { 2982684f6988SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 29831575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 2984684f6988SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 2985684f6988SStefano Zampini } 2986684f6988SStefano Zampini if (dirichlet) { /* Dirichlet */ 29870fccc4e9SStefano Zampini ierr = VecSetRandom(pcis->vec1_D,NULL);CHKERRQ(ierr); 29880fccc4e9SStefano Zampini ierr = MatMult(pcis->A_II,pcis->vec1_D,pcis->vec2_D);CHKERRQ(ierr); 29890fccc4e9SStefano Zampini ierr = KSPSolve(pcbddc->ksp_D,pcis->vec2_D,pcis->vec2_D);CHKERRQ(ierr); 29900fccc4e9SStefano Zampini ierr = VecAXPY(pcis->vec1_D,m_one,pcis->vec2_D);CHKERRQ(ierr); 29910fccc4e9SStefano Zampini ierr = VecNorm(pcis->vec1_D,NORM_INFINITY,&value);CHKERRQ(ierr); 2992304d26faSStefano Zampini /* need to be adapted? */ 2993b8ffe317SStefano Zampini use_exact = (PetscAbsReal(value) > 1.e-4 ? PETSC_FALSE : PETSC_TRUE); 2994b8ffe317SStefano Zampini ierr = MPI_Allreduce(&use_exact,&use_exact_reduced,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 2995b8ffe317SStefano Zampini ierr = PCBDDCSetUseExactDirichlet(pc,use_exact_reduced);CHKERRQ(ierr); 2996304d26faSStefano Zampini /* print info */ 2997304d26faSStefano Zampini if (pcbddc->dbg_flag) { 2998e604994aSStefano 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); 2999304d26faSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 3000304d26faSStefano Zampini } 3001b8ffe317SStefano Zampini if (pcbddc->NullSpace && !use_exact_reduced && !pcbddc->switch_static) { 3002298c0119SStefano Zampini ierr = PCBDDCNullSpaceAssembleCorrection(pc,PETSC_TRUE,pcis->is_I_local);CHKERRQ(ierr); 3003304d26faSStefano Zampini } 3004684f6988SStefano Zampini } 3005684f6988SStefano Zampini if (neumann) { /* Neumann */ 30060fccc4e9SStefano Zampini ierr = VecSetRandom(pcbddc->vec1_R,NULL);CHKERRQ(ierr); 30070fccc4e9SStefano Zampini ierr = MatMult(A_RR,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 30080fccc4e9SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec2_R,pcbddc->vec2_R);CHKERRQ(ierr); 30090fccc4e9SStefano Zampini ierr = VecAXPY(pcbddc->vec1_R,m_one,pcbddc->vec2_R);CHKERRQ(ierr); 30100fccc4e9SStefano Zampini ierr = VecNorm(pcbddc->vec1_R,NORM_INFINITY,&value);CHKERRQ(ierr); 3011304d26faSStefano Zampini /* need to be adapted? */ 3012b8ffe317SStefano Zampini use_exact = (PetscAbsReal(value) > 1.e-4 ? PETSC_FALSE : PETSC_TRUE); 3013b8ffe317SStefano Zampini ierr = MPI_Allreduce(&use_exact,&use_exact_reduced,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 3014304d26faSStefano Zampini /* print info */ 3015304d26faSStefano Zampini if (pcbddc->dbg_flag) { 3016e604994aSStefano 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); 3017304d26faSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 3018304d26faSStefano Zampini } 3019b8ffe317SStefano Zampini if (pcbddc->NullSpace && !use_exact_reduced) { /* is it the right logic? */ 3020298c0119SStefano Zampini ierr = PCBDDCNullSpaceAssembleCorrection(pc,PETSC_FALSE,pcbddc->is_R_local);CHKERRQ(ierr); 3021304d26faSStefano Zampini } 30220fccc4e9SStefano Zampini } 3023684f6988SStefano Zampini } 30245cbda25cSStefano Zampini /* free Neumann problem's matrix */ 30255cbda25cSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 3026304d26faSStefano Zampini PetscFunctionReturn(0); 3027304d26faSStefano Zampini } 3028304d26faSStefano Zampini 3029304d26faSStefano Zampini #undef __FUNCT__ 3030ba15a52eSStefano Zampini #define __FUNCT__ "PCBDDCSolveSubstructureCorrection" 303180677318SStefano Zampini static PetscErrorCode PCBDDCSolveSubstructureCorrection(PC pc, Vec inout_B, Vec inout_D, PetscBool applytranspose) 3032674ae819SStefano Zampini { 3033674ae819SStefano Zampini PetscErrorCode ierr; 3034674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 3035be83ff47SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 3036674ae819SStefano Zampini 3037674ae819SStefano Zampini PetscFunctionBegin; 3038df4d28bfSStefano Zampini if (!sub_schurs->reuse_solver) { 303980677318SStefano Zampini ierr = VecSet(pcbddc->vec1_R,0.);CHKERRQ(ierr); 304020c7b377SStefano Zampini } 304180677318SStefano Zampini if (!pcbddc->switch_static) { 304280677318SStefano Zampini if (applytranspose && pcbddc->local_auxmat1) { 304380677318SStefano Zampini ierr = MatMultTranspose(pcbddc->local_auxmat2,inout_B,pcbddc->vec1_C);CHKERRQ(ierr); 304480677318SStefano Zampini ierr = MatMultTransposeAdd(pcbddc->local_auxmat1,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr); 304520c7b377SStefano Zampini } 3046df4d28bfSStefano Zampini if (!sub_schurs->reuse_solver) { 304780677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 304880677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 304920c7b377SStefano Zampini } else { 3050df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 3051be83ff47SStefano Zampini 3052df4d28bfSStefano Zampini ierr = VecScatterBegin(reuse_solver->correction_scatter_B,inout_B,reuse_solver->rhs_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 3053df4d28bfSStefano Zampini ierr = VecScatterEnd(reuse_solver->correction_scatter_B,inout_B,reuse_solver->rhs_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 305420c7b377SStefano Zampini } 3055be83ff47SStefano Zampini } else { 305680677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 305780677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 305880677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,inout_D,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 305980677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,inout_D,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 306080677318SStefano Zampini if (applytranspose && pcbddc->local_auxmat1) { 306180677318SStefano Zampini ierr = MatMultTranspose(pcbddc->local_auxmat2,pcbddc->vec1_R,pcbddc->vec1_C);CHKERRQ(ierr); 306280677318SStefano Zampini ierr = MatMultTransposeAdd(pcbddc->local_auxmat1,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr); 306380677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 306480677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 3065674ae819SStefano Zampini } 3066674ae819SStefano Zampini } 306772b8c272SStefano Zampini if (!sub_schurs->reuse_solver || pcbddc->switch_static) { 306880677318SStefano Zampini if (applytranspose) { 306980677318SStefano Zampini ierr = KSPSolveTranspose(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 307080677318SStefano Zampini } else { 307180677318SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 307280677318SStefano Zampini } 3073be83ff47SStefano Zampini } else { 3074df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 3075be83ff47SStefano Zampini 3076be83ff47SStefano Zampini if (applytranspose) { 3077df4d28bfSStefano Zampini ierr = MatFactorSolveSchurComplementTranspose(reuse_solver->F,reuse_solver->rhs_B,reuse_solver->sol_B);CHKERRQ(ierr); 3078be83ff47SStefano Zampini } else { 3079df4d28bfSStefano Zampini ierr = MatFactorSolveSchurComplement(reuse_solver->F,reuse_solver->rhs_B,reuse_solver->sol_B);CHKERRQ(ierr); 3080be83ff47SStefano Zampini } 3081be83ff47SStefano Zampini } 308280677318SStefano Zampini ierr = VecSet(inout_B,0.);CHKERRQ(ierr); 308380677318SStefano Zampini if (!pcbddc->switch_static) { 3084df4d28bfSStefano Zampini if (!sub_schurs->reuse_solver) { 308580677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 308680677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 3087be83ff47SStefano Zampini } else { 3088df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 3089be83ff47SStefano Zampini 3090df4d28bfSStefano Zampini ierr = VecScatterBegin(reuse_solver->correction_scatter_B,reuse_solver->sol_B,inout_B,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 3091df4d28bfSStefano Zampini ierr = VecScatterEnd(reuse_solver->correction_scatter_B,reuse_solver->sol_B,inout_B,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 3092be83ff47SStefano Zampini } 309380677318SStefano Zampini if (!applytranspose && pcbddc->local_auxmat1) { 309480677318SStefano Zampini ierr = MatMult(pcbddc->local_auxmat1,inout_B,pcbddc->vec1_C);CHKERRQ(ierr); 309580677318SStefano Zampini ierr = MatMultAdd(pcbddc->local_auxmat2,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr); 309680677318SStefano Zampini } 309780677318SStefano Zampini } else { 309880677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 309980677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 310080677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 310180677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 310280677318SStefano Zampini if (!applytranspose && pcbddc->local_auxmat1) { 310380677318SStefano Zampini ierr = MatMult(pcbddc->local_auxmat1,inout_B,pcbddc->vec1_C);CHKERRQ(ierr); 310480677318SStefano Zampini ierr = MatMultAdd(pcbddc->local_auxmat2,pcbddc->vec1_C,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 310580677318SStefano Zampini } 310680677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 310780677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 310880677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 310980677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 3110674ae819SStefano Zampini } 3111674ae819SStefano Zampini PetscFunctionReturn(0); 3112674ae819SStefano Zampini } 3113674ae819SStefano Zampini 3114dc359a40SStefano Zampini /* parameter apply transpose determines if the interface preconditioner should be applied transposed or not */ 3115674ae819SStefano Zampini #undef __FUNCT__ 3116674ae819SStefano Zampini #define __FUNCT__ "PCBDDCApplyInterfacePreconditioner" 3117dc359a40SStefano Zampini PetscErrorCode PCBDDCApplyInterfacePreconditioner(PC pc, PetscBool applytranspose) 3118674ae819SStefano Zampini { 3119674ae819SStefano Zampini PetscErrorCode ierr; 3120674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 3121674ae819SStefano Zampini PC_IS* pcis = (PC_IS*) (pc->data); 3122674ae819SStefano Zampini const PetscScalar zero = 0.0; 3123674ae819SStefano Zampini 3124674ae819SStefano Zampini PetscFunctionBegin; 3125dc359a40SStefano Zampini /* Application of PSI^T or PHI^T (depending on applytranspose, see comment above) */ 31264fee134fSStefano Zampini if (!pcbddc->benign_apply_coarse_only) { 3127dc359a40SStefano Zampini if (applytranspose) { 3128674ae819SStefano Zampini ierr = MatMultTranspose(pcbddc->coarse_phi_B,pcis->vec1_B,pcbddc->vec1_P);CHKERRQ(ierr); 31298eeda7d8SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultTransposeAdd(pcbddc->coarse_phi_D,pcis->vec1_D,pcbddc->vec1_P,pcbddc->vec1_P);CHKERRQ(ierr); } 3130dc359a40SStefano Zampini } else { 3131674ae819SStefano Zampini ierr = MatMultTranspose(pcbddc->coarse_psi_B,pcis->vec1_B,pcbddc->vec1_P);CHKERRQ(ierr); 3132674ae819SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultTransposeAdd(pcbddc->coarse_psi_D,pcis->vec1_D,pcbddc->vec1_P,pcbddc->vec1_P);CHKERRQ(ierr); } 313315aaf578SStefano Zampini } 31344fee134fSStefano Zampini } else { 31354fee134fSStefano Zampini ierr = VecSet(pcbddc->vec1_P,zero);CHKERRQ(ierr); 31364fee134fSStefano Zampini } 3137efc2fbd9SStefano Zampini 3138efc2fbd9SStefano Zampini /* add p0 to the last value of vec1_P holding the coarse dof relative to p0 */ 31394f1b2e48SStefano Zampini if (pcbddc->benign_n) { 3140efc2fbd9SStefano Zampini PetscScalar *array; 31414f1b2e48SStefano Zampini PetscInt j; 3142efc2fbd9SStefano Zampini 3143efc2fbd9SStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 31444f1b2e48SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) array[pcbddc->local_primal_size-pcbddc->benign_n+j] += pcbddc->benign_p0[j]; 3145efc2fbd9SStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 3146efc2fbd9SStefano Zampini } 3147efc2fbd9SStefano Zampini 314812edc857SStefano Zampini /* start communications from local primal nodes to rhs of coarse solver */ 314912edc857SStefano Zampini ierr = VecSet(pcbddc->coarse_vec,zero);CHKERRQ(ierr); 315012edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataBegin(pc,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 315112edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataEnd(pc,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 315212edc857SStefano Zampini 31539f00e9b4SStefano Zampini /* Coarse solution -> rhs and sol updated inside PCBDDCScattarCoarseDataBegin/End */ 315412edc857SStefano Zampini /* TODO remove null space when doing multilevel */ 315512edc857SStefano Zampini if (pcbddc->coarse_ksp) { 315651694757SStefano Zampini Mat coarse_mat; 3157964fefecSStefano Zampini Vec rhs,sol; 315851694757SStefano Zampini MatNullSpace nullsp; 3159964fefecSStefano Zampini 3160964fefecSStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&rhs);CHKERRQ(ierr); 3161964fefecSStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&sol);CHKERRQ(ierr); 316251694757SStefano Zampini ierr = KSPGetOperators(pcbddc->coarse_ksp,&coarse_mat,NULL);CHKERRQ(ierr); 316351694757SStefano Zampini ierr = MatGetNullSpace(coarse_mat,&nullsp);CHKERRQ(ierr); 316451694757SStefano Zampini if (nullsp) { 316551694757SStefano Zampini ierr = MatNullSpaceRemove(nullsp,rhs);CHKERRQ(ierr); 316651694757SStefano Zampini } 316712edc857SStefano Zampini if (applytranspose) { 3168964fefecSStefano Zampini ierr = KSPSolveTranspose(pcbddc->coarse_ksp,rhs,sol);CHKERRQ(ierr); 316912edc857SStefano Zampini } else { 3170964fefecSStefano Zampini ierr = KSPSolve(pcbddc->coarse_ksp,rhs,sol);CHKERRQ(ierr); 317112edc857SStefano Zampini } 317251694757SStefano Zampini if (nullsp) { 317351694757SStefano Zampini ierr = MatNullSpaceRemove(nullsp,sol);CHKERRQ(ierr); 317451694757SStefano Zampini } 317512edc857SStefano Zampini } 3176674ae819SStefano Zampini 3177674ae819SStefano Zampini /* Local solution on R nodes */ 31784fee134fSStefano Zampini if (pcis->n && !pcbddc->benign_apply_coarse_only) { 317980677318SStefano Zampini ierr = PCBDDCSolveSubstructureCorrection(pc,pcis->vec1_B,pcis->vec1_D,applytranspose);CHKERRQ(ierr); 31809f00e9b4SStefano Zampini } 31819f00e9b4SStefano Zampini /* communications from coarse sol to local primal nodes */ 31829f00e9b4SStefano Zampini ierr = PCBDDCScatterCoarseDataBegin(pc,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 318312edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataEnd(pc,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 3184674ae819SStefano Zampini 31854fee134fSStefano Zampini /* Sum contributions from the two levels */ 31864fee134fSStefano Zampini if (!pcbddc->benign_apply_coarse_only) { 3187dc359a40SStefano Zampini if (applytranspose) { 3188dc359a40SStefano Zampini ierr = MatMultAdd(pcbddc->coarse_psi_B,pcbddc->vec1_P,pcis->vec1_B,pcis->vec1_B);CHKERRQ(ierr); 3189dc359a40SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultAdd(pcbddc->coarse_psi_D,pcbddc->vec1_P,pcis->vec1_D,pcis->vec1_D);CHKERRQ(ierr); } 3190dc359a40SStefano Zampini } else { 3191674ae819SStefano Zampini ierr = MatMultAdd(pcbddc->coarse_phi_B,pcbddc->vec1_P,pcis->vec1_B,pcis->vec1_B);CHKERRQ(ierr); 31928eeda7d8SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultAdd(pcbddc->coarse_phi_D,pcbddc->vec1_P,pcis->vec1_D,pcis->vec1_D);CHKERRQ(ierr); } 3193dc359a40SStefano Zampini } 3194efc2fbd9SStefano Zampini /* store p0 */ 31954f1b2e48SStefano Zampini if (pcbddc->benign_n) { 3196efc2fbd9SStefano Zampini PetscScalar *array; 31974f1b2e48SStefano Zampini PetscInt j; 3198efc2fbd9SStefano Zampini 3199efc2fbd9SStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 32004f1b2e48SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) pcbddc->benign_p0[j] = array[pcbddc->local_primal_size-pcbddc->benign_n+j]; 3201efc2fbd9SStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 3202efc2fbd9SStefano Zampini } 32034fee134fSStefano Zampini } else { /* expand the coarse solution */ 32044fee134fSStefano Zampini if (applytranspose) { 32054fee134fSStefano Zampini ierr = MatMult(pcbddc->coarse_psi_B,pcbddc->vec1_P,pcis->vec1_B);CHKERRQ(ierr); 32064fee134fSStefano Zampini } else { 32074fee134fSStefano Zampini ierr = MatMult(pcbddc->coarse_phi_B,pcbddc->vec1_P,pcis->vec1_B);CHKERRQ(ierr); 32084fee134fSStefano Zampini } 32094fee134fSStefano Zampini } 3210674ae819SStefano Zampini PetscFunctionReturn(0); 3211674ae819SStefano Zampini } 3212674ae819SStefano Zampini 3213674ae819SStefano Zampini #undef __FUNCT__ 3214674ae819SStefano Zampini #define __FUNCT__ "PCBDDCScatterCoarseDataBegin" 321512edc857SStefano Zampini PetscErrorCode PCBDDCScatterCoarseDataBegin(PC pc,InsertMode imode, ScatterMode smode) 3216674ae819SStefano Zampini { 3217674ae819SStefano Zampini PetscErrorCode ierr; 3218674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 321958da7f69SStefano Zampini PetscScalar *array; 322012edc857SStefano Zampini Vec from,to; 3221674ae819SStefano Zampini 3222674ae819SStefano Zampini PetscFunctionBegin; 322312edc857SStefano Zampini if (smode == SCATTER_REVERSE) { /* from global to local -> get data from coarse solution */ 322412edc857SStefano Zampini from = pcbddc->coarse_vec; 322512edc857SStefano Zampini to = pcbddc->vec1_P; 322612edc857SStefano Zampini if (pcbddc->coarse_ksp) { /* get array from coarse processes */ 322712edc857SStefano Zampini Vec tvec; 322858da7f69SStefano Zampini 322958da7f69SStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 323058da7f69SStefano Zampini ierr = VecResetArray(tvec);CHKERRQ(ierr); 323112edc857SStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 323258da7f69SStefano Zampini ierr = VecGetArray(tvec,&array);CHKERRQ(ierr); 323358da7f69SStefano Zampini ierr = VecPlaceArray(from,array);CHKERRQ(ierr); 323458da7f69SStefano Zampini ierr = VecRestoreArray(tvec,&array);CHKERRQ(ierr); 323512edc857SStefano Zampini } 323612edc857SStefano Zampini } else { /* from local to global -> put data in coarse right hand side */ 323712edc857SStefano Zampini from = pcbddc->vec1_P; 323812edc857SStefano Zampini to = pcbddc->coarse_vec; 323912edc857SStefano Zampini } 324012edc857SStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,from,to,imode,smode);CHKERRQ(ierr); 3241674ae819SStefano Zampini PetscFunctionReturn(0); 3242674ae819SStefano Zampini } 3243674ae819SStefano Zampini 3244674ae819SStefano Zampini #undef __FUNCT__ 3245674ae819SStefano Zampini #define __FUNCT__ "PCBDDCScatterCoarseDataEnd" 324612edc857SStefano Zampini PetscErrorCode PCBDDCScatterCoarseDataEnd(PC pc, InsertMode imode, ScatterMode smode) 3247674ae819SStefano Zampini { 3248674ae819SStefano Zampini PetscErrorCode ierr; 3249674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 325058da7f69SStefano Zampini PetscScalar *array; 325112edc857SStefano Zampini Vec from,to; 3252674ae819SStefano Zampini 3253674ae819SStefano Zampini PetscFunctionBegin; 325412edc857SStefano Zampini if (smode == SCATTER_REVERSE) { /* from global to local -> get data from coarse solution */ 325512edc857SStefano Zampini from = pcbddc->coarse_vec; 325612edc857SStefano Zampini to = pcbddc->vec1_P; 325712edc857SStefano Zampini } else { /* from local to global -> put data in coarse right hand side */ 325812edc857SStefano Zampini from = pcbddc->vec1_P; 325912edc857SStefano Zampini to = pcbddc->coarse_vec; 326012edc857SStefano Zampini } 326112edc857SStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,from,to,imode,smode);CHKERRQ(ierr); 326212edc857SStefano Zampini if (smode == SCATTER_FORWARD) { 326312edc857SStefano Zampini if (pcbddc->coarse_ksp) { /* get array from coarse processes */ 326412edc857SStefano Zampini Vec tvec; 326558da7f69SStefano Zampini 326612edc857SStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 326758da7f69SStefano Zampini ierr = VecGetArray(to,&array);CHKERRQ(ierr); 326858da7f69SStefano Zampini ierr = VecPlaceArray(tvec,array);CHKERRQ(ierr); 326958da7f69SStefano Zampini ierr = VecRestoreArray(to,&array);CHKERRQ(ierr); 327058da7f69SStefano Zampini } 327158da7f69SStefano Zampini } else { 327258da7f69SStefano Zampini if (pcbddc->coarse_ksp) { /* restore array of pcbddc->coarse_vec */ 327358da7f69SStefano Zampini ierr = VecResetArray(from);CHKERRQ(ierr); 327412edc857SStefano Zampini } 327512edc857SStefano Zampini } 3276674ae819SStefano Zampini PetscFunctionReturn(0); 3277674ae819SStefano Zampini } 3278674ae819SStefano Zampini 3279984c4197SStefano Zampini /* uncomment for testing purposes */ 3280984c4197SStefano Zampini /* #define PETSC_MISSING_LAPACK_GESVD 1 */ 3281674ae819SStefano Zampini #undef __FUNCT__ 3282674ae819SStefano Zampini #define __FUNCT__ "PCBDDCConstraintsSetUp" 3283674ae819SStefano Zampini PetscErrorCode PCBDDCConstraintsSetUp(PC pc) 3284674ae819SStefano Zampini { 3285674ae819SStefano Zampini PetscErrorCode ierr; 3286674ae819SStefano Zampini PC_IS* pcis = (PC_IS*)(pc->data); 3287674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 3288674ae819SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 3289984c4197SStefano Zampini /* one and zero */ 3290984c4197SStefano Zampini PetscScalar one=1.0,zero=0.0; 3291984c4197SStefano Zampini /* space to store constraints and their local indices */ 32929162d606SStefano Zampini PetscScalar *constraints_data; 32939162d606SStefano Zampini PetscInt *constraints_idxs,*constraints_idxs_B; 32949162d606SStefano Zampini PetscInt *constraints_idxs_ptr,*constraints_data_ptr; 32959162d606SStefano Zampini PetscInt *constraints_n; 3296984c4197SStefano Zampini /* iterators */ 3297b3d85658SStefano Zampini PetscInt i,j,k,total_counts,total_counts_cc,cum; 3298984c4197SStefano Zampini /* BLAS integers */ 3299e310c8b4SStefano Zampini PetscBLASInt lwork,lierr; 3300e310c8b4SStefano Zampini PetscBLASInt Blas_N,Blas_M,Blas_K,Blas_one=1; 3301c4303822SStefano Zampini PetscBLASInt Blas_LDA,Blas_LDB,Blas_LDC; 3302727cdba6SStefano Zampini /* reuse */ 33030e6343abSStefano Zampini PetscInt olocal_primal_size,olocal_primal_size_cc; 33040e6343abSStefano Zampini PetscInt *olocal_primal_ref_node,*olocal_primal_ref_mult; 3305984c4197SStefano Zampini /* change of basis */ 3306b3d85658SStefano Zampini PetscBool qr_needed; 33079162d606SStefano Zampini PetscBT change_basis,qr_needed_idx; 3308984c4197SStefano Zampini /* auxiliary stuff */ 330964efe560SStefano Zampini PetscInt *nnz,*is_indices; 33108a0068c3SStefano Zampini PetscInt ncc; 3311984c4197SStefano Zampini /* some quantities */ 331245a1bb75SStefano Zampini PetscInt n_vertices,total_primal_vertices,valid_constraints; 3313a58a30b4SStefano Zampini PetscInt size_of_constraint,max_size_of_constraint=0,max_constraints,temp_constraints; 3314984c4197SStefano Zampini 3315674ae819SStefano Zampini PetscFunctionBegin; 33168e61c736SStefano Zampini /* Destroy Mat objects computed previously */ 33178e61c736SStefano Zampini ierr = MatDestroy(&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 33188e61c736SStefano Zampini ierr = MatDestroy(&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 3319088faed8SStefano Zampini /* save info on constraints from previous setup (if any) */ 3320088faed8SStefano Zampini olocal_primal_size = pcbddc->local_primal_size; 33210e6343abSStefano Zampini olocal_primal_size_cc = pcbddc->local_primal_size_cc; 33220e6343abSStefano Zampini ierr = PetscMalloc2(olocal_primal_size_cc,&olocal_primal_ref_node,olocal_primal_size_cc,&olocal_primal_ref_mult);CHKERRQ(ierr); 33230e6343abSStefano Zampini ierr = PetscMemcpy(olocal_primal_ref_node,pcbddc->local_primal_ref_node,olocal_primal_size_cc*sizeof(PetscInt));CHKERRQ(ierr); 33240e6343abSStefano Zampini ierr = PetscMemcpy(olocal_primal_ref_mult,pcbddc->local_primal_ref_mult,olocal_primal_size_cc*sizeof(PetscInt));CHKERRQ(ierr); 33250e6343abSStefano Zampini ierr = PetscFree2(pcbddc->local_primal_ref_node,pcbddc->local_primal_ref_mult);CHKERRQ(ierr); 3326088faed8SStefano Zampini ierr = PetscFree(pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 3327cf5a6209SStefano Zampini 3328cf5a6209SStefano Zampini /* print some info */ 3329cf5a6209SStefano Zampini if (pcbddc->dbg_flag) { 3330cf5a6209SStefano Zampini IS vertices; 3331cf5a6209SStefano Zampini PetscInt nv,nedges,nfaces; 3332cf5a6209SStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,&nfaces,NULL,&nedges,NULL,&vertices);CHKERRQ(ierr); 3333cf5a6209SStefano Zampini ierr = ISGetSize(vertices,&nv);CHKERRQ(ierr); 3334cf5a6209SStefano Zampini ierr = ISDestroy(&vertices);CHKERRQ(ierr); 33351575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 3336cf5a6209SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"--------------------------------------------------------------\n");CHKERRQ(ierr); 3337cf5a6209SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate vertices (%d)\n",PetscGlobalRank,nv,pcbddc->use_vertices);CHKERRQ(ierr); 3338fd14bc51SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate edges (%d)\n",PetscGlobalRank,nedges,pcbddc->use_edges);CHKERRQ(ierr); 3339fd14bc51SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate faces (%d)\n",PetscGlobalRank,nfaces,pcbddc->use_faces);CHKERRQ(ierr); 3340cf5a6209SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 33411575c14dSBarry Smith ierr = PetscViewerASCIIPopSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 3342cf5a6209SStefano Zampini } 3343cf5a6209SStefano Zampini 3344cf5a6209SStefano Zampini if (!pcbddc->adaptive_selection) { 33459162d606SStefano Zampini IS ISForVertices,*ISForFaces,*ISForEdges; 3346cf5a6209SStefano Zampini MatNullSpace nearnullsp; 3347cf5a6209SStefano Zampini const Vec *nearnullvecs; 3348cf5a6209SStefano Zampini Vec *localnearnullsp; 3349cf5a6209SStefano Zampini PetscScalar *array; 3350cf5a6209SStefano Zampini PetscInt n_ISForFaces,n_ISForEdges,nnsp_size; 3351cf5a6209SStefano Zampini PetscBool nnsp_has_cnst; 3352674ae819SStefano Zampini /* LAPACK working arrays for SVD or POD */ 3353b3d85658SStefano Zampini PetscBool skip_lapack,boolforchange; 3354674ae819SStefano Zampini PetscScalar *work; 3355674ae819SStefano Zampini PetscReal *singular_vals; 3356674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX) 3357674ae819SStefano Zampini PetscReal *rwork; 3358674ae819SStefano Zampini #endif 3359674ae819SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 3360674ae819SStefano Zampini PetscScalar *temp_basis,*correlation_mat; 3361674ae819SStefano Zampini #else 3362964fefecSStefano Zampini PetscBLASInt dummy_int=1; 3363964fefecSStefano Zampini PetscScalar dummy_scalar=1.; 3364674ae819SStefano Zampini #endif 3365674ae819SStefano Zampini 3366674ae819SStefano Zampini /* Get index sets for faces, edges and vertices from graph */ 3367d06fc5fdSStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,&n_ISForFaces,&ISForFaces,&n_ISForEdges,&ISForEdges,&ISForVertices);CHKERRQ(ierr); 3368d06fc5fdSStefano Zampini /* free unneeded index sets */ 3369d06fc5fdSStefano Zampini if (!pcbddc->use_vertices) { 3370d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForVertices);CHKERRQ(ierr); 3371674ae819SStefano Zampini } 3372d06fc5fdSStefano Zampini if (!pcbddc->use_edges) { 3373d06fc5fdSStefano Zampini for (i=0;i<n_ISForEdges;i++) { 3374d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForEdges[i]);CHKERRQ(ierr); 3375d06fc5fdSStefano Zampini } 3376d06fc5fdSStefano Zampini ierr = PetscFree(ISForEdges);CHKERRQ(ierr); 3377d06fc5fdSStefano Zampini n_ISForEdges = 0; 3378d06fc5fdSStefano Zampini } 3379d06fc5fdSStefano Zampini if (!pcbddc->use_faces) { 3380d06fc5fdSStefano Zampini for (i=0;i<n_ISForFaces;i++) { 3381d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForFaces[i]);CHKERRQ(ierr); 3382d06fc5fdSStefano Zampini } 3383d06fc5fdSStefano Zampini ierr = PetscFree(ISForFaces);CHKERRQ(ierr); 3384d06fc5fdSStefano Zampini n_ISForFaces = 0; 3385d06fc5fdSStefano Zampini } 338670022509SStefano Zampini 338770022509SStefano Zampini #if defined(PETSC_USE_DEBUG) 338870022509SStefano Zampini /* HACK: when solving singular problems not using vertices, a change of basis is mandatory. 