11cf9b237SStefano Zampini #include <../src/mat/impls/aij/seq/aij.h> 2ab5c6b0cSJed Brown #include <../src/ksp/pc/impls/bddc/bddc.h> 3ab5c6b0cSJed Brown #include <../src/ksp/pc/impls/bddc/bddcprivate.h> 4674ae819SStefano Zampini #include <petscblaslapack.h> 5674ae819SStefano Zampini 6906d46d4SStefano Zampini static PetscErrorCode PCBDDCMatMultTranspose_Private(Mat A, Vec x, Vec y); 7906d46d4SStefano Zampini static PetscErrorCode PCBDDCMatMult_Private(Mat A, Vec x, Vec y); 8906d46d4SStefano Zampini 9a3df083aSStefano Zampini #undef __FUNCT__ 10a3df083aSStefano Zampini #define __FUNCT__ "PCBDDCBenignMatMult_Private_Private" 11a3df083aSStefano Zampini PetscErrorCode PCBDDCBenignMatMult_Private_Private(Mat A, Vec x, Vec y, PetscBool transpose) 12a3df083aSStefano Zampini { 13a3df083aSStefano Zampini PCBDDCBenignMatMult_ctx ctx; 14a3df083aSStefano Zampini PetscErrorCode ierr; 15a3df083aSStefano Zampini PetscBool apply_right,apply_left,reset_x; 16a3df083aSStefano Zampini 17a3df083aSStefano Zampini PetscFunctionBegin; 18a3df083aSStefano Zampini ierr = MatShellGetContext(A,&ctx);CHKERRQ(ierr); 19a3df083aSStefano Zampini if (transpose) { 20a3df083aSStefano Zampini apply_right = ctx->apply_left; 21a3df083aSStefano Zampini apply_left = ctx->apply_right; 22a3df083aSStefano Zampini } else { 23a3df083aSStefano Zampini apply_right = ctx->apply_right; 24a3df083aSStefano Zampini apply_left = ctx->apply_left; 25a3df083aSStefano Zampini } 26a3df083aSStefano Zampini reset_x = PETSC_FALSE; 27a3df083aSStefano Zampini if (apply_right) { 28a3df083aSStefano Zampini const PetscScalar *ax; 29a3df083aSStefano Zampini PetscInt nl,i; 30a3df083aSStefano Zampini 31a3df083aSStefano Zampini ierr = VecGetLocalSize(x,&nl);CHKERRQ(ierr); 32a3df083aSStefano Zampini ierr = VecGetArrayRead(x,&ax);CHKERRQ(ierr); 33a3df083aSStefano Zampini ierr = PetscMemcpy(ctx->work,ax,nl*sizeof(PetscScalar));CHKERRQ(ierr); 34a3df083aSStefano Zampini ierr = VecRestoreArrayRead(x,&ax);CHKERRQ(ierr); 35a3df083aSStefano Zampini for (i=0;i<ctx->benign_n;i++) { 36a3df083aSStefano Zampini PetscScalar sum,val; 37a3df083aSStefano Zampini const PetscInt *idxs; 38a3df083aSStefano Zampini PetscInt nz,j; 39a3df083aSStefano Zampini ierr = ISGetLocalSize(ctx->benign_zerodiag_subs[i],&nz);CHKERRQ(ierr); 40a3df083aSStefano Zampini ierr = ISGetIndices(ctx->benign_zerodiag_subs[i],&idxs);CHKERRQ(ierr); 41a3df083aSStefano Zampini sum = 0.; 42a3df083aSStefano Zampini if (ctx->apply_p0) { 43a3df083aSStefano Zampini val = ctx->work[idxs[nz-1]]; 44a3df083aSStefano Zampini for (j=0;j<nz-1;j++) { 45a3df083aSStefano Zampini sum += ctx->work[idxs[j]]; 46a3df083aSStefano Zampini ctx->work[idxs[j]] += val; 47a3df083aSStefano Zampini } 48a3df083aSStefano Zampini } else { 49a3df083aSStefano Zampini for (j=0;j<nz-1;j++) { 50a3df083aSStefano Zampini sum += ctx->work[idxs[j]]; 51a3df083aSStefano Zampini } 52a3df083aSStefano Zampini } 53a3df083aSStefano Zampini ctx->work[idxs[nz-1]] -= sum; 54a3df083aSStefano Zampini ierr = ISRestoreIndices(ctx->benign_zerodiag_subs[i],&idxs);CHKERRQ(ierr); 55a3df083aSStefano Zampini } 56a3df083aSStefano