338970022509SStefano Zampini Also use_change_of_basis should be consistent among processors */ 339070022509SStefano Zampini if (pcbddc->NullSpace) { 339170022509SStefano Zampini PetscBool tbool[2],gbool[2]; 339270022509SStefano Zampini 339370022509SStefano Zampini if (!ISForVertices && !pcbddc->user_ChangeOfBasisMatrix) { 3394b8ffe317SStefano Zampini pcbddc->use_change_of_basis = PETSC_TRUE; 3395d06fc5fdSStefano Zampini if (!ISForEdges) { 3396d06fc5fdSStefano Zampini pcbddc->use_change_on_faces = PETSC_TRUE; 3397d06fc5fdSStefano Zampini } 3398b8ffe317SStefano Zampini } 3399d06fc5fdSStefano Zampini tbool[0] = pcbddc->use_change_of_basis; 3400d06fc5fdSStefano Zampini tbool[1] = pcbddc->use_change_on_faces; 3401d06fc5fdSStefano Zampini ierr = MPI_Allreduce(tbool,gbool,2,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 3402d06fc5fdSStefano Zampini pcbddc->use_change_of_basis = gbool[0]; 3403d06fc5fdSStefano Zampini pcbddc->use_change_on_faces = gbool[1]; 340498a51de6SStefano Zampini } 340570022509SStefano Zampini #endif 340608122e43SStefano Zampini 3407674ae819SStefano Zampini /* check if near null space is attached to global mat */ 3408674ae819SStefano Zampini ierr = MatGetNearNullSpace(pc->pmat,&nearnullsp);CHKERRQ(ierr); 3409674ae819SStefano Zampini if (nearnullsp) { 3410674ae819SStefano Zampini ierr = MatNullSpaceGetVecs(nearnullsp,&nnsp_has_cnst,&nnsp_size,&nearnullvecs);CHKERRQ(ierr); 3411f4ddd8eeSStefano Zampini /* remove any stored info */ 3412f4ddd8eeSStefano Zampini ierr = MatNullSpaceDestroy(&pcbddc->onearnullspace);CHKERRQ(ierr); 3413f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->onearnullvecs_state);CHKERRQ(ierr); 3414f4ddd8eeSStefano Zampini /* store information for BDDC solver reuse */ 3415f4ddd8eeSStefano Zampini ierr = PetscObjectReference((PetscObject)nearnullsp);CHKERRQ(ierr); 3416f4ddd8eeSStefano Zampini pcbddc->onearnullspace = nearnullsp; 3417473ba861SJed Brown ierr = PetscMalloc1(nnsp_size,&pcbddc->onearnullvecs_state);CHKERRQ(ierr); 3418f4ddd8eeSStefano Zampini for (i=0;i<nnsp_size;i++) { 3419f4ddd8eeSStefano Zampini ierr = PetscObjectStateGet((PetscObject)nearnullvecs[i],&pcbddc->onearnullvecs_state[i]);CHKERRQ(ierr); 3420f4ddd8eeSStefano Zampini } 3421984c4197SStefano Zampini } else { /* if near null space is not provided BDDC uses constants by default */ 3422984c4197SStefano Zampini nnsp_size = 0; 3423674ae819SStefano Zampini nnsp_has_cnst = PETSC_TRUE; 3424674ae819SStefano Zampini } 3425984c4197SStefano Zampini /* get max number of constraints on a single cc */ 3426984c4197SStefano Zampini max_constraints = nnsp_size; 3427984c4197SStefano Zampini if (nnsp_has_cnst) max_constraints++; 3428984c4197SStefano Zampini 3429674ae819SStefano Zampini /* 3430674ae819SStefano Zampini Evaluate maximum storage size needed by the procedure 34319162d606SStefano Zampini - Indices for connected component i stored at "constraints_idxs + constraints_idxs_ptr[i]" 34329162d606SStefano Zampini - Values for constraints on connected component i stored at "constraints_data + constraints_data_ptr[i]" 34339162d606SStefano Zampini There can be multiple constraints per connected component 3434674ae819SStefano Zampini */ 3435674ae819SStefano Zampini n_vertices = 0; 3436674ae819SStefano Zampini if (ISForVertices) { 3437674ae819SStefano Zampini ierr = ISGetSize(ISForVertices,&n_vertices);CHKERRQ(ierr); 3438674ae819SStefano Zampini } 34399162d606SStefano Zampini ncc = n_vertices+n_ISForFaces+n_ISForEdges; 34409162d606SStefano Zampini ierr = PetscMalloc3(ncc+1,&constraints_idxs_ptr,ncc+1,&constraints_data_ptr,ncc,&constraints_n);CHKERRQ(ierr); 34419162d606SStefano Zampini 34429162d606SStefano Zampini total_counts = n_ISForFaces+n_ISForEdges; 34439162d606SStefano Zampini total_counts *= max_constraints; 3444674ae819SStefano Zampini total_counts += n_vertices; 34454641a718SStefano Zampini ierr = PetscBTCreate(total_counts,&change_basis);CHKERRQ(ierr); 34469162d606SStefano Zampini 3447674ae819SStefano Zampini total_counts = 0; 3448674ae819SStefano Zampini max_size_of_constraint = 0; 3449674ae819SStefano Zampini for (i=0;i<n_ISForEdges+n_ISForFaces;i++) { 34509162d606SStefano Zampini IS used_is; 3451674ae819SStefano Zampini if (i<n_ISForEdges) { 34529162d606SStefano Zampini used_is = ISForEdges[i]; 3453674ae819SStefano Zampini } else { 34549162d606SStefano Zampini used_is = ISForFaces[i-n_ISForEdges]; 3455674ae819SStefano Zampini } 34569162d606SStefano Zampini ierr = ISGetSize(used_is,&j);CHKERRQ(ierr); 3457674ae819SStefano Zampini total_counts += j; 3458674ae819SStefano Zampini max_size_of_constraint = PetscMax(j,max_size_of_constraint); 3459674ae819SStefano Zampini } 34609162d606SStefano 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); 34619162d606SStefano Zampini 3462984c4197SStefano Zampini /* get local part of global near null space vectors */ 3463785e854fSJed Brown ierr = PetscMalloc1(nnsp_size,&localnearnullsp);CHKERRQ(ierr); 3464984c4197SStefano Zampini for (k=0;k<nnsp_size;k++) { 3465984c4197SStefano Zampini ierr = VecDuplicate(pcis->vec1_N,&localnearnullsp[k]);CHKERRQ(ierr); 3466e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,nearnullvecs[k],localnearnullsp[k],INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 3467e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,nearnullvecs[k],localnearnullsp[k],INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 3468984c4197SStefano Zampini } 3469674ae819SStefano Zampini 3470242a89d7SStefano Zampini /* whether or not to skip lapack calls */ 3471242a89d7SStefano Zampini skip_lapack = PETSC_TRUE; 3472a773dcb8SStefano Zampini if (n_ISForFaces+n_ISForEdges && max_constraints > 1 && !pcbddc->use_nnsp_true) skip_lapack = PETSC_FALSE; 3473242a89d7SStefano Zampini 3474984c4197SStefano Zampini /* First we issue queries to allocate optimal workspace for LAPACKgesvd (or LAPACKsyev if SVD is missing) */ 3475a773dcb8SStefano Zampini if (!skip_lapack) { 3476674ae819SStefano Zampini PetscScalar temp_work; 3477911cabfeSStefano Zampini 3478674ae819SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 3479984c4197SStefano Zampini /* Proper Orthogonal Decomposition (POD) using the snapshot method */ 3480785e854fSJed Brown ierr = PetscMalloc1(max_constraints*max_constraints,&correlation_mat);CHKERRQ(ierr); 3481785e854fSJed Brown ierr = PetscMalloc1(max_constraints,&singular_vals);CHKERRQ(ierr); 3482785e854fSJed Brown ierr = PetscMalloc1(max_size_of_constraint*max_constraints,&temp_basis);CHKERRQ(ierr); 3483674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX) 3484785e854fSJed Brown ierr = PetscMalloc1(3*max_constraints,&rwork);CHKERRQ(ierr); 3485674ae819SStefano Zampini #endif 3486674ae819SStefano Zampini /* now we evaluate the optimal workspace using query with lwork=-1 */ 3487c8244a33SStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_N);CHKERRQ(ierr); 3488c8244a33SStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_LDA);CHKERRQ(ierr); 3489674ae819SStefano Zampini lwork = -1; 3490674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 3491674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 3492c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,&temp_work,&lwork,&lierr)); 3493674ae819SStefano Zampini #else 3494c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,&temp_work,&lwork,rwork,&lierr)); 3495674ae819SStefano Zampini #endif 3496674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 3497984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to SYEV Lapack routine %d",(int)lierr); 3498674ae819SStefano Zampini #else /* on missing GESVD */ 3499674ae819SStefano Zampini /* SVD */ 3500674ae819SStefano Zampini PetscInt max_n,min_n; 3501674ae819SStefano Zampini max_n = max_size_of_constraint; 3502984c4197SStefano Zampini min_n = max_constraints; 3503984c4197SStefano Zampini if (max_size_of_constraint < max_constraints) { 3504674ae819SStefano Zampini min_n = max_size_of_constraint; 3505984c4197SStefano Zampini max_n = max_constraints; 3506674ae819SStefano Zampini } 3507785e854fSJed Brown ierr = PetscMalloc1(min_n,&singular_vals);CHKERRQ(ierr); 3508674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX) 3509785e854fSJed Brown ierr = PetscMalloc1(5*min_n,&rwork);CHKERRQ(ierr); 3510674ae819SStefano Zampini #endif 3511674ae819SStefano Zampini /* now we evaluate the optimal workspace using query with lwork=-1 */ 3512674ae819SStefano Zampini lwork = -1; 3513e310c8b4SStefano Zampini ierr = PetscBLASIntCast(max_n,&Blas_M);CHKERRQ(ierr); 3514e310c8b4SStefano Zampini ierr = PetscBLASIntCast(min_n,&Blas_N);CHKERRQ(ierr); 3515b7d8b9f8SStefano Zampini ierr = PetscBLASIntCast(max_n,&Blas_LDA);CHKERRQ(ierr); 3516674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 3517674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 35189162d606SStefano 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)); 3519674ae819SStefano Zampini #else 35209162d606SStefano 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)); 3521674ae819SStefano Zampini #endif 3522674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 3523984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to GESVD Lapack routine %d",(int)lierr); 3524984c4197SStefano Zampini #endif /* on missing GESVD */ 3525674ae819SStefano Zampini /* Allocate optimal workspace */ 3526674ae819SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(temp_work),&lwork);CHKERRQ(ierr); 3527854ce69bSBarry Smith ierr = PetscMalloc1(lwork,&work);CHKERRQ(ierr); 3528674ae819SStefano Zampini } 3529674ae819SStefano Zampini /* Now we can loop on constraining sets */ 3530674ae819SStefano Zampini total_counts = 0; 35319162d606SStefano Zampini constraints_idxs_ptr[0] = 0; 35329162d606SStefano Zampini constraints_data_ptr[0] = 0; 3533674ae819SStefano Zampini /* vertices */ 35349162d606SStefano Zampini if (n_vertices) { 3535674ae819SStefano Zampini ierr = ISGetIndices(ISForVertices,(const PetscInt**)&is_indices);CHKERRQ(ierr); 35369162d606SStefano Zampini ierr = PetscMemcpy(constraints_idxs,is_indices,n_vertices*sizeof(PetscInt));CHKERRQ(ierr); 3537674ae819SStefano Zampini for (i=0;i<n_vertices;i++) { 35389162d606SStefano Zampini constraints_n[total_counts] = 1; 35399162d606SStefano Zampini constraints_data[total_counts] = 1.0; 35409162d606SStefano Zampini constraints_idxs_ptr[total_counts+1] = constraints_idxs_ptr[total_counts]+1; 35419162d606SStefano Zampini constraints_data_ptr[total_counts+1] = constraints_data_ptr[total_counts]+1; 3542674ae819SStefano Zampini total_counts++; 3543674ae819SStefano Zampini } 3544674ae819SStefano Zampini ierr = ISRestoreIndices(ISForVertices,(const PetscInt**)&is_indices);CHKERRQ(ierr); 3545674ae819SStefano Zampini n_vertices = total_counts; 3546674ae819SStefano Zampini } 3547984c4197SStefano Zampini 3548674ae819SStefano Zampini /* edges and faces */ 35499162d606SStefano Zampini total_counts_cc = total_counts; 3550911cabfeSStefano Zampini for (ncc=0;ncc<n_ISForEdges+n_ISForFaces;ncc++) { 35519162d606SStefano Zampini IS used_is; 35529162d606SStefano Zampini PetscBool idxs_copied = PETSC_FALSE; 35539162d606SStefano Zampini 3554911cabfeSStefano Zampini if (ncc<n_ISForEdges) { 35559162d606SStefano Zampini used_is = ISForEdges[ncc]; 3556984c4197SStefano Zampini boolforchange = pcbddc->use_change_of_basis; /* change or not the basis on the edge */ 3557674ae819SStefano Zampini } else { 35589162d606SStefano Zampini used_is = ISForFaces[ncc-n_ISForEdges]; 3559984c4197SStefano Zampini boolforchange = (PetscBool)(pcbddc->use_change_of_basis && pcbddc->use_change_on_faces); /* change or not the basis on the face */ 3560674ae819SStefano Zampini } 3561674ae819SStefano Zampini temp_constraints = 0; /* zero the number of constraints I have on this conn comp */ 35629162d606SStefano Zampini 35639162d606SStefano Zampini ierr = ISGetSize(used_is,&size_of_constraint);CHKERRQ(ierr); 35649162d606SStefano Zampini ierr = ISGetIndices(used_is,(const PetscInt**)&is_indices);CHKERRQ(ierr); 3565984c4197SStefano Zampini /* change of basis should not be performed on local periodic nodes */ 3566984c4197SStefano Zampini if (pcbddc->mat_graph->mirrors && pcbddc->mat_graph->mirrors[is_indices[0]]) boolforchange = PETSC_FALSE; 3567674ae819SStefano Zampini if (nnsp_has_cnst) { 35685b08dc53SStefano Zampini PetscScalar quad_value; 35699162d606SStefano Zampini 35709162d606SStefano Zampini ierr = PetscMemcpy(constraints_idxs + constraints_idxs_ptr[total_counts_cc],is_indices,size_of_constraint*sizeof(PetscInt));CHKERRQ(ierr); 35719162d606SStefano Zampini idxs_copied = PETSC_TRUE; 35729162d606SStefano Zampini 3573a773dcb8SStefano Zampini if (!pcbddc->use_nnsp_true) { 3574674ae819SStefano Zampini quad_value = (PetscScalar)(1.0/PetscSqrtReal((PetscReal)size_of_constraint)); 3575a773dcb8SStefano Zampini } else { 3576a773dcb8SStefano Zampini quad_value = 1.0; 3577a773dcb8SStefano Zampini } 3578674ae819SStefano Zampini for (j=0;j<size_of_constraint;j++) { 35799162d606SStefano Zampini constraints_data[constraints_data_ptr[total_counts_cc]+j] = quad_value; 3580674ae819SStefano Zampini } 35819162d606SStefano Zampini temp_constraints++; 3582674ae819SStefano Zampini total_counts++; 3583674ae819SStefano Zampini } 3584674ae819SStefano Zampini for (k=0;k<nnsp_size;k++) { 3585984c4197SStefano Zampini PetscReal real_value; 35869162d606SStefano Zampini PetscScalar *ptr_to_data; 35879162d606SStefano Zampini 3588984c4197SStefano Zampini ierr = VecGetArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 35899162d606SStefano Zampini ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]+temp_constraints*size_of_constraint]; 3590674ae819SStefano Zampini for (j=0;j<size_of_constraint;j++) { 35919162d606SStefano Zampini ptr_to_data[j] = array[is_indices[j]]; 3592674ae819SStefano Zampini } 3593984c4197SStefano Zampini ierr = VecRestoreArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 3594984c4197SStefano Zampini /* check if array is null on the connected component */ 3595e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 35969162d606SStefano Zampini PetscStackCallBLAS("BLASasum",real_value = BLASasum_(&Blas_N,ptr_to_data,&Blas_one)); 35975b08dc53SStefano Zampini if (real_value > 0.0) { /* keep indices and values */ 3598674ae819SStefano Zampini temp_constraints++; 3599674ae819SStefano Zampini total_counts++; 36009162d606SStefano Zampini if (!idxs_copied) { 36019162d606SStefano Zampini ierr = PetscMemcpy(constraints_idxs + constraints_idxs_ptr[total_counts_cc],is_indices,size_of_constraint*sizeof(PetscInt));CHKERRQ(ierr); 36029162d606SStefano Zampini idxs_copied = PETSC_TRUE; 3603674ae819SStefano Zampini } 3604674ae819SStefano Zampini } 36059162d606SStefano Zampini } 36069162d606SStefano Zampini ierr = ISRestoreIndices(used_is,(const PetscInt**)&is_indices);CHKERRQ(ierr); 360745a1bb75SStefano Zampini valid_constraints = temp_constraints; 3608eb97c9d2SStefano Zampini if (!pcbddc->use_nnsp_true && temp_constraints) { 3609a773dcb8SStefano Zampini if (temp_constraints == 1) { /* just normalize the constraint */ 36109162d606SStefano Zampini PetscScalar norm,*ptr_to_data; 36119162d606SStefano Zampini 36129162d606SStefano Zampini ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]]; 3613a773dcb8SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 36149162d606SStefano Zampini PetscStackCallBLAS("BLASdot",norm = BLASdot_(&Blas_N,ptr_to_data,&Blas_one,ptr_to_data,&Blas_one)); 3615a773dcb8SStefano Zampini norm = 1.0/PetscSqrtReal(PetscRealPart(norm)); 36169162d606SStefano Zampini PetscStackCallBLAS("BLASscal",BLASscal_(&Blas_N,&norm,ptr_to_data,&Blas_one)); 3617a773dcb8SStefano Zampini } else { /* perform SVD */ 3618984c4197SStefano Zampini PetscReal tol = 1.0e-8; /* tolerance for retaining eigenmodes */ 36199162d606SStefano Zampini PetscScalar *ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]]; 3620674ae819SStefano Zampini 3621674ae819SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 3622984c4197SStefano Zampini /* SVD: Y = U*S*V^H -> U (eigenvectors of Y*Y^H) = Y*V*(S)^\dag 3623984c4197SStefano Zampini POD: Y^H*Y = V*D*V^H, D = S^H*S -> U = Y*V*D^(-1/2) 3624984c4197SStefano Zampini -> When PETSC_USE_COMPLEX and PETSC_MISSING_LAPACK_GESVD are defined 3625984c4197SStefano Zampini the constraints basis will differ (by a complex factor with absolute value equal to 1) 3626984c4197SStefano Zampini from that computed using LAPACKgesvd 3627984c4197SStefano Zampini -> This is due to a different computation of eigenvectors in LAPACKheev 3628984c4197SStefano Zampini -> The quality of the POD-computed basis will be the same */ 3629984c4197SStefano Zampini ierr = PetscMemzero(correlation_mat,temp_constraints*temp_constraints*sizeof(PetscScalar));CHKERRQ(ierr); 3630674ae819SStefano Zampini /* Store upper triangular part of correlation matrix */ 3631e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 3632984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 3633674ae819SStefano Zampini for (j=0;j<temp_constraints;j++) { 3634674ae819SStefano Zampini for (k=0;k<j+1;k++) { 36359162d606SStefano 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)); 3636674ae819SStefano Zampini } 3637674ae819SStefano Zampini } 3638e310c8b4SStefano Zampini /* compute eigenvalues and eigenvectors of correlation matrix */ 3639e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 3640e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_LDA);CHKERRQ(ierr); 3641674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 3642c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,work,&lwork,&lierr)); 3643674ae819SStefano Zampini #else 3644c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,work,&lwork,rwork,&lierr)); 3645674ae819SStefano Zampini #endif 3646674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 3647984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYEV Lapack routine %d",(int)lierr); 3648984c4197SStefano Zampini /* retain eigenvalues greater than tol: note that LAPACKsyev gives eigs in ascending order */ 3649674ae819SStefano Zampini j = 0; 3650984c4197SStefano Zampini while (j < temp_constraints && singular_vals[j] < tol) j++; 3651674ae819SStefano Zampini total_counts = total_counts-j; 365245a1bb75SStefano Zampini valid_constraints = temp_constraints-j; 3653e310c8b4SStefano Zampini /* scale and copy POD basis into used quadrature memory */ 3654c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 3655c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 3656c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_K);CHKERRQ(ierr); 3657c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 3658c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_LDB);CHKERRQ(ierr); 3659c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDC);CHKERRQ(ierr); 3660674ae819SStefano Zampini if (j<temp_constraints) { 3661984c4197SStefano Zampini PetscInt ii; 3662984c4197SStefano Zampini for (k=j;k<temp_constraints;k++) singular_vals[k] = 1.0/PetscSqrtReal(singular_vals[k]); 3663674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 36649162d606SStefano 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)); 3665674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 3666984c4197SStefano Zampini for (k=0;k<temp_constraints-j;k++) { 3667674ae819SStefano Zampini for (ii=0;ii<size_of_constraint;ii++) { 36689162d606SStefano 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]; 3669674ae819SStefano Zampini } 3670674ae819SStefano Zampini } 3671674ae819SStefano Zampini } 3672674ae819SStefano Zampini #else /* on missing GESVD */ 3673e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 3674e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 3675b7d8b9f8SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 3676674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 3677674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 36789162d606SStefano 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)); 3679674ae819SStefano Zampini #else 36809162d606SStefano 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)); 3681674ae819SStefano Zampini #endif 3682984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GESVD Lapack routine %d",(int)lierr); 3683674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 3684984c4197SStefano Zampini /* retain eigenvalues greater than tol: note that LAPACKgesvd gives eigs in descending order */ 3685e310c8b4SStefano Zampini k = temp_constraints; 3686e310c8b4SStefano Zampini if (k > size_of_constraint) k = size_of_constraint; 3687674ae819SStefano Zampini j = 0; 3688e310c8b4SStefano Zampini while (j < k && singular_vals[k-j-1] < tol) j++; 368945a1bb75SStefano Zampini valid_constraints = k-j; 3690911cabfeSStefano Zampini total_counts = total_counts-temp_constraints+valid_constraints; 3691984c4197SStefano Zampini #endif /* on missing GESVD */ 3692674ae819SStefano Zampini } 3693a773dcb8SStefano Zampini } 36949162d606SStefano Zampini /* update pointers information */ 36959162d606SStefano Zampini if (valid_constraints) { 36969162d606SStefano Zampini constraints_n[total_counts_cc] = valid_constraints; 36979162d606SStefano Zampini constraints_idxs_ptr[total_counts_cc+1] = constraints_idxs_ptr[total_counts_cc]+size_of_constraint; 36989162d606SStefano Zampini constraints_data_ptr[total_counts_cc+1] = constraints_data_ptr[total_counts_cc]+size_of_constraint*valid_constraints; 36999162d606SStefano Zampini /* set change_of_basis flag */ 370045a1bb75SStefano Zampini if (boolforchange) { 3701b3d85658SStefano Zampini PetscBTSet(change_basis,total_counts_cc); 37029162d606SStefano Zampini } 3703b3d85658SStefano Zampini total_counts_cc++; 370445a1bb75SStefano Zampini } 370545a1bb75SStefano Zampini } 3706984c4197SStefano Zampini /* free workspace */ 37078f1c130eSStefano Zampini if (!skip_lapack) { 3708984c4197SStefano Zampini ierr = PetscFree(work);CHKERRQ(ierr); 3709984c4197SStefano Zampini #if defined(PETSC_USE_COMPLEX) 3710984c4197SStefano Zampini ierr = PetscFree(rwork);CHKERRQ(ierr); 3711984c4197SStefano Zampini #endif 3712984c4197SStefano Zampini ierr = PetscFree(singular_vals);CHKERRQ(ierr); 3713984c4197SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 3714984c4197SStefano Zampini ierr = PetscFree(correlation_mat);CHKERRQ(ierr); 3715984c4197SStefano Zampini ierr = PetscFree(temp_basis);CHKERRQ(ierr); 3716984c4197SStefano Zampini #endif 3717984c4197SStefano Zampini } 3718984c4197SStefano Zampini for (k=0;k<nnsp_size;k++) { 3719984c4197SStefano Zampini ierr = VecDestroy(&localnearnullsp[k]);CHKERRQ(ierr); 3720984c4197SStefano Zampini } 3721984c4197SStefano Zampini ierr = PetscFree(localnearnullsp);CHKERRQ(ierr); 3722cf5a6209SStefano Zampini /* free index sets of faces, edges and vertices */ 3723cf5a6209SStefano Zampini for (i=0;i<n_ISForFaces;i++) { 3724cf5a6209SStefano Zampini ierr = ISDestroy(&ISForFaces[i]);CHKERRQ(ierr); 3725cf5a6209SStefano Zampini } 3726cf5a6209SStefano Zampini if (n_ISForFaces) { 3727cf5a6209SStefano Zampini ierr = PetscFree(ISForFaces);CHKERRQ(ierr); 3728cf5a6209SStefano Zampini } 3729cf5a6209SStefano Zampini for (i=0;i<n_ISForEdges;i++) { 3730cf5a6209SStefano Zampini ierr = ISDestroy(&ISForEdges[i]);CHKERRQ(ierr); 3731cf5a6209SStefano Zampini } 3732cf5a6209SStefano Zampini if (n_ISForEdges) { 3733cf5a6209SStefano Zampini ierr = PetscFree(ISForEdges);CHKERRQ(ierr); 3734cf5a6209SStefano Zampini } 3735cf5a6209SStefano Zampini ierr = ISDestroy(&ISForVertices);CHKERRQ(ierr); 373608122e43SStefano Zampini } else { 373708122e43SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 3738984c4197SStefano Zampini 373908122e43SStefano Zampini total_counts = 0; 374008122e43SStefano Zampini n_vertices = 0; 3741d62866d3SStefano Zampini if (sub_schurs->is_vertices && pcbddc->use_vertices) { 3742d62866d3SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_vertices,&n_vertices);CHKERRQ(ierr); 374308122e43SStefano Zampini } 374408122e43SStefano Zampini max_constraints = 0; 37459162d606SStefano Zampini total_counts_cc = 0; 374608122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs+n_vertices;i++) { 374708122e43SStefano Zampini total_counts += pcbddc->adaptive_constraints_n[i]; 37489162d606SStefano Zampini if (pcbddc->adaptive_constraints_n[i]) total_counts_cc++; 374908122e43SStefano Zampini max_constraints = PetscMax(max_constraints,pcbddc->adaptive_constraints_n[i]); 375008122e43SStefano Zampini } 37519162d606SStefano Zampini constraints_idxs_ptr = pcbddc->adaptive_constraints_idxs_ptr; 37529162d606SStefano Zampini constraints_data_ptr = pcbddc->adaptive_constraints_data_ptr; 37539162d606SStefano Zampini constraints_idxs = pcbddc->adaptive_constraints_idxs; 37549162d606SStefano Zampini constraints_data = pcbddc->adaptive_constraints_data; 375574d5cdf7SStefano Zampini /* constraints_n differs from pcbddc->adaptive_constraints_n */ 37569162d606SStefano Zampini ierr = PetscMalloc1(total_counts_cc,&constraints_n);CHKERRQ(ierr); 37579162d606SStefano Zampini total_counts_cc = 0; 37589162d606SStefano Zampini for (i=0;i<sub_schurs->n_subs+n_vertices;i++) { 37599162d606SStefano Zampini if (pcbddc->adaptive_constraints_n[i]) { 37609162d606SStefano Zampini constraints_n[total_counts_cc++] = pcbddc->adaptive_constraints_n[i]; 376108122e43SStefano Zampini } 376208122e43SStefano Zampini } 37639162d606SStefano Zampini #if 0 37649162d606SStefano Zampini printf("Found %d totals (%d)\n",total_counts_cc,total_counts); 37659162d606SStefano Zampini for (i=0;i<total_counts_cc;i++) { 37669162d606SStefano Zampini printf("const %d, start %d",i,constraints_idxs_ptr[i]); 37679162d606SStefano Zampini printf(" end %d:\n",constraints_idxs_ptr[i+1]); 37689162d606SStefano Zampini for (j=constraints_idxs_ptr[i];j<constraints_idxs_ptr[i+1];j++) { 37699162d606SStefano Zampini printf(" %d",constraints_idxs[j]); 37709162d606SStefano Zampini } 37719162d606SStefano Zampini printf("\n"); 37729162d606SStefano Zampini printf("number of cc: %d\n",constraints_n[i]); 37739162d606SStefano Zampini } 37741b968477SStefano Zampini for (i=0;i<n_vertices;i++) { 37758bec7fa6SStefano Zampini PetscPrintf(PETSC_COMM_SELF,"[%d] vertex %d, n %d\n",PetscGlobalRank,i,pcbddc->adaptive_constraints_n[i]); 37761b968477SStefano Zampini } 37771b968477SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 37788bec7fa6SStefano 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]); 37791b968477SStefano Zampini } 378008122e43SStefano Zampini #endif 378108122e43SStefano Zampini 37828bec7fa6SStefano Zampini max_size_of_constraint = 0; 37839162d606SStefano 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]); 37849162d606SStefano Zampini ierr = PetscMalloc1(constraints_idxs_ptr[total_counts_cc],&constraints_idxs_B);CHKERRQ(ierr); 378508122e43SStefano Zampini /* Change of basis */ 3786b3d85658SStefano Zampini ierr = PetscBTCreate(total_counts_cc,&change_basis);CHKERRQ(ierr); 378708122e43SStefano Zampini if (pcbddc->use_change_of_basis) { 378808122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 378908122e43SStefano Zampini if (PetscBTLookup(sub_schurs->is_edge,i) || pcbddc->use_change_on_faces) { 3790b3d85658SStefano Zampini ierr = PetscBTSet(change_basis,i+n_vertices);CHKERRQ(ierr); 379108122e43SStefano Zampini } 379208122e43SStefano Zampini } 379308122e43SStefano Zampini } 379408122e43SStefano Zampini } 3795984c4197SStefano Zampini pcbddc->local_primal_size = total_counts; 37964f1b2e48SStefano Zampini ierr = PetscMalloc1(pcbddc->local_primal_size+pcbddc->benign_n,&pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 379708122e43SStefano Zampini 37989162d606SStefano Zampini /* map constraints_idxs in boundary numbering */ 37999162d606SStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,constraints_idxs_ptr[total_counts_cc],constraints_idxs,&i,constraints_idxs_B);CHKERRQ(ierr); 38009162d606SStefano Zampini if (i != constraints_idxs_ptr[total_counts_cc]) { 38019162d606SStefano 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); 380208122e43SStefano Zampini } 3803674ae819SStefano Zampini 3804674ae819SStefano Zampini /* Create constraint matrix */ 3805674ae819SStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 380616f15bc4SStefano Zampini ierr = MatSetType(pcbddc->ConstraintMatrix,MATAIJ);CHKERRQ(ierr); 3807984c4197SStefano Zampini ierr = MatSetSizes(pcbddc->ConstraintMatrix,pcbddc->local_primal_size,pcis->n,pcbddc->local_primal_size,pcis->n);CHKERRQ(ierr); 3808984c4197SStefano Zampini 3809984c4197SStefano Zampini /* find primal_dofs: subdomain corners plus dofs selected as primal after change of basis */ 3810a717540cSStefano Zampini /* determine if a QR strategy is needed for change of basis */ 3811a717540cSStefano Zampini qr_needed = PETSC_FALSE; 381274d5cdf7SStefano Zampini ierr = PetscBTCreate(total_counts_cc,&qr_needed_idx);CHKERRQ(ierr); 3813984c4197SStefano Zampini total_primal_vertices=0; 3814b3d85658SStefano Zampini pcbddc->local_primal_size_cc = 0; 38159162d606SStefano Zampini for (i=0;i<total_counts_cc;i++) { 38169162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 381772b8c272SStefano Zampini if (size_of_constraint == 1 && pcbddc->mat_graph->custom_minimal_size) { 38189162d606SStefano Zampini pcbddc->primal_indices_local_idxs[total_primal_vertices++] = constraints_idxs[constraints_idxs_ptr[i]]; 3819b3d85658SStefano Zampini pcbddc->local_primal_size_cc += 1; 382064efe560SStefano Zampini } else if (PetscBTLookup(change_basis,i)) { 38219162d606SStefano Zampini for (k=0;k<constraints_n[i];k++) { 38229162d606SStefano Zampini pcbddc->primal_indices_local_idxs[total_primal_vertices++] = constraints_idxs[constraints_idxs_ptr[i]+k]; 3823a717540cSStefano Zampini } 3824b3d85658SStefano Zampini pcbddc->local_primal_size_cc += constraints_n[i]; 382574d5cdf7SStefano Zampini if (constraints_n[i] > 1 || pcbddc->use_qr_single || pcbddc->faster_deluxe) { 3826a717540cSStefano Zampini PetscBTSet(qr_needed_idx,i); 3827a717540cSStefano Zampini qr_needed = PETSC_TRUE; 3828a717540cSStefano Zampini } 3829fa434743SStefano Zampini } else { 3830b3d85658SStefano Zampini pcbddc->local_primal_size_cc += 1; 3831fa434743SStefano Zampini } 3832a717540cSStefano Zampini } 3833b371cd4fSStefano Zampini /* note that the local variable n_vertices used below stores the number of pointwise constraints */ 3834b371cd4fSStefano Zampini pcbddc->n_vertices = total_primal_vertices; 3835674ae819SStefano Zampini /* permute indices in order to have a sorted set of vertices */ 383670022509SStefano Zampini ierr = PetscSortInt(total_primal_vertices,pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 3837b3d85658SStefano Zampini 38384f1b2e48SStefano 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); 38390e6343abSStefano Zampini ierr = PetscMemcpy(pcbddc->local_primal_ref_node,pcbddc->primal_indices_local_idxs,total_primal_vertices*sizeof(PetscInt));CHKERRQ(ierr); 38400e6343abSStefano Zampini for (i=0;i<total_primal_vertices;i++) pcbddc->local_primal_ref_mult[i] = 1; 3841984c4197SStefano Zampini 3842984c4197SStefano Zampini /* nonzero structure of constraint matrix */ 384374d5cdf7SStefano Zampini /* and get reference dof for local constraints */ 3844785e854fSJed Brown ierr = PetscMalloc1(pcbddc->local_primal_size,&nnz);CHKERRQ(ierr); 3845984c4197SStefano Zampini for (i=0;i<total_primal_vertices;i++) nnz[i] = 1; 384674d5cdf7SStefano Zampini 3847984c4197SStefano Zampini j = total_primal_vertices; 384874d5cdf7SStefano Zampini total_counts = total_primal_vertices; 3849b3d85658SStefano Zampini cum = total_primal_vertices; 38509162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 38514641a718SStefano Zampini if (!PetscBTLookup(change_basis,i)) { 3852b3d85658SStefano Zampini pcbddc->local_primal_ref_node[cum] = constraints_idxs[constraints_idxs_ptr[i]]; 3853b3d85658SStefano Zampini pcbddc->local_primal_ref_mult[cum] = constraints_n[i]; 3854b3d85658SStefano Zampini cum++; 38559162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 385674d5cdf7SStefano Zampini for (k=0;k<constraints_n[i];k++) { 385774d5cdf7SStefano Zampini pcbddc->primal_indices_local_idxs[total_counts++] = constraints_idxs[constraints_idxs_ptr[i]+k]; 385874d5cdf7SStefano Zampini nnz[j+k] = size_of_constraint; 385974d5cdf7SStefano Zampini } 38609162d606SStefano Zampini j += constraints_n[i]; 3861674ae819SStefano Zampini } 3862674ae819SStefano Zampini } 3863674ae819SStefano Zampini ierr = MatSeqAIJSetPreallocation(pcbddc->ConstraintMatrix,0,nnz);CHKERRQ(ierr); 3864674ae819SStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 3865088faed8SStefano Zampini 3866674ae819SStefano Zampini /* set values in constraint matrix */ 3867984c4197SStefano Zampini for (i=0;i<total_primal_vertices;i++) { 38680e6343abSStefano Zampini ierr = MatSetValue(pcbddc->ConstraintMatrix,i,pcbddc->local_primal_ref_node[i],1.0,INSERT_VALUES);CHKERRQ(ierr); 3869674ae819SStefano Zampini } 3870984c4197SStefano Zampini total_counts = total_primal_vertices; 38719162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 38724641a718SStefano Zampini if (!PetscBTLookup(change_basis,i)) { 38739162d606SStefano Zampini PetscInt *cols; 38749162d606SStefano Zampini 38759162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 38769162d606SStefano Zampini cols = constraints_idxs+constraints_idxs_ptr[i]; 38779162d606SStefano Zampini for (k=0;k<constraints_n[i];k++) { 38789162d606SStefano Zampini PetscInt row = total_counts+k; 38799162d606SStefano Zampini PetscScalar *vals; 38809162d606SStefano Zampini 38819162d606SStefano Zampini vals = constraints_data+constraints_data_ptr[i]+k*size_of_constraint; 38829162d606SStefano Zampini