Zampini ierr = VecPlaceArray(x,ctx->work);CHKERRQ(ierr); 57a3df083aSStefano Zampini reset_x = PETSC_TRUE; 58a3df083aSStefano Zampini } 59a3df083aSStefano Zampini if (transpose) { 60a3df083aSStefano Zampini ierr = MatMultTranspose(ctx->A,x,y);CHKERRQ(ierr); 61a3df083aSStefano Zampini } else { 62a3df083aSStefano Zampini ierr = MatMult(ctx->A,x,y);CHKERRQ(ierr); 63a3df083aSStefano Zampini } 64a3df083aSStefano Zampini if (reset_x) { 65a3df083aSStefano Zampini ierr = VecResetArray(x);CHKERRQ(ierr); 66a3df083aSStefano Zampini } 67a3df083aSStefano Zampini if (apply_left) { 68a3df083aSStefano Zampini PetscScalar *ay; 69a3df083aSStefano Zampini PetscInt i; 70a3df083aSStefano Zampini 71a3df083aSStefano Zampini ierr = VecGetArray(y,&ay);CHKERRQ(ierr); 72a3df083aSStefano Zampini for (i=0;i<ctx->benign_n;i++) { 73a3df083aSStefano Zampini PetscScalar sum,val; 74a3df083aSStefano Zampini const PetscInt *idxs; 75a3df083aSStefano Zampini PetscInt nz,j; 76a3df083aSStefano Zampini ierr = ISGetLocalSize(ctx->benign_zerodiag_subs[i],&nz);CHKERRQ(ierr); 77a3df083aSStefano Zampini ierr = ISGetIndices(ctx->benign_zerodiag_subs[i],&idxs);CHKERRQ(ierr); 78a3df083aSStefano Zampini val = -ay[idxs[nz-1]]; 79a3df083aSStefano Zampini if (ctx->apply_p0) { 80a3df083aSStefano Zampini sum = 0.; 81a3df083aSStefano Zampini for (j=0;j<nz-1;j++) { 82a3df083aSStefano Zampini sum += ay[idxs[j]]; 83a3df083aSStefano Zampini ay[idxs[j]] += val; 84a3df083aSStefano Zampini } 85a3df083aSStefano Zampini ay[idxs[nz-1]] += sum; 86a3df083aSStefano Zampini } else { 87a3df083aSStefano Zampini for (j=0;j<nz-1;j++) { 88a3df083aSStefano Zampini ay[idxs[j]] += val; 89a3df083aSStefano Zampini } 90a3df083aSStefano Zampini ay[idxs[nz-1]] = 0.; 91a3df083aSStefano Zampini } 92a3df083aSStefano Zampini ierr = ISRestoreIndices(ctx->benign_zerodiag_subs[i],&idxs);CHKERRQ(ierr); 93a3df083aSStefano Zampini } 94a3df083aSStefano Zampini ierr = VecRestoreArray(y,&ay);CHKERRQ(ierr); 95a3df083aSStefano Zampini } 96a3df083aSStefano Zampini PetscFunctionReturn(0); 97a3df083aSStefano Zampini } 98a3df083aSStefano Zampini 99a3df083aSStefano Zampini #undef __FUNCT__ 100a3df083aSStefano Zampini #define __FUNCT__ "PCBDDCBenignMatMultTranspose_Private" 101a3df083aSStefano Zampini PetscErrorCode PCBDDCBenignMatMultTranspose_Private(Mat A, Vec x, Vec y) 102a3df083aSStefano Zampini { 103a3df083aSStefano Zampini PetscErrorCode ierr; 104a3df083aSStefano Zampini 105a3df083aSStefano Zampini PetscFunctionBegin; 106a3df083aSStefano Zampini ierr = PCBDDCBenignMatMult_Private_Private(A,x,y,PETSC_TRUE);CHKERRQ(ierr); 107a3df083aSStefano Zampini PetscFunctionReturn(0); 108a3df083aSStefano Zampini } 109a3df083aSStefano Zampini 110a3df083aSStefano Zampini #undef __FUNCT__ 111a3df083aSStefano Zampini #define __FUNCT__ "PCBDDCBenignMatMult_Private" 112a3df083aSStefano Zampini PetscErrorCode PCBDDCBenignMatMult_Private(Mat A, Vec x, Vec y) 113a3df083aSStefano Zampini { 114a3df083aSStefano Zampini PetscErrorCode ierr; 115a3df083aSStefano Zampini 116a3df083aSStefano Zampini PetscFunctionBegin; 117a3df083aSStefano Zampini ierr = PCBDDCBenignMatMult_Private_Private(A,x,y,PETSC_FALSE);CHKERRQ(ierr); 118a3df083aSStefano Zampini PetscFunctionReturn(0); 119a3df083aSStefano Zampini } 120a3df083aSStefano Zampini 121a3df083aSStefano Zampini #undef __FUNCT__ 122a3df083aSStefano Zampini #define __FUNCT__ "PCBDDCBenignShellMat" 123a3df083aSStefano Zampini PetscErrorCode PCBDDCBenignShellMat(PC pc, PetscBool restore) 124a3df083aSStefano Zampini { 125a3df083aSStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 126a3df083aSStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 127a3df083aSStefano Zampini Mat_IS *matis = (Mat_IS*)pc->pmat->data; 128a3df083aSStefano Zampini PCBDDCBenignMatMult_ctx ctx; 129a3df083aSStefano Zampini PetscErrorCode ierr; 130a3df083aSStefano Zampini 131a3df083aSStefano Zampini PetscFunctionBegin; 132a3df083aSStefano Zampini if (!restore) { 133a3df083aSStefano Zampini Mat A,A_IB,A_BI; 134a3df083aSStefano Zampini PetscScalar *work; 135df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse = pcbddc->sub_schurs->reuse_solver; 136a3df083aSStefano Zampini 137a3df083aSStefano Zampini /* local mat */ 138a3df083aSStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&A);CHKERRQ(ierr); 139a3df083aSStefano Zampini ierr = MatSetSizes(A,pcis->n,pcis->n,PETSC_DECIDE,PETSC_DECIDE);CHKERRQ(ierr); 140a3df083aSStefano Zampini ierr = MatSetType(A,MATSHELL);CHKERRQ(ierr); 141a3df083aSStefano Zampini ierr = MatShellSetOperation(A,MATOP_MULT,(void (*)(void))PCBDDCBenignMatMult_Private);CHKERRQ(ierr); 142a3df083aSStefano Zampini ierr = MatShellSetOperation(A,MATOP_MULT_TRANSPOSE,(void (*)(void))PCBDDCBenignMatMultTranspose_Private);CHKERRQ(ierr); 143a3df083aSStefano Zampini ierr = PetscNew(&ctx);CHKERRQ(ierr); 144a3df083aSStefano Zampini ierr = MatShellSetContext(A,ctx);CHKERRQ(ierr); 145a3df083aSStefano Zampini ctx->apply_left = PETSC_TRUE; 146a3df083aSStefano Zampini ctx->apply_right = PETSC_TRUE; 147a3df083aSStefano Zampini ctx->apply_p0 = PETSC_TRUE; 148a3df083aSStefano Zampini ctx->benign_n = pcbddc->benign_n; 149a3df083aSStefano Zampini ctx->benign_zerodiag_subs = pcbddc->benign_zerodiag_subs; 150a3df083aSStefano Zampini ctx->A = matis->A; 151a3df083aSStefano Zampini ierr = PetscMalloc1(pcis->n,&work);CHKERRQ(ierr); 152a3df083aSStefano Zampini ctx->work = work; 153a3df083aSStefano Zampini ierr = MatSetUp(A);CHKERRQ(ierr); 154a3df083aSStefano Zampini ierr = MatAssemblyBegin(A,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 155a3df083aSStefano Zampini ierr = MatAssemblyEnd(A,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 156a3df083aSStefano Zampini matis->A = A; 157a3df083aSStefano Zampini 158a3df083aSStefano Zampini /* A_IB */ 159a3df083aSStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&A_IB);CHKERRQ(ierr); 160a3df083aSStefano Zampini ierr = MatSetSizes(A_IB,pcis->n-pcis->n_B,pcis->n_B,PETSC_DECIDE,PETSC_DECIDE);CHKERRQ(ierr); 161a3df083aSStefano Zampini ierr = MatSetType(A_IB,MATSHELL);CHKERRQ(ierr); 162a3df083aSStefano Zampini ierr = MatShellSetOperation(A_IB,MATOP_MULT,(void (*)(void))PCBDDCBenignMatMult_Private);CHKERRQ(ierr); 163a3df083aSStefano