ierr = MatSetValues(pcbddc->ConstraintMatrix,1,&row,size_of_constraint,cols,vals,INSERT_VALUES);CHKERRQ(ierr); 38839162d606SStefano Zampini } 38849162d606SStefano Zampini total_counts += constraints_n[i]; 3885674ae819SStefano Zampini } 3886674ae819SStefano Zampini } 3887674ae819SStefano Zampini /* assembling */ 3888674ae819SStefano Zampini ierr = MatAssemblyBegin(pcbddc->ConstraintMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3889674ae819SStefano Zampini ierr = MatAssemblyEnd(pcbddc->ConstraintMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3890088faed8SStefano Zampini 3891984c4197SStefano Zampini /* 389245a1bb75SStefano Zampini ierr = PetscViewerSetFormat(PETSC_VIEWER_STDOUT_SELF,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 3893984c4197SStefano Zampini ierr = MatView(pcbddc->ConstraintMatrix,(PetscViewer)0);CHKERRQ(ierr); 3894984c4197SStefano Zampini */ 3895674ae819SStefano Zampini /* Create matrix for change of basis. We don't need it in case pcbddc->use_change_of_basis is FALSE */ 3896674ae819SStefano Zampini if (pcbddc->use_change_of_basis) { 3897026de310SStefano Zampini /* dual and primal dofs on a single cc */ 3898984c4197SStefano Zampini PetscInt dual_dofs,primal_dofs; 3899984c4197SStefano Zampini /* working stuff for GEQRF */ 390081d9aea3SBarry Smith PetscScalar *qr_basis,*qr_tau = NULL,*qr_work,lqr_work_t; 3901984c4197SStefano Zampini PetscBLASInt lqr_work; 3902984c4197SStefano Zampini /* working stuff for UNGQR */ 3903984c4197SStefano Zampini PetscScalar *gqr_work,lgqr_work_t; 3904984c4197SStefano Zampini PetscBLASInt lgqr_work; 3905984c4197SStefano Zampini /* working stuff for TRTRS */ 3906984c4197SStefano Zampini PetscScalar *trs_rhs; 39073f08241aSStefano Zampini PetscBLASInt Blas_NRHS; 3908984c4197SStefano Zampini /* pointers for values insertion into change of basis matrix */ 3909984c4197SStefano Zampini PetscInt *start_rows,*start_cols; 3910984c4197SStefano Zampini PetscScalar *start_vals; 3911984c4197SStefano Zampini /* working stuff for values insertion */ 39124641a718SStefano Zampini PetscBT is_primal; 391364efe560SStefano Zampini PetscInt *aux_primal_numbering_B; 3914906d46d4SStefano Zampini /* matrix sizes */ 3915906d46d4SStefano Zampini PetscInt global_size,local_size; 3916906d46d4SStefano Zampini /* temporary change of basis */ 3917906d46d4SStefano Zampini Mat localChangeOfBasisMatrix; 3918cf5a6209SStefano Zampini /* extra space for debugging */ 3919cf5a6209SStefano Zampini PetscScalar *dbg_work; 3920984c4197SStefano Zampini 3921906d46d4SStefano Zampini /* local temporary change of basis acts on local interfaces -> dimension is n_B x n_B */ 3922906d46d4SStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&localChangeOfBasisMatrix);CHKERRQ(ierr); 392316f15bc4SStefano Zampini ierr = MatSetType(localChangeOfBasisMatrix,MATAIJ);CHKERRQ(ierr); 3924bbb9e6c6SStefano Zampini ierr = MatSetSizes(localChangeOfBasisMatrix,pcis->n,pcis->n,pcis->n,pcis->n);CHKERRQ(ierr); 3925906d46d4SStefano Zampini /* nonzeros for local mat */ 3926bbb9e6c6SStefano Zampini ierr = PetscMalloc1(pcis->n,&nnz);CHKERRQ(ierr); 3927bbb9e6c6SStefano Zampini for (i=0;i<pcis->n;i++) nnz[i]=1; 39289162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 3929a717540cSStefano Zampini if (PetscBTLookup(change_basis,i)) { 39309162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 3931a717540cSStefano Zampini if (PetscBTLookup(qr_needed_idx,i)) { 39329162d606SStefano Zampini for (j=0;j<size_of_constraint;j++) nnz[constraints_idxs[constraints_idxs_ptr[i]+j]] = size_of_constraint; 3933a717540cSStefano Zampini } else { 39349162d606SStefano Zampini nnz[constraints_idxs[constraints_idxs_ptr[i]]] = size_of_constraint; 39359162d606SStefano Zampini for (j=1;j<size_of_constraint;j++) nnz[constraints_idxs[constraints_idxs_ptr[i]+j]] = 2; 3936a717540cSStefano Zampini } 3937a717540cSStefano Zampini } 3938a717540cSStefano Zampini } 3939906d46d4SStefano Zampini ierr = MatSeqAIJSetPreallocation(localChangeOfBasisMatrix,0,nnz);CHKERRQ(ierr); 3940bbb9e6c6SStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 3941a717540cSStefano Zampini /* Set initial identity in the matrix */ 3942bbb9e6c6SStefano Zampini for (i=0;i<pcis->n;i++) { 3943906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,i,i,1.0,INSERT_VALUES);CHKERRQ(ierr); 3944a717540cSStefano Zampini } 3945a717540cSStefano Zampini 3946a717540cSStefano Zampini if (pcbddc->dbg_flag) { 3947a717540cSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"--------------------------------------------------------------\n");CHKERRQ(ierr); 3948a717540cSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Checking change of basis computation for subdomain %04d\n",PetscGlobalRank);CHKERRQ(ierr); 3949a717540cSStefano Zampini } 3950a717540cSStefano Zampini 3951a717540cSStefano Zampini 3952a717540cSStefano Zampini /* Now we loop on the constraints which need a change of basis */ 3953a717540cSStefano Zampini /* 3954a717540cSStefano Zampini Change of basis matrix is evaluated similarly to the FIRST APPROACH in 3955a717540cSStefano Zampini Klawonn and Widlund, Dual-primal FETI-DP methods for linear elasticity, (see Sect 6.2.1) 3956a717540cSStefano Zampini 3957a6b551f4SStefano Zampini Basic blocks of change of basis matrix T computed by 3958a717540cSStefano Zampini 3959a6b551f4SStefano 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) 3960a6b551f4SStefano Zampini 3961a6b551f4SStefano Zampini | 1 0 ... 0 s_1/S | 3962a6b551f4SStefano Zampini | 0 1 ... 0 s_2/S | 3963a717540cSStefano Zampini | ... | 3964a6b551f4SStefano Zampini | 0 ... 1 s_{n-1}/S | 3965a6b551f4SStefano Zampini | -s_1/s_n ... -s_{n-1}/s_n s_n/S | 3966a717540cSStefano Zampini 3967a6b551f4SStefano Zampini with S = \sum_{i=1}^n s_i^2 3968a6b551f4SStefano Zampini NOTE: in the above example, the primal dof is the last one of the edge in LOCAL ordering 3969a6b551f4SStefano Zampini in the current implementation, the primal dof is the first one of the edge in GLOBAL ordering 3970a6b551f4SStefano Zampini 3971a6b551f4SStefano Zampini - QR decomposition of constraints otherwise 3972a717540cSStefano Zampini */ 3973a717540cSStefano Zampini if (qr_needed) { 3974984c4197SStefano Zampini /* space to store Q */ 3975854ce69bSBarry Smith ierr = PetscMalloc1(max_size_of_constraint*max_size_of_constraint,&qr_basis);CHKERRQ(ierr); 3976984c4197SStefano Zampini /* first we issue queries for optimal work */ 39773f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_M);CHKERRQ(ierr); 39783f08241aSStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_N);CHKERRQ(ierr); 39793f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 3980984c4197SStefano Zampini lqr_work = -1; 39813f08241aSStefano Zampini PetscStackCallBLAS("LAPACKgeqrf",LAPACKgeqrf_(&Blas_M,&Blas_N,qr_basis,&Blas_LDA,qr_tau,&lqr_work_t,&lqr_work,&lierr)); 3982984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to GEQRF Lapack routine %d",(int)lierr); 3983984c4197SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lqr_work_t),&lqr_work);CHKERRQ(ierr); 3984785e854fSJed Brown ierr = PetscMalloc1((PetscInt)PetscRealPart(lqr_work_t),&qr_work);CHKERRQ(ierr); 3985984c4197SStefano Zampini lgqr_work = -1; 39863f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_M);CHKERRQ(ierr); 39873f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_N);CHKERRQ(ierr); 39883f08241aSStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_K);CHKERRQ(ierr); 39893f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 39903f08241aSStefano Zampini if (Blas_K>Blas_M) Blas_K=Blas_M; /* adjust just for computing optimal work */ 39913f08241aSStefano Zampini PetscStackCallBLAS("LAPACKungqr",LAPACKungqr_(&Blas_M,&Blas_N,&Blas_K,qr_basis,&Blas_LDA,qr_tau,&lgqr_work_t,&lgqr_work,&lierr)); 3992984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to UNGQR Lapack routine %d",(int)lierr); 3993984c4197SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lgqr_work_t),&lgqr_work);CHKERRQ(ierr); 3994785e854fSJed Brown ierr = PetscMalloc1((PetscInt)PetscRealPart(lgqr_work_t),&gqr_work);CHKERRQ(ierr); 3995984c4197SStefano Zampini /* array to store scaling factors for reflectors */ 3996785e854fSJed Brown ierr = PetscMalloc1(max_constraints,&qr_tau);CHKERRQ(ierr); 3997984c4197SStefano Zampini /* array to store rhs and solution of triangular solver */ 3998785e854fSJed Brown ierr = PetscMalloc1(max_constraints*max_constraints,&trs_rhs);CHKERRQ(ierr); 3999a717540cSStefano Zampini /* allocating workspace for check */ 4000a717540cSStefano Zampini if (pcbddc->dbg_flag) { 4001cf5a6209SStefano Zampini ierr = PetscMalloc1(max_size_of_constraint*(max_constraints+max_size_of_constraint),&dbg_work);CHKERRQ(ierr); 4002a717540cSStefano Zampini } 4003a717540cSStefano Zampini } 4004984c4197SStefano Zampini /* array to store whether a node is primal or not */ 40054641a718SStefano Zampini ierr = PetscBTCreate(pcis->n_B,&is_primal);CHKERRQ(ierr); 4006473ba861SJed Brown ierr = PetscMalloc1(total_primal_vertices,&aux_primal_numbering_B);CHKERRQ(ierr); 40070e6343abSStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,total_primal_vertices,pcbddc->local_primal_ref_node,&i,aux_primal_numbering_B);CHKERRQ(ierr); 400839e2fb2aSStefano Zampini if (i != total_primal_vertices) { 400939e2fb2aSStefano Zampini SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_SUP,"Error in boundary numbering for BDDC vertices! %d != %d\n",total_primal_vertices,i); 40104641a718SStefano Zampini } 401139e2fb2aSStefano Zampini for (i=0;i<total_primal_vertices;i++) { 401239e2fb2aSStefano Zampini ierr = PetscBTSet(is_primal,aux_primal_numbering_B[i]);CHKERRQ(ierr); 401339e2fb2aSStefano Zampini } 401439e2fb2aSStefano Zampini ierr = PetscFree(aux_primal_numbering_B);CHKERRQ(ierr); 4015984c4197SStefano Zampini 4016a717540cSStefano Zampini /* loop on constraints and see whether or not they need a change of basis and compute it */ 40179162d606SStefano Zampini for (total_counts=n_vertices;total_counts<total_counts_cc;total_counts++) { 40189162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[total_counts+1]-constraints_idxs_ptr[total_counts]; 40194641a718SStefano Zampini if (PetscBTLookup(change_basis,total_counts)) { 4020984c4197SStefano Zampini /* get constraint info */ 40219162d606SStefano Zampini primal_dofs = constraints_n[total_counts]; 4022984c4197SStefano Zampini dual_dofs = size_of_constraint-primal_dofs; 4023984c4197SStefano Zampini 4024984c4197SStefano Zampini if (pcbddc->dbg_flag) { 40259162d606SStefano 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); 4026674ae819SStefano Zampini } 4027984c4197SStefano Zampini 4028fa434743SStefano Zampini if (PetscBTLookup(qr_needed_idx,total_counts)) { /* QR */ 4029a717540cSStefano Zampini 4030a717540cSStefano Zampini /* copy quadrature constraints for change of basis check */ 4031a717540cSStefano Zampini if (pcbddc->dbg_flag) { 40329162d606SStefano Zampini ierr = PetscMemcpy(dbg_work,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 4033a717540cSStefano Zampini } 4034984c4197SStefano Zampini /* copy temporary constraints into larger work vector (in order to store all columns of Q) */ 40359162d606SStefano Zampini ierr = PetscMemcpy(qr_basis,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 4036984c4197SStefano Zampini 4037984c4197SStefano Zampini /* compute QR decomposition of constraints */ 40383f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 40393f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 40403f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 4041674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 40423f08241aSStefano Zampini PetscStackCallBLAS("LAPACKgeqrf",LAPACKgeqrf_(&Blas_M,&Blas_N,qr_basis,&Blas_LDA,qr_tau,qr_work,&lqr_work,&lierr)); 4043984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GEQRF Lapack routine %d",(int)lierr); 4044674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 4045984c4197SStefano Zampini 4046984c4197SStefano Zampini /* explictly compute R^-T */ 4047984c4197SStefano Zampini ierr = PetscMemzero(trs_rhs,primal_dofs*primal_dofs*sizeof(*trs_rhs));CHKERRQ(ierr); 4048984c4197SStefano Zampini for (j=0;j<primal_dofs;j++) trs_rhs[j*(primal_dofs+1)] = 1.0; 40493f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 40503f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_NRHS);CHKERRQ(ierr); 40513f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 40523f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDB);CHKERRQ(ierr); 4053984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 40543f08241aSStefano Zampini PetscStackCallBLAS("LAPACKtrtrs",LAPACKtrtrs_("U","T","N",&Blas_N,&Blas_NRHS,qr_basis,&Blas_LDA,trs_rhs,&Blas_LDB,&lierr)); 4055984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in TRTRS Lapack routine %d",(int)lierr); 4056984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 4057984c4197SStefano Zampini 4058a717540cSStefano Zampini /* explicitly compute all columns of Q (Q = [Q1 | Q2] ) overwriting QR factorization in qr_basis */ 40593f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 40603f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 40613f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_K);CHKERRQ(ierr); 40623f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 4063984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 40643f08241aSStefano Zampini PetscStackCallBLAS("LAPACKungqr",LAPACKungqr_(&Blas_M,&Blas_N,&Blas_K,qr_basis,&Blas_LDA,qr_tau,gqr_work,&lgqr_work,&lierr)); 4065984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in UNGQR Lapack routine %d",(int)lierr); 4066984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 4067984c4197SStefano Zampini 4068984c4197SStefano Zampini /* first primal_dofs columns of Q need to be re-scaled in order to be unitary w.r.t constraints 4069984c4197SStefano Zampini i.e. C_{pxn}*Q_{nxn} should be equal to [I_pxp | 0_pxd] (see check below) 4070984c4197SStefano Zampini where n=size_of_constraint, p=primal_dofs, d=dual_dofs (n=p+d), I and 0 identity and null matrix resp. */ 40713f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 40723f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 40733f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_K);CHKERRQ(ierr); 40743f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 40753f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDB);CHKERRQ(ierr); 40763f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDC);CHKERRQ(ierr); 4077984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 40789162d606SStefano 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)); 4079984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 40809162d606SStefano Zampini ierr = PetscMemcpy(qr_basis,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 4081984c4197SStefano Zampini 4082984c4197SStefano Zampini /* insert values in change of basis matrix respecting global ordering of new primal dofs */ 40839162d606SStefano Zampini start_rows = &constraints_idxs[constraints_idxs_ptr[total_counts]]; 4084984c4197SStefano Zampini /* insert cols for primal dofs */ 4085984c4197SStefano Zampini for (j=0;j<primal_dofs;j++) { 4086984c4197SStefano Zampini start_vals = &qr_basis[j*size_of_constraint]; 40879162d606SStefano Zampini start_cols = &constraints_idxs[constraints_idxs_ptr[total_counts]+j]; 4088906d46d4SStefano Zampini ierr = MatSetValues(localChangeOfBasisMatrix,size_of_constraint,start_rows,1,start_cols,start_vals,INSERT_VALUES);CHKERRQ(ierr); 4089984c4197SStefano Zampini } 4090984c4197SStefano Zampini /* insert cols for dual dofs */ 4091984c4197SStefano Zampini for (j=0,k=0;j<dual_dofs;k++) { 40929162d606SStefano Zampini if (!PetscBTLookup(is_primal,constraints_idxs_B[constraints_idxs_ptr[total_counts]+k])) { 4093984c4197SStefano Zampini start_vals = &qr_basis[(primal_dofs+j)*size_of_constraint]; 40949162d606SStefano Zampini start_cols = &constraints_idxs[constraints_idxs_ptr[total_counts]+k]; 4095906d46d4SStefano Zampini ierr = MatSetValues(localChangeOfBasisMatrix,size_of_constraint,start_rows,1,start_cols,start_vals,INSERT_VALUES);CHKERRQ(ierr); 4096984c4197SStefano Zampini j++; 4097674ae819SStefano Zampini } 4098674ae819SStefano Zampini } 4099984c4197SStefano Zampini 4100984c4197SStefano Zampini /* check change of basis */ 4101984c4197SStefano Zampini if (pcbddc->dbg_flag) { 4102984c4197SStefano Zampini PetscInt ii,jj; 4103984c4197SStefano Zampini PetscBool valid_qr=PETSC_TRUE; 4104c4303822SStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_M);CHKERRQ(ierr); 4105c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 4106c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_K);CHKERRQ(ierr); 4107c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 4108c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDB);CHKERRQ(ierr); 4109c4303822SStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDC);CHKERRQ(ierr); 4110984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 4111cf5a6209SStefano 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)); 4112984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 4113984c4197SStefano Zampini for (jj=0;jj<size_of_constraint;jj++) { 4114984c4197SStefano Zampini for (ii=0;ii<primal_dofs;ii++) { 4115cf5a6209SStefano Zampini if (ii != jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]) > 1.e-12) valid_qr = PETSC_FALSE; 4116cf5a6209SStefano 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; 4117674ae819SStefano Zampini } 4118674ae819SStefano Zampini } 4119984c4197SStefano Zampini if (!valid_qr) { 412022d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> wrong change of basis!\n");CHKERRQ(ierr); 4121984c4197SStefano Zampini for (jj=0;jj<size_of_constraint;jj++) { 4122984c4197SStefano Zampini for (ii=0;ii<primal_dofs;ii++) { 4123cf5a6209SStefano Zampini if (ii != jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]) > 1.e-12) { 4124cf5a6209SStefano 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])); 4125674ae819SStefano Zampini } 4126cf5a6209SStefano Zampini if (ii == jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]-1.0) > 1.e-12) { 4127cf5a6209SStefano 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])); 4128984c4197SStefano Zampini } 4129984c4197SStefano Zampini } 4130984c4197SStefano Zampini } 4131674ae819SStefano Zampini } else { 413222d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> right change of basis!\n");CHKERRQ(ierr); 4133674ae819SStefano Zampini } 4134674ae819SStefano Zampini } 4135a717540cSStefano Zampini } else { /* simple transformation block */ 4136a717540cSStefano Zampini PetscInt row,col; 4137a6b551f4SStefano Zampini PetscScalar val,norm; 4138a6b551f4SStefano Zampini 4139a6b551f4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 41409162d606SStefano 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)); 4141a717540cSStefano Zampini for (j=0;j<size_of_constraint;j++) { 41429162d606SStefano Zampini PetscInt row_B = constraints_idxs_B[constraints_idxs_ptr[total_counts]+j]; 41439162d606SStefano Zampini row = constraints_idxs[constraints_idxs_ptr[total_counts]+j]; 4144bbb9e6c6SStefano Zampini if (!PetscBTLookup(is_primal,row_B)) { 41459162d606SStefano Zampini col = constraints_idxs[constraints_idxs_ptr[total_counts]]; 4146906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,row,1.0,INSERT_VALUES);CHKERRQ(ierr); 41479162d606SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,col,constraints_data[constraints_data_ptr[total_counts]+j]/norm,INSERT_VALUES);CHKERRQ(ierr); 4148a717540cSStefano Zampini } else { 4149a717540cSStefano Zampini for (k=0;k<size_of_constraint;k++) { 41509162d606SStefano Zampini col = constraints_idxs[constraints_idxs_ptr[total_counts]+k]; 4151a717540cSStefano Zampini if (row != col) { 41529162d606SStefano Zampini val = -constraints_data[constraints_data_ptr[total_counts]+k]/constraints_data[constraints_data_ptr[total_counts]]; 4153a717540cSStefano Zampini } else { 41549162d606SStefano Zampini val = constraints_data[constraints_data_ptr[total_counts]]/norm; 4155a717540cSStefano Zampini } 4156906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,col,val,INSERT_VALUES);CHKERRQ(ierr); 4157a717540cSStefano Zampini } 4158a717540cSStefano Zampini } 4159a717540cSStefano Zampini } 416098a51de6SStefano Zampini if (pcbddc->dbg_flag) { 416122d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> using standard change of basis\n");CHKERRQ(ierr); 4162a717540cSStefano Zampini } 4163674ae819SStefano Zampini } 4164984c4197SStefano Zampini } else { 4165984c4197SStefano Zampini if (pcbddc->dbg_flag) { 41669162d606SStefano 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); 4167674ae819SStefano Zampini } 4168674ae819SStefano Zampini } 4169674ae819SStefano Zampini } 4170a717540cSStefano Zampini 4171a717540cSStefano Zampini /* free workspace */ 4172a717540cSStefano Zampini if (qr_needed) { 4173984c4197SStefano Zampini if (pcbddc->dbg_flag) { 4174cf5a6209SStefano Zampini ierr = PetscFree(dbg_work);CHKERRQ(ierr); 4175984c4197SStefano Zampini } 4176984c4197SStefano Zampini ierr = PetscFree(trs_rhs);CHKERRQ(ierr); 4177984c4197SStefano Zampini ierr = PetscFree(qr_tau);CHKERRQ(ierr); 4178984c4197SStefano Zampini ierr = PetscFree(qr_work);CHKERRQ(ierr); 4179984c4197SStefano Zampini ierr = PetscFree(gqr_work);CHKERRQ(ierr); 4180984c4197SStefano Zampini ierr = PetscFree(qr_basis);CHKERRQ(ierr); 4181674ae819SStefano Zampini } 4182a717540cSStefano Zampini ierr = PetscBTDestroy(&is_primal);CHKERRQ(ierr); 4183906d46d4SStefano Zampini ierr = MatAssemblyBegin(localChangeOfBasisMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4184906d46d4SStefano Zampini ierr = MatAssemblyEnd(localChangeOfBasisMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4185906d46d4SStefano Zampini 4186906d46d4SStefano Zampini /* assembling of global change of variable */ 418788c03ad3SStefano Zampini if (!pcbddc->fake_change) { 4188bbb9e6c6SStefano Zampini Mat tmat; 418916f15bc4SStefano Zampini PetscInt bs; 419016f15bc4SStefano Zampini 4191906d46d4SStefano Zampini ierr = VecGetSize(pcis->vec1_global,&global_size);CHKERRQ(ierr); 4192906d46d4SStefano Zampini ierr = VecGetLocalSize(pcis->vec1_global,&local_size);CHKERRQ(ierr); 4193bbb9e6c6SStefano Zampini ierr = MatDuplicate(pc->pmat,MAT_DO_NOT_COPY_VALUES,&tmat);CHKERRQ(ierr); 4194bbb9e6c6SStefano Zampini ierr = MatISSetLocalMat(tmat,localChangeOfBasisMatrix);CHKERRQ(ierr); 4195bbb9e6c6SStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)pc),&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 4196bbb9e6c6SStefano Zampini ierr = MatSetType(pcbddc->ChangeOfBasisMatrix,MATAIJ);CHKERRQ(ierr); 419716f15bc4SStefano Zampini ierr = MatGetBlockSize(pc->pmat,&bs);CHKERRQ(ierr); 419816f15bc4SStefano Zampini ierr = MatSetBlockSize(pcbddc->ChangeOfBasisMatrix,bs);CHKERRQ(ierr); 4199906d46d4SStefano Zampini ierr = MatSetSizes(pcbddc->ChangeOfBasisMatrix,local_size,local_size,global_size,global_size);CHKERRQ(ierr); 4200bbb9e6c6SStefano Zampini ierr = MatISSetMPIXAIJPreallocation_Private(tmat,pcbddc->ChangeOfBasisMatrix,PETSC_TRUE);CHKERRQ(ierr); 4201bbb9e6c6SStefano Zampini ierr = MatISGetMPIXAIJ(tmat,MAT_REUSE_MATRIX,&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 4202bbb9e6c6SStefano Zampini ierr = MatDestroy(&tmat);CHKERRQ(ierr); 4203bbb9e6c6SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 4204bbb9e6c6SStefano Zampini ierr = VecSet(pcis->vec1_N,1.0);CHKERRQ(ierr); 4205e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4206e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4207bbb9e6c6SStefano Zampini ierr = VecReciprocal(pcis->vec1_global);CHKERRQ(ierr); 4208bbb9e6c6SStefano Zampini ierr = MatDiagonalScale(pcbddc->ChangeOfBasisMatrix,pcis->vec1_global,NULL);CHKERRQ(ierr); 420988c03ad3SStefano Zampini 4210906d46d4SStefano Zampini /* check */ 4211906d46d4SStefano Zampini if (pcbddc->dbg_flag) { 4212906d46d4SStefano Zampini PetscReal error; 4213906d46d4SStefano Zampini Vec x,x_change; 4214906d46d4SStefano Zampini 4215906d46d4SStefano Zampini ierr = VecDuplicate(pcis->vec1_global,&x);CHKERRQ(ierr); 4216906d46d4SStefano Zampini ierr = VecDuplicate(pcis->vec1_global,&x_change);CHKERRQ(ierr); 4217906d46d4SStefano Zampini ierr = VecSetRandom(x,NULL);CHKERRQ(ierr); 4218906d46d4SStefano Zampini ierr = VecCopy(x,pcis->vec1_global);CHKERRQ(ierr); 4219e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,x,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4220e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,x,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4221bbb9e6c6SStefano Zampini ierr = MatMult(localChangeOfBasisMatrix,pcis->vec1_N,pcis->vec2_N);CHKERRQ(ierr); 4222e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec2_N,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4223e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec2_N,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4224906d46d4SStefano Zampini ierr = MatMult(pcbddc->ChangeOfBasisMatrix,pcis->vec1_global,x_change);CHKERRQ(ierr); 4225906d46d4SStefano Zampini ierr = VecAXPY(x,-1.0,x_change);CHKERRQ(ierr); 4226906d46d4SStefano Zampini ierr = VecNorm(x,NORM_INFINITY,&error);CHKERRQ(ierr); 4227906d46d4SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4228bbb9e6c6SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Error global vs local change: %1.6e\n",error);CHKERRQ(ierr); 4229906d46d4SStefano Zampini ierr = VecDestroy(&x);CHKERRQ(ierr); 4230906d46d4SStefano Zampini ierr = VecDestroy(&x_change);CHKERRQ(ierr); 4231906d46d4SStefano Zampini } 4232b96c3477SStefano Zampini /* adapt sub_schurs computed (if any) */ 4233b96c3477SStefano Zampini if (pcbddc->use_deluxe_scaling) { 4234b96c3477SStefano Zampini PCBDDCSubSchurs sub_schurs=pcbddc->sub_schurs; 4235bf3a8328SStefano Zampini 4236bf3a8328SStefano Zampini if (pcbddc->use_change_of_basis && pcbddc->adaptive_userdefined) { 4237bf3a8328SStefano Zampini SETERRQ(PetscObjectComm((PetscObject)pc),PETSC_ERR_SUP,"Cannot mix automatic change of basis, adaptive selection and user-defined constraints");CHKERRQ(ierr); 4238bf3a8328SStefano Zampini } 4239b96c3477SStefano Zampini if (sub_schurs->S_Ej_all) { 4240ac632422SStefano Zampini Mat S_new,tmat; 4241bf3a8328SStefano Zampini IS is_all_N,is_V_Sall = NULL; 4242bbb9e6c6SStefano Zampini 4243bbb9e6c6SStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(pcis->BtoNmap,sub_schurs->is_Ej_all,&is_all_N);CHKERRQ(ierr); 42446816873aSStefano Zampini ierr = MatGetSubMatrix(localChangeOfBasisMatrix,is_all_N,is_all_N,MAT_INITIAL_MATRIX,&tmat);CHKERRQ(ierr); 4245bf3a8328SStefano Zampini if (pcbddc->deluxe_zerorows) { 4246bf3a8328SStefano Zampini ISLocalToGlobalMapping NtoSall; 4247bf3a8328SStefano Zampini IS is_V; 4248b087196eSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,pcbddc->n_vertices,pcbddc->local_primal_ref_node,PETSC_COPY_VALUES,&is_V);CHKERRQ(ierr); 4249b087196eSStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(is_all_N,&NtoSall);CHKERRQ(ierr); 4250b087196eSStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(NtoSall,IS_GTOLM_DROP,is_V,&is_V_Sall);CHKERRQ(ierr); 4251b087196eSStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&NtoSall);CHKERRQ(ierr); 4252b087196eSStefano Zampini ierr = ISDestroy(&is_V);CHKERRQ(ierr); 4253bf3a8328SStefano Zampini } 4254bf3a8328SStefano Zampini ierr = ISDestroy(&is_all_N);CHKERRQ(ierr); 4255ac632422SStefano Zampini ierr = MatPtAP(sub_schurs->S_Ej_all,tmat,MAT_INITIAL_MATRIX,1.0,&S_new);CHKERRQ(ierr); 4256b96c3477SStefano Zampini ierr = MatDestroy(&sub_schurs->S_Ej_all);CHKERRQ(ierr); 4257ac632422SStefano Zampini ierr = PetscObjectReference((PetscObject)S_new);CHKERRQ(ierr); 4258bf3a8328SStefano Zampini if (pcbddc->deluxe_zerorows) { 4259bf3a8328SStefano Zampini const PetscScalar *array; 4260bf3a8328SStefano Zampini const PetscInt *idxs_V,*idxs_all; 4261bf3a8328SStefano Zampini PetscInt i,n_V; 4262bf3a8328SStefano Zampini 4263b087196eSStefano Zampini ierr = MatZeroRowsColumnsIS(S_new,is_V_Sall,1.,NULL,NULL);CHKERRQ(ierr); 4264b087196eSStefano Zampini ierr = ISGetLocalSize(is_V_Sall,&n_V);CHKERRQ(ierr); 4265b087196eSStefano Zampini ierr = ISGetIndices(is_V_Sall,&idxs_V);CHKERRQ(ierr); 4266b087196eSStefano Zampini ierr = ISGetIndices(sub_schurs->is_Ej_all,&idxs_all);CHKERRQ(ierr); 4267b087196eSStefano Zampini ierr = VecGetArrayRead(pcis->D,&array);CHKERRQ(ierr); 4268b087196eSStefano Zampini for (i=0;i<n_V;i++) { 4269b087196eSStefano Zampini PetscScalar val; 4270b087196eSStefano Zampini PetscInt idx; 4271b087196eSStefano Zampini 4272b087196eSStefano Zampini idx = idxs_V[i]; 4273b087196eSStefano Zampini val = array[idxs_all[idxs_V[i]]]; 4274b087196eSStefano Zampini ierr = MatSetValue(S_new,idx,idx,val,INSERT_VALUES);CHKERRQ(ierr); 4275b087196eSStefano Zampini } 4276b087196eSStefano Zampini ierr = MatAssemblyBegin(S_new,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4277b087196eSStefano Zampini ierr = MatAssemblyEnd(S_new,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4278bf3a8328SStefano Zampini ierr = VecRestoreArrayRead(pcis->D,&array);CHKERRQ(ierr); 4279bf3a8328SStefano Zampini ierr = ISRestoreIndices(sub_schurs->is_Ej_all,&idxs_all);CHKERRQ(ierr); 4280bf3a8328SStefano Zampini ierr = ISRestoreIndices(is_V_Sall,&idxs_V);CHKERRQ(ierr); 4281bf3a8328SStefano Zampini } 4282ac632422SStefano Zampini sub_schurs->S_Ej_all = S_new; 4283ac632422SStefano Zampini ierr = MatDestroy(&S_new);CHKERRQ(ierr); 4284ac632422SStefano Zampini if (sub_schurs->sum_S_Ej_all) { 4285ac632422SStefano Zampini ierr = MatPtAP(sub_schurs->sum_S_Ej_all,tmat,MAT_INITIAL_MATRIX,1.0,&S_new);CHKERRQ(ierr); 4286b96c3477SStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_all);CHKERRQ(ierr); 4287ac632422SStefano Zampini ierr = PetscObjectReference((PetscObject)S_new);CHKERRQ(ierr); 4288bf3a8328SStefano Zampini if (pcbddc->deluxe_zerorows) { 4289b087196eSStefano Zampini ierr = MatZeroRowsColumnsIS(S_new,is_V_Sall,1.,NULL,NULL);CHKERRQ(ierr); 4290bf3a8328SStefano Zampini } 4291ac632422SStefano Zampini sub_schurs->sum_S_Ej_all = S_new; 4292ac632422SStefano Zampini ierr = MatDestroy(&S_new);CHKERRQ(ierr); 4293ac632422SStefano Zampini } 4294b087196eSStefano Zampini ierr = ISDestroy(&is_V_Sall);CHKERRQ(ierr); 429588c03ad3SStefano Zampini ierr = MatDestroy(&tmat);CHKERRQ(ierr); 4296b96c3477SStefano Zampini } 4297c9db6a07SStefano Zampini /* destroy any change of basis context in sub_schurs */ 4298c9db6a07SStefano Zampini if (sub_schurs->change) { 4299c9db6a07SStefano Zampini PetscInt i; 4300c9db6a07SStefano Zampini 4301c9db6a07SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 4302c9db6a07SStefano Zampini ierr = KSPDestroy(&sub_schurs->change[i]);CHKERRQ(ierr); 4303c9db6a07SStefano Zampini } 4304c9db6a07SStefano Zampini ierr = PetscFree(sub_schurs->change);CHKERRQ(ierr); 4305c9db6a07SStefano Zampini } 4306b96c3477SStefano Zampini } 4307906d46d4SStefano Zampini ierr = MatDestroy(&localChangeOfBasisMatrix);CHKERRQ(ierr); 430872b8c272SStefano Zampini } else { /* fake change (get back change of basis into ConstraintMatrix and info on qr) */ 430972b8c272SStefano Zampini ierr = MatDestroy(&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 431072b8c272SStefano Zampini pcbddc->ConstraintMatrix = localChangeOfBasisMatrix; 431172b8c272SStefano Zampini pcbddc->use_qr_single = qr_needed; 431272b8c272SStefano Zampini } 4313906d46d4SStefano Zampini } else if (pcbddc->user_ChangeOfBasisMatrix) { 4314b9b85e73SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->user_ChangeOfBasisMatrix);CHKERRQ(ierr); 4315b9b85e73SStefano Zampini pcbddc->ChangeOfBasisMatrix = pcbddc->user_ChangeOfBasisMatrix; 4316b9b85e73SStefano Zampini } 4317906d46d4SStefano Zampini 4318906d46d4SStefano Zampini /* set up change of basis context */ 4319906d46d4SStefano Zampini if (pcbddc->ChangeOfBasisMatrix) { 4320906d46d4SStefano Zampini PCBDDCChange_ctx change_ctx; 4321906d46d4SStefano Zampini 4322906d46d4SStefano Zampini if (!pcbddc->new_global_mat) { 4323906d46d4SStefano Zampini PetscInt global_size,local_size; 4324906d46d4SStefano Zampini 4325906d46d4SStefano Zampini ierr = VecGetSize(pcis->vec1_global,&global_size);CHKERRQ(ierr); 4326906d46d4SStefano Zampini ierr = VecGetLocalSize(pcis->vec1_global,&local_size);CHKERRQ(ierr); 4327906d46d4SStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)pc),&pcbddc->new_global_mat);CHKERRQ(ierr); 4328906d46d4SStefano Zampini ierr = MatSetSizes(pcbddc->new_global_mat,local_size,local_size,global_size,global_size);CHKERRQ(ierr); 4329906d46d4SStefano Zampini ierr = MatSetType(pcbddc->new_global_mat,MATSHELL);CHKERRQ(ierr); 4330906d46d4SStefano Zampini ierr = MatShellSetOperation(pcbddc->new_global_mat,MATOP_MULT,(void (*)(void))PCBDDCMatMult_Private);CHKERRQ(ierr); 4331906d46d4SStefano Zampini ierr = MatShellSetOperation(pcbddc->new_global_mat,MATOP_MULT_TRANSPOSE,(void (*)(void))PCBDDCMatMultTranspose_Private);CHKERRQ(ierr); 4332906d46d4SStefano Zampini ierr = PetscNew(&change_ctx);CHKERRQ(ierr); 4333906d46d4SStefano Zampini ierr = MatShellSetContext(pcbddc->new_global_mat,change_ctx);CHKERRQ(ierr); 4334906d46d4SStefano Zampini } else { 4335906d46d4SStefano Zampini ierr = MatShellGetContext(pcbddc->new_global_mat,&change_ctx);CHKERRQ(ierr); 4336906d46d4SStefano Zampini ierr = MatDestroy(&change_ctx->global_change);CHKERRQ(ierr); 4337906d46d4SStefano Zampini ierr = VecDestroyVecs(2,&change_ctx->work);CHKERRQ(ierr); 4338906d46d4SStefano Zampini } 4339906d46d4SStefano Zampini if (!pcbddc->user_ChangeOfBasisMatrix) { 4340906d46d4SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 4341906d46d4SStefano Zampini change_ctx->global_change = pcbddc->ChangeOfBasisMatrix; 4342906d46d4SStefano Zampini } else { 4343906d46d4SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->user_ChangeOfBasisMatrix);CHKERRQ(ierr); 4344906d46d4SStefano Zampini change_ctx->global_change = pcbddc->user_ChangeOfBasisMatrix; 4345906d46d4SStefano Zampini } 4346906d46d4SStefano Zampini ierr = VecDuplicateVecs(pcis->vec1_global,2,&change_ctx->work);CHKERRQ(ierr); 4347906d46d4SStefano Zampini ierr = MatSetUp(pcbddc->new_global_mat);CHKERRQ(ierr); 4348906d46d4SStefano Zampini ierr = MatAssemblyBegin(pcbddc->new_global_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4349906d46d4SStefano Zampini ierr = MatAssemblyEnd(pcbddc->new_global_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4350a8661815SStefano Zampini } else { 4351a8661815SStefano Zampini ierr = MatDestroy(&pcbddc->new_global_mat);CHKERRQ(ierr); 4352b9b85e73SStefano Zampini } 4353a717540cSStefano Zampini 435472b8c272SStefano Zampini if (!pcbddc->fake_change) { 43554f1b2e48SStefano Zampini /* add pressure dofs to set of primal nodes for numbering purposes */ 43564f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 43574f1b2e48SStefano Zampini pcbddc->local_primal_ref_node[pcbddc->local_primal_size_cc] = pcbddc->benign_p0_lidx[i]; 43584f1b2e48SStefano Zampini pcbddc->primal_indices_local_idxs[pcbddc->local_primal_size] = pcbddc->benign_p0_lidx[i]; 4359019a44ceSStefano Zampini pcbddc->local_primal_ref_mult[pcbddc->local_primal_size_cc] = 1; 4360019a44ceSStefano Zampini pcbddc->local_primal_size_cc++; 4361019a44ceSStefano Zampini pcbddc->local_primal_size++; 4362019a44ceSStefano Zampini } 4363019a44ceSStefano Zampini 4364019a44ceSStefano Zampini /* check if a new primal space has been introduced (also take into account benign trick) */ 4365727cdba6SStefano Zampini pcbddc->new_primal_space_local = PETSC_TRUE; 4366727cdba6SStefano Zampini if (olocal_primal_size == pcbddc->local_primal_size) { 43679f47a83aSStefano 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); 4368c1c8e736SStefano Zampini pcbddc->new_primal_space_local = (PetscBool)(!pcbddc->new_primal_space_local); 43690e6343abSStefano Zampini if (!pcbddc->new_primal_space_local) { 43709f47a83aSStefano 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); 4371727cdba6SStefano Zampini pcbddc->new_primal_space_local = (PetscBool)(!pcbddc->new_primal_space_local); 4372727cdba6SStefano Zampini } 43730e6343abSStefano Zampini } 4374727cdba6SStefano Zampini /* new_primal_space will be used for numbering of coarse dofs, so it should be the same across all subdomains */ 4375727cdba6SStefano Zampini ierr = MPI_Allreduce(&pcbddc->new_primal_space_local,&pcbddc->new_primal_space,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 437672b8c272SStefano Zampini } 437772b8c272SStefano Zampini ierr = PetscFree2(olocal_primal_ref_node,olocal_primal_ref_mult);CHKERRQ(ierr); 4378727cdba6SStefano Zampini 4379a717540cSStefano Zampini /* flush dbg viewer */ 4380b8ffe317SStefano Zampini if (pcbddc->dbg_flag) { 4381b8ffe317SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4382b8ffe317SStefano Zampini } 4383a717540cSStefano Zampini 4384e310c8b4SStefano Zampini /* free workspace */ 4385a717540cSStefano Zampini ierr = PetscBTDestroy(&qr_needed_idx);CHKERRQ(ierr); 43864641a718SStefano Zampini ierr = PetscBTDestroy(&change_basis);CHKERRQ(ierr); 438708122e43SStefano Zampini if (!pcbddc->adaptive_selection) { 43889162d606SStefano Zampini ierr = PetscFree3(constraints_idxs_ptr,constraints_data_ptr,constraints_n);CHKERRQ(ierr); 43899162d606SStefano Zampini ierr = PetscFree3(constraints_data,constraints_idxs,constraints_idxs_B);CHKERRQ(ierr); 439008122e43SStefano Zampini } else { 43919162d606SStefano Zampini ierr = PetscFree5(pcbddc->adaptive_constraints_n, 43929162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr, 43939162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr, 439408122e43SStefano Zampini pcbddc->adaptive_constraints_idxs, 439508122e43SStefano Zampini pcbddc->adaptive_constraints_data);CHKERRQ(ierr); 43969162d606SStefano Zampini ierr = PetscFree(constraints_n);CHKERRQ(ierr); 43979162d606SStefano Zampini ierr = PetscFree(constraints_idxs_B);CHKERRQ(ierr); 439808122e43SStefano Zampini } 4399674ae819SStefano Zampini PetscFunctionReturn(0); 4400674ae819SStefano Zampini } 4401674ae819SStefano Zampini 4402674ae819SStefano Zampini #undef __FUNCT__ 4403674ae819SStefano Zampini #define __FUNCT__ "PCBDDCAnalyzeInterface" 4404674ae819SStefano Zampini PetscErrorCode PCBDDCAnalyzeInterface(PC pc) 4405674ae819SStefano Zampini { 4406674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 4407674ae819SStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 4408674ae819SStefano Zampini Mat_IS *matis = (Mat_IS*)pc->pmat->data; 440914f95afaSStefano Zampini PetscInt ierr,i,N; 4410674ae819SStefano Zampini PetscViewer viewer=pcbddc->dbg_viewer; 4411674ae819SStefano Zampini 4412674ae819SStefano Zampini PetscFunctionBegin; 44138e61c736SStefano Zampini /* Reset previously computed graph */ 44148e61c736SStefano Zampini ierr = PCBDDCGraphReset(pcbddc->mat_graph);CHKERRQ(ierr); 4415674ae819SStefano Zampini /* Init local Graph struct */ 44167fb0e2dbSStefano Zampini ierr = MatGetSize(pc->pmat,&N,NULL);CHKERRQ(ierr); 44173bbff08aSStefano Zampini ierr = PCBDDCGraphInit(pcbddc->mat_graph,pcis->mapping,N);CHKERRQ(ierr); 4418674ae819SStefano Zampini 4419575ad6abSStefano Zampini /* Check validity of the csr graph passed in by the user */ 44205099eff2SStefano Zampini if (pcbddc->mat_graph->nvtxs_csr && pcbddc->mat_graph->nvtxs_csr != pcbddc->mat_graph->nvtxs) { 44215099eff2SStefano 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); 4422575ad6abSStefano Zampini } 44239577ea80SStefano Zampini 4424674ae819SStefano Zampini /* Set default CSR adjacency of local dofs if not provided by the user with PCBDDCSetLocalAdjacencyGraph */ 4425d4d8cf7bSStefano Zampini if ( (!pcbddc->mat_graph->xadj || !pcbddc->mat_graph->adjncy) && pcbddc->use_local_adj) { 44264d379d7bSStefano Zampini PetscInt *xadj,*adjncy; 44274d379d7bSStefano Zampini PetscInt nvtxs; 4428e496cd5dSStefano Zampini PetscBool flg_row=PETSC_FALSE; 4429674ae819SStefano Zampini 44302fffb893SStefano Zampini ierr = MatGetRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 44312fffb893SStefano Zampini if (flg_row) { 44324d379d7bSStefano Zampini ierr = PCBDDCSetLocalAdjacencyGraph(pc,nvtxs,xadj,adjncy,PETSC_COPY_VALUES);CHKERRQ(ierr); 4433b96c3477SStefano Zampini pcbddc->computed_rowadj = PETSC_TRUE; 44342fffb893SStefano Zampini } 44352fffb893SStefano Zampini ierr = MatRestoreRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 4436674ae819SStefano Zampini } 44379b28b941SStefano Zampini if (pcbddc->dbg_flag) { 44389b28b941SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4439674ae819SStefano Zampini } 4440674ae819SStefano Zampini 444163602bcaSStefano Zampini /* Set default dofs' splitting if no information has been provided by the user with PCBDDCSetDofsSplitting or PCBDDCSetDofsSplittingLocal */ 444263602bcaSStefano Zampini if (pcbddc->user_provided_isfordofs) { 444363602bcaSStefano Zampini if (pcbddc->n_ISForDofs) { /* need to convert from global to local and remove references to global dofs splitting */ 444495ecbf38SStefano Zampini ierr = PetscMalloc1(pcbddc->n_ISForDofs,&pcbddc->ISForDofsLocal);CHKERRQ(ierr); 444563602bcaSStefano Zampini for (i=0;i<pcbddc->n_ISForDofs;i++) { 4446e176bc59SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->rctx,pcis->vec1_global,pcis->vec1_N,pcbddc->ISForDofs[i],&pcbddc->ISForDofsLocal[i]);CHKERRQ(ierr); 444763602bcaSStefano Zampini ierr = ISDestroy(&pcbddc->ISForDofs[i]);CHKERRQ(ierr); 4448674ae819SStefano Zampini } 444963602bcaSStefano Zampini pcbddc->n_ISForDofsLocal = pcbddc->n_ISForDofs; 445063602bcaSStefano Zampini pcbddc->n_ISForDofs = 0; 445163602bcaSStefano Zampini ierr = PetscFree(pcbddc->ISForDofs);CHKERRQ(ierr); 4452674ae819SStefano Zampini } 445363602bcaSStefano Zampini } else { 445463602bcaSStefano Zampini if (!pcbddc->n_ISForDofsLocal) { /* field split not present, create it in local ordering */ 445563602bcaSStefano Zampini ierr = MatGetBlockSize(pc->pmat,&pcbddc->n_ISForDofsLocal);CHKERRQ(ierr); 4456854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->n_ISForDofsLocal,&pcbddc->ISForDofsLocal);CHKERRQ(ierr); 445763602bcaSStefano Zampini for (i=0;i<pcbddc->n_ISForDofsLocal;i++) { 445863602bcaSStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)pc),pcis->n/pcbddc->n_ISForDofsLocal,i,pcbddc->n_ISForDofsLocal,&pcbddc->ISForDofsLocal[i]);CHKERRQ(ierr); 445963602bcaSStefano Zampini } 446063602bcaSStefano Zampini } 4461674ae819SStefano Zampini } 4462674ae819SStefano Zampini 4463674ae819SStefano Zampini /* Setup of Graph */ 4464785d1243SStefano Zampini if (!pcbddc->DirichletBoundariesLocal && pcbddc->DirichletBoundaries) { /* need to convert from global to local */ 4465e176bc59SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->rctx,pcis->vec1_global,pcis->vec1_N,pcbddc->DirichletBoundaries,&pcbddc->DirichletBoundariesLocal);CHKERRQ(ierr); 4466785d1243SStefano Zampini } 4467785d1243SStefano Zampini if (!pcbddc->NeumannBoundariesLocal && pcbddc->NeumannBoundaries) { /* need to convert from global to local */ 4468e176bc59SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->rctx,pcis->vec1_global,pcis->vec1_N,pcbddc->NeumannBoundaries,&pcbddc->NeumannBoundariesLocal);CHKERRQ(ierr); 4469785d1243SStefano Zampini } 447030368db7SStefano Zampini if (!pcbddc->user_primal_vertices_local && pcbddc->user_primal_vertices) { /* need to convert from global to local */ 447130368db7SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->rctx,pcis->vec1_global,pcis->vec1_N,pcbddc->user_primal_vertices,&pcbddc->user_primal_vertices_local);CHKERRQ(ierr); 447230368db7SStefano Zampini } 447314f95afaSStefano 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); 4474674ae819SStefano Zampini 44754f1b2e48SStefano Zampini /* attach info on disconnected subdomains if present */ 44764f1b2e48SStefano Zampini if (pcbddc->n_local_subs) { 44774f1b2e48SStefano Zampini PetscInt *local_subs; 44784f1b2e48SStefano Zampini 44794f1b2e48SStefano Zampini ierr = PetscMalloc1(N,&local_subs);CHKERRQ(ierr); 44804f1b2e48SStefano Zampini for (i=0;i<pcbddc->n_local_subs;i++) { 44814f1b2e48SStefano Zampini const PetscInt *idxs; 44824f1b2e48SStefano Zampini PetscInt nl,j; 44834f1b2e48SStefano Zampini 44844f1b2e48SStefano Zampini ierr = ISGetLocalSize(pcbddc->local_subs[i],&nl);CHKERRQ(ierr); 44854f1b2e48SStefano Zampini ierr = ISGetIndices(pcbddc->local_subs[i],&idxs);CHKERRQ(ierr); 44864f1b2e48SStefano Zampini for (j=0;j<nl;j++) { 44874f1b2e48SStefano Zampini local_subs[idxs[j]] = i; 44884f1b2e48SStefano Zampini } 44894f1b2e48SStefano Zampini ierr = ISRestoreIndices(pcbddc->local_subs[i],&idxs);CHKERRQ(ierr); 44904f1b2e48SStefano Zampini } 44914f1b2e48SStefano Zampini pcbddc->mat_graph->n_local_subs = pcbddc->n_local_subs; 44924f1b2e48SStefano Zampini pcbddc->mat_graph->local_subs = local_subs; 44934f1b2e48SStefano Zampini } 44944f1b2e48SStefano Zampini 4495674ae819SStefano Zampini /* Graph's connected components analysis */ 4496674ae819SStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(pcbddc->mat_graph);CHKERRQ(ierr); 4497674ae819SStefano Zampini 4498674ae819SStefano Zampini /* print some info to stdout */ 4499674ae819SStefano Zampini if (pcbddc->dbg_flag) { 4500302440fdSBarry Smith ierr = PCBDDCGraphASCIIView(pcbddc->mat_graph,pcbddc->dbg_flag,viewer);CHKERRQ(ierr); 4501674ae819SStefano Zampini } 4502674ae819SStefano Zampini PetscFunctionReturn(0); 4503674ae819SStefano Zampini } 4504674ae819SStefano Zampini 4505dc456d91SStefano Zampini /* given an index sets possibly with holes, renumbers the indexes removing the holes */ 4506674ae819SStefano Zampini #undef __FUNCT__ 4507674ae819SStefano Zampini #define __FUNCT__ "PCBDDCSubsetNumbering" 4508dc456d91SStefano Zampini PetscErrorCode PCBDDCSubsetNumbering(IS subset, IS subset_mult, PetscInt *N_n, IS *subset_n) 4509674ae819SStefano Zampini { 4510dc456d91SStefano Zampini PetscSF sf; 4511dc456d91SStefano Zampini PetscLayout map; 4512dc456d91SStefano Zampini const PetscInt *idxs; 4513dc456d91SStefano Zampini PetscInt *leaf_data,*root_data,*gidxs; 4514dc456d91SStefano Zampini PetscInt N,n,i,lbounds[2],gbounds[2],Nl; 4515dc456d91SStefano Zampini PetscInt n_n,nlocals,start,first_index; 4516dc456d91SStefano Zampini PetscMPIInt commsize; 4517674ae819SStefano Zampini PetscBool first_found; 4518674ae819SStefano Zampini PetscErrorCode ierr; 4519674ae819SStefano Zampini 4520674ae819SStefano Zampini PetscFunctionBegin; 4521dc456d91SStefano Zampini ierr = ISGetLocalSize(subset,&n);CHKERRQ(ierr); 4522dc456d91SStefano Zampini if (subset_mult) { 4523dc456d91SStefano Zampini PetscCheckSameComm(subset,1,subset_mult,2); 4524dc456d91SStefano Zampini ierr = ISGetLocalSize(subset,&i);CHKERRQ(ierr); 4525dc456d91SStefano Zampini if (i != n) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Local subset and multiplicity sizes don't match! %d != %d",n,i); 4526674ae819SStefano Zampini } 4527dc456d91SStefano Zampini /* create workspace layout for computing global indices of subset */ 4528dc456d91SStefano Zampini ierr = ISGetIndices(subset,&idxs);CHKERRQ(ierr); 4529dc456d91SStefano Zampini lbounds[0] = lbounds[1] = 0; 4530dc456d91SStefano Zampini for (i=0;i<n;i++) { 4531dc456d91SStefano Zampini if (idxs[i] < lbounds[0]) lbounds[0] = idxs[i]; 4532dc456d91SStefano Zampini else if (idxs[i] > lbounds[1]) lbounds[1] = idxs[i]; 4533674ae819SStefano Zampini } 4534dc456d91SStefano Zampini lbounds[0] = -lbounds[0]; 4535dc456d91SStefano Zampini ierr = MPI_Allreduce(lbounds,gbounds,2,MPIU_INT,MPI_MAX,PetscObjectComm((PetscObject)subset));CHKERRQ(ierr); 4536dc456d91SStefano Zampini gbounds[0] = -gbounds[0]; 4537dc456d91SStefano Zampini N = gbounds[1] - gbounds[0] + 1; 4538dc456d91SStefano Zampini ierr = PetscLayoutCreate(PetscObjectComm((PetscObject)subset),&map);CHKERRQ(ierr); 4539dc456d91SStefano Zampini ierr = PetscLayoutSetBlockSize(map,1);CHKERRQ(ierr); 4540dc456d91SStefano Zampini ierr = PetscLayoutSetSize(map,N);CHKERRQ(ierr); 4541dc456d91SStefano Zampini ierr = PetscLayoutSetUp(map);CHKERRQ(ierr); 4542dc456d91SStefano Zampini ierr = PetscLayoutGetLocalSize(map,&Nl);CHKERRQ(ierr); 4543dc456d91SStefano Zampini 4544dc456d91SStefano Zampini /* create sf : leaf_data == multiplicity of indexes, root data == global index in layout */ 4545dc456d91SStefano Zampini ierr = PetscMalloc2(n,&leaf_data,Nl,&root_data);CHKERRQ(ierr); 4546dc456d91SStefano Zampini if (subset_mult) { 4547dc456d91SStefano Zampini const PetscInt* idxs_mult; 4548dc456d91SStefano Zampini 4549dc456d91SStefano Zampini ierr = ISGetIndices(subset_mult,&idxs_mult);CHKERRQ(ierr); 4550dc456d91SStefano Zampini ierr = PetscMemcpy(leaf_data,idxs_mult,n*sizeof(PetscInt));CHKERRQ(ierr); 4551dc456d91SStefano Zampini ierr = ISRestoreIndices(subset_mult,&idxs_mult);CHKERRQ(ierr); 4552674ae819SStefano Zampini } else { 4553dc456d91SStefano Zampini for (i=0;i<n;i++) leaf_data[i] = 1; 4554674ae819SStefano Zampini } 4555dc456d91SStefano Zampini /* local size of new subset */ 4556dc456d91SStefano Zampini n_n = 0; 4557dc456d91SStefano Zampini for (i=0;i<n;i++) n_n += leaf_data[i]; 4558dc456d91SStefano Zampini 4559dc456d91SStefano Zampini /* global indexes in layout */ 4560dc456d91SStefano Zampini ierr = PetscMalloc1(n_n,&gidxs);CHKERRQ(ierr); /* allocating possibly extra space in gidxs which will be used later */ 4561dc456d91SStefano Zampini for (i=0;i<n;i++) gidxs[i] = idxs[i] - gbounds[0]; 4562dc456d91SStefano Zampini ierr = ISRestoreIndices(subset,&idxs);CHKERRQ(ierr); 4563dc456d91SStefano Zampini ierr = PetscSFCreate(PetscObjectComm((PetscObject)subset),&sf);CHKERRQ(ierr); 4564dc456d91SStefano Zampini ierr = PetscSFSetGraphLayout(sf,map,n,NULL,PETSC_COPY_VALUES,gidxs);CHKERRQ(ierr); 4565dc456d91SStefano Zampini ierr = PetscLayoutDestroy(&map);CHKERRQ(ierr); 4566dc456d91SStefano Zampini 4567dc456d91SStefano Zampini /* reduce from leaves to roots */ 4568dc456d91SStefano Zampini ierr = PetscMemzero(root_data,Nl*sizeof(PetscInt));CHKERRQ(ierr); 456964a8e5bcSStefano Zampini ierr = PetscSFReduceBegin(sf,MPIU_INT,leaf_data,root_data,MPI_MAX);CHKERRQ(ierr); 457064a8e5bcSStefano Zampini ierr = PetscSFReduceEnd(sf,MPIU_INT,leaf_data,root_data,MPI_MAX);CHKERRQ(ierr); 4571dc456d91SStefano Zampini 4572dc456d91SStefano Zampini /* count indexes in local part of layout */ 4573674ae819SStefano Zampini nlocals = 0; 4574674ae819SStefano Zampini first_index = -1; 4575674ae819SStefano Zampini first_found = PETSC_FALSE; 4576dc456d91SStefano Zampini for (i=0;i<Nl;i++) { 4577dc456d91SStefano Zampini if (!first_found && root_data[i]) { 4578674ae819SStefano Zampini first_found = PETSC_TRUE; 4579674ae819SStefano Zampini first_index = i; 4580674ae819SStefano Zampini } 4581dc456d91SStefano Zampini nlocals += root_data[i]; 4582674ae819SStefano Zampini } 4583dc456d91SStefano Zampini 4584dc456d91SStefano Zampini /* cumulative of number of indexes and size of subset without holes */ 45855fa240b1SStefano Zampini #if defined(PETSC_HAVE_MPI_EXSCAN) 4586dc456d91SStefano Zampini start = 0; 458764a8e5bcSStefano Zampini ierr = MPI_Exscan(&nlocals,&start,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)subset));CHKERRQ(ierr); 45885fa240b1SStefano Zampini #else 458964a8e5bcSStefano Zampini ierr = MPI_Scan(&nlocals,&start,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)subset));CHKERRQ(ierr); 45905fa240b1SStefano Zampini start = start-nlocals; 45915fa240b1SStefano Zampini #endif 45925fa240b1SStefano Zampini 4593dc456d91SStefano Zampini if (N_n) { /* compute total size of new subset if requested */ 4594dc456d91SStefano Zampini *N_n = start + nlocals; 4595dc456d91SStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)subset),&commsize);CHKERRQ(ierr); 4596dc456d91SStefano Zampini ierr = MPI_Bcast(N_n,1,MPIU_INT,commsize-1,PetscObjectComm((PetscObject)subset));CHKERRQ(ierr); 4597674ae819SStefano Zampini } 45985fa240b1SStefano Zampini 45995fa240b1SStefano Zampini /* adapt root data with cumulative */ 4600674ae819SStefano Zampini if (first_found) { 4601dc456d91SStefano Zampini PetscInt old_index; 4602dc456d91SStefano Zampini 4603dc456d91SStefano Zampini root_data[first_index] += start; 4604674ae819SStefano Zampini old_index = first_index; 4605dc456d91SStefano Zampini for (i=first_index+1;i<Nl;i++) { 4606dc456d91SStefano Zampini if (root_data[i]) { 4607dc456d91SStefano Zampini root_data[i] += root_data[old_index]; 4608674ae819SStefano Zampini old_index = i; 4609674ae819SStefano Zampini } 4610674ae819SStefano Zampini } 4611674ae819SStefano Zampini } 4612dc456d91SStefano Zampini 4613dc456d91SStefano Zampini /* from roots to leaves */ 4614dc456d91SStefano Zampini ierr = PetscSFBcastBegin(sf,MPIU_INT,root_data,leaf_data);CHKERRQ(ierr); 4615dc456d91SStefano Zampini ierr = PetscSFBcastEnd(sf,MPIU_INT,root_data,leaf_data);CHKERRQ(ierr); 4616dc456d91SStefano Zampini ierr = PetscSFDestroy(&sf);CHKERRQ(ierr); 4617dc456d91SStefano Zampini 4618dc456d91SStefano Zampini /* create new IS with global indexes without holes */ 4619dc456d91SStefano Zampini if (subset_mult) { 4620dc456d91SStefano Zampini const PetscInt* idxs_mult; 4621dc456d91SStefano Zampini PetscInt cum; 4622dc456d91SStefano Zampini 4623dc456d91SStefano Zampini cum = 0; 4624dc456d91SStefano Zampini ierr = ISGetIndices(subset_mult,&idxs_mult);CHKERRQ(ierr); 4625dc456d91SStefano Zampini for (i=0;i<n;i++) { 4626dc456d91SStefano Zampini PetscInt j; 4627dc456d91SStefano Zampini for (j=0;j<idxs_mult[i];j++) gidxs[cum++] = leaf_data[i] - idxs_mult[i] + j; 4628674ae819SStefano Zampini } 4629dc456d91SStefano Zampini ierr = ISRestoreIndices(subset_mult,&idxs_mult);CHKERRQ(ierr); 4630674ae819SStefano Zampini } else { 4631dc456d91SStefano Zampini for (i=0;i<n;i++) { 4632dc456d91SStefano Zampini gidxs[i] = leaf_data[i]-1; 4633674ae819SStefano Zampini } 4634674ae819SStefano Zampini } 4635dc456d91SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)subset),n_n,gidxs,PETSC_OWN_POINTER,subset_n);CHKERRQ(ierr); 4636dc456d91SStefano Zampini ierr = PetscFree2(leaf_data,root_data);CHKERRQ(ierr); 4637674ae819SStefano Zampini PetscFunctionReturn(0); 4638674ae819SStefano Zampini } 46399a7d3425SStefano Zampini 46409a7d3425SStefano Zampini #undef __FUNCT__ 46419a7d3425SStefano Zampini #define __FUNCT__ "PCBDDCOrthonormalizeVecs" 46429a7d3425SStefano Zampini PetscErrorCode PCBDDCOrthonormalizeVecs(PetscInt n, Vec vecs[]) 46439a7d3425SStefano Zampini { 46449a7d3425SStefano Zampini PetscInt i,j; 46459a7d3425SStefano Zampini PetscScalar *alphas; 46469a7d3425SStefano Zampini PetscErrorCode ierr; 46479a7d3425SStefano Zampini 46489a7d3425SStefano Zampini PetscFunctionBegin; 46499a7d3425SStefano Zampini /* this implements stabilized Gram-Schmidt */ 4650785e854fSJed Brown ierr = PetscMalloc1(n,&alphas);CHKERRQ(ierr); 46519a7d3425SStefano Zampini for (i=0;i<n;i++) { 46529a7d3425SStefano Zampini ierr = VecNormalize(vecs[i],NULL);CHKERRQ(ierr); 46539a7d3425SStefano Zampini if (i<n) { ierr = VecMDot(vecs[i],n-i-1,&vecs[i+1],&alphas[i+1]);CHKERRQ(ierr); } 46549a7d3425SStefano Zampini for (j=i+1;j<n;j++) { ierr = VecAXPY(vecs[j],PetscConj(-alphas[j]),vecs[i]);CHKERRQ(ierr); } 46559a7d3425SStefano Zampini } 46569a7d3425SStefano Zampini ierr = PetscFree(alphas);CHKERRQ(ierr); 46579a7d3425SStefano Zampini PetscFunctionReturn(0); 46589a7d3425SStefano Zampini } 46599a7d3425SStefano Zampini 4660e7931f94SStefano Zampini #undef __FUNCT__ 466170cf5478SStefano Zampini #define __FUNCT__ "MatISGetSubassemblingPattern" 4662b0c7d250SStefano Zampini PetscErrorCode MatISGetSubassemblingPattern(Mat mat, PetscInt n_subdomains, PetscInt redprocs, IS* is_sends) 4663e7931f94SStefano Zampini { 466452e5ac9dSStefano Zampini IS ranks_send_to; 4665e7931f94SStefano Zampini PetscInt n_neighs,*neighs,*n_shared,**shared; 4666e7931f94SStefano Zampini PetscMPIInt size,rank,color; 466752e5ac9dSStefano Zampini PetscInt *xadj,*adjncy; 466852e5ac9dSStefano Zampini PetscInt *adjncy_wgt,*v_wgt,*ranks_send_to_idx; 46693837a79fSStefano Zampini PetscInt i,local_size,threshold=0; 46702b510759SStefano Zampini PetscBool use_vwgt=PETSC_FALSE,use_square=PETSC_FALSE; 4671e7931f94SStefano Zampini PetscSubcomm subcomm; 467252e5ac9dSStefano Zampini PetscErrorCode ierr; 4673a57a6d2fSStefano Zampini 4674e7931f94SStefano Zampini PetscFunctionBegin; 46752b510759SStefano Zampini ierr = PetscOptionsGetBool(NULL,"-matis_partitioning_use_square",&use_square,NULL);CHKERRQ(ierr); 46762b510759SStefano Zampini ierr = PetscOptionsGetBool(NULL,"-matis_partitioning_use_vwgt",&use_vwgt,NULL);CHKERRQ(ierr); 46772b510759SStefano Zampini ierr = PetscOptionsGetInt(NULL,"-matis_partitioning_threshold",&threshold,NULL);CHKERRQ(ierr); 4678e7931f94SStefano Zampini 4679e7931f94SStefano Zampini /* Get info on mapping */ 46803bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetSize(mat->rmap->mapping,&local_size);CHKERRQ(ierr); 46813bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetInfo(mat->rmap->mapping,&n_neighs,&neighs,&n_shared,&shared);CHKERRQ(ierr); 4682e7931f94SStefano Zampini 4683e7931f94SStefano Zampini /* build local CSR graph of subdomains' connectivity */ 4684785e854fSJed Brown ierr = PetscMalloc1(2,&xadj);CHKERRQ(ierr); 4685e7931f94SStefano Zampini xadj[0] = 0; 4686e7931f94SStefano Zampini xadj[1] = PetscMax(n_neighs-1,0); 4687785e854fSJed Brown ierr = PetscMalloc1(xadj[1],&adjncy);CHKERRQ(ierr); 4688785e854fSJed Brown ierr = PetscMalloc1(xadj[1],&adjncy_wgt);CHKERRQ(ierr); 4689e7931f94SStefano Zampini 46902b510759SStefano Zampini if (threshold) { 4691d023bfaeSStefano Zampini PetscInt xadj_count = 0; 46922b510759SStefano Zampini for (i=1;i<n_neighs;i++) { 4693d023bfaeSStefano Zampini if (n_shared[i] > threshold) { 4694d023bfaeSStefano Zampini adjncy[xadj_count] = neighs[i]; 4695d023bfaeSStefano Zampini adjncy_wgt[xadj_count] = n_shared[i]; 4696d023bfaeSStefano Zampini xadj_count++; 4697e7931f94SStefano Zampini } 4698e7931f94SStefano Zampini } 4699d023bfaeSStefano Zampini xadj[1] = xadj_count; 4700c8587f34SStefano Zampini } else { 4701e7931f94SStefano Zampini if (xadj[1]) { 4702e7931f94SStefano Zampini ierr = PetscMemcpy(adjncy,&neighs[1],xadj[1]*sizeof(*adjncy));CHKERRQ(ierr); 4703e7931f94SStefano Zampini ierr = PetscMemcpy(adjncy_wgt,&n_shared[1],xadj[1]*sizeof(*adjncy_wgt));CHKERRQ(ierr); 4704c8587f34SStefano Zampini } 4705e7931f94SStefano Zampini } 47063bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingRestoreInfo(mat->rmap->mapping,&n_neighs,&neighs,&n_shared,&shared);CHKERRQ(ierr); 4707e7931f94SStefano Zampini if (use_square) { 4708e7931f94SStefano Zampini for (i=0;i<xadj[1];i++) { 4709e7931f94SStefano Zampini adjncy_wgt[i] = adjncy_wgt[i]*adjncy_wgt[i]; 4710e7931f94SStefano Zampini } 4711e7931f94SStefano Zampini } 4712e7931f94SStefano Zampini ierr = PetscSortIntWithArray(xadj[1],adjncy,adjncy_wgt);CHKERRQ(ierr); 4713e7931f94SStefano Zampini 47143837a79fSStefano Zampini ierr = PetscMalloc1(1,&ranks_send_to_idx);CHKERRQ(ierr); 4715e7931f94SStefano Zampini 4716e7931f94SStefano Zampini /* 4717e7931f94SStefano Zampini Restrict work on active processes only. 4718e7931f94SStefano Zampini */ 4719e7931f94SStefano Zampini ierr = PetscSubcommCreate(PetscObjectComm((PetscObject)mat),&subcomm);CHKERRQ(ierr); 4720e7931f94SStefano Zampini ierr = PetscSubcommSetNumber(subcomm,2);CHKERRQ(ierr); /* 2 groups, active process and not active processes */ 4721e7931f94SStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)mat),&rank);CHKERRQ(ierr); 47222b510759SStefano Zampini ierr = PetscMPIIntCast(!local_size,&color);CHKERRQ(ierr); 4723d3531aaaSJed Brown ierr = PetscSubcommSetTypeGeneral(subcomm,color,rank);CHKERRQ(ierr); 4724e7931f94SStefano Zampini if (color) { 4725e7931f94SStefano Zampini ierr = PetscFree(xadj);CHKERRQ(ierr); 4726e7931f94SStefano Zampini ierr = PetscFree(adjncy);CHKERRQ(ierr); 4727e7931f94SStefano Zampini ierr = PetscFree(adjncy_wgt);CHKERRQ(ierr); 4728c8587f34SStefano Zampini } else { 472952e5ac9dSStefano Zampini Mat subdomain_adj; 473052e5ac9dSStefano Zampini IS new_ranks,new_ranks_contig; 473152e5ac9dSStefano Zampini MatPartitioning partitioner; 473252e5ac9dSStefano Zampini PetscInt prank,rstart=0,rend=0; 473352e5ac9dSStefano Zampini PetscInt *is_indices,*oldranks; 4734b0c7d250SStefano Zampini PetscBool aggregate; 4735b0c7d250SStefano Zampini 4736306c2d5bSBarry Smith ierr = MPI_Comm_size(PetscSubcommChild(subcomm),&size);CHKERRQ(ierr); 4737785e854fSJed Brown ierr = PetscMalloc1(size,&oldranks);CHKERRQ(ierr); 4738e7931f94SStefano Zampini prank = rank; 4739306c2d5bSBarry Smith ierr = MPI_Allgather(&prank,1,MPIU_INT,oldranks,1,MPIU_INT,PetscSubcommChild(subcomm));CHKERRQ(ierr); 47408002ef2cSStefano Zampini /* 4741e7931f94SStefano Zampini for (i=0;i<size;i++) { 4742e7931f94SStefano Zampini PetscPrintf(subcomm->comm,"oldranks[%d] = %d\n",i,oldranks[i]); 4743c8587f34SStefano Zampini } 47448002ef2cSStefano Zampini */ 4745e7931f94SStefano Zampini for (i=0;i<xadj[1];i++) { 4746e7931f94SStefano Zampini ierr = PetscFindInt(adjncy[i],size,oldranks,&adjncy[i]);CHKERRQ(ierr); 4747c8587f34SStefano Zampini } 4748e7931f94SStefano Zampini ierr = PetscSortIntWithArray(xadj[1],adjncy,adjncy_wgt);CHKERRQ(ierr); 4749b0c7d250SStefano Zampini aggregate = ((redprocs > 0 && redprocs < size) ? PETSC_TRUE : PETSC_FALSE); 4750b0c7d250SStefano Zampini if (aggregate) { 4751b0c7d250SStefano Zampini PetscInt lrows,row,ncols,*cols; 4752b0c7d250SStefano Zampini PetscMPIInt nrank; 4753b0c7d250SStefano Zampini PetscScalar *vals; 4754b0c7d250SStefano Zampini 4755b0c7d250SStefano Zampini ierr = MPI_Comm_rank(PetscSubcommChild(subcomm),&nrank);CHKERRQ(ierr); 4756b0c7d250SStefano Zampini lrows = 0; 4757b0c7d250SStefano Zampini if (nrank<redprocs) { 4758b0c7d250SStefano Zampini lrows = size/redprocs; 4759b0c7d250SStefano Zampini if (nrank<size%redprocs) lrows++; 4760b0c7d250SStefano Zampini } 47615fa240b1SStefano Zampini ierr = MatCreateAIJ(PetscSubcommChild(subcomm),lrows,lrows,size,size,50,NULL,50,NULL,&subdomain_adj);CHKERRQ(ierr); 4762b0c7d250SStefano Zampini ierr = MatGetOwnershipRange(subdomain_adj,&rstart,&rend);CHKERRQ(ierr); 4763b0c7d250SStefano Zampini ierr = MatSetOption(subdomain_adj,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_FALSE);CHKERRQ(ierr); 4764b0c7d250SStefano Zampini ierr = MatSetOption(subdomain_adj,MAT_NEW_NONZERO_ALLOCATION_ERR,PETSC_FALSE);CHKERRQ(ierr); 4765b0c7d250SStefano Zampini row = nrank; 4766b0c7d250SStefano Zampini ncols = xadj[1]-xadj[0]; 4767b0c7d250SStefano Zampini cols = adjncy; 4768b0c7d250SStefano Zampini ierr = PetscMalloc1(ncols,&vals);CHKERRQ(ierr); 4769b0c7d250SStefano Zampini for (i=0;i<ncols;i++) vals[i] = adjncy_wgt[i]; 4770b0c7d250SStefano Zampini ierr = MatSetValues(subdomain_adj,1,&row,ncols,cols,vals,INSERT_VALUES);CHKERRQ(ierr); 4771b0c7d250SStefano Zampini ierr = MatAssemblyBegin(subdomain_adj,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4772b0c7d250SStefano Zampini ierr = MatAssemblyEnd(subdomain_adj,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 477352e5ac9dSStefano Zampini ierr = PetscFree(xadj);CHKERRQ(ierr); 477452e5ac9dSStefano Zampini ierr = PetscFree(adjncy);CHKERRQ(ierr); 477552e5ac9dSStefano Zampini ierr = PetscFree(adjncy_wgt);CHKERRQ(ierr); 4776b0c7d250SStefano Zampini ierr = PetscFree(vals);CHKERRQ(ierr); 4777b0c7d250SStefano Zampini } else { 4778306c2d5bSBarry Smith ierr = MatCreateMPIAdj(PetscSubcommChild(subcomm),1,(PetscInt)size,xadj,adjncy,adjncy_wgt,&subdomain_adj);CHKERRQ(ierr); 4779b0c7d250SStefano Zampini } 478022b6e8a2SStefano Zampini /* ierr = MatView(subdomain_adj,0);CHKERRQ(ierr); */ 4781e7931f94SStefano Zampini 4782e7931f94SStefano Zampini /* Partition */ 4783306c2d5bSBarry Smith ierr = MatPartitioningCreate(PetscSubcommChild(subcomm),&partitioner);CHKERRQ(ierr); 4784e7931f94SStefano Zampini ierr = MatPartitioningSetAdjacency(partitioner,subdomain_adj);CHKERRQ(ierr); 4785e7931f94SStefano Zampini if (use_vwgt) { 47863837a79fSStefano Zampini ierr = PetscMalloc1(1,&v_wgt);CHKERRQ(ierr); 4787e7931f94SStefano Zampini v_wgt[0] = local_size; 4788e7931f94SStefano Zampini ierr = MatPartitioningSetVertexWeights(partitioner,v_wgt);CHKERRQ(ierr); 4789c8587f34SStefano Zampini } 479028143c3dSStefano Zampini n_subdomains = PetscMin((PetscInt)size,n_subdomains); 479128143c3dSStefano Zampini ierr = MatPartitioningSetNParts(partitioner,n_subdomains);CHKERRQ(ierr); 4792e7931f94SStefano Zampini ierr = MatPartitioningSetFromOptions(partitioner);CHKERRQ(ierr); 4793e7931f94SStefano Zampini ierr = MatPartitioningApply(partitioner,&new_ranks);CHKERRQ(ierr); 479422b6e8a2SStefano Zampini /* ierr = MatPartitioningView(partitioner,0);CHKERRQ(ierr); */ 4795e7931f94SStefano Zampini 479652e5ac9dSStefano Zampini /* renumber new_ranks to avoid "holes" in new set of processors */ 479752e5ac9dSStefano Zampini ierr = PCBDDCSubsetNumbering(new_ranks,NULL,NULL,&new_ranks_contig);CHKERRQ(ierr); 479852e5ac9dSStefano Zampini ierr = ISDestroy(&new_ranks);CHKERRQ(ierr); 479952e5ac9dSStefano Zampini ierr = ISGetIndices(new_ranks_contig,(const PetscInt**)&is_indices);CHKERRQ(ierr); 4800b0c7d250SStefano Zampini if (!redprocs) { 4801b0c7d250SStefano Zampini ranks_send_to_idx[0] = oldranks[is_indices[0]]; 480228143c3dSStefano Zampini } else { 4803b0c7d250SStefano Zampini PetscInt idxs[1]; 4804b0c7d250SStefano Zampini PetscMPIInt tag; 4805b0c7d250SStefano Zampini MPI_Request *reqs; 4806b0c7d250SStefano Zampini 4807b0c7d250SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)subdomain_adj,&tag);CHKERRQ(ierr); 4808b0c7d250SStefano Zampini ierr = PetscMalloc1(rend-rstart,&reqs);CHKERRQ(ierr); 4809b0c7d250SStefano Zampini for (i=rstart;i<rend;i++) { 4810b0c7d250SStefano Zampini ierr = MPI_Isend(is_indices+i-rstart,1,MPIU_INT,i,tag,PetscSubcommChild(subcomm),&reqs[i-rstart]);CHKERRQ(ierr); 481128143c3dSStefano Zampini } 4812b0c7d250SStefano Zampini ierr = MPI_Recv(idxs,1,MPIU_INT,MPI_ANY_SOURCE,tag,PetscSubcommChild(subcomm),MPI_STATUS_IGNORE);CHKERRQ(ierr); 4813b0c7d250SStefano Zampini ierr = MPI_Waitall(rend-rstart,reqs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 4814b0c7d250SStefano Zampini ierr = PetscFree(reqs);CHKERRQ(ierr); 4815b0c7d250SStefano Zampini ranks_send_to_idx[0] = oldranks[idxs[0]]; 4816e7931f94SStefano Zampini } 481752e5ac9dSStefano Zampini ierr = ISRestoreIndices(new_ranks_contig,(const PetscInt**)&is_indices);CHKERRQ(ierr); 4818e7931f94SStefano Zampini /* clean up */ 4819e7931f94SStefano Zampini ierr = PetscFree(oldranks);CHKERRQ(ierr); 482052e5ac9dSStefano Zampini ierr = ISDestroy(&new_ranks_contig);CHKERRQ(ierr); 4821e7931f94SStefano Zampini ierr = MatDestroy(&subdomain_adj);CHKERRQ(ierr); 4822e7931f94SStefano Zampini ierr = MatPartitioningDestroy(&partitioner);CHKERRQ(ierr); 4823e7931f94SStefano Zampini } 4824e7931f94SStefano Zampini ierr = PetscSubcommDestroy(&subcomm);CHKERRQ(ierr); 4825e7931f94SStefano Zampini 4826e7931f94SStefano Zampini /* assemble parallel IS for sends */ 4827e7931f94SStefano Zampini i = 1; 4828e7931f94SStefano Zampini if (color) i=0; 4829e7931f94SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)mat),i,ranks_send_to_idx,PETSC_OWN_POINTER,&ranks_send_to);CHKERRQ(ierr); 4830e7931f94SStefano Zampini /* get back IS */ 4831e7931f94SStefano Zampini *is_sends = ranks_send_to; 4832e7931f94SStefano Zampini PetscFunctionReturn(0); 4833e7931f94SStefano Zampini } 4834e7931f94SStefano Zampini 4835e7931f94SStefano Zampini typedef enum {MATDENSE_PRIVATE=0,MATAIJ_PRIVATE,MATBAIJ_PRIVATE,MATSBAIJ_PRIVATE}MatTypePrivate; 4836e7931f94SStefano Zampini 4837e7931f94SStefano Zampini #undef __FUNCT__ 4838e7931f94SStefano Zampini #define __FUNCT__ "MatISSubassemble" 483953a05cb3SStefano 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[]) 4840e7931f94SStefano Zampini { 484170cf5478SStefano Zampini Mat local_mat; 4842e7931f94SStefano Zampini IS is_sends_internal; 48439d30be91SStefano Zampini PetscInt rows,cols,new_local_rows; 484428143c3dSStefano Zampini PetscInt i,bs,buf_size_idxs,buf_size_idxs_is,buf_size_vals; 48459d30be91SStefano Zampini PetscBool ismatis,isdense,newisdense,destroy_mat; 4846e7931f94SStefano Zampini ISLocalToGlobalMapping l2gmap; 4847e7931f94SStefano Zampini PetscInt* l2gmap_indices; 4848e7931f94SStefano Zampini const PetscInt* is_indices; 4849e7931f94SStefano Zampini MatType new_local_type; 4850e7931f94SStefano Zampini /* buffers */ 4851e7931f94SStefano Zampini PetscInt *ptr_idxs,*send_buffer_idxs,*recv_buffer_idxs; 485228143c3dSStefano Zampini PetscInt *ptr_idxs_is,*send_buffer_idxs_is,*recv_buffer_idxs_is; 48539d30be91SStefano Zampini PetscInt *recv_buffer_idxs_local; 4854e7931f94SStefano Zampini PetscScalar *ptr_vals,*send_buffer_vals,*recv_buffer_vals; 4855e7931f94SStefano Zampini /* MPI */ 485628143c3dSStefano Zampini MPI_Comm comm,comm_n; 485728143c3dSStefano Zampini PetscSubcomm subcomm; 4858e7931f94SStefano Zampini PetscMPIInt n_sends,n_recvs,commsize; 485928143c3dSStefano Zampini PetscMPIInt *iflags,*ilengths_idxs,*ilengths_vals,*ilengths_idxs_is; 486028143c3dSStefano Zampini PetscMPIInt *onodes,*onodes_is,*olengths_idxs,*olengths_idxs_is,*olengths_vals; 486128143c3dSStefano Zampini PetscMPIInt len,tag_idxs,tag_idxs_is,tag_vals,source_dest; 486228143c3dSStefano Zampini MPI_Request *send_req_idxs,*send_req_idxs_is,*send_req_vals; 486328143c3dSStefano Zampini MPI_Request *recv_req_idxs,*recv_req_idxs_is,*recv_req_vals; 4864e7931f94SStefano Zampini PetscErrorCode ierr; 4865e7931f94SStefano Zampini 4866e7931f94SStefano Zampini PetscFunctionBegin; 486728143c3dSStefano Zampini /* TODO: add missing checks */ 486828143c3dSStefano Zampini PetscValidLogicalCollectiveInt(mat,n_subdomains,3); 486928143c3dSStefano Zampini PetscValidLogicalCollectiveBool(mat,restrict_comm,4); 487028143c3dSStefano Zampini PetscValidLogicalCollectiveEnum(mat,reuse,5); 487128143c3dSStefano Zampini PetscValidLogicalCollectiveInt(mat,nis,7); 4872e7931f94SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)mat,MATIS,&ismatis);CHKERRQ(ierr); 487328143c3dSStefano Zampini if (!ismatis) SETERRQ1(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot use %s on a matrix object which is not of type MATIS",__FUNCT__); 4874e7931f94SStefano Zampini ierr = MatISGetLocalMat(mat,&local_mat);CHKERRQ(ierr); 4875e7931f94SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)local_mat,MATSEQDENSE,&isdense);CHKERRQ(ierr); 4876e7931f94SStefano Zampini if (!isdense) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Currently cannot subassemble MATIS when local matrix type is not of type SEQDENSE"); 4877e7931f94SStefano Zampini ierr = MatGetSize(local_mat,&rows,&cols);CHKERRQ(ierr); 4878e7931f94SStefano Zampini if (rows != cols) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Local MATIS matrices should be square"); 487928143c3dSStefano Zampini if (reuse == MAT_REUSE_MATRIX && *mat_n) { 488070cf5478SStefano Zampini PetscInt mrows,mcols,mnrows,mncols; 488170cf5478SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)*mat_n,MATIS,&ismatis);CHKERRQ(ierr); 488228143c3dSStefano Zampini if (!ismatis) SETERRQ(PetscObjectComm((PetscObject)*mat_n),PETSC_ERR_SUP,"Cannot reuse a matrix which is not of type MATIS"); 488370cf5478SStefano Zampini ierr = MatGetSize(mat,&mrows,&mcols);CHKERRQ(ierr); 488470cf5478SStefano Zampini ierr = MatGetSize(*mat_n,&mnrows,&mncols);CHKERRQ(ierr); 488570cf5478SStefano Zampini if (mrows != mnrows) SETERRQ2(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot reuse matrix! Wrong number of rows %D != %D",mrows,mnrows); 488670cf5478SStefano Zampini if (mcols != mncols) SETERRQ2(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot reuse matrix! Wrong number of cols %D != %D",mcols,mncols); 488770cf5478SStefano Zampini } 4888e7931f94SStefano Zampini ierr = MatGetBlockSize(local_mat,&bs);CHKERRQ(ierr); 4889e7931f94SStefano Zampini PetscValidLogicalCollectiveInt(mat,bs,0); 4890e7931f94SStefano Zampini /* prepare IS for sending if not provided */ 4891e7931f94SStefano Zampini if (!is_sends) { 489228143c3dSStefano Zampini if (!n_subdomains) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"You should specify either an IS or a target number of subdomains"); 4893b0c7d250SStefano Zampini ierr = MatISGetSubassemblingPattern(mat,n_subdomains,0,&is_sends_internal);CHKERRQ(ierr); 4894c8587f34SStefano Zampini } else { 4895e7931f94SStefano Zampini ierr = PetscObjectReference((PetscObject)is_sends);CHKERRQ(ierr); 4896e7931f94SStefano Zampini is_sends_internal = is_sends; 4897c8587f34SStefano Zampini } 4898e7931f94SStefano Zampini 4899e7931f94SStefano Zampini /* get comm */ 4900a316fed8SStefano Zampini ierr = PetscObjectGetComm((PetscObject)mat,&comm);CHKERRQ(ierr); 4901e7931f94SStefano Zampini 4902e7931f94SStefano Zampini /* compute number of sends */ 4903e7931f94SStefano Zampini ierr = ISGetLocalSize(is_sends_internal,&i);CHKERRQ(ierr); 4904e7931f94SStefano Zampini ierr = PetscMPIIntCast(i,&n_sends);CHKERRQ(ierr); 4905e7931f94SStefano Zampini 4906e7931f94SStefano Zampini /* compute number of receives */ 4907e7931f94SStefano Zampini ierr = MPI_Comm_size(comm,&commsize);CHKERRQ(ierr); 4908785e854fSJed Brown ierr = PetscMalloc1(commsize,&iflags);CHKERRQ(ierr); 4909e7931f94SStefano Zampini ierr = PetscMemzero(iflags,commsize*sizeof(*iflags));CHKERRQ(ierr); 4910e7931f94SStefano Zampini ierr = ISGetIndices(is_sends_internal,&is_indices);CHKERRQ(ierr); 4911e7931f94SStefano Zampini for (i=0;i<n_sends;i++) iflags[is_indices[i]] = 1; 4912e7931f94SStefano Zampini ierr = PetscGatherNumberOfMessages(comm,iflags,NULL,&n_recvs);CHKERRQ(ierr); 4913e7931f94SStefano Zampini ierr = PetscFree(iflags);CHKERRQ(ierr); 4914e7931f94SStefano Zampini 491528143c3dSStefano Zampini /* restrict comm if requested */ 491628143c3dSStefano Zampini subcomm = 0; 491728143c3dSStefano Zampini destroy_mat = PETSC_FALSE; 491828143c3dSStefano Zampini if (restrict_comm) { 4919779c1cceSStefano Zampini PetscMPIInt color,subcommsize; 4920779c1cceSStefano Zampini 492128143c3dSStefano Zampini color = 0; 492253a05cb3SStefano Zampini if (restrict_full) { 492353a05cb3SStefano Zampini if (!n_recvs) color = 1; /* processes not receiving anything will not partecipate in new comm (full restriction) */ 492453a05cb3SStefano Zampini } else { 492553a05cb3SStefano Zampini if (!n_recvs && n_sends) color = 1; /* just those processes that are sending but not receiving anything will not partecipate in new comm */ 492653a05cb3SStefano Zampini } 492728143c3dSStefano Zampini ierr = MPI_Allreduce(&color,&subcommsize,1,MPI_INT,MPI_SUM,comm);CHKERRQ(ierr); 492828143c3dSStefano Zampini subcommsize = commsize - subcommsize; 492928143c3dSStefano Zampini /* check if reuse has been requested */ 493028143c3dSStefano Zampini if (reuse == MAT_REUSE_MATRIX) { 493128143c3dSStefano Zampini if (*mat_n) { 493228143c3dSStefano Zampini PetscMPIInt subcommsize2; 493328143c3dSStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)*mat_n),&subcommsize2);CHKERRQ(ierr); 493428143c3dSStefano Zampini if (subcommsize != subcommsize2) SETERRQ2(PetscObjectComm((PetscObject)*mat_n),PETSC_ERR_PLIB,"Cannot reuse matrix! wrong subcomm size %d != %d",subcommsize,subcommsize2); 493528143c3dSStefano Zampini comm_n = PetscObjectComm((PetscObject)*mat_n); 493628143c3dSStefano Zampini } else { 493728143c3dSStefano Zampini comm_n = PETSC_COMM_SELF; 493828143c3dSStefano Zampini } 493928143c3dSStefano Zampini } else { /* MAT_INITIAL_MATRIX */ 4940779c1cceSStefano Zampini PetscMPIInt rank; 4941779c1cceSStefano Zampini 4942779c1cceSStefano Zampini ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 494328143c3dSStefano Zampini ierr = PetscSubcommCreate(comm,&subcomm);CHKERRQ(ierr); 494428143c3dSStefano Zampini ierr = PetscSubcommSetNumber(subcomm,2);CHKERRQ(ierr); 494528143c3dSStefano Zampini ierr = PetscSubcommSetTypeGeneral(subcomm,color,rank);CHKERRQ(ierr); 4946306c2d5bSBarry Smith comm_n = PetscSubcommChild(subcomm); 494728143c3dSStefano Zampini } 494828143c3dSStefano Zampini /* flag to destroy *mat_n if not significative */ 494928143c3dSStefano Zampini if (color) destroy_mat = PETSC_TRUE; 495028143c3dSStefano Zampini } else { 495128143c3dSStefano Zampini comm_n = comm; 495228143c3dSStefano Zampini } 495328143c3dSStefano Zampini 4954e7931f94SStefano Zampini /* prepare send/receive buffers */ 4955785e854fSJed Brown ierr = PetscMalloc1(commsize,&ilengths_idxs);CHKERRQ(ierr); 4956e7931f94SStefano Zampini ierr = PetscMemzero(ilengths_idxs,commsize*sizeof(*ilengths_idxs));CHKERRQ(ierr); 4957785e854fSJed Brown ierr = PetscMalloc1(commsize,&ilengths_vals);CHKERRQ(ierr); 4958e7931f94SStefano Zampini ierr = PetscMemzero(ilengths_vals,commsize*sizeof(*ilengths_vals));CHKERRQ(ierr); 495928143c3dSStefano Zampini if (nis) { 4960854ce69bSBarry Smith ierr = PetscCalloc1(commsize,&ilengths_idxs_is);CHKERRQ(ierr); 496128143c3dSStefano Zampini } 4962e7931f94SStefano Zampini 496328143c3dSStefano Zampini /* Get data from local matrices */ 4964e7931f94SStefano Zampini if (!isdense) { 4965a26c9d0eSStefano Zampini SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Subassembling of AIJ local matrices not yet implemented"); 4966e7931f94SStefano Zampini /* TODO: See below some guidelines on how to prepare the local buffers */ 4967e7931f94SStefano Zampini /* 4968e7931f94SStefano Zampini send_buffer_vals should contain the raw values of the local matrix 4969e7931f94SStefano Zampini send_buffer_idxs should contain: 4970e7931f94SStefano Zampini - MatType_PRIVATE type 4971e7931f94SStefano Zampini - PetscInt size_of_l2gmap 4972e7931f94SStefano Zampini - PetscInt global_row_indices[size_of_l2gmap] 4973e7931f94SStefano Zampini - PetscInt all_other_info_which_is_needed_to_compute_preallocation_and_set_values 4974e7931f94SStefano Zampini */ 4975e7931f94SStefano Zampini } else { 4976e7931f94SStefano Zampini ierr = MatDenseGetArray(local_mat,&send_buffer_vals);CHKERRQ(ierr); 49773bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetSize(mat->rmap->mapping,&i);CHKERRQ(ierr); 4978854ce69bSBarry Smith ierr = PetscMalloc1(i+2,&send_buffer_idxs);CHKERRQ(ierr); 4979e7931f94SStefano Zampini send_buffer_idxs[0] = (PetscInt)MATDENSE_PRIVATE; 4980e7931f94SStefano Zampini send_buffer_idxs[1] = i; 49813bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetIndices(mat->rmap->mapping,(const PetscInt**)&ptr_idxs);CHKERRQ(ierr); 4982e7931f94SStefano Zampini ierr = PetscMemcpy(&send_buffer_idxs[2],ptr_idxs,i*sizeof(PetscInt));CHKERRQ(ierr); 49833bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingRestoreIndices(mat->rmap->mapping,(const PetscInt**)&ptr_idxs);CHKERRQ(ierr); 4984e7931f94SStefano Zampini ierr = PetscMPIIntCast(i,&len);CHKERRQ(ierr); 4985e7931f94SStefano Zampini for (i=0;i<n_sends;i++) { 4986e7931f94SStefano Zampini ilengths_vals[is_indices[i]] = len*len; 4987e7931f94SStefano Zampini ilengths_idxs[is_indices[i]] = len+2; 4988c8587f34SStefano Zampini } 4989c8587f34SStefano Zampini } 4990e7931f94SStefano Zampini ierr = PetscGatherMessageLengths2(comm,n_sends,n_recvs,ilengths_idxs,ilengths_vals,&onodes,&olengths_idxs,&olengths_vals);CHKERRQ(ierr); 499128143c3dSStefano Zampini /* additional is (if any) */ 499228143c3dSStefano Zampini if (nis) { 499328143c3dSStefano Zampini PetscMPIInt psum; 499428143c3dSStefano Zampini PetscInt j; 499528143c3dSStefano Zampini for (j=0,psum=0;j<nis;j++) { 499628143c3dSStefano Zampini PetscInt plen; 499728143c3dSStefano Zampini ierr = ISGetLocalSize(isarray[j],&plen);CHKERRQ(ierr); 499828143c3dSStefano Zampini ierr = PetscMPIIntCast(plen,&len);CHKERRQ(ierr); 499928143c3dSStefano Zampini psum += len+1; /* indices + lenght */ 500028143c3dSStefano Zampini } 5001854ce69bSBarry Smith ierr = PetscMalloc1(psum,&send_buffer_idxs_is);CHKERRQ(ierr); 500228143c3dSStefano Zampini for (j=0,psum=0;j<nis;j++) { 500328143c3dSStefano Zampini PetscInt plen; 500428143c3dSStefano Zampini const PetscInt *is_array_idxs; 500528143c3dSStefano Zampini ierr = ISGetLocalSize(isarray[j],&plen);CHKERRQ(ierr); 500628143c3dSStefano Zampini send_buffer_idxs_is[psum] = plen; 500728143c3dSStefano Zampini ierr = ISGetIndices(isarray[j],&is_array_idxs);CHKERRQ(ierr); 500828143c3dSStefano Zampini ierr = PetscMemcpy(&send_buffer_idxs_is[psum+1],is_array_idxs,plen*sizeof(PetscInt));CHKERRQ(ierr); 500928143c3dSStefano Zampini ierr = ISRestoreIndices(isarray[j],&is_array_idxs);CHKERRQ(ierr); 501028143c3dSStefano Zampini psum += plen+1; /* indices + lenght */ 501128143c3dSStefano Zampini } 501228143c3dSStefano Zampini for (i=0;i<n_sends;i++) { 501328143c3dSStefano Zampini ilengths_idxs_is[is_indices[i]] = psum; 501428143c3dSStefano Zampini } 501528143c3dSStefano Zampini ierr = PetscGatherMessageLengths(comm,n_sends,n_recvs,ilengths_idxs_is,&onodes_is,&olengths_idxs_is);CHKERRQ(ierr); 501628143c3dSStefano Zampini } 501728143c3dSStefano Zampini 5018e7931f94SStefano Zampini buf_size_idxs = 0; 5019e7931f94SStefano Zampini buf_size_vals = 0; 502028143c3dSStefano Zampini buf_size_idxs_is = 0; 5021e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 5022e7931f94SStefano Zampini buf_size_idxs += (PetscInt)olengths_idxs[i]; 5023e7931f94SStefano Zampini buf_size_vals += (PetscInt)olengths_vals[i]; 502428143c3dSStefano Zampini if (nis) buf_size_idxs_is += (PetscInt)olengths_idxs_is[i]; 5025e7931f94SStefano Zampini } 5026785e854fSJed Brown ierr = PetscMalloc1(buf_size_idxs,&recv_buffer_idxs);CHKERRQ(ierr); 5027785e854fSJed Brown ierr = PetscMalloc1(buf_size_vals,&recv_buffer_vals);CHKERRQ(ierr); 502895ecbf38SStefano Zampini ierr = PetscMalloc1(buf_size_idxs_is,&recv_buffer_idxs_is);CHKERRQ(ierr); 5029e7931f94SStefano Zampini 5030e7931f94SStefano Zampini /* get new tags for clean communications */ 5031e7931f94SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_idxs);CHKERRQ(ierr); 5032e7931f94SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_vals);CHKERRQ(ierr); 503328143c3dSStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_idxs_is);CHKERRQ(ierr); 5034e7931f94SStefano Zampini 5035e7931f94SStefano Zampini /* allocate for requests */ 5036785e854fSJed Brown ierr = PetscMalloc1(n_sends,&send_req_idxs);CHKERRQ(ierr); 5037785e854fSJed Brown ierr = PetscMalloc1(n_sends,&send_req_vals);CHKERRQ(ierr); 503895ecbf38SStefano Zampini ierr = PetscMalloc1(n_sends,&send_req_idxs_is);CHKERRQ(ierr); 5039785e854fSJed Brown ierr = PetscMalloc1(n_recvs,&recv_req_idxs);CHKERRQ(ierr); 5040785e854fSJed Brown ierr = PetscMalloc1(n_recvs,&recv_req_vals);CHKERRQ(ierr); 504195ecbf38SStefano Zampini ierr = PetscMalloc1(n_recvs,&recv_req_idxs_is);CHKERRQ(ierr); 5042e7931f94SStefano Zampini 5043e7931f94SStefano Zampini /* communications */ 5044e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 5045e7931f94SStefano Zampini ptr_vals = recv_buffer_vals; 504628143c3dSStefano Zampini ptr_idxs_is = recv_buffer_idxs_is; 5047e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 5048e7931f94SStefano Zampini source_dest = onodes[i]; 5049e7931f94SStefano Zampini ierr = MPI_Irecv(ptr_idxs,olengths_idxs[i],MPIU_INT,source_dest,tag_idxs,comm,&recv_req_idxs[i]);CHKERRQ(ierr); 5050e7931f94SStefano Zampini ierr = MPI_Irecv(ptr_vals,olengths_vals[i],MPIU_SCALAR,source_dest,tag_vals,comm,&recv_req_vals[i]);CHKERRQ(ierr); 5051e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 5052e7931f94SStefano Zampini ptr_vals += olengths_vals[i]; 505328143c3dSStefano Zampini if (nis) { 505428143c3dSStefano 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); 505528143c3dSStefano Zampini ptr_idxs_is += olengths_idxs_is[i]; 505628143c3dSStefano Zampini } 5057e7931f94SStefano Zampini } 5058e7931f94SStefano Zampini for (i=0;i<n_sends;i++) { 5059e7931f94SStefano Zampini ierr = PetscMPIIntCast(is_indices[i],&source_dest);CHKERRQ(ierr); 5060e7931f94SStefano Zampini ierr = MPI_Isend(send_buffer_idxs,ilengths_idxs[source_dest],MPIU_INT,source_dest,tag_idxs,comm,&send_req_idxs[i]);CHKERRQ(ierr); 5061e7931f94SStefano Zampini ierr = MPI_Isend(send_buffer_vals,ilengths_vals[source_dest],MPIU_SCALAR,source_dest,tag_vals,comm,&send_req_vals[i]);CHKERRQ(ierr); 506228143c3dSStefano Zampini if (nis) { 506328143c3dSStefano 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); 506428143c3dSStefano Zampini } 5065e7931f94SStefano Zampini } 5066e7931f94SStefano Zampini ierr = ISRestoreIndices(is_sends_internal,&is_indices);CHKERRQ(ierr); 5067e7931f94SStefano Zampini ierr = ISDestroy(&is_sends_internal);CHKERRQ(ierr); 5068e7931f94SStefano Zampini 5069e7931f94SStefano Zampini /* assemble new l2g map */ 5070e7931f94SStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_idxs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 5071e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 50729d30be91SStefano Zampini new_local_rows = 0; 5073e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 50749d30be91SStefano Zampini new_local_rows += *(ptr_idxs+1); /* second element is the local size of the l2gmap */ 5075e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 5076e7931f94SStefano Zampini } 50779d30be91SStefano Zampini ierr = PetscMalloc1(new_local_rows,&l2gmap_indices);CHKERRQ(ierr); 5078e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 50799d30be91SStefano Zampini new_local_rows = 0; 5080e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 50819d30be91SStefano Zampini ierr = PetscMemcpy(&l2gmap_indices[new_local_rows],ptr_idxs+2,(*(ptr_idxs+1))*sizeof(PetscInt));CHKERRQ(ierr); 50829d30be91SStefano Zampini new_local_rows += *(ptr_idxs+1); /* second element is the local size of the l2gmap */ 5083e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 5084e7931f94SStefano Zampini } 50859d30be91SStefano Zampini ierr = PetscSortRemoveDupsInt(&new_local_rows,l2gmap_indices);CHKERRQ(ierr); 50869d30be91SStefano Zampini ierr = ISLocalToGlobalMappingCreate(comm_n,1,new_local_rows,l2gmap_indices,PETSC_COPY_VALUES,&l2gmap);CHKERRQ(ierr); 5087e7931f94SStefano Zampini ierr = PetscFree(l2gmap_indices);CHKERRQ(ierr); 5088e7931f94SStefano Zampini 5089e7931f94SStefano Zampini /* infer new local matrix type from received local matrices type */ 5090e7931f94SStefano Zampini /* currently if all local matrices are of type X, then the resulting matrix will be of type X, except for the dense case */ 5091e7931f94SStefano 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) */ 5092e7931f94SStefano Zampini if (n_recvs) { 509328143c3dSStefano Zampini MatTypePrivate new_local_type_private = (MatTypePrivate)send_buffer_idxs[0]; 5094e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 5095e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 5096e7931f94SStefano Zampini if ((PetscInt)new_local_type_private != *ptr_idxs) { 5097e7931f94SStefano Zampini new_local_type_private = MATAIJ_PRIVATE; 5098e7931f94SStefano Zampini break; 5099e7931f94SStefano Zampini } 5100e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 5101e7931f94SStefano Zampini } 5102e7931f94SStefano Zampini switch (new_local_type_private) { 510328143c3dSStefano Zampini case MATDENSE_PRIVATE: 510428143c3dSStefano Zampini if (n_recvs>1) { /* subassembling of dense matrices does not give a dense matrix! */ 5105e7931f94SStefano Zampini new_local_type = MATSEQAIJ; 5106e7931f94SStefano Zampini bs = 1; 510728143c3dSStefano Zampini } else { /* if I receive only 1 dense matrix */ 510828143c3dSStefano Zampini new_local_type = MATSEQDENSE; 510928143c3dSStefano Zampini bs = 1; 511028143c3dSStefano Zampini } 5111e7931f94SStefano Zampini break; 5112e7931f94SStefano Zampini case MATAIJ_PRIVATE: 5113e7931f94SStefano Zampini new_local_type = MATSEQAIJ; 5114e7931f94SStefano Zampini bs = 1; 5115e7931f94SStefano Zampini break; 5116e7931f94SStefano Zampini case MATBAIJ_PRIVATE: 5117e7931f94SStefano Zampini new_local_type = MATSEQBAIJ; 5118e7931f94SStefano Zampini break; 5119e7931f94SStefano Zampini case MATSBAIJ_PRIVATE: 5120e7931f94SStefano Zampini new_local_type = MATSEQSBAIJ; 5121e7931f94SStefano Zampini break; 5122e7931f94SStefano Zampini default: 51239d30be91SStefano Zampini SETERRQ2(comm,PETSC_ERR_SUP,"Unsupported private type %d in %s",new_local_type_private,__FUNCT__); 5124e7931f94SStefano Zampini break; 5125e7931f94SStefano Zampini } 512628143c3dSStefano Zampini } else { /* by default, new_local_type is seqdense */ 512728143c3dSStefano Zampini new_local_type = MATSEQDENSE; 512828143c3dSStefano Zampini bs = 1; 5129e7931f94SStefano Zampini } 5130e7931f94SStefano Zampini 513170cf5478SStefano Zampini /* create MATIS object if needed */ 513270cf5478SStefano Zampini if (reuse == MAT_INITIAL_MATRIX) { 5133e7931f94SStefano Zampini ierr = MatGetSize(mat,&rows,&cols);CHKERRQ(ierr); 5134e176bc59SStefano Zampini ierr = MatCreateIS(comm_n,bs,PETSC_DECIDE,PETSC_DECIDE,rows,cols,l2gmap,NULL,mat_n);CHKERRQ(ierr); 513570cf5478SStefano Zampini } else { 513670cf5478SStefano Zampini /* it also destroys the local matrices */ 513770cf5478SStefano Zampini ierr = MatSetLocalToGlobalMapping(*mat_n,l2gmap,l2gmap);CHKERRQ(ierr); 513870cf5478SStefano Zampini } 513970cf5478SStefano Zampini ierr = MatISGetLocalMat(*mat_n,&local_mat);CHKERRQ(ierr); 5140e7931f94SStefano Zampini ierr = MatSetType(local_mat,new_local_type);CHKERRQ(ierr); 51419d30be91SStefano Zampini 51429d30be91SStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_vals,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 51439d30be91SStefano Zampini 51449d30be91SStefano Zampini /* Global to local map of received indices */ 51459d30be91SStefano Zampini ierr = PetscMalloc1(buf_size_idxs,&recv_buffer_idxs_local);CHKERRQ(ierr); /* needed for values insertion */ 51469d30be91SStefano Zampini ierr = ISGlobalToLocalMappingApply(l2gmap,IS_GTOLM_MASK,buf_size_idxs,recv_buffer_idxs,&i,recv_buffer_idxs_local);CHKERRQ(ierr); 51479d30be91SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&l2gmap);CHKERRQ(ierr); 51489d30be91SStefano Zampini 51499d30be91SStefano Zampini /* restore attributes -> type of incoming data and its size */ 51509d30be91SStefano Zampini buf_size_idxs = 0; 51519d30be91SStefano Zampini for (i=0;i<n_recvs;i++) { 51529d30be91SStefano Zampini recv_buffer_idxs_local[buf_size_idxs] = recv_buffer_idxs[buf_size_idxs]; 51539d30be91SStefano Zampini recv_buffer_idxs_local[buf_size_idxs+1] = recv_buffer_idxs[buf_size_idxs+1]; 51549d30be91SStefano Zampini buf_size_idxs += (PetscInt)olengths_idxs[i]; 51559d30be91SStefano Zampini } 51569d30be91SStefano Zampini ierr = PetscFree(recv_buffer_idxs);CHKERRQ(ierr); 51579d30be91SStefano Zampini 51589d30be91SStefano Zampini /* set preallocation */ 51599d30be91SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)local_mat,MATSEQDENSE,&newisdense);CHKERRQ(ierr); 51609d30be91SStefano Zampini if (!newisdense) { 51619d30be91SStefano Zampini PetscInt *new_local_nnz=0; 51629d30be91SStefano Zampini 51639d30be91SStefano Zampini ptr_vals = recv_buffer_vals; 51649d30be91SStefano Zampini ptr_idxs = recv_buffer_idxs_local; 51659d30be91SStefano Zampini if (n_recvs) { 51669d30be91SStefano Zampini ierr = PetscCalloc1(new_local_rows,&new_local_nnz);CHKERRQ(ierr); 51679d30be91SStefano Zampini } 51689d30be91SStefano Zampini for (i=0;i<n_recvs;i++) { 51699d30be91SStefano Zampini PetscInt j; 51709d30be91SStefano Zampini if (*ptr_idxs == (PetscInt)MATDENSE_PRIVATE) { /* preallocation provided for dense case only */ 51719d30be91SStefano Zampini for (j=0;j<*(ptr_idxs+1);j++) { 51729d30be91SStefano Zampini new_local_nnz[*(ptr_idxs+2+j)] += *(ptr_idxs+1); 51739d30be91SStefano Zampini } 51749d30be91SStefano Zampini } else { 51759d30be91SStefano Zampini /* TODO */ 51769d30be91SStefano Zampini } 51779d30be91SStefano Zampini ptr_idxs += olengths_idxs[i]; 51789d30be91SStefano Zampini } 51799d30be91SStefano Zampini if (new_local_nnz) { 51809d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] = PetscMin(new_local_nnz[i],new_local_rows); 51819d30be91SStefano Zampini ierr = MatSeqAIJSetPreallocation(local_mat,0,new_local_nnz);CHKERRQ(ierr); 51829d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] /= bs; 51839d30be91SStefano Zampini ierr = MatSeqBAIJSetPreallocation(local_mat,bs,0,new_local_nnz);CHKERRQ(ierr); 51849d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] = PetscMax(new_local_nnz[i]-i,0); 51859d30be91SStefano Zampini ierr = MatSeqSBAIJSetPreallocation(local_mat,bs,0,new_local_nnz);CHKERRQ(ierr); 51869d30be91SStefano Zampini } else { 51879d30be91SStefano Zampini ierr = MatSetUp(local_mat);CHKERRQ(ierr); 51889d30be91SStefano Zampini } 51899d30be91SStefano Zampini ierr = PetscFree(new_local_nnz);CHKERRQ(ierr); 51909d30be91SStefano Zampini } else { 51919d30be91SStefano Zampini ierr = MatSetUp(local_mat);CHKERRQ(ierr); 51929d30be91SStefano Zampini } 5193e7931f94SStefano Zampini 5194e7931f94SStefano Zampini /* set values */ 5195e7931f94SStefano Zampini ptr_vals = recv_buffer_vals; 51969d30be91SStefano Zampini ptr_idxs = recv_buffer_idxs_local; 5197e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 5198e7931f94SStefano Zampini if (*ptr_idxs == (PetscInt)MATDENSE_PRIVATE) { /* values insertion provided for dense case only */ 5199e7931f94SStefano Zampini ierr = MatSetOption(local_mat,MAT_ROW_ORIENTED,PETSC_FALSE);CHKERRQ(ierr); 52009d30be91SStefano Zampini ierr = MatSetValues(local_mat,*(ptr_idxs+1),ptr_idxs+2,*(ptr_idxs+1),ptr_idxs+2,ptr_vals,ADD_VALUES);CHKERRQ(ierr); 5201e7931f94SStefano Zampini ierr = MatAssemblyBegin(local_mat,MAT_FLUSH_ASSEMBLY);CHKERRQ(ierr); 5202e7931f94SStefano Zampini ierr = MatAssemblyEnd(local_mat,MAT_FLUSH_ASSEMBLY);CHKERRQ(ierr); 5203e7931f94SStefano Zampini ierr = MatSetOption(local_mat,MAT_ROW_ORIENTED,PETSC_TRUE);CHKERRQ(ierr); 520428143c3dSStefano Zampini } else { 520528143c3dSStefano Zampini /* TODO */ 5206e7931f94SStefano Zampini } 5207e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 5208e7931f94SStefano Zampini ptr_vals += olengths_vals[i]; 5209e7931f94SStefano Zampini } 5210e7931f94SStefano Zampini ierr = MatAssemblyBegin(local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5211e7931f94SStefano Zampini ierr = MatAssemblyEnd(local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 521270cf5478SStefano Zampini ierr = MatAssemblyBegin(*mat_n,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 521370cf5478SStefano Zampini ierr = MatAssemblyEnd(*mat_n,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 52149d30be91SStefano Zampini ierr = PetscFree(recv_buffer_vals);CHKERRQ(ierr); 52159d30be91SStefano Zampini ierr = PetscFree(recv_buffer_idxs_local);CHKERRQ(ierr); 5216e7931f94SStefano Zampini 5217dfd14d43SStefano Zampini #if 0 521828143c3dSStefano Zampini if (!restrict_comm) { /* check */ 5219e7931f94SStefano Zampini Vec lvec,rvec; 5220e7931f94SStefano Zampini PetscReal infty_error; 5221e7931f94SStefano Zampini 52222a7a6963SBarry Smith ierr = MatCreateVecs(mat,&rvec,&lvec);CHKERRQ(ierr); 5223e7931f94SStefano Zampini ierr = VecSetRandom(rvec,NULL);CHKERRQ(ierr); 5224e7931f94SStefano Zampini ierr = MatMult(mat,rvec,lvec);CHKERRQ(ierr); 5225e7931f94SStefano Zampini ierr = VecScale(lvec,-1.0);CHKERRQ(ierr); 522670cf5478SStefano Zampini ierr = MatMultAdd(*mat_n,rvec,lvec,lvec);CHKERRQ(ierr); 5227e7931f94SStefano Zampini ierr = VecNorm(lvec,NORM_INFINITY,&infty_error);CHKERRQ(ierr); 5228e7931f94SStefano Zampini ierr = PetscPrintf(PetscObjectComm((PetscObject)mat),"Infinity error subassembling %1.6e\n",infty_error); 5229e7931f94SStefano Zampini ierr = VecDestroy(&rvec);CHKERRQ(ierr); 5230e7931f94SStefano Zampini ierr = VecDestroy(&lvec);CHKERRQ(ierr); 5231e7931f94SStefano Zampini } 523228143c3dSStefano Zampini #endif 5233e7931f94SStefano Zampini 523428143c3dSStefano Zampini /* assemble new additional is (if any) */ 523528143c3dSStefano Zampini if (nis) { 523628143c3dSStefano Zampini PetscInt **temp_idxs,*count_is,j,psum; 523728143c3dSStefano Zampini 523828143c3dSStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_idxs_is,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 5239854ce69bSBarry Smith ierr = PetscCalloc1(nis,&count_is);CHKERRQ(ierr); 524028143c3dSStefano Zampini ptr_idxs = recv_buffer_idxs_is; 524128143c3dSStefano Zampini psum = 0; 524228143c3dSStefano Zampini for (i=0;i<n_recvs;i++) { 524328143c3dSStefano Zampini for (j=0;j<nis;j++) { 524428143c3dSStefano Zampini PetscInt plen = *(ptr_idxs); /* first element is the local size of IS's indices */ 524528143c3dSStefano Zampini count_is[j] += plen; /* increment counting of buffer for j-th IS */ 524628143c3dSStefano Zampini psum += plen; 524728143c3dSStefano Zampini ptr_idxs += plen+1; /* shift pointer to received data */ 524828143c3dSStefano Zampini } 524928143c3dSStefano Zampini } 5250854ce69bSBarry Smith ierr = PetscMalloc1(nis,&temp_idxs);CHKERRQ(ierr); 5251854ce69bSBarry Smith ierr = PetscMalloc1(psum,&temp_idxs[0]);CHKERRQ(ierr); 525228143c3dSStefano Zampini for (i=1;i<nis;i++) { 525328143c3dSStefano Zampini temp_idxs[i] = temp_idxs[i-1]+count_is[i-1]; 525428143c3dSStefano Zampini } 525528143c3dSStefano Zampini ierr = PetscMemzero(count_is,nis*sizeof(PetscInt));CHKERRQ(ierr); 525628143c3dSStefano Zampini ptr_idxs = recv_buffer_idxs_is; 525728143c3dSStefano Zampini for (i=0;i<n_recvs;i++) { 525828143c3dSStefano Zampini for (j=0;j<nis;j++) { 525928143c3dSStefano Zampini PetscInt plen = *(ptr_idxs); /* first element is the local size of IS's indices */ 526028143c3dSStefano Zampini ierr = PetscMemcpy(&temp_idxs[j][count_is[j]],ptr_idxs+1,plen*sizeof(PetscInt));CHKERRQ(ierr); 526128143c3dSStefano Zampini count_is[j] += plen; /* increment starting point of buffer for j-th IS */ 526228143c3dSStefano Zampini ptr_idxs += plen+1; /* shift pointer to received data */ 526328143c3dSStefano Zampini } 526428143c3dSStefano Zampini } 526528143c3dSStefano Zampini for (i=0;i<nis;i++) { 526628143c3dSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 526728143c3dSStefano Zampini ierr = PetscSortRemoveDupsInt(&count_is[i],temp_idxs[i]);CHKERRQ(ierr);CHKERRQ(ierr); 526828143c3dSStefano Zampini ierr = ISCreateGeneral(comm_n,count_is[i],temp_idxs[i],PETSC_COPY_VALUES,&isarray[i]);CHKERRQ(ierr); 526928143c3dSStefano Zampini } 527028143c3dSStefano Zampini ierr = PetscFree(count_is);CHKERRQ(ierr); 527128143c3dSStefano Zampini ierr = PetscFree(temp_idxs[0]);CHKERRQ(ierr); 527228143c3dSStefano Zampini ierr = PetscFree(temp_idxs);CHKERRQ(ierr); 527328143c3dSStefano Zampini } 5274e7931f94SStefano Zampini /* free workspace */ 527528143c3dSStefano Zampini ierr = PetscFree(recv_buffer_idxs_is);CHKERRQ(ierr); 5276e7931f94SStefano Zampini ierr = MPI_Waitall(n_sends,send_req_idxs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 5277e7931f94SStefano Zampini ierr = PetscFree(send_buffer_idxs);CHKERRQ(ierr); 5278e7931f94SStefano Zampini ierr = MPI_Waitall(n_sends,send_req_vals,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 5279e7931f94SStefano Zampini if (isdense) { 5280e7931f94SStefano Zampini ierr = MatISGetLocalMat(mat,&local_mat);CHKERRQ(ierr); 5281e7931f94SStefano Zampini ierr = MatDenseRestoreArray(local_mat,&send_buffer_vals);CHKERRQ(ierr); 5282e7931f94SStefano Zampini } else { 5283e7931f94SStefano Zampini /* ierr = PetscFree(send_buffer_vals);CHKERRQ(ierr); */ 5284e7931f94SStefano Zampini } 528528143c3dSStefano Zampini if (nis) { 528628143c3dSStefano Zampini ierr = MPI_Waitall(n_sends,send_req_idxs_is,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 528728143c3dSStefano Zampini ierr = PetscFree(send_buffer_idxs_is);CHKERRQ(ierr); 528828143c3dSStefano Zampini } 5289e7931f94SStefano Zampini ierr = PetscFree(recv_req_idxs);CHKERRQ(ierr); 5290e7931f94SStefano Zampini ierr = PetscFree(recv_req_vals);CHKERRQ(ierr); 529128143c3dSStefano Zampini ierr = PetscFree(recv_req_idxs_is);CHKERRQ(ierr); 5292e7931f94SStefano Zampini ierr = PetscFree(send_req_idxs);CHKERRQ(ierr); 5293e7931f94SStefano Zampini ierr = PetscFree(send_req_vals);CHKERRQ(ierr); 529428143c3dSStefano Zampini ierr = PetscFree(send_req_idxs_is);CHKERRQ(ierr); 5295e7931f94SStefano Zampini ierr = PetscFree(ilengths_vals);CHKERRQ(ierr); 5296e7931f94SStefano Zampini ierr = PetscFree(ilengths_idxs);CHKERRQ(ierr); 5297e7931f94SStefano Zampini ierr = PetscFree(olengths_vals);CHKERRQ(ierr); 5298e7931f94SStefano Zampini ierr = PetscFree(olengths_idxs);CHKERRQ(ierr); 5299e7931f94SStefano Zampini ierr = PetscFree(onodes);CHKERRQ(ierr); 530028143c3dSStefano Zampini if (nis) { 530128143c3dSStefano Zampini ierr = PetscFree(ilengths_idxs_is);CHKERRQ(ierr); 530228143c3dSStefano Zampini ierr = PetscFree(olengths_idxs_is);CHKERRQ(ierr); 530328143c3dSStefano Zampini ierr = PetscFree(onodes_is);CHKERRQ(ierr); 530428143c3dSStefano Zampini } 530528143c3dSStefano Zampini ierr = PetscSubcommDestroy(&subcomm);CHKERRQ(ierr); 530628143c3dSStefano Zampini if (destroy_mat) { /* destroy mat is true only if restrict comm is true and process will not partecipate */ 530728143c3dSStefano Zampini ierr = MatDestroy(mat_n);CHKERRQ(ierr); 530828143c3dSStefano Zampini for (i=0;i<nis;i++) { 530928143c3dSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 531028143c3dSStefano Zampini } 531153a05cb3SStefano Zampini *mat_n = NULL; 531228143c3dSStefano Zampini } 5313e7931f94SStefano Zampini PetscFunctionReturn(0); 5314e7931f94SStefano Zampini } 5315a57a6d2fSStefano Zampini 531612edc857SStefano Zampini /* temporary hack into ksp private data structure */ 5317af0996ceSBarry Smith #include <petsc/private/kspimpl.h> 531812edc857SStefano Zampini 5319c8587f34SStefano Zampini #undef __FUNCT__ 5320c8587f34SStefano Zampini #define __FUNCT__ "PCBDDCSetUpCoarseSolver" 5321c8587f34SStefano Zampini PetscErrorCode PCBDDCSetUpCoarseSolver(PC pc,PetscScalar* coarse_submat_vals) 5322c8587f34SStefano Zampini { 5323c8587f34SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 5324c8587f34SStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 532520a2ab83SStefano Zampini Mat coarse_mat,coarse_mat_is,coarse_submat_dense; 53269881197aSStefano Zampini MatNullSpace CoarseNullSpace=NULL; 532720a2ab83SStefano Zampini ISLocalToGlobalMapping coarse_islg; 53286e683305SStefano Zampini IS coarse_is,*isarray; 53296e683305SStefano Zampini PetscInt i,im_active=-1,active_procs=-1; 533030368db7SStefano Zampini PetscInt nis,nisdofs,nisneu,nisvert; 5331f9eb5b7dSStefano Zampini PC pc_temp; 5332c8587f34SStefano Zampini PCType coarse_pc_type; 5333c8587f34SStefano Zampini KSPType coarse_ksp_type; 5334f9eb5b7dSStefano Zampini PetscBool multilevel_requested,multilevel_allowed; 53354f3a063dSStefano Zampini PetscBool isredundant,isbddc,isnn,coarse_reuse; 53366e683305SStefano Zampini Mat t_coarse_mat_is; 53376e683305SStefano Zampini PetscInt void_procs,ncoarse_ml,ncoarse_ds,ncoarse; 53386e683305SStefano Zampini PetscMPIInt all_procs; 533974e2c79eSStefano Zampini PetscBool csin_ml,csin_ds,csin,csin_type_simple,redist; 534068457ee5SStefano Zampini PetscBool compute_vecs = PETSC_FALSE; 534122bc73bbSStefano Zampini PetscScalar *array; 53429881197aSStefano Zampini PetscErrorCode ierr; 5343fdc09c96SStefano Zampini 5344c8587f34SStefano Zampini PetscFunctionBegin; 5345c8587f34SStefano Zampini /* Assign global numbering to coarse dofs */ 534668457ee5SStefano 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 */ 5347fa7f1dd8SStefano Zampini PetscInt ocoarse_size; 53485a75c04eSSatish Balay compute_vecs = PETSC_TRUE; 5349fa7f1dd8SStefano Zampini ocoarse_size = pcbddc->coarse_size; 5350f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->global_primal_indices);CHKERRQ(ierr); 5351f4ddd8eeSStefano Zampini ierr = PCBDDCComputePrimalNumbering(pc,&pcbddc->coarse_size,&pcbddc->global_primal_indices);CHKERRQ(ierr); 5352f4ddd8eeSStefano Zampini /* see if we can avoid some work */ 5353fa7f1dd8SStefano Zampini if (pcbddc->coarse_ksp) { /* coarse ksp has already been created */ 535451bea450SStefano Zampini /* if the coarse size is different or we are using adaptive selection, better to not reuse the coarse matrix */ 535551bea450SStefano Zampini if (ocoarse_size != pcbddc->coarse_size || pcbddc->adaptive_selection) { 5356dc4bcba2SStefano Zampini PC pc; 5357dc4bcba2SStefano Zampini PetscBool isbddc; 5358dc4bcba2SStefano Zampini 5359dc4bcba2SStefano Zampini /* temporary workaround since PCBDDC does not have a reset method so far */ 5360dc4bcba2SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&pc);CHKERRQ(ierr); 5361dc4bcba2SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc,PCBDDC,&isbddc);CHKERRQ(ierr); 5362dc4bcba2SStefano Zampini if (isbddc) { 536363c961adSStefano Zampini ierr = KSPDestroy(&pcbddc->coarse_ksp);CHKERRQ(ierr); 536463c961adSStefano Zampini } else { 5365727cdba6SStefano Zampini ierr = KSPReset(pcbddc->coarse_ksp);CHKERRQ(ierr); 536663c961adSStefano Zampini } 5367fa7f1dd8SStefano Zampini coarse_reuse = PETSC_FALSE; 5368fa7f1dd8SStefano Zampini } else { /* we can safely reuse already computed coarse matrix */ 5369fa7f1dd8SStefano Zampini coarse_reuse = PETSC_TRUE; 5370f4ddd8eeSStefano Zampini } 5371fa7f1dd8SStefano Zampini } else { /* there's no coarse ksp, so we need to create the coarse matrix too */ 5372fa7f1dd8SStefano Zampini coarse_reuse = PETSC_FALSE; 5373f4ddd8eeSStefano Zampini } 537470cf5478SStefano Zampini /* reset any subassembling information */ 537570cf5478SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling);CHKERRQ(ierr); 53766e683305SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 53776e683305SStefano Zampini } else { /* primal space is unchanged, so we can reuse coarse matrix */ 5378fa7f1dd8SStefano Zampini coarse_reuse = PETSC_TRUE; 5379f4ddd8eeSStefano Zampini } 5380c8587f34SStefano Zampini 53816e683305SStefano Zampini /* count "active" (i.e. with positive local size) and "void" processes */ 53822b510759SStefano Zampini im_active = !!(pcis->n); 53832b510759SStefano Zampini ierr = MPI_Allreduce(&im_active,&active_procs,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 53846e683305SStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)pc),&all_procs);CHKERRQ(ierr); 53856e683305SStefano Zampini void_procs = all_procs-active_procs; 53866e683305SStefano Zampini csin_type_simple = PETSC_TRUE; 538774e2c79eSStefano Zampini redist = PETSC_FALSE; 538822bc73bbSStefano Zampini if (pcbddc->current_level && void_procs) { 53896e683305SStefano Zampini csin_ml = PETSC_TRUE; 53906e683305SStefano Zampini ncoarse_ml = void_procs; 5391779c1cceSStefano Zampini /* it has no sense to redistribute on a set of processors larger than the number of active processes */ 5392779c1cceSStefano Zampini if (pcbddc->redistribute_coarse > 0 && pcbddc->redistribute_coarse < active_procs) { 53936e683305SStefano Zampini csin_ds = PETSC_TRUE; 539418a45a71SStefano Zampini ncoarse_ds = pcbddc->redistribute_coarse; 539518a45a71SStefano Zampini redist = PETSC_TRUE; 539618a45a71SStefano Zampini } else { 53976e683305SStefano Zampini csin_ds = PETSC_TRUE; 5398779c1cceSStefano Zampini ncoarse_ds = active_procs; 5399779c1cceSStefano Zampini redist = PETSC_TRUE; 540018a45a71SStefano Zampini } 54016e683305SStefano Zampini } else { 54026e683305SStefano Zampini csin_ml = PETSC_FALSE; 54036e683305SStefano Zampini ncoarse_ml = all_procs; 54046e683305SStefano Zampini if (void_procs) { 54056e683305SStefano Zampini csin_ds = PETSC_TRUE; 54066e683305SStefano Zampini ncoarse_ds = void_procs; 54076e683305SStefano Zampini csin_type_simple = PETSC_FALSE; 54086e683305SStefano Zampini } else { 5409779c1cceSStefano Zampini if (pcbddc->redistribute_coarse > 0 && pcbddc->redistribute_coarse < all_procs) { 541074e2c79eSStefano Zampini csin_ds = PETSC_TRUE; 541174e2c79eSStefano Zampini ncoarse_ds = pcbddc->redistribute_coarse; 541274e2c79eSStefano Zampini redist = PETSC_TRUE; 541374e2c79eSStefano Zampini } else { 54146e683305SStefano Zampini csin_ds = PETSC_FALSE; 54156e683305SStefano Zampini ncoarse_ds = all_procs; 54166e683305SStefano Zampini } 54176e683305SStefano Zampini } 541874e2c79eSStefano Zampini } 54196e683305SStefano Zampini 54206e683305SStefano Zampini /* 54216e683305SStefano Zampini test if we can go multilevel: three conditions must be satisfied: 54226e683305SStefano Zampini - we have not exceeded the number of levels requested 54236e683305SStefano Zampini - we can actually subassemble the active processes 54246e683305SStefano Zampini - we can find a suitable number of MPI processes where we can place the subassembled problem 54256e683305SStefano Zampini */ 54266e683305SStefano Zampini multilevel_allowed = PETSC_FALSE; 54276e683305SStefano Zampini multilevel_requested = PETSC_FALSE; 54286e683305SStefano Zampini if (pcbddc->current_level < pcbddc->max_levels) { 54296e683305SStefano Zampini multilevel_requested = PETSC_TRUE; 54306e683305SStefano Zampini if (active_procs/pcbddc->coarsening_ratio < 2 || ncoarse_ml/pcbddc->coarsening_ratio < 2) { 5431f9eb5b7dSStefano Zampini multilevel_allowed = PETSC_FALSE; 54322b510759SStefano Zampini } else { 5433f9eb5b7dSStefano Zampini multilevel_allowed = PETSC_TRUE; 5434c8587f34SStefano Zampini } 5435c8587f34SStefano Zampini } 54366e683305SStefano Zampini /* determine number of process partecipating to coarse solver */ 54376e683305SStefano Zampini if (multilevel_allowed) { 54386e683305SStefano Zampini ncoarse = ncoarse_ml; 54396e683305SStefano Zampini csin = csin_ml; 544058da7f69SStefano Zampini redist = PETSC_FALSE; 54416e683305SStefano Zampini } else { 54426e683305SStefano Zampini ncoarse = ncoarse_ds; 54436e683305SStefano Zampini csin = csin_ds; 54446e683305SStefano Zampini } 5445e7931f94SStefano Zampini 5446abbbba34SStefano Zampini /* creates temporary l2gmap and IS for coarse indexes */ 5447abbbba34SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),pcbddc->local_primal_size,pcbddc->global_primal_indices,PETSC_COPY_VALUES,&coarse_is);CHKERRQ(ierr); 5448abbbba34SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(coarse_is,&coarse_islg);CHKERRQ(ierr); 5449abbbba34SStefano Zampini 5450abbbba34SStefano Zampini /* creates temporary MATIS object for coarse matrix */ 545122bc73bbSStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,pcbddc->local_primal_size,pcbddc->local_primal_size,NULL,&coarse_submat_dense);CHKERRQ(ierr); 545222bc73bbSStefano Zampini ierr = MatDenseGetArray(coarse_submat_dense,&array);CHKERRQ(ierr); 545322bc73bbSStefano Zampini ierr = PetscMemcpy(array,coarse_submat_vals,sizeof(*coarse_submat_vals)*pcbddc->local_primal_size*pcbddc->local_primal_size);CHKERRQ(ierr); 545422bc73bbSStefano Zampini ierr = MatDenseRestoreArray(coarse_submat_dense,&array);CHKERRQ(ierr); 5455e176bc59SStefano 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); 54566e683305SStefano Zampini ierr = MatISSetLocalMat(t_coarse_mat_is,coarse_submat_dense);CHKERRQ(ierr); 54576e683305SStefano Zampini ierr = MatAssemblyBegin(t_coarse_mat_is,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 54586e683305SStefano Zampini ierr = MatAssemblyEnd(t_coarse_mat_is,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5459abbbba34SStefano Zampini ierr = MatDestroy(&coarse_submat_dense);CHKERRQ(ierr); 5460abbbba34SStefano Zampini 54616e683305SStefano Zampini /* compute dofs splitting and neumann boundaries for coarse dofs */ 546230368db7SStefano Zampini if (multilevel_allowed && (pcbddc->n_ISForDofsLocal || pcbddc->NeumannBoundariesLocal || (pcbddc->benign_saddle_point && pcbddc->user_primal_vertices_local))) { /* protects from unneded computations */ 54636e683305SStefano Zampini PetscInt *tidxs,*tidxs2,nout,tsize,i; 54646e683305SStefano Zampini const PetscInt *idxs; 54656e683305SStefano Zampini ISLocalToGlobalMapping tmap; 54666e683305SStefano Zampini 54676e683305SStefano Zampini /* create map between primal indices (in local representative ordering) and local primal numbering */ 54680be93d54SStefano Zampini ierr = ISLocalToGlobalMappingCreate(PETSC_COMM_SELF,1,pcbddc->local_primal_size,pcbddc->primal_indices_local_idxs,PETSC_COPY_VALUES,&tmap);CHKERRQ(ierr); 54696e683305SStefano Zampini /* allocate space for temporary storage */ 5470854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->local_primal_size,&tidxs);CHKERRQ(ierr); 5471854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->local_primal_size,&tidxs2);CHKERRQ(ierr); 54726e683305SStefano Zampini /* allocate for IS array */ 54736e683305SStefano Zampini nisdofs = pcbddc->n_ISForDofsLocal; 54746e683305SStefano Zampini nisneu = !!pcbddc->NeumannBoundariesLocal; 547530368db7SStefano Zampini nisvert = 0; 547630368db7SStefano Zampini if (pcbddc->benign_saddle_point && pcbddc->user_primal_vertices_local) { 547730368db7SStefano Zampini nisvert = 1; 547830368db7SStefano Zampini } 547930368db7SStefano Zampini nis = nisdofs + nisneu + nisvert; 5480854ce69bSBarry Smith ierr = PetscMalloc1(nis,&isarray);CHKERRQ(ierr); 54816e683305SStefano Zampini /* dofs splitting */ 54826e683305SStefano Zampini for (i=0;i<nisdofs;i++) { 54836e683305SStefano Zampini /* ierr = ISView(pcbddc->ISForDofsLocal[i],0);CHKERRQ(ierr); */ 54846e683305SStefano Zampini ierr = ISGetLocalSize(pcbddc->ISForDofsLocal[i],&tsize);CHKERRQ(ierr); 54856e683305SStefano Zampini ierr = ISGetIndices(pcbddc->ISForDofsLocal[i],&idxs);CHKERRQ(ierr); 54866e683305SStefano Zampini ierr = ISGlobalToLocalMappingApply(tmap,IS_GTOLM_DROP,tsize,idxs,&nout,tidxs);CHKERRQ(ierr); 54876e683305SStefano Zampini ierr = ISRestoreIndices(pcbddc->ISForDofsLocal[i],&idxs);CHKERRQ(ierr); 54886e683305SStefano Zampini ierr = ISLocalToGlobalMappingApply(coarse_islg,nout,tidxs,tidxs2);CHKERRQ(ierr); 548930368db7SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),nout,tidxs2,PETSC_COPY_VALUES,&isarray[i]);CHKERRQ(ierr); 54906e683305SStefano Zampini /* ierr = ISView(isarray[i],0);CHKERRQ(ierr); */ 54916e683305SStefano Zampini } 54926e683305SStefano Zampini /* neumann boundaries */ 54936e683305SStefano Zampini if (pcbddc->NeumannBoundariesLocal) { 54946e683305SStefano Zampini /* ierr = ISView(pcbddc->NeumannBoundariesLocal,0);CHKERRQ(ierr); */ 54956e683305SStefano Zampini ierr = ISGetLocalSize(pcbddc->NeumannBoundariesLocal,&tsize);CHKERRQ(ierr); 54966e683305SStefano Zampini ierr = ISGetIndices(pcbddc->NeumannBoundariesLocal,&idxs);CHKERRQ(ierr); 54976e683305SStefano Zampini ierr = ISGlobalToLocalMappingApply(tmap,IS_GTOLM_DROP,tsize,idxs,&nout,tidxs);CHKERRQ(ierr); 54986e683305SStefano Zampini ierr = ISRestoreIndices(pcbddc->NeumannBoundariesLocal,&idxs);CHKERRQ(ierr); 54996e683305SStefano Zampini ierr = ISLocalToGlobalMappingApply(coarse_islg,nout,tidxs,tidxs2);CHKERRQ(ierr); 550030368db7SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),nout,tidxs2,PETSC_COPY_VALUES,&isarray[nisdofs]);CHKERRQ(ierr); 55016e683305SStefano Zampini /* ierr = ISView(isarray[nisdofs],0);CHKERRQ(ierr); */ 55026e683305SStefano Zampini } 550330368db7SStefano Zampini /* primal vertices (benign) */ 550430368db7SStefano Zampini if (pcbddc->benign_saddle_point && pcbddc->user_primal_vertices_local) { 550530368db7SStefano Zampini ierr = ISGetLocalSize(pcbddc->user_primal_vertices_local,&tsize);CHKERRQ(ierr); 550630368db7SStefano Zampini ierr = ISGetIndices(pcbddc->user_primal_vertices_local,&idxs);CHKERRQ(ierr); 550730368db7SStefano Zampini ierr = ISGlobalToLocalMappingApply(tmap,IS_GTOLM_DROP,tsize,idxs,&nout,tidxs);CHKERRQ(ierr); 550830368db7SStefano Zampini ierr = ISRestoreIndices(pcbddc->user_primal_vertices_local,&idxs);CHKERRQ(ierr); 550930368db7SStefano Zampini ierr = ISLocalToGlobalMappingApply(coarse_islg,nout,tidxs,tidxs2);CHKERRQ(ierr); 551030368db7SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),nout,tidxs2,PETSC_COPY_VALUES,&isarray[nis-1]);CHKERRQ(ierr); 551130368db7SStefano Zampini } 55126e683305SStefano Zampini /* free memory */ 55136e683305SStefano Zampini ierr = PetscFree(tidxs);CHKERRQ(ierr); 55146e683305SStefano Zampini ierr = PetscFree(tidxs2);CHKERRQ(ierr); 55156e683305SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&tmap);CHKERRQ(ierr); 55166e683305SStefano Zampini } else { 55176e683305SStefano Zampini nis = 0; 55186e683305SStefano Zampini nisdofs = 0; 55196e683305SStefano Zampini nisneu = 0; 552030368db7SStefano Zampini nisvert = 0; 55216e683305SStefano Zampini isarray = NULL; 55226e683305SStefano Zampini } 55236e683305SStefano Zampini /* destroy no longer needed map */ 55246e683305SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&coarse_islg);CHKERRQ(ierr); 55256e683305SStefano Zampini 55266e683305SStefano Zampini /* restrict on coarse candidates (if needed) */ 55276e683305SStefano Zampini coarse_mat_is = NULL; 55286e683305SStefano Zampini if (csin) { 55296e683305SStefano Zampini if (!pcbddc->coarse_subassembling_init ) { /* creates subassembling init pattern if not present */ 553074e2c79eSStefano Zampini if (redist) { 553174e2c79eSStefano Zampini PetscMPIInt rank; 5532779c1cceSStefano Zampini PetscInt spc,n_spc_p1,dest[1],destsize; 553374e2c79eSStefano Zampini 553474e2c79eSStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)pc),&rank);CHKERRQ(ierr); 553558da7f69SStefano Zampini spc = active_procs/ncoarse; 553658da7f69SStefano Zampini n_spc_p1 = active_procs%ncoarse; 5537779c1cceSStefano Zampini if (im_active) { 5538779c1cceSStefano Zampini destsize = 1; 553974e2c79eSStefano Zampini if (rank > n_spc_p1*(spc+1)-1) { 554074e2c79eSStefano Zampini dest[0] = n_spc_p1+(rank-(n_spc_p1*(spc+1)))/spc; 554174e2c79eSStefano Zampini } else { 554274e2c79eSStefano Zampini dest[0] = rank/(spc+1); 554374e2c79eSStefano Zampini } 554474e2c79eSStefano Zampini } else { 5545779c1cceSStefano Zampini destsize = 0; 55466e683305SStefano Zampini } 5547779c1cceSStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),destsize,dest,PETSC_COPY_VALUES,&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 5548779c1cceSStefano Zampini } else if (csin_type_simple) { 55496e683305SStefano Zampini PetscMPIInt rank; 55506e683305SStefano Zampini PetscInt issize,isidx; 5551779c1cceSStefano Zampini 55526e683305SStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)pc),&rank);CHKERRQ(ierr); 55536e683305SStefano Zampini if (im_active) { 55546e683305SStefano Zampini issize = 1; 55556e683305SStefano Zampini isidx = (PetscInt)rank; 55566e683305SStefano Zampini } else { 55576e683305SStefano Zampini issize = 0; 55586e683305SStefano Zampini isidx = -1; 55596e683305SStefano Zampini } 55606e683305SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),issize,&isidx,PETSC_COPY_VALUES,&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 5561779c1cceSStefano Zampini } else { /* get a suitable subassembling pattern from MATIS code */ 5562b0c7d250SStefano Zampini ierr = MatISGetSubassemblingPattern(t_coarse_mat_is,ncoarse,pcbddc->coarse_adj_red,&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 55636e683305SStefano Zampini } 5564779c1cceSStefano Zampini 5565779c1cceSStefano Zampini /* we need to shift on coarse candidates either if we are not redistributing or we are redistributing and we have enough void processes */ 5566779c1cceSStefano Zampini if (!redist || ncoarse <= void_procs) { 5567779c1cceSStefano Zampini PetscInt ncoarse_cand,tissize,*nisindices; 5568779c1cceSStefano Zampini PetscInt *coarse_candidates; 5569779c1cceSStefano Zampini const PetscInt* tisindices; 5570779c1cceSStefano Zampini 5571779c1cceSStefano Zampini /* get coarse candidates' ranks in pc communicator */ 5572779c1cceSStefano Zampini ierr = PetscMalloc1(all_procs,&coarse_candidates);CHKERRQ(ierr); 5573779c1cceSStefano Zampini ierr = MPI_Allgather(&im_active,1,MPIU_INT,coarse_candidates,1,MPIU_INT,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 5574779c1cceSStefano Zampini for (i=0,ncoarse_cand=0;i<all_procs;i++) { 5575779c1cceSStefano Zampini if (!coarse_candidates[i]) { 5576779c1cceSStefano Zampini coarse_candidates[ncoarse_cand++]=i; 5577779c1cceSStefano Zampini } 5578779c1cceSStefano Zampini } 5579779c1cceSStefano Zampini if (ncoarse_cand < ncoarse) SETERRQ2(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"This should not happen! %d < %d",ncoarse_cand,ncoarse); 5580779c1cceSStefano Zampini 5581779c1cceSStefano Zampini 55826e683305SStefano Zampini if (pcbddc->dbg_flag) { 55836e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 55846e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Subassembling pattern init (before shift)\n");CHKERRQ(ierr); 55856e683305SStefano Zampini ierr = ISView(pcbddc->coarse_subassembling_init,pcbddc->dbg_viewer);CHKERRQ(ierr); 55866e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Coarse candidates\n");CHKERRQ(ierr); 5587779c1cceSStefano Zampini for (i=0;i<ncoarse_cand;i++) { 55886e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"%d ",coarse_candidates[i]);CHKERRQ(ierr); 55896e683305SStefano Zampini } 55906e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"\n");CHKERRQ(ierr); 55916e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 55926e683305SStefano Zampini } 55936e683305SStefano Zampini /* shift the pattern on coarse candidates */ 55946e683305SStefano Zampini ierr = ISGetLocalSize(pcbddc->coarse_subassembling_init,&tissize);CHKERRQ(ierr); 55956e683305SStefano Zampini ierr = ISGetIndices(pcbddc->coarse_subassembling_init,&tisindices);CHKERRQ(ierr); 5596854ce69bSBarry Smith ierr = PetscMalloc1(tissize,&nisindices);CHKERRQ(ierr); 55976e683305SStefano Zampini for (i=0;i<tissize;i++) nisindices[i] = coarse_candidates[tisindices[i]]; 55986e683305SStefano Zampini ierr = ISRestoreIndices(pcbddc->coarse_subassembling_init,&tisindices);CHKERRQ(ierr); 55996e683305SStefano Zampini ierr = ISGeneralSetIndices(pcbddc->coarse_subassembling_init,tissize,nisindices,PETSC_OWN_POINTER);CHKERRQ(ierr); 56006e683305SStefano Zampini ierr = PetscFree(coarse_candidates);CHKERRQ(ierr); 56016e683305SStefano Zampini } 56026e683305SStefano Zampini if (pcbddc->dbg_flag) { 56036e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 56046e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Subassembling pattern init\n");CHKERRQ(ierr); 56056e683305SStefano Zampini ierr = ISView(pcbddc->coarse_subassembling_init,pcbddc->dbg_viewer);CHKERRQ(ierr); 56066e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 56076e683305SStefano Zampini } 5608779c1cceSStefano Zampini } 56096e683305SStefano Zampini /* get temporary coarse mat in IS format restricted on coarse procs (plus additional index sets of isarray) */ 561053a05cb3SStefano Zampini if (multilevel_allowed) { /* we need to keep tracking of void processes for future placements */ 561153a05cb3SStefano 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); 561253a05cb3SStefano Zampini } else { /* this is the last level, so use just receiving processes in subcomm */ 561353a05cb3SStefano 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); 561453a05cb3SStefano Zampini } 56156e683305SStefano Zampini } else { 56166e683305SStefano Zampini if (pcbddc->dbg_flag) { 56176e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 56186e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Subassembling pattern init not needed\n");CHKERRQ(ierr); 56196e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 56206e683305SStefano Zampini } 56216e683305SStefano Zampini ierr = PetscObjectReference((PetscObject)t_coarse_mat_is);CHKERRQ(ierr); 56226e683305SStefano Zampini coarse_mat_is = t_coarse_mat_is; 56236e683305SStefano Zampini } 56246e683305SStefano Zampini 56256e683305SStefano Zampini /* create local to global scatters for coarse problem */ 562668457ee5SStefano Zampini if (compute_vecs) { 56276e683305SStefano Zampini PetscInt lrows; 56286e683305SStefano Zampini ierr = VecDestroy(&pcbddc->coarse_vec);CHKERRQ(ierr); 56296e683305SStefano Zampini if (coarse_mat_is) { 56306e683305SStefano Zampini ierr = MatGetLocalSize(coarse_mat_is,&lrows,NULL);CHKERRQ(ierr); 56316e683305SStefano Zampini } else { 56326e683305SStefano Zampini lrows = 0; 56336e683305SStefano Zampini } 56346e683305SStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pc),&pcbddc->coarse_vec);CHKERRQ(ierr); 56356e683305SStefano Zampini ierr = VecSetSizes(pcbddc->coarse_vec,lrows,PETSC_DECIDE);CHKERRQ(ierr); 56366e683305SStefano Zampini ierr = VecSetType(pcbddc->coarse_vec,VECSTANDARD);CHKERRQ(ierr); 56376e683305SStefano Zampini ierr = VecScatterDestroy(&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 56386e683305SStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_P,NULL,pcbddc->coarse_vec,coarse_is,&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 56396e683305SStefano Zampini } 56406e683305SStefano Zampini ierr = ISDestroy(&coarse_is);CHKERRQ(ierr); 56416e683305SStefano Zampini ierr = MatDestroy(&t_coarse_mat_is);CHKERRQ(ierr); 5642c8587f34SStefano Zampini 5643f9eb5b7dSStefano Zampini /* set defaults for coarse KSP and PC */ 5644f9eb5b7dSStefano Zampini if (multilevel_allowed) { 5645f9eb5b7dSStefano Zampini coarse_ksp_type = KSPRICHARDSON; 5646f9eb5b7dSStefano Zampini coarse_pc_type = PCBDDC; 5647f9eb5b7dSStefano Zampini } else { 5648f9eb5b7dSStefano Zampini coarse_ksp_type = KSPPREONLY; 5649f9eb5b7dSStefano Zampini coarse_pc_type = PCREDUNDANT; 5650c8587f34SStefano Zampini } 5651c8587f34SStefano Zampini 56526e683305SStefano Zampini /* print some info if requested */ 56536e683305SStefano Zampini if (pcbddc->dbg_flag) { 56546e683305SStefano Zampini if (!multilevel_allowed) { 56556e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 56566e683305SStefano Zampini if (multilevel_requested) { 56576e683305SStefano 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); 56586e683305SStefano Zampini } else if (pcbddc->max_levels) { 56596e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Maximum number of requested levels reached (%d)\n",pcbddc->max_levels);CHKERRQ(ierr); 56606e683305SStefano Zampini } 56616e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 56626e683305SStefano Zampini } 56636e683305SStefano Zampini } 56646e683305SStefano Zampini 5665f9eb5b7dSStefano Zampini /* create the coarse KSP object only once with defaults */ 56666e683305SStefano Zampini if (coarse_mat_is) { 56676e683305SStefano Zampini MatReuse coarse_mat_reuse; 56686a1308c2SStefano Zampini PetscViewer dbg_viewer = NULL; 56696e683305SStefano Zampini if (pcbddc->dbg_flag) { 56706e683305SStefano Zampini dbg_viewer = PETSC_VIEWER_STDOUT_(PetscObjectComm((PetscObject)coarse_mat_is)); 56716e683305SStefano Zampini ierr = PetscViewerASCIIAddTab(dbg_viewer,2*pcbddc->current_level);CHKERRQ(ierr); 56726e683305SStefano Zampini } 5673f9eb5b7dSStefano Zampini if (!pcbddc->coarse_ksp) { 5674312be037SStefano Zampini char prefix[256],str_level[16]; 5675e604994aSStefano Zampini size_t len; 56766e683305SStefano Zampini ierr = KSPCreate(PetscObjectComm((PetscObject)coarse_mat_is),&pcbddc->coarse_ksp);CHKERRQ(ierr); 5677422a814eSBarry Smith ierr = KSPSetErrorIfNotConverged(pcbddc->coarse_ksp,pc->erroriffailure);CHKERRQ(ierr); 5678c8587f34SStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->coarse_ksp,(PetscObject)pc,1);CHKERRQ(ierr); 5679f9eb5b7dSStefano Zampini ierr = KSPSetTolerances(pcbddc->coarse_ksp,PETSC_DEFAULT,PETSC_DEFAULT,PETSC_DEFAULT,1);CHKERRQ(ierr); 56805f76c7aeSStefano Zampini ierr = KSPSetOperators(pcbddc->coarse_ksp,coarse_mat_is,coarse_mat_is);CHKERRQ(ierr); 5681c8587f34SStefano Zampini ierr = KSPSetType(pcbddc->coarse_ksp,coarse_ksp_type);CHKERRQ(ierr); 56826e683305SStefano Zampini ierr = KSPSetNormType(pcbddc->coarse_ksp,KSP_NORM_NONE);CHKERRQ(ierr); 5683c8587f34SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&pc_temp);CHKERRQ(ierr); 5684c8587f34SStefano Zampini ierr = PCSetType(pc_temp,coarse_pc_type);CHKERRQ(ierr); 5685e604994aSStefano Zampini /* prefix */ 5686e604994aSStefano Zampini ierr = PetscStrcpy(prefix,"");CHKERRQ(ierr); 5687e604994aSStefano Zampini ierr = PetscStrcpy(str_level,"");CHKERRQ(ierr); 5688e604994aSStefano Zampini if (!pcbddc->current_level) { 5689e604994aSStefano Zampini ierr = PetscStrcpy(prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 5690e604994aSStefano Zampini ierr = PetscStrcat(prefix,"pc_bddc_coarse_");CHKERRQ(ierr); 5691c8587f34SStefano Zampini } else { 5692e604994aSStefano Zampini ierr = PetscStrlen(((PetscObject)pc)->prefix,&len);CHKERRQ(ierr); 5693312be037SStefano Zampini if (pcbddc->current_level>1) len -= 3; /* remove "lX_" with X level number */ 5694312be037SStefano Zampini if (pcbddc->current_level>10) len -= 1; /* remove another char from level number */ 569534d6797cSStefano Zampini ierr = PetscStrncpy(prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 5696312be037SStefano Zampini sprintf(str_level,"l%d_",(int)(pcbddc->current_level)); 5697e604994aSStefano Zampini ierr = PetscStrcat(prefix,str_level);CHKERRQ(ierr); 5698e604994aSStefano Zampini } 5699e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->coarse_ksp,prefix);CHKERRQ(ierr); 57003e3c6dadSStefano Zampini /* propagate BDDC info to the next level (these are dummy calls if pc_temp is not of type PCBDDC) */ 57013e3c6dadSStefano Zampini ierr = PCBDDCSetLevel(pc_temp,pcbddc->current_level+1);CHKERRQ(ierr); 57023e3c6dadSStefano Zampini ierr = PCBDDCSetCoarseningRatio(pc_temp,pcbddc->coarsening_ratio);CHKERRQ(ierr); 57033e3c6dadSStefano Zampini ierr = PCBDDCSetLevels(pc_temp,pcbddc->max_levels);CHKERRQ(ierr); 5704f9eb5b7dSStefano Zampini /* allow user customization */ 5705f9eb5b7dSStefano Zampini ierr = KSPSetFromOptions(pcbddc->coarse_ksp);CHKERRQ(ierr); 57063e3c6dadSStefano Zampini } 57073e3c6dadSStefano Zampini /* propagate BDDC info to the next level (these are dummy calls if pc_temp is not of type PCBDDC) */ 570851bea450SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&pc_temp);CHKERRQ(ierr); 57093e3c6dadSStefano Zampini if (nisdofs) { 57103e3c6dadSStefano Zampini ierr = PCBDDCSetDofsSplitting(pc_temp,nisdofs,isarray);CHKERRQ(ierr); 57113e3c6dadSStefano Zampini for (i=0;i<nisdofs;i++) { 57123e3c6dadSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 57133e3c6dadSStefano Zampini } 57143e3c6dadSStefano Zampini } 57153e3c6dadSStefano Zampini if (nisneu) { 57163e3c6dadSStefano Zampini ierr = PCBDDCSetNeumannBoundaries(pc_temp,isarray[nisdofs]);CHKERRQ(ierr); 57173e3c6dadSStefano Zampini ierr = ISDestroy(&isarray[nisdofs]);CHKERRQ(ierr); 5718312be037SStefano Zampini } 571930368db7SStefano Zampini if (nisvert) { 572030368db7SStefano Zampini ierr = PCBDDCSetPrimalVerticesIS(pc_temp,isarray[nis-1]);CHKERRQ(ierr); 572130368db7SStefano Zampini ierr = ISDestroy(&isarray[nis-1]);CHKERRQ(ierr); 572230368db7SStefano Zampini } 5723f9eb5b7dSStefano Zampini 5724f9eb5b7dSStefano Zampini /* get some info after set from options */ 5725f9eb5b7dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCNN,&isnn);CHKERRQ(ierr); 5726f9eb5b7dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCBDDC,&isbddc);CHKERRQ(ierr); 57274f3a063dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCREDUNDANT,&isredundant);CHKERRQ(ierr); 57286e683305SStefano Zampini if (isbddc && !multilevel_allowed) { /* multilevel can only be requested via pc_bddc_set_levels */ 5729f9eb5b7dSStefano Zampini ierr = PCSetType(pc_temp,coarse_pc_type);CHKERRQ(ierr); 5730f9eb5b7dSStefano Zampini isbddc = PETSC_FALSE; 5731f9eb5b7dSStefano Zampini } 573239f0f02cSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 57334f3a063dSStefano Zampini if (isredundant) { 57344f3a063dSStefano Zampini KSP inner_ksp; 57354f3a063dSStefano Zampini PC inner_pc; 57364f3a063dSStefano Zampini ierr = PCRedundantGetKSP(pc_temp,&inner_ksp);CHKERRQ(ierr); 57374f3a063dSStefano Zampini ierr = KSPGetPC(inner_ksp,&inner_pc);CHKERRQ(ierr); 57384f3a063dSStefano Zampini ierr = PCFactorSetReuseFill(inner_pc,PETSC_TRUE);CHKERRQ(ierr); 57394f3a063dSStefano Zampini } 5740f9eb5b7dSStefano Zampini 5741f9eb5b7dSStefano Zampini /* assemble coarse matrix */ 5742fa7f1dd8SStefano Zampini if (coarse_reuse) { 574381d9aea3SBarry Smith ierr = KSPGetOperators(pcbddc->coarse_ksp,&coarse_mat,NULL);CHKERRQ(ierr); 5744fa7f1dd8SStefano Zampini ierr = PetscObjectReference((PetscObject)coarse_mat);CHKERRQ(ierr); 57456e683305SStefano Zampini coarse_mat_reuse = MAT_REUSE_MATRIX; 5746fa7f1dd8SStefano Zampini } else { 57476e683305SStefano Zampini coarse_mat_reuse = MAT_INITIAL_MATRIX; 5748fa7f1dd8SStefano Zampini } 5749c8587f34SStefano Zampini if (isbddc || isnn) { 5750720d30f9SStefano Zampini if (isbddc) { /* currently there are no APIs for these options */ 5751720d30f9SStefano Zampini PC_BDDC* pcbddc_coarse = (PC_BDDC*)pc_temp->data; 5752720d30f9SStefano Zampini pcbddc_coarse->detect_disconnected = PETSC_TRUE; 5753720d30f9SStefano Zampini pcbddc_coarse->benign_saddle_point = pcbddc->benign_saddle_point; 5754720d30f9SStefano Zampini pcbddc_coarse->benign_compute_nonetflux = pcbddc->benign_compute_nonetflux; 5755720d30f9SStefano Zampini if (pcbddc_coarse->benign_compute_nonetflux) { 5756720d30f9SStefano Zampini pcbddc_coarse->adaptive_userdefined = PETSC_TRUE; 5757720d30f9SStefano Zampini } 5758d4d8cf7bSStefano Zampini } 575922bc73bbSStefano Zampini if (pcbddc->coarsening_ratio > 1) { 576070cf5478SStefano Zampini if (!pcbddc->coarse_subassembling) { /* subassembling info is not present */ 5761b0c7d250SStefano Zampini ierr = MatISGetSubassemblingPattern(coarse_mat_is,active_procs/pcbddc->coarsening_ratio,pcbddc->coarse_adj_red,&pcbddc->coarse_subassembling);CHKERRQ(ierr); 576222b6e8a2SStefano Zampini if (pcbddc->dbg_flag) { 57636e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 57646e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Subassembling pattern\n");CHKERRQ(ierr); 57656e683305SStefano Zampini ierr = ISView(pcbddc->coarse_subassembling,dbg_viewer);CHKERRQ(ierr); 57666e683305SStefano Zampini ierr = PetscViewerFlush(dbg_viewer);CHKERRQ(ierr); 576722b6e8a2SStefano Zampini } 576870cf5478SStefano Zampini } 576953a05cb3SStefano Zampini ierr = MatISSubassemble(coarse_mat_is,pcbddc->coarse_subassembling,0,PETSC_FALSE,PETSC_FALSE,coarse_mat_reuse,&coarse_mat,0,NULL);CHKERRQ(ierr); 577070cf5478SStefano Zampini } else { 577122bc73bbSStefano Zampini ierr = PetscObjectReference((PetscObject)coarse_mat_is);CHKERRQ(ierr); 577222bc73bbSStefano Zampini coarse_mat = coarse_mat_is; 577322bc73bbSStefano Zampini } 577422bc73bbSStefano Zampini } else { 57752e1e5fa4SStefano Zampini ierr = MatISGetMPIXAIJ(coarse_mat_is,coarse_mat_reuse,&coarse_mat);CHKERRQ(ierr); 5776c8587f34SStefano Zampini } 5777c8587f34SStefano Zampini ierr = MatDestroy(&coarse_mat_is);CHKERRQ(ierr); 5778c8587f34SStefano Zampini 57793301b35fSStefano Zampini /* propagate symmetry info of coarse matrix */ 57805a16e3a0SStefano Zampini ierr = MatSetOption(coarse_mat,MAT_STRUCTURALLY_SYMMETRIC,PETSC_TRUE);CHKERRQ(ierr); 57813301b35fSStefano Zampini if (pc->pmat->symmetric_set) { 57823301b35fSStefano Zampini ierr = MatSetOption(coarse_mat,MAT_SYMMETRIC,pc->pmat->symmetric);CHKERRQ(ierr); 57833301b35fSStefano Zampini } 57843301b35fSStefano Zampini if (pc->pmat->hermitian_set) { 57853301b35fSStefano Zampini ierr = MatSetOption(coarse_mat,MAT_HERMITIAN,pc->pmat->hermitian);CHKERRQ(ierr); 57863301b35fSStefano Zampini } 57873301b35fSStefano Zampini if (pc->pmat->spd_set) { 57883301b35fSStefano Zampini ierr = MatSetOption(coarse_mat,MAT_SPD,pc->pmat->spd);CHKERRQ(ierr); 57893301b35fSStefano Zampini } 57906e683305SStefano Zampini /* set operators */ 57915f76c7aeSStefano Zampini ierr = KSPSetOperators(pcbddc->coarse_ksp,coarse_mat,coarse_mat);CHKERRQ(ierr); 57926e683305SStefano Zampini if (pcbddc->dbg_flag) { 57936e683305SStefano Zampini ierr = PetscViewerASCIISubtractTab(dbg_viewer,2*pcbddc->current_level);CHKERRQ(ierr); 57946e683305SStefano Zampini } 57956e683305SStefano Zampini } else { /* processes non partecipating to coarse solver (if any) */ 57966e683305SStefano Zampini coarse_mat = 0; 57976e683305SStefano Zampini } 57986e683305SStefano Zampini ierr = PetscFree(isarray);CHKERRQ(ierr); 5799b1ecc7b1SStefano Zampini #if 0 5800b9b85e73SStefano Zampini { 5801b9b85e73SStefano Zampini PetscViewer viewer; 5802b9b85e73SStefano Zampini char filename[256]; 5803b1ecc7b1SStefano Zampini sprintf(filename,"coarse_mat_level%d.m",pcbddc->current_level); 5804b1ecc7b1SStefano Zampini ierr = PetscViewerASCIIOpen(PetscObjectComm((PetscObject)coarse_mat),filename,&viewer);CHKERRQ(ierr); 5805b9b85e73SStefano Zampini ierr = PetscViewerSetFormat(viewer,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 5806b9b85e73SStefano Zampini ierr = MatView(coarse_mat,viewer);CHKERRQ(ierr); 5807b9b85e73SStefano Zampini ierr = PetscViewerDestroy(&viewer);CHKERRQ(ierr); 5808b9b85e73SStefano Zampini } 5809b9b85e73SStefano Zampini #endif 5810c8587f34SStefano Zampini 5811c8587f34SStefano Zampini /* Compute coarse null space (special handling by BDDC only) */ 5812298c0119SStefano Zampini #if 0 5813c8587f34SStefano Zampini if (pcbddc->NullSpace) { 5814c8587f34SStefano Zampini ierr = PCBDDCNullSpaceAssembleCoarse(pc,coarse_mat,&CoarseNullSpace);CHKERRQ(ierr); 581598a51de6SStefano Zampini } 5816298c0119SStefano Zampini #endif 5817b0f5fe93SStefano Zampini /* hack */ 581898a51de6SStefano Zampini if (pcbddc->coarse_ksp) { 581998a51de6SStefano Zampini Vec crhs,csol; 582004708bb6SStefano Zampini 5821f347579bSStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&csol);CHKERRQ(ierr); 5822f347579bSStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&crhs);CHKERRQ(ierr); 5823f347579bSStefano Zampini if (!csol) { 58242a7a6963SBarry Smith ierr = MatCreateVecs(coarse_mat,&((pcbddc->coarse_ksp)->vec_sol),NULL);CHKERRQ(ierr); 5825f9eb5b7dSStefano Zampini } 5826f347579bSStefano Zampini if (!crhs) { 58272a7a6963SBarry Smith ierr = MatCreateVecs(coarse_mat,NULL,&((pcbddc->coarse_ksp)->vec_rhs));CHKERRQ(ierr); 5828f347579bSStefano Zampini } 5829b0f5fe93SStefano Zampini } 5830b0f5fe93SStefano Zampini 5831b0f5fe93SStefano Zampini /* compute null space for coarse solver if the benign trick has been requested */ 5832b0f5fe93SStefano Zampini if (pcbddc->benign_null) { 5833b0f5fe93SStefano Zampini 5834b0f5fe93SStefano Zampini ierr = VecSet(pcbddc->vec1_P,0.);CHKERRQ(ierr); 58354f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 58364f1b2e48SStefano Zampini ierr = VecSetValue(pcbddc->vec1_P,pcbddc->local_primal_size-pcbddc->benign_n+i,1.0,INSERT_VALUES);CHKERRQ(ierr); 58374f1b2e48SStefano Zampini } 5838b0f5fe93SStefano Zampini ierr = VecAssemblyBegin(pcbddc->vec1_P);CHKERRQ(ierr); 5839b0f5fe93SStefano Zampini ierr = VecAssemblyEnd(pcbddc->vec1_P);CHKERRQ(ierr); 5840b0f5fe93SStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 5841b0f5fe93SStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 5842b0f5fe93SStefano Zampini if (coarse_mat) { 5843b0f5fe93SStefano Zampini Vec nullv; 5844b0f5fe93SStefano Zampini PetscScalar *array,*array2; 5845b0f5fe93SStefano Zampini PetscInt nl; 5846b0f5fe93SStefano Zampini 5847b0f5fe93SStefano Zampini ierr = MatCreateVecs(coarse_mat,&nullv,NULL);CHKERRQ(ierr); 5848b0f5fe93SStefano Zampini ierr = VecGetLocalSize(nullv,&nl);CHKERRQ(ierr); 5849b0f5fe93SStefano Zampini ierr = VecGetArrayRead(pcbddc->coarse_vec,(const PetscScalar**)&array);CHKERRQ(ierr); 5850b0f5fe93SStefano Zampini ierr = VecGetArray(nullv,&array2);CHKERRQ(ierr); 5851b0f5fe93SStefano Zampini ierr = PetscMemcpy(array2,array,nl*sizeof(*array));CHKERRQ(ierr); 5852b0f5fe93SStefano Zampini ierr = VecRestoreArray(nullv,&array2);CHKERRQ(ierr); 5853b0f5fe93SStefano Zampini ierr = VecRestoreArrayRead(pcbddc->coarse_vec,(const PetscScalar**)&array);CHKERRQ(ierr); 5854b0f5fe93SStefano Zampini ierr = VecNormalize(nullv,NULL);CHKERRQ(ierr); 5855b0f5fe93SStefano Zampini ierr = MatNullSpaceCreate(PetscObjectComm((PetscObject)coarse_mat),PETSC_FALSE,1,&nullv,&CoarseNullSpace);CHKERRQ(ierr); 5856b0f5fe93SStefano Zampini ierr = VecDestroy(&nullv);CHKERRQ(ierr); 5857b0f5fe93SStefano Zampini } 5858b0f5fe93SStefano Zampini } 5859b0f5fe93SStefano Zampini 5860b0f5fe93SStefano Zampini if (pcbddc->coarse_ksp) { 5861b0f5fe93SStefano Zampini PetscBool ispreonly; 5862b0f5fe93SStefano Zampini 5863b0f5fe93SStefano Zampini if (CoarseNullSpace) { 5864b0f5fe93SStefano Zampini PetscBool isnull; 5865b0f5fe93SStefano Zampini ierr = MatNullSpaceTest(CoarseNullSpace,coarse_mat,&isnull);CHKERRQ(ierr); 5866d4d8cf7bSStefano Zampini if (0) { 586730368db7SStefano Zampini if (isbddc && !pcbddc->benign_saddle_point) { 5868b0f5fe93SStefano Zampini ierr = PCBDDCSetNullSpace(pc_temp,CoarseNullSpace);CHKERRQ(ierr); 5869b0f5fe93SStefano Zampini } else { 5870b0f5fe93SStefano Zampini ierr = MatSetNullSpace(coarse_mat,CoarseNullSpace);CHKERRQ(ierr); 5871b0f5fe93SStefano Zampini } 5872b0f5fe93SStefano Zampini } else { 5873b0f5fe93SStefano Zampini ierr = MatNullSpaceDestroy(&CoarseNullSpace);CHKERRQ(ierr); 5874b0f5fe93SStefano Zampini } 5875b0f5fe93SStefano Zampini } 5876b0f5fe93SStefano Zampini /* setup coarse ksp */ 5877b0f5fe93SStefano Zampini ierr = KSPSetUp(pcbddc->coarse_ksp);CHKERRQ(ierr); 5878cbcc2c2aSStefano Zampini /* Check coarse problem if in debug mode or if solving with an iterative method */ 5879cbcc2c2aSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcbddc->coarse_ksp,KSPPREONLY,&ispreonly);CHKERRQ(ierr); 58806e683305SStefano Zampini if (pcbddc->dbg_flag || (!ispreonly && pcbddc->use_coarse_estimates) ) { 5881c8587f34SStefano Zampini KSP check_ksp; 58822b510759SStefano Zampini KSPType check_ksp_type; 5883c8587f34SStefano Zampini PC check_pc; 58846e683305SStefano Zampini Vec check_vec,coarse_vec; 58856a1308c2SStefano Zampini PetscReal abs_infty_error,infty_error,lambda_min=1.0,lambda_max=1.0; 58862b510759SStefano Zampini PetscInt its; 58876e683305SStefano Zampini PetscBool compute_eigs; 58886e683305SStefano Zampini PetscReal *eigs_r,*eigs_c; 58896e683305SStefano Zampini PetscInt neigs; 58908e185a42SStefano Zampini const char *prefix; 5891c8587f34SStefano Zampini 58922b510759SStefano Zampini /* Create ksp object suitable for estimation of extreme eigenvalues */ 58936e683305SStefano Zampini ierr = KSPCreate(PetscObjectComm((PetscObject)pcbddc->coarse_ksp),&check_ksp);CHKERRQ(ierr); 5894422a814eSBarry Smith ierr = KSPSetErrorIfNotConverged(pcbddc->coarse_ksp,pc->erroriffailure);CHKERRQ(ierr); 589523ee1639SBarry Smith ierr = KSPSetOperators(check_ksp,coarse_mat,coarse_mat);CHKERRQ(ierr); 5896f4ddd8eeSStefano Zampini ierr = KSPSetTolerances(check_ksp,1.e-12,1.e-12,PETSC_DEFAULT,pcbddc->coarse_size);CHKERRQ(ierr); 58972b510759SStefano Zampini if (ispreonly) { 58982b510759SStefano Zampini check_ksp_type = KSPPREONLY; 58996e683305SStefano Zampini compute_eigs = PETSC_FALSE; 59002b510759SStefano Zampini } else { 5901cbcc2c2aSStefano Zampini check_ksp_type = KSPGMRES; 59026e683305SStefano Zampini compute_eigs = PETSC_TRUE; 5903c8587f34SStefano Zampini } 5904c8587f34SStefano Zampini ierr = KSPSetType(check_ksp,check_ksp_type);CHKERRQ(ierr); 59056e683305SStefano Zampini ierr = KSPSetComputeSingularValues(check_ksp,compute_eigs);CHKERRQ(ierr); 59066e683305SStefano Zampini ierr = KSPSetComputeEigenvalues(check_ksp,compute_eigs);CHKERRQ(ierr); 59076e683305SStefano Zampini ierr = KSPGMRESSetRestart(check_ksp,pcbddc->coarse_size+1);CHKERRQ(ierr); 5908a7dc3881SStefano Zampini ierr = KSPGetOptionsPrefix(pcbddc->coarse_ksp,&prefix);CHKERRQ(ierr); 5909a7dc3881SStefano Zampini ierr = KSPSetOptionsPrefix(check_ksp,prefix);CHKERRQ(ierr); 5910a7dc3881SStefano Zampini ierr = KSPAppendOptionsPrefix(check_ksp,"check_");CHKERRQ(ierr); 5911a7dc3881SStefano Zampini ierr = KSPSetFromOptions(check_ksp);CHKERRQ(ierr); 5912c8587f34SStefano Zampini ierr = KSPSetUp(check_ksp);CHKERRQ(ierr); 5913c8587f34SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&check_pc);CHKERRQ(ierr); 5914c8587f34SStefano Zampini ierr = KSPSetPC(check_ksp,check_pc);CHKERRQ(ierr); 5915c8587f34SStefano Zampini /* create random vec */ 59166e683305SStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&coarse_vec);CHKERRQ(ierr); 59176e683305SStefano Zampini ierr = VecDuplicate(coarse_vec,&check_vec);CHKERRQ(ierr); 5918c8587f34SStefano Zampini ierr = VecSetRandom(check_vec,NULL);CHKERRQ(ierr); 5919c8587f34SStefano Zampini if (CoarseNullSpace) { 5920c8587f34SStefano Zampini ierr = MatNullSpaceRemove(CoarseNullSpace,check_vec);CHKERRQ(ierr); 5921c8587f34SStefano Zampini } 59226e683305SStefano Zampini ierr = MatMult(coarse_mat,check_vec,coarse_vec);CHKERRQ(ierr); 5923c8587f34SStefano Zampini /* solve coarse problem */ 59246e683305SStefano Zampini ierr = KSPSolve(check_ksp,coarse_vec,coarse_vec);CHKERRQ(ierr); 5925c8587f34SStefano Zampini if (CoarseNullSpace) { 59266e683305SStefano Zampini ierr = MatNullSpaceRemove(CoarseNullSpace,coarse_vec);CHKERRQ(ierr); 5927c8587f34SStefano Zampini } 5928cbcc2c2aSStefano Zampini /* set eigenvalue estimation if preonly has not been requested */ 59296e683305SStefano Zampini if (compute_eigs) { 5930854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->coarse_size+1,&eigs_r);CHKERRQ(ierr); 5931854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->coarse_size+1,&eigs_c);CHKERRQ(ierr); 59326e683305SStefano Zampini ierr = KSPComputeEigenvalues(check_ksp,pcbddc->coarse_size+1,eigs_r,eigs_c,&neigs);CHKERRQ(ierr); 59336e683305SStefano Zampini lambda_max = eigs_r[neigs-1]; 59346e683305SStefano Zampini lambda_min = eigs_r[0]; 59356e683305SStefano Zampini if (pcbddc->use_coarse_estimates) { 59366e683305SStefano Zampini if (lambda_max>lambda_min) { 5937cbcc2c2aSStefano Zampini ierr = KSPChebyshevSetEigenvalues(pcbddc->coarse_ksp,lambda_max,lambda_min);CHKERRQ(ierr); 5938cbcc2c2aSStefano Zampini ierr = KSPRichardsonSetScale(pcbddc->coarse_ksp,2.0/(lambda_max+lambda_min));CHKERRQ(ierr); 5939cbcc2c2aSStefano Zampini } 5940c8587f34SStefano Zampini } 5941c8587f34SStefano Zampini } 5942cbcc2c2aSStefano Zampini 5943c8587f34SStefano Zampini /* check coarse problem residual error */ 59446e683305SStefano Zampini if (pcbddc->dbg_flag) { 59456e683305SStefano Zampini PetscViewer dbg_viewer = PETSC_VIEWER_STDOUT_(PetscObjectComm((PetscObject)pcbddc->coarse_ksp)); 59466e683305SStefano Zampini ierr = PetscViewerASCIIAddTab(dbg_viewer,2*(pcbddc->current_level+1));CHKERRQ(ierr); 59476e683305SStefano Zampini ierr = VecAXPY(check_vec,-1.0,coarse_vec);CHKERRQ(ierr); 5948c8587f34SStefano Zampini ierr = VecNorm(check_vec,NORM_INFINITY,&infty_error);CHKERRQ(ierr); 59496e683305SStefano Zampini ierr = MatMult(coarse_mat,check_vec,coarse_vec);CHKERRQ(ierr); 59506e683305SStefano Zampini ierr = VecNorm(coarse_vec,NORM_INFINITY,&abs_infty_error);CHKERRQ(ierr); 5951c8587f34SStefano Zampini ierr = VecDestroy(&check_vec);CHKERRQ(ierr); 5952779c1cceSStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem details (use estimates %d)\n",pcbddc->use_coarse_estimates);CHKERRQ(ierr); 59536e683305SStefano Zampini ierr = PetscObjectPrintClassNamePrefixType((PetscObject)(pcbddc->coarse_ksp),dbg_viewer);CHKERRQ(ierr); 59546e683305SStefano Zampini ierr = PetscObjectPrintClassNamePrefixType((PetscObject)(check_pc),dbg_viewer);CHKERRQ(ierr); 59556e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem exact infty_error : %1.6e\n",infty_error);CHKERRQ(ierr); 59566e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem residual infty_error: %1.6e\n",abs_infty_error);CHKERRQ(ierr); 5957b0f5fe93SStefano Zampini if (CoarseNullSpace) { 5958b0f5fe93SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem is singular\n");CHKERRQ(ierr); 5959b0f5fe93SStefano Zampini } 59606e683305SStefano Zampini if (compute_eigs) { 59616e683305SStefano Zampini PetscReal lambda_max_s,lambda_min_s; 5962deec49d1SStefano Zampini ierr = KSPGetType(check_ksp,&check_ksp_type);CHKERRQ(ierr); 5963c8587f34SStefano Zampini ierr = KSPGetIterationNumber(check_ksp,&its);CHKERRQ(ierr); 59646e683305SStefano Zampini ierr = KSPComputeExtremeSingularValues(check_ksp,&lambda_max_s,&lambda_min_s);CHKERRQ(ierr); 59656e683305SStefano 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); 59666e683305SStefano Zampini for (i=0;i<neigs;i++) { 59676e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"%1.6e %1.6ei\n",eigs_r[i],eigs_c[i]);CHKERRQ(ierr); 5968c8587f34SStefano Zampini } 59696e683305SStefano Zampini } 59706e683305SStefano Zampini ierr = PetscViewerFlush(dbg_viewer);CHKERRQ(ierr); 59716e683305SStefano Zampini ierr = PetscViewerASCIISubtractTab(dbg_viewer,2*(pcbddc->current_level+1));CHKERRQ(ierr); 59726e683305SStefano Zampini } 5973c8587f34SStefano Zampini ierr = KSPDestroy(&check_ksp);CHKERRQ(ierr); 59746e683305SStefano Zampini if (compute_eigs) { 59756e683305SStefano Zampini ierr = PetscFree(eigs_r);CHKERRQ(ierr); 59766e683305SStefano Zampini ierr = PetscFree(eigs_c);CHKERRQ(ierr); 5977c8587f34SStefano Zampini } 59786e683305SStefano Zampini } 59796e683305SStefano Zampini } 5980cbcc2c2aSStefano Zampini /* print additional info */ 5981cbcc2c2aSStefano Zampini if (pcbddc->dbg_flag) { 59826e683305SStefano Zampini /* waits until all processes reaches this point */ 59836e683305SStefano Zampini ierr = PetscBarrier((PetscObject)pc);CHKERRQ(ierr); 5984cbcc2c2aSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Coarse solver setup completed at level %d\n",pcbddc->current_level);CHKERRQ(ierr); 5985cbcc2c2aSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 5986cbcc2c2aSStefano Zampini } 5987cbcc2c2aSStefano Zampini 59882b510759SStefano Zampini /* free memory */ 5989c8587f34SStefano Zampini ierr = MatNullSpaceDestroy(&CoarseNullSpace);CHKERRQ(ierr); 5990fdc635d7SStefano Zampini ierr = MatDestroy(&coarse_mat);CHKERRQ(ierr); 5991c8587f34SStefano Zampini PetscFunctionReturn(0); 5992c8587f34SStefano Zampini } 5993674ae819SStefano Zampini 5994f34684f1SStefano Zampini #undef __FUNCT__ 5995f34684f1SStefano Zampini #define __FUNCT__ "PCBDDCComputePrimalNumbering" 5996f34684f1SStefano Zampini PetscErrorCode PCBDDCComputePrimalNumbering(PC pc,PetscInt* coarse_size_n,PetscInt** local_primal_indices_n) 5997f34684f1SStefano Zampini { 5998f34684f1SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 5999f34684f1SStefano Zampini PC_IS* pcis = (PC_IS*)pc->data; 6000f34684f1SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 6001dc456d91SStefano Zampini IS subset,subset_mult,subset_n; 6002dc456d91SStefano Zampini PetscInt local_size,coarse_size=0; 600373be2a3aSStefano Zampini PetscInt *local_primal_indices=NULL; 6004dc456d91SStefano Zampini const PetscInt *t_local_primal_indices; 6005f34684f1SStefano Zampini PetscErrorCode ierr; 6006f34684f1SStefano Zampini 6007f34684f1SStefano Zampini PetscFunctionBegin; 6008f34684f1SStefano Zampini /* Compute global number of coarse dofs */ 60090e6343abSStefano Zampini if (pcbddc->local_primal_size && !pcbddc->local_primal_ref_node) { 60100e6343abSStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"BDDC ConstraintsSetUp should be called first"); 6011727cdba6SStefano Zampini } 6012dc456d91SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)(pc->pmat)),pcbddc->local_primal_size_cc,pcbddc->local_primal_ref_node,PETSC_COPY_VALUES,&subset_n);CHKERRQ(ierr); 60133bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(pcis->mapping,subset_n,&subset);CHKERRQ(ierr); 6014dc456d91SStefano Zampini ierr = ISDestroy(&subset_n);CHKERRQ(ierr); 6015dc456d91SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)(pc->pmat)),pcbddc->local_primal_size_cc,pcbddc->local_primal_ref_mult,PETSC_COPY_VALUES,&subset_mult);CHKERRQ(ierr); 6016dc456d91SStefano Zampini ierr = PCBDDCSubsetNumbering(subset,subset_mult,&coarse_size,&subset_n);CHKERRQ(ierr); 6017dc456d91SStefano Zampini ierr = ISDestroy(&subset);CHKERRQ(ierr); 6018dc456d91SStefano Zampini ierr = ISDestroy(&subset_mult);CHKERRQ(ierr); 6019dc456d91SStefano Zampini ierr = ISGetLocalSize(subset_n,&local_size);CHKERRQ(ierr); 60200e6343abSStefano Zampini if (local_size != pcbddc->local_primal_size) { 60210e6343abSStefano Zampini SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Invalid number of local primal indices computed %d != %d",local_size,pcbddc->local_primal_size); 60220e6343abSStefano Zampini } 6023dc456d91SStefano Zampini ierr = PetscMalloc1(local_size,&local_primal_indices);CHKERRQ(ierr); 6024dc456d91SStefano Zampini ierr = ISGetIndices(subset_n,&t_local_primal_indices);CHKERRQ(ierr); 6025dc456d91SStefano Zampini ierr = PetscMemcpy(local_primal_indices,t_local_primal_indices,local_size*sizeof(PetscInt));CHKERRQ(ierr); 6026dc456d91SStefano Zampini ierr = ISRestoreIndices(subset_n,&t_local_primal_indices);CHKERRQ(ierr); 6027dc456d91SStefano Zampini ierr = ISDestroy(&subset_n);CHKERRQ(ierr); 6028f34684f1SStefano Zampini 6029f34684f1SStefano Zampini /* check numbering */ 6030f34684f1SStefano Zampini if (pcbddc->dbg_flag) { 6031019a44ceSStefano Zampini PetscScalar coarsesum,*array,*array2; 6032dc456d91SStefano Zampini PetscInt i; 6033b9b85e73SStefano Zampini PetscBool set_error = PETSC_FALSE,set_error_reduced = PETSC_FALSE; 6034f34684f1SStefano Zampini 6035f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 6036f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 6037f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Check coarse indices\n");CHKERRQ(ierr); 60381575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 6039019a44ceSStefano Zampini /* counter */ 6040019a44ceSStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 6041019a44ceSStefano Zampini ierr = VecSet(pcis->vec1_N,1.0);CHKERRQ(ierr); 6042019a44ceSStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6043019a44ceSStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6044019a44ceSStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_global,pcis->vec2_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6045019a44ceSStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_global,pcis->vec2_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6046f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_N,0.0);CHKERRQ(ierr); 6047f34684f1SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 6048727cdba6SStefano Zampini ierr = VecSetValue(pcis->vec1_N,pcbddc->primal_indices_local_idxs[i],1.0,INSERT_VALUES);CHKERRQ(ierr); 6049f34684f1SStefano Zampini } 6050f34684f1SStefano Zampini ierr = VecAssemblyBegin(pcis->vec1_N);CHKERRQ(ierr); 6051f34684f1SStefano Zampini ierr = VecAssemblyEnd(pcis->vec1_N);CHKERRQ(ierr); 6052f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 6053e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6054e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6055e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_global,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6056e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_global,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6057f34684f1SStefano Zampini ierr = VecGetArray(pcis->vec1_N,&array);CHKERRQ(ierr); 6058019a44ceSStefano Zampini ierr = VecGetArray(pcis->vec2_N,&array2);CHKERRQ(ierr); 6059f34684f1SStefano Zampini for (i=0;i<pcis->n;i++) { 6060019a44ceSStefano Zampini if (array[i] != 0.0 && array[i] != array2[i]) { 606175c01103SStefano Zampini PetscInt owned = (PetscInt)PetscRealPart(array[i]); 606275c01103SStefano Zampini PetscInt neigh = (PetscInt)PetscRealPart(array2[i]); 6063b9b85e73SStefano Zampini set_error = PETSC_TRUE; 6064019a44ceSStefano 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); 6065f34684f1SStefano Zampini } 6066f34684f1SStefano Zampini } 6067019a44ceSStefano Zampini ierr = VecRestoreArray(pcis->vec2_N,&array2);CHKERRQ(ierr); 6068b9b85e73SStefano Zampini ierr = MPI_Allreduce(&set_error,&set_error_reduced,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 6069f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 6070f34684f1SStefano Zampini for (i=0;i<pcis->n;i++) { 6071f34684f1SStefano Zampini if (PetscRealPart(array[i]) > 0.0) array[i] = 1.0/PetscRealPart(array[i]); 6072f34684f1SStefano Zampini } 6073f34684f1SStefano Zampini ierr = VecRestoreArray(pcis->vec1_N,&array);CHKERRQ(ierr); 6074f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 6075e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6076e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6077f34684f1SStefano Zampini ierr = VecSum(pcis->vec1_global,&coarsesum);CHKERRQ(ierr); 6078f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Size of coarse problem is %d (%lf)\n",coarse_size,PetscRealPart(coarsesum));CHKERRQ(ierr); 6079b9b85e73SStefano Zampini if (pcbddc->dbg_flag > 1 || set_error_reduced) { 6080ca8b9ea9SStefano Zampini PetscInt *gidxs; 6081ca8b9ea9SStefano Zampini 6082ca8b9ea9SStefano Zampini ierr = PetscMalloc1(pcbddc->local_primal_size,&gidxs);CHKERRQ(ierr); 60833bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingApply(pcis->mapping,pcbddc->local_primal_size,pcbddc->primal_indices_local_idxs,gidxs);CHKERRQ(ierr); 6084f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Distribution of local primal indices\n");CHKERRQ(ierr); 6085f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 6086f34684f1SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d\n",PetscGlobalRank);CHKERRQ(ierr); 6087f34684f1SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 60884bc2dc4bSStefano 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); 6089f34684f1SStefano Zampini } 6090f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 6091ca8b9ea9SStefano Zampini ierr = PetscFree(gidxs);CHKERRQ(ierr); 6092f34684f1SStefano Zampini } 6093f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 60941575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 6095302440fdSBarry Smith if (set_error_reduced) SETERRQ(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"BDDC Numbering of coarse dofs failed"); 6096f34684f1SStefano Zampini } 60978bec7fa6SStefano Zampini /* ierr = PetscPrintf(PetscObjectComm((PetscObject)pc),"Size of coarse problem is %d\n",coarse_size);CHKERRQ(ierr); */ 6098f34684f1SStefano Zampini /* get back data */ 6099f34684f1SStefano Zampini *coarse_size_n = coarse_size; 6100f34684f1SStefano Zampini *local_primal_indices_n = local_primal_indices; 6101674ae819SStefano Zampini PetscFunctionReturn(0); 6102674ae819SStefano Zampini } 6103674ae819SStefano Zampini 6104e456f2a8SStefano Zampini #undef __FUNCT__ 6105e456f2a8SStefano Zampini #define __FUNCT__ "PCBDDCGlobalToLocal" 6106a7dc3881SStefano Zampini PetscErrorCode PCBDDCGlobalToLocal(VecScatter g2l_ctx,Vec gwork, Vec lwork, IS globalis, IS* localis) 6107e456f2a8SStefano Zampini { 6108e456f2a8SStefano Zampini IS localis_t; 6109a7dc3881SStefano Zampini PetscInt i,lsize,*idxs,n; 6110e456f2a8SStefano Zampini PetscScalar *vals; 6111e456f2a8SStefano Zampini PetscErrorCode ierr; 6112e456f2a8SStefano Zampini 6113e456f2a8SStefano Zampini PetscFunctionBegin; 6114a7dc3881SStefano Zampini /* get indices in local ordering exploiting local to global map */ 6115e456f2a8SStefano Zampini ierr = ISGetLocalSize(globalis,&lsize);CHKERRQ(ierr); 6116854ce69bSBarry Smith ierr = PetscMalloc1(lsize,&vals);CHKERRQ(ierr); 6117e456f2a8SStefano Zampini for (i=0;i<lsize;i++) vals[i] = 1.0; 6118e456f2a8SStefano Zampini ierr = ISGetIndices(globalis,(const PetscInt**)&idxs);CHKERRQ(ierr); 6119a7dc3881SStefano Zampini ierr = VecSet(gwork,0.0);CHKERRQ(ierr); 6120a7dc3881SStefano Zampini ierr = VecSet(lwork,0.0);CHKERRQ(ierr); 61211035eff8SStefano Zampini if (idxs) { /* multilevel guard */ 6122a7dc3881SStefano Zampini ierr = VecSetValues(gwork,lsize,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 61231035eff8SStefano Zampini } 6124a7dc3881SStefano Zampini ierr = VecAssemblyBegin(gwork);CHKERRQ(ierr); 6125e456f2a8SStefano Zampini ierr = ISRestoreIndices(globalis,(const PetscInt**)&idxs);CHKERRQ(ierr); 6126e456f2a8SStefano Zampini ierr = PetscFree(vals);CHKERRQ(ierr); 6127a7dc3881SStefano Zampini ierr = VecAssemblyEnd(gwork);CHKERRQ(ierr); 6128a7dc3881SStefano Zampini /* now compute set in local ordering */ 6129a7dc3881SStefano Zampini ierr = VecScatterBegin(g2l_ctx,gwork,lwork,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6130a7dc3881SStefano Zampini ierr = VecScatterEnd(g2l_ctx,gwork,lwork,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6131a7dc3881SStefano Zampini ierr = VecGetArrayRead(lwork,(const PetscScalar**)&vals);CHKERRQ(ierr); 6132a7dc3881SStefano Zampini ierr = VecGetSize(lwork,&n);CHKERRQ(ierr); 6133a7dc3881SStefano Zampini for (i=0,lsize=0;i<n;i++) { 6134ff92baa0SMatthew G. Knepley if (PetscRealPart(vals[i]) > 0.5) { 6135e456f2a8SStefano Zampini lsize++; 6136e456f2a8SStefano Zampini } 6137e456f2a8SStefano Zampini } 6138854ce69bSBarry Smith ierr = PetscMalloc1(lsize,&idxs);CHKERRQ(ierr); 6139a7dc3881SStefano Zampini for (i=0,lsize=0;i<n;i++) { 6140ff92baa0SMatthew G. Knepley if (PetscRealPart(vals[i]) > 0.5) { 6141e456f2a8SStefano Zampini idxs[lsize++] = i; 6142e456f2a8SStefano Zampini } 6143e456f2a8SStefano Zampini } 6144a7dc3881SStefano Zampini ierr = VecRestoreArrayRead(lwork,(const PetscScalar**)&vals);CHKERRQ(ierr); 6145a7dc3881SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)gwork),lsize,idxs,PETSC_OWN_POINTER,&localis_t);CHKERRQ(ierr); 6146e456f2a8SStefano Zampini *localis = localis_t; 6147e456f2a8SStefano Zampini PetscFunctionReturn(0); 6148e456f2a8SStefano Zampini } 6149906d46d4SStefano Zampini 6150906d46d4SStefano Zampini /* the next two functions will be called in KSPMatMult if a change of basis has been requested */ 6151906d46d4SStefano Zampini #undef __FUNCT__ 6152906d46d4SStefano Zampini #define __FUNCT__ "PCBDDCMatMult_Private" 6153906d46d4SStefano Zampini static PetscErrorCode PCBDDCMatMult_Private(Mat A, Vec x, Vec y) 6154906d46d4SStefano Zampini { 6155906d46d4SStefano Zampini PCBDDCChange_ctx change_ctx; 6156906d46d4SStefano Zampini PetscErrorCode ierr; 6157906d46d4SStefano Zampini 6158906d46d4SStefano Zampini PetscFunctionBegin; 6159906d46d4SStefano Zampini ierr = MatShellGetContext(A,&change_ctx);CHKERRQ(ierr); 6160906d46d4SStefano Zampini ierr = MatMult(change_ctx->global_change,x,change_ctx->work[0]);CHKERRQ(ierr); 6161906d46d4SStefano Zampini ierr = MatMult(change_ctx->original_mat,change_ctx->work[0],change_ctx->work[1]);CHKERRQ(ierr); 6162906d46d4SStefano Zampini ierr = MatMultTranspose(change_ctx->global_change,change_ctx->work[1],y);CHKERRQ(ierr); 6163906d46d4SStefano Zampini PetscFunctionReturn(0); 6164906d46d4SStefano Zampini } 6165906d46d4SStefano Zampini 6166906d46d4SStefano Zampini #undef __FUNCT__ 6167906d46d4SStefano Zampini #define __FUNCT__ "PCBDDCMatMultTranspose_Private" 6168906d46d4SStefano Zampini static PetscErrorCode PCBDDCMatMultTranspose_Private(Mat A, Vec x, Vec y) 6169906d46d4SStefano Zampini { 6170906d46d4SStefano Zampini PCBDDCChange_ctx change_ctx; 6171906d46d4SStefano Zampini PetscErrorCode ierr; 6172906d46d4SStefano Zampini 6173906d46d4SStefano Zampini PetscFunctionBegin; 6174906d46d4SStefano Zampini ierr = MatShellGetContext(A,&change_ctx);CHKERRQ(ierr); 6175906d46d4SStefano Zampini ierr = MatMult(change_ctx->global_change,x,change_ctx->work[0]);CHKERRQ(ierr); 6176906d46d4SStefano Zampini ierr = MatMultTranspose(change_ctx->original_mat,change_ctx->work[0],change_ctx->work[1]);CHKERRQ(ierr); 6177906d46d4SStefano Zampini ierr = MatMultTranspose(change_ctx->global_change,change_ctx->work[1],y);CHKERRQ(ierr); 6178906d46d4SStefano Zampini PetscFunctionReturn(0); 6179906d46d4SStefano Zampini } 6180b96c3477SStefano Zampini 6181b96c3477SStefano Zampini #undef __FUNCT__ 6182b96c3477SStefano Zampini #define __FUNCT__ "PCBDDCSetUpSubSchurs" 618308122e43SStefano Zampini PetscErrorCode PCBDDCSetUpSubSchurs(PC pc) 6184b96c3477SStefano Zampini { 6185a64f4aa4SStefano Zampini PC_IS *pcis=(PC_IS*)pc->data; 6186b96c3477SStefano Zampini PC_BDDC *pcbddc=(PC_BDDC*)pc->data; 6187b96c3477SStefano Zampini PCBDDCSubSchurs sub_schurs=pcbddc->sub_schurs; 6188a64f4aa4SStefano Zampini Mat S_j; 6189b96c3477SStefano Zampini PetscInt *used_xadj,*used_adjncy; 6190b96c3477SStefano Zampini PetscBool free_used_adj; 6191b96c3477SStefano Zampini PetscErrorCode ierr; 6192b96c3477SStefano Zampini 6193b96c3477SStefano Zampini PetscFunctionBegin; 6194b96c3477SStefano Zampini /* decide the adjacency to be used for determining internal problems for local schur on subsets */ 6195b96c3477SStefano Zampini free_used_adj = PETSC_FALSE; 619608122e43SStefano Zampini if (pcbddc->sub_schurs_layers == -1) { 6197b96c3477SStefano Zampini used_xadj = NULL; 6198b96c3477SStefano Zampini used_adjncy = NULL; 6199b96c3477SStefano Zampini } else { 620008122e43SStefano Zampini if (pcbddc->sub_schurs_use_useradj && pcbddc->mat_graph->xadj) { 620108122e43SStefano Zampini used_xadj = pcbddc->mat_graph->xadj; 620208122e43SStefano Zampini used_adjncy = pcbddc->mat_graph->adjncy; 620308122e43SStefano Zampini } else if (pcbddc->computed_rowadj) { 6204b96c3477SStefano Zampini used_xadj = pcbddc->mat_graph->xadj; 6205b96c3477SStefano Zampini used_adjncy = pcbddc->mat_graph->adjncy; 6206b96c3477SStefano Zampini } else { 62072fffb893SStefano Zampini PetscBool flg_row=PETSC_FALSE; 6208b96c3477SStefano Zampini const PetscInt *xadj,*adjncy; 6209b96c3477SStefano Zampini PetscInt nvtxs; 6210b96c3477SStefano Zampini 62112fffb893SStefano Zampini ierr = MatGetRowIJ(pcbddc->local_mat,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 62122fffb893SStefano Zampini if (flg_row) { 6213b96c3477SStefano Zampini ierr = PetscMalloc2(nvtxs+1,&used_xadj,xadj[nvtxs],&used_adjncy);CHKERRQ(ierr); 6214b96c3477SStefano Zampini ierr = PetscMemcpy(used_xadj,xadj,(nvtxs+1)*sizeof(*xadj));CHKERRQ(ierr); 6215b96c3477SStefano Zampini ierr = PetscMemcpy(used_adjncy,adjncy,(xadj[nvtxs])*sizeof(*adjncy));CHKERRQ(ierr); 6216b96c3477SStefano Zampini free_used_adj = PETSC_TRUE; 62172fffb893SStefano Zampini } else { 62182fffb893SStefano Zampini pcbddc->sub_schurs_layers = -1; 62192fffb893SStefano Zampini used_xadj = NULL; 62202fffb893SStefano Zampini used_adjncy = NULL; 62212fffb893SStefano Zampini } 62222fffb893SStefano Zampini ierr = MatRestoreRowIJ(pcbddc->local_mat,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 6223b96c3477SStefano Zampini } 6224b96c3477SStefano Zampini } 6225d5574798SStefano Zampini 6226d5574798SStefano Zampini /* setup sub_schurs data */ 6227a64f4aa4SStefano Zampini ierr = MatCreateSchurComplement(pcis->A_II,pcis->A_II,pcis->A_IB,pcis->A_BI,pcis->A_BB,&S_j);CHKERRQ(ierr); 6228df4d28bfSStefano Zampini if (!sub_schurs->schur_explicit) { 6229df4d28bfSStefano Zampini /* pcbddc->ksp_D up to date only if not using MatFactor with Schur complement support */ 6230a64f4aa4SStefano Zampini ierr = MatSchurComplementSetKSP(S_j,pcbddc->ksp_D);CHKERRQ(ierr); 623172b8c272SStefano 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); 6232a64f4aa4SStefano Zampini } else { 62336816873aSStefano Zampini PetscBool reuse_solvers = (PetscBool)!pcbddc->use_change_of_basis; 623404708bb6SStefano Zampini PetscBool isseqaij; 6235a3df083aSStefano Zampini PetscInt benign_n; 623672b8c272SStefano Zampini Mat change = NULL; 623772b8c272SStefano Zampini IS change_primal = NULL; 6238a3df083aSStefano Zampini 62395feab87aSStefano Zampini if (!pcbddc->use_vertices && reuse_solvers) { 62405feab87aSStefano Zampini PetscInt n_vertices; 62415feab87aSStefano Zampini 62425feab87aSStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_vertices,&n_vertices);CHKERRQ(ierr); 62432034aafcSStefano Zampini reuse_solvers = (PetscBool)!n_vertices; 62445feab87aSStefano Zampini } 624504708bb6SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcbddc->local_mat,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 624604708bb6SStefano Zampini if (!isseqaij) { 624704708bb6SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 624804708bb6SStefano Zampini if (matis->A == pcbddc->local_mat) { 624904708bb6SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 625004708bb6SStefano Zampini ierr = MatConvert(matis->A,MATSEQAIJ,MAT_INITIAL_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 625104708bb6SStefano Zampini } else { 625204708bb6SStefano Zampini ierr = MatConvert(pcbddc->local_mat,MATSEQAIJ,MAT_REUSE_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 625304708bb6SStefano Zampini } 625404708bb6SStefano Zampini } 6255a3df083aSStefano Zampini if (!pcbddc->benign_change_explicit) { 6256a3df083aSStefano Zampini benign_n = pcbddc->benign_n; 6257ca92afb2SStefano Zampini } else { 6258a3df083aSStefano Zampini benign_n = 0; 6259ca92afb2SStefano Zampini } 6260bf3a8328SStefano Zampini /* If the user defines additional constraints, we import them here. 6261bf3a8328SStefano 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 */ 626272b8c272SStefano Zampini if (pcbddc->adaptive_userdefined || (pcbddc->deluxe_zerorows && !pcbddc->use_change_of_basis)) { 626388c03ad3SStefano Zampini PC_IS *pcisf; 626488c03ad3SStefano Zampini PC_BDDC *pcbddcf; 626588c03ad3SStefano Zampini PC pcf; 626688c03ad3SStefano Zampini 626788c03ad3SStefano Zampini if (pcbddc->sub_schurs_rebuild) { 626888c03ad3SStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Cannot compute change of basis with a different graph"); 626988c03ad3SStefano Zampini } 627088c03ad3SStefano Zampini ierr = PCCreate(PetscObjectComm((PetscObject)pc),&pcf);CHKERRQ(ierr); 627188c03ad3SStefano Zampini ierr = PCSetOperators(pcf,pc->mat,pc->pmat);CHKERRQ(ierr); 627288c03ad3SStefano Zampini ierr = PCSetType(pcf,PCBDDC);CHKERRQ(ierr); 627388c03ad3SStefano Zampini /* hacks */ 627488c03ad3SStefano Zampini pcisf = (PC_IS*)pcf->data; 627572b8c272SStefano Zampini pcisf->is_B_local = pcis->is_B_local; 627672b8c272SStefano Zampini pcisf->vec1_N = pcis->vec1_N; 627772b8c272SStefano Zampini pcisf->BtoNmap = pcis->BtoNmap; 627872b8c272SStefano Zampini pcisf->n = pcis->n; 627972b8c272SStefano Zampini pcisf->n_B = pcis->n_B; 628088c03ad3SStefano Zampini pcbddcf = (PC_BDDC*)pcf->data; 628188c03ad3SStefano Zampini ierr = PetscFree(pcbddcf->mat_graph);CHKERRQ(ierr); 628288c03ad3SStefano Zampini pcbddcf->mat_graph = pcbddc->mat_graph; 628388c03ad3SStefano Zampini pcbddcf->use_faces = PETSC_TRUE; 628488c03ad3SStefano Zampini pcbddcf->use_change_of_basis = PETSC_TRUE; 628588c03ad3SStefano Zampini pcbddcf->use_change_on_faces = PETSC_TRUE; 628672b8c272SStefano Zampini pcbddcf->use_qr_single = PETSC_TRUE; 628788c03ad3SStefano Zampini pcbddcf->fake_change = PETSC_TRUE; 628888c03ad3SStefano Zampini ierr = PCBDDCConstraintsSetUp(pcf);CHKERRQ(ierr); 628972b8c272SStefano Zampini /* store information on primal vertices and change of basis (in local numbering) */ 629072b8c272SStefano Zampini sub_schurs->change_with_qr = pcbddcf->use_qr_single; 629172b8c272SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,pcbddcf->n_vertices,pcbddcf->local_primal_ref_node,PETSC_COPY_VALUES,&change_primal);CHKERRQ(ierr); 629272b8c272SStefano Zampini change = pcbddcf->ConstraintMatrix; 629372b8c272SStefano Zampini pcbddcf->ConstraintMatrix = NULL; 629488c03ad3SStefano Zampini /* free unneeded memory allocated in PCBDDCConstraintsSetUp */ 629572b8c272SStefano Zampini ierr = PetscFree(pcbddcf->sub_schurs);CHKERRQ(ierr); 629688c03ad3SStefano Zampini ierr = MatNullSpaceDestroy(&pcbddcf->onearnullspace);CHKERRQ(ierr); 629788c03ad3SStefano Zampini ierr = PetscFree2(pcbddcf->local_primal_ref_node,pcbddcf->local_primal_ref_mult);CHKERRQ(ierr); 629888c03ad3SStefano Zampini ierr = PetscFree(pcbddcf->primal_indices_local_idxs);CHKERRQ(ierr); 629988c03ad3SStefano Zampini ierr = PetscFree(pcbddcf->onearnullvecs_state);CHKERRQ(ierr); 630088c03ad3SStefano Zampini ierr = PetscFree(pcf->data);CHKERRQ(ierr); 630188c03ad3SStefano Zampini pcf->ops->destroy = NULL; 630288c03ad3SStefano Zampini ierr = PCDestroy(&pcf);CHKERRQ(ierr); 630388c03ad3SStefano Zampini } 630472b8c272SStefano 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); 630572b8c272SStefano Zampini ierr = MatDestroy(&change);CHKERRQ(ierr); 630672b8c272SStefano Zampini ierr = ISDestroy(&change_primal);CHKERRQ(ierr); 6307ca92afb2SStefano Zampini } 6308d12d3064SStefano Zampini ierr = MatDestroy(&S_j);CHKERRQ(ierr); 6309b96c3477SStefano Zampini 6310b96c3477SStefano Zampini /* free adjacency */ 6311b96c3477SStefano Zampini if (free_used_adj) { 6312b96c3477SStefano Zampini ierr = PetscFree2(used_xadj,used_adjncy);CHKERRQ(ierr); 6313b96c3477SStefano Zampini } 6314b96c3477SStefano Zampini PetscFunctionReturn(0); 6315b96c3477SStefano Zampini } 6316b96c3477SStefano Zampini 6317b96c3477SStefano Zampini #undef __FUNCT__ 6318b96c3477SStefano Zampini #define __FUNCT__ "PCBDDCInitSubSchurs" 631908122e43SStefano Zampini PetscErrorCode PCBDDCInitSubSchurs(PC pc) 6320b96c3477SStefano Zampini { 6321b96c3477SStefano Zampini PC_IS *pcis=(PC_IS*)pc->data; 6322b96c3477SStefano Zampini PC_BDDC *pcbddc=(PC_BDDC*)pc->data; 6323b96c3477SStefano Zampini PCBDDCSubSchurs sub_schurs=pcbddc->sub_schurs; 6324b96c3477SStefano Zampini PCBDDCGraph graph; 6325b96c3477SStefano Zampini PetscErrorCode ierr; 6326b96c3477SStefano Zampini 6327b96c3477SStefano Zampini PetscFunctionBegin; 6328b96c3477SStefano Zampini /* attach interface graph for determining subsets */ 632908122e43SStefano Zampini if (pcbddc->sub_schurs_rebuild) { /* in case rebuild has been requested, it uses a graph generated only by the neighbouring information */ 63303301b35fSStefano Zampini IS verticesIS,verticescomm; 63313301b35fSStefano Zampini PetscInt vsize,*idxs; 6332b96c3477SStefano Zampini 6333b96c3477SStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,NULL,NULL,NULL,NULL,&verticesIS);CHKERRQ(ierr); 63343301b35fSStefano Zampini ierr = ISGetSize(verticesIS,&vsize);CHKERRQ(ierr); 63353301b35fSStefano Zampini ierr = ISGetIndices(verticesIS,(const PetscInt**)&idxs);CHKERRQ(ierr); 63363301b35fSStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),vsize,idxs,PETSC_COPY_VALUES,&verticescomm);CHKERRQ(ierr); 63373301b35fSStefano Zampini ierr = ISRestoreIndices(verticesIS,(const PetscInt**)&idxs);CHKERRQ(ierr); 63383301b35fSStefano Zampini ierr = ISDestroy(&verticesIS);CHKERRQ(ierr); 6339b96c3477SStefano Zampini ierr = PCBDDCGraphCreate(&graph);CHKERRQ(ierr); 63407fb0e2dbSStefano Zampini ierr = PCBDDCGraphInit(graph,pcbddc->mat_graph->l2gmap,pcbddc->mat_graph->nvtxs_global);CHKERRQ(ierr); 63413301b35fSStefano Zampini ierr = PCBDDCGraphSetUp(graph,0,NULL,pcbddc->DirichletBoundariesLocal,0,NULL,verticescomm);CHKERRQ(ierr); 63423301b35fSStefano Zampini ierr = ISDestroy(&verticescomm);CHKERRQ(ierr); 6343b96c3477SStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(graph);CHKERRQ(ierr); 6344b96c3477SStefano Zampini /* 6345b96c3477SStefano Zampini if (pcbddc->dbg_flag) { 6346b96c3477SStefano Zampini ierr = PCBDDCGraphASCIIView(graph,pcbddc->dbg_flag,pcbddc->dbg_viewer);CHKERRQ(ierr); 6347b96c3477SStefano Zampini } 6348b96c3477SStefano Zampini */ 6349b96c3477SStefano Zampini } else { 6350b96c3477SStefano Zampini graph = pcbddc->mat_graph; 6351b96c3477SStefano Zampini } 6352b96c3477SStefano Zampini 6353b96c3477SStefano Zampini /* sub_schurs init */ 6354a64f4aa4SStefano Zampini ierr = PCBDDCSubSchursInit(sub_schurs,pcis->is_I_local,pcis->is_B_local,graph,pcis->BtoNmap);CHKERRQ(ierr); 6355a64f4aa4SStefano Zampini 6356b96c3477SStefano Zampini /* free graph struct */ 635708122e43SStefano Zampini if (pcbddc->sub_schurs_rebuild) { 6358b96c3477SStefano Zampini ierr = PCBDDCGraphDestroy(&graph);CHKERRQ(ierr); 6359b96c3477SStefano Zampini } 6360b96c3477SStefano Zampini PetscFunctionReturn(0); 6361b96c3477SStefano Zampini } 6362fa34dd3eSStefano Zampini 6363fa34dd3eSStefano Zampini #undef __FUNCT__ 6364fa34dd3eSStefano Zampini #define __FUNCT__ "PCBDDCCheckOperator" 6365fa34dd3eSStefano Zampini PetscErrorCode PCBDDCCheckOperator(PC pc) 6366fa34dd3eSStefano Zampini { 6367fa34dd3eSStefano Zampini PC_IS *pcis=(PC_IS*)pc->data; 6368fa34dd3eSStefano Zampini PC_BDDC *pcbddc=(PC_BDDC*)pc->data; 6369fa34dd3eSStefano Zampini PetscErrorCode ierr; 6370fa34dd3eSStefano Zampini 6371fa34dd3eSStefano Zampini PetscFunctionBegin; 6372fa34dd3eSStefano Zampini if (pcbddc->n_vertices == pcbddc->local_primal_size) { 6373fa34dd3eSStefano Zampini IS zerodiag = NULL; 63744f1b2e48SStefano Zampini Mat S_j,B0_B=NULL; 6375fa34dd3eSStefano Zampini Vec dummy_vec=NULL,vec_check_B,vec_scale_P; 63764f1b2e48SStefano Zampini PetscScalar *p0_check,*array,*array2; 637775c01103SStefano Zampini PetscReal norm; 6378fa34dd3eSStefano Zampini PetscInt i; 6379fa34dd3eSStefano Zampini 6380fa34dd3eSStefano Zampini /* B0 and B0_B */ 6381fa34dd3eSStefano Zampini if (zerodiag) { 6382fa34dd3eSStefano Zampini IS dummy; 6383fa34dd3eSStefano Zampini 63844f1b2e48SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcbddc->benign_n,0,1,&dummy);CHKERRQ(ierr); 63854f1b2e48SStefano Zampini ierr = MatGetSubMatrix(pcbddc->benign_B0,dummy,pcis->is_B_local,MAT_INITIAL_MATRIX,&B0_B);CHKERRQ(ierr); 6386fa34dd3eSStefano Zampini ierr = MatCreateVecs(B0_B,NULL,&dummy_vec);CHKERRQ(ierr); 6387fa34dd3eSStefano Zampini ierr = ISDestroy(&dummy);CHKERRQ(ierr); 6388fa34dd3eSStefano Zampini } 6389fa34dd3eSStefano Zampini /* I need a primal vector to scale primal nodes since BDDC sums contibutions */ 6390fa34dd3eSStefano Zampini ierr = VecDuplicate(pcbddc->vec1_P,&vec_scale_P);CHKERRQ(ierr); 6391fa34dd3eSStefano Zampini ierr = VecSet(pcbddc->vec1_P,1.0);CHKERRQ(ierr); 6392fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6393fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6394fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,vec_scale_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6395fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,vec_scale_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6396fa34dd3eSStefano Zampini ierr = VecReciprocal(vec_scale_P);CHKERRQ(ierr); 6397fa34dd3eSStefano Zampini /* S_j */ 6398fa34dd3eSStefano Zampini ierr = MatCreateSchurComplement(pcis->A_II,pcis->A_II,pcis->A_IB,pcis->A_BI,pcis->A_BB,&S_j);CHKERRQ(ierr); 6399fa34dd3eSStefano Zampini ierr = MatSchurComplementSetKSP(S_j,pcbddc->ksp_D);CHKERRQ(ierr); 6400fa34dd3eSStefano Zampini 6401fa34dd3eSStefano Zampini /* mimic vector in \widetilde{W}_\Gamma */ 6402fa34dd3eSStefano Zampini ierr = VecSetRandom(pcis->vec1_N,NULL);CHKERRQ(ierr); 6403fa34dd3eSStefano Zampini /* continuous in primal space */ 6404fa34dd3eSStefano Zampini ierr = VecSetRandom(pcbddc->coarse_vec,NULL);CHKERRQ(ierr); 6405fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,pcbddc->vec1_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6406fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,pcbddc->vec1_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6407fa34dd3eSStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 64084f1b2e48SStefano Zampini ierr = PetscCalloc1(pcbddc->benign_n,&p0_check);CHKERRQ(ierr); 64094f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) p0_check[i] = array[pcbddc->local_primal_size-pcbddc->benign_n+i]; 6410fa34dd3eSStefano Zampini ierr = VecSetValues(pcis->vec1_N,pcbddc->local_primal_size,pcbddc->local_primal_ref_node,array,INSERT_VALUES);CHKERRQ(ierr); 6411fa34dd3eSStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 6412fa34dd3eSStefano Zampini ierr = VecAssemblyBegin(pcis->vec1_N);CHKERRQ(ierr); 6413fa34dd3eSStefano Zampini ierr = VecAssemblyEnd(pcis->vec1_N);CHKERRQ(ierr); 6414fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcis->N_to_B,pcis->vec1_N,pcis->vec2_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6415fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcis->N_to_B,pcis->vec1_N,pcis->vec2_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6416fa34dd3eSStefano Zampini ierr = VecDuplicate(pcis->vec2_B,&vec_check_B);CHKERRQ(ierr); 6417fa34dd3eSStefano Zampini ierr = VecCopy(pcis->vec2_B,vec_check_B);CHKERRQ(ierr); 6418fa34dd3eSStefano Zampini 6419fa34dd3eSStefano Zampini /* assemble rhs for coarse problem */ 6420fa34dd3eSStefano Zampini /* widetilde{S}_\Gamma w_\Gamma + \widetilde{B0}^T_B p0 */ 6421fa34dd3eSStefano Zampini /* local with Schur */ 6422fa34dd3eSStefano Zampini ierr = MatMult(S_j,pcis->vec2_B,pcis->vec1_B);CHKERRQ(ierr); 6423fa34dd3eSStefano Zampini if (zerodiag) { 6424fa34dd3eSStefano Zampini ierr = VecGetArray(dummy_vec,&array);CHKERRQ(ierr); 64254f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) array[i] = p0_check[i]; 6426fa34dd3eSStefano Zampini ierr = VecRestoreArray(dummy_vec,&array);CHKERRQ(ierr); 6427fa34dd3eSStefano Zampini ierr = MatMultTransposeAdd(B0_B,dummy_vec,pcis->vec1_B,pcis->vec1_B);CHKERRQ(ierr); 6428fa34dd3eSStefano Zampini } 6429fa34dd3eSStefano Zampini /* sum on primal nodes the local contributions */ 6430fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcis->N_to_B,pcis->vec1_B,pcis->vec1_N,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6431fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcis->N_to_B,pcis->vec1_B,pcis->vec1_N,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6432fa34dd3eSStefano Zampini ierr = VecGetArray(pcis->vec1_N,&array);CHKERRQ(ierr); 6433fa34dd3eSStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array2);CHKERRQ(ierr); 6434fa34dd3eSStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) array2[i] = array[pcbddc->local_primal_ref_node[i]]; 6435fa34dd3eSStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array2);CHKERRQ(ierr); 6436fa34dd3eSStefano Zampini ierr = VecRestoreArray(pcis->vec1_N,&array);CHKERRQ(ierr); 6437fa34dd3eSStefano Zampini ierr = VecSet(pcbddc->coarse_vec,0.);CHKERRQ(ierr); 6438fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6439fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6440fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,pcbddc->vec1_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6441fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,pcbddc->vec1_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6442fa34dd3eSStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 6443fa34dd3eSStefano Zampini /* scale primal nodes (BDDC sums contibutions) */ 6444fa34dd3eSStefano Zampini ierr = VecPointwiseMult(pcbddc->vec1_P,vec_scale_P,pcbddc->vec1_P);CHKERRQ(ierr); 6445fa34dd3eSStefano Zampini ierr = VecSetValues(pcis->vec1_N,pcbddc->local_primal_size,pcbddc->local_primal_ref_node,array,INSERT_VALUES);CHKERRQ(ierr); 6446fa34dd3eSStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 6447fa34dd3eSStefano Zampini ierr = VecAssemblyBegin(pcis->vec1_N);CHKERRQ(ierr); 6448fa34dd3eSStefano Zampini ierr = VecAssemblyEnd(pcis->vec1_N);CHKERRQ(ierr); 6449fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcis->N_to_B,pcis->vec1_N,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6450fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcis->N_to_B,pcis->vec1_N,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6451fa34dd3eSStefano Zampini /* global: \widetilde{B0}_B w_\Gamma */ 6452fa34dd3eSStefano Zampini if (zerodiag) { 6453fa34dd3eSStefano Zampini ierr = MatMult(B0_B,pcis->vec2_B,dummy_vec);CHKERRQ(ierr); 6454fa34dd3eSStefano Zampini ierr = VecGetArray(dummy_vec,&array);CHKERRQ(ierr); 64554f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) pcbddc->benign_p0[i] = array[i]; 6456fa34dd3eSStefano Zampini ierr = VecRestoreArray(dummy_vec,&array);CHKERRQ(ierr); 6457fa34dd3eSStefano Zampini } 6458fa34dd3eSStefano Zampini /* BDDC */ 6459fa34dd3eSStefano Zampini ierr = VecSet(pcis->vec1_D,0.);CHKERRQ(ierr); 6460fa34dd3eSStefano Zampini ierr = PCBDDCApplyInterfacePreconditioner(pc,PETSC_FALSE);CHKERRQ(ierr); 6461fa34dd3eSStefano Zampini 6462fa34dd3eSStefano Zampini ierr = VecCopy(pcis->vec1_B,pcis->vec2_B);CHKERRQ(ierr); 6463fa34dd3eSStefano Zampini ierr = VecAXPY(pcis->vec1_B,-1.0,vec_check_B);CHKERRQ(ierr); 6464fa34dd3eSStefano Zampini ierr = VecNorm(pcis->vec1_B,NORM_INFINITY,&norm);CHKERRQ(ierr); 6465fa34dd3eSStefano Zampini PetscPrintf(PETSC_COMM_SELF,"[%d] BDDC local error is %1.4e\n",PetscGlobalRank,norm); 64664f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 64674f1b2e48SStefano Zampini PetscPrintf(PETSC_COMM_SELF,"[%d] BDDC p0[%d] error is %1.4e\n",PetscGlobalRank,i,PetscAbsScalar(pcbddc->benign_p0[i]-p0_check[i])); 6468fa34dd3eSStefano Zampini } 64694f1b2e48SStefano Zampini ierr = PetscFree(p0_check);CHKERRQ(ierr); 6470fa34dd3eSStefano Zampini ierr = VecDestroy(&vec_scale_P);CHKERRQ(ierr); 6471fa34dd3eSStefano Zampini ierr = VecDestroy(&vec_check_B);CHKERRQ(ierr); 6472fa34dd3eSStefano Zampini ierr = VecDestroy(&dummy_vec);CHKERRQ(ierr); 6473fa34dd3eSStefano Zampini ierr = MatDestroy(&S_j);CHKERRQ(ierr); 6474fa34dd3eSStefano Zampini ierr = MatDestroy(&B0_B);CHKERRQ(ierr); 6475fa34dd3eSStefano Zampini } 6476fa34dd3eSStefano Zampini PetscFunctionReturn(0); 6477fa34dd3eSStefano Zampini } 6478