Zampini ierr = MatShellSetOperation(A_IB,MATOP_MULT_TRANSPOSE,(void (*)(void))PCBDDCBenignMatMultTranspose_Private);CHKERRQ(ierr); 164a3df083aSStefano Zampini ierr = PetscNew(&ctx);CHKERRQ(ierr); 165a3df083aSStefano Zampini ierr = MatShellSetContext(A_IB,ctx);CHKERRQ(ierr); 166a3df083aSStefano Zampini ctx->apply_left = PETSC_TRUE; 167a3df083aSStefano Zampini ctx->apply_right = PETSC_FALSE; 168a3df083aSStefano Zampini ctx->apply_p0 = PETSC_FALSE; 169a3df083aSStefano Zampini ctx->benign_n = pcbddc->benign_n; 170059032f7SStefano Zampini if (reuse) { 171a3df083aSStefano Zampini ctx->benign_zerodiag_subs = reuse->benign_zerodiag_subs; 172059032f7SStefano Zampini } else { /* TODO: could be optimized for successive solves */ 173059032f7SStefano Zampini ISLocalToGlobalMapping N_to_D; 174059032f7SStefano Zampini PetscInt i; 175059032f7SStefano Zampini 176059032f7SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(pcis->is_I_local,&N_to_D);CHKERRQ(ierr); 177059032f7SStefano Zampini ierr = PetscMalloc1(pcbddc->benign_n,&ctx->benign_zerodiag_subs);CHKERRQ(ierr); 178059032f7SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 179059032f7SStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(N_to_D,IS_GTOLM_DROP,pcbddc->benign_zerodiag_subs[i],&ctx->benign_zerodiag_subs[i]);CHKERRQ(ierr); 180059032f7SStefano Zampini } 181059032f7SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&N_to_D);CHKERRQ(ierr); 182059032f7SStefano Zampini } 183a3df083aSStefano Zampini ctx->A = pcis->A_IB; 184a3df083aSStefano Zampini ctx->work = work; 185a3df083aSStefano Zampini ierr = MatSetUp(A_IB);CHKERRQ(ierr); 186a3df083aSStefano Zampini ierr = MatAssemblyBegin(A_IB,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 187a3df083aSStefano Zampini ierr = MatAssemblyEnd(A_IB,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 188a3df083aSStefano Zampini pcis->A_IB = A_IB; 189a3df083aSStefano Zampini 190a3df083aSStefano Zampini /* A_BI as A_IB^T */ 191a3df083aSStefano Zampini ierr = MatCreateTranspose(A_IB,&A_BI);CHKERRQ(ierr); 192a3df083aSStefano Zampini pcbddc->benign_original_mat = pcis->A_BI; 193a3df083aSStefano Zampini pcis->A_BI = A_BI; 194a3df083aSStefano Zampini } else { 195059032f7SStefano Zampini PCBDDCReuseSolvers reuse = pcbddc->sub_schurs->reuse_solver; 196059032f7SStefano Zampini 197a3df083aSStefano Zampini ierr = MatShellGetContext(matis->A,&ctx);CHKERRQ(ierr); 198a3df083aSStefano Zampini ierr = MatDestroy(&matis->A);CHKERRQ(ierr); 199a3df083aSStefano Zampini matis->A = ctx->A; 200a3df083aSStefano Zampini ierr = PetscFree(ctx);CHKERRQ(ierr); 201a3df083aSStefano Zampini ierr = MatShellGetContext(pcis->A_IB,&ctx);CHKERRQ(ierr); 202a3df083aSStefano Zampini ierr = MatDestroy(&pcis->A_IB);CHKERRQ(ierr); 203a3df083aSStefano Zampini pcis->A_IB = ctx->A; 204059032f7SStefano Zampini if (!reuse) { 205059032f7SStefano Zampini PetscInt i; 206059032f7SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 207059032f7SStefano Zampini ierr = ISDestroy(&ctx->benign_zerodiag_subs[i]);CHKERRQ(ierr); 208059032f7SStefano Zampini } 209059032f7SStefano Zampini ierr = PetscFree(ctx->benign_zerodiag_subs);CHKERRQ(ierr); 210059032f7SStefano Zampini } 211a3df083aSStefano Zampini ierr = PetscFree(ctx->work);CHKERRQ(ierr); 212a3df083aSStefano Zampini ierr = PetscFree(ctx);CHKERRQ(ierr); 213a3df083aSStefano Zampini ierr = MatDestroy(&pcis->A_BI);CHKERRQ(ierr); 214a3df083aSStefano Zampini pcis->A_BI = pcbddc->benign_original_mat; 215a3df083aSStefano Zampini pcbddc->benign_original_mat = NULL; 216a3df083aSStefano Zampini } 217a3df083aSStefano Zampini PetscFunctionReturn(0); 218a3df083aSStefano Zampini } 219a3df083aSStefano Zampini 220a3df083aSStefano Zampini /* used just in bddc debug mode */ 221a3df083aSStefano Zampini #undef __FUNCT__ 222a3df083aSStefano Zampini #define __FUNCT__ "PCBDDCBenignProject" 223a3df083aSStefano Zampini PetscErrorCode PCBDDCBenignProject(PC pc, IS is1, IS is2, Mat *B) 224a3df083aSStefano Zampini { 225a3df083aSStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 226a3df083aSStefano Zampini Mat_IS *matis = (Mat_IS*)pc->pmat->data; 227a3df083aSStefano Zampini Mat An; 228a3df083aSStefano Zampini PetscErrorCode ierr; 229a3df083aSStefano Zampini 230a3df083aSStefano Zampini PetscFunctionBegin; 231a3df083aSStefano Zampini ierr = MatPtAP(matis->A,pcbddc->benign_change,MAT_INITIAL_MATRIX,2.0,&An);CHKERRQ(ierr); 232a3df083aSStefano Zampini ierr = MatZeroRowsColumns(An,pcbddc->benign_n,pcbddc->benign_p0_lidx,1.0,NULL,NULL);CHKERRQ(ierr); 233a3df083aSStefano Zampini if (is1) { 234a3df083aSStefano Zampini ierr = MatGetSubMatrix(An,is1,is2,MAT_INITIAL_MATRIX,B);CHKERRQ(ierr); 235a3df083aSStefano Zampini ierr = MatDestroy(&An);CHKERRQ(ierr); 236a3df083aSStefano Zampini } else { 237a3df083aSStefano Zampini *B = An; 238a3df083aSStefano Zampini } 239a3df083aSStefano Zampini PetscFunctionReturn(0); 240a3df083aSStefano Zampini } 241a3df083aSStefano Zampini 2421cf9b237SStefano Zampini /* TODO: add reuse flag */ 2431cf9b237SStefano Zampini #undef __FUNCT__ 2441cf9b237SStefano Zampini #define __FUNCT__ "MatSeqAIJCompress" 2451cf9b237SStefano Zampini PetscErrorCode MatSeqAIJCompress(Mat A, Mat *B) 2461cf9b237SStefano Zampini { 2471cf9b237SStefano Zampini Mat Bt; 2481cf9b237SStefano Zampini PetscScalar *a,*bdata; 2491cf9b237SStefano Zampini const PetscInt *ii,*ij; 2501cf9b237SStefano Zampini PetscInt m,n,i,nnz,*bii,*bij; 2511cf9b237SStefano Zampini PetscBool flg_row; 2521cf9b237SStefano Zampini PetscErrorCode ierr; 2531cf9b237SStefano Zampini 2541cf9b237SStefano Zampini PetscFunctionBegin; 2551cf9b237SStefano Zampini ierr = MatGetSize(A,&n,&m);CHKERRQ(ierr); 2561cf9b237SStefano Zampini ierr = MatGetRowIJ(A,0,PETSC_FALSE,PETSC_FALSE,&n,&ii,&ij,&flg_row);CHKERRQ(ierr); 2571cf9b237SStefano Zampini ierr = MatSeqAIJGetArray(A,&a);CHKERRQ(ierr); 2581cf9b237SStefano Zampini nnz = n; 2591cf9b237SStefano Zampini for (i=0;i<ii[n];i++) { 2601cf9b237SStefano Zampini if (PetscLikely(PetscAbsScalar(a[i]) > PETSC_SMALL)) nnz++; 2611cf9b237SStefano Zampini } 2621cf9b237SStefano Zampini ierr = PetscMalloc1(n+1,&bii);CHKERRQ(ierr); 2631cf9b237SStefano Zampini ierr = PetscMalloc1(nnz,&bij);CHKERRQ(ierr); 2641cf9b237SStefano Zampini ierr = PetscMalloc1(nnz,&bdata);CHKERRQ(ierr); 2651cf9b237SStefano Zampini nnz = 0; 2661cf9b237SStefano Zampini bii[0] = 0; 2671cf9b237SStefano Zampini for (i=0;i<n;i++) { 2681cf9b237SStefano Zampini PetscInt j; 2691cf9b237SStefano Zampini for (j=ii[i];j<ii[i+1];j++) { 2701cf9b237SStefano Zampini PetscScalar entry = a[j]; 2711cf9b237SStefano Zampini if (PetscLikely(PetscAbsScalar(entry) > PETSC_SMALL) || ij[j] == i) { 2721cf9b237SStefano Zampini bij[nnz] = ij[j]; 2731cf9b237SStefano Zampini bdata[nnz] = entry; 2741cf9b237SStefano Zampini nnz++; 2751cf9b237SStefano Zampini } 2761cf9b237SStefano Zampini } 2771cf9b237SStefano Zampini bii[i+1] = nnz; 2781cf9b237SStefano Zampini } 2791cf9b237SStefano Zampini ierr = MatSeqAIJRestoreArray(A,&a);CHKERRQ(ierr); 2801cf9b237SStefano Zampini ierr = MatCreateSeqAIJWithArrays(PetscObjectComm((PetscObject)A),n,m,bii,bij,bdata,&Bt);CHKERRQ(ierr); 2811cf9b237SStefano Zampini ierr = MatRestoreRowIJ(A,0,PETSC_FALSE,PETSC_FALSE,&n,&ii,&ij,&flg_row);CHKERRQ(ierr); 2821cf9b237SStefano Zampini { 2831cf9b237SStefano Zampini Mat_SeqAIJ *b = (Mat_SeqAIJ*)(Bt->data); 2841cf9b237SStefano Zampini b->free_a = PETSC_TRUE; 2851cf9b237SStefano Zampini b->free_ij = PETSC_TRUE; 2861cf9b237SStefano Zampini } 2871cf9b237SStefano Zampini *B = Bt; 2881cf9b237SStefano Zampini PetscFunctionReturn(0); 2891cf9b237SStefano Zampini } 2901cf9b237SStefano Zampini 291674ae819SStefano Zampini #undef __FUNCT__ 2924f1b2e48SStefano Zampini #define __FUNCT__ "MatDetectDisconnectedComponents" 2934f1b2e48SStefano Zampini PetscErrorCode MatDetectDisconnectedComponents(Mat A, PetscBool filter, PetscInt *ncc, IS* cc[]) 2944f1b2e48SStefano Zampini { 2954f1b2e48SStefano Zampini Mat B; 2964f1b2e48SStefano Zampini IS is_dummy,*cc_n; 2974f1b2e48SStefano Zampini ISLocalToGlobalMapping l2gmap_dummy; 2984f1b2e48SStefano Zampini PCBDDCGraph graph; 2994f1b2e48SStefano Zampini PetscInt i,n; 3004f1b2e48SStefano Zampini PetscInt *xadj,*adjncy; 3014f1b2e48SStefano Zampini PetscInt *xadj_filtered,*adjncy_filtered; 3024f1b2e48SStefano Zampini PetscBool flg_row,isseqaij; 3034f1b2e48SStefano Zampini PetscErrorCode ierr; 3044f1b2e48SStefano Zampini 3054f1b2e48SStefano Zampini PetscFunctionBegin; 30663c961adSStefano Zampini if (!A->rmap->N || !A->cmap->N) { 30763c961adSStefano Zampini *ncc = 0; 30863c961adSStefano Zampini *cc = NULL; 30963c961adSStefano Zampini PetscFunctionReturn(0); 31063c961adSStefano Zampini } 3114f1b2e48SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)A,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 3124f1b2e48SStefano Zampini if (!isseqaij && filter) { 3131cf9b237SStefano Zampini PetscBool isseqdense; 3141cf9b237SStefano Zampini 3151cf9b237SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)A,MATSEQDENSE,&isseqdense);CHKERRQ(ierr); 3161cf9b237SStefano Zampini if (!isseqdense) { 3174f1b2e48SStefano Zampini ierr = MatConvert(A,MATSEQAIJ,MAT_INITIAL_MATRIX,&B);CHKERRQ(ierr); 3181cf9b237SStefano Zampini } else { /* TODO: rectangular case and LDA */ 3191cf9b237SStefano Zampini PetscScalar *array; 3201cf9b237SStefano Zampini PetscReal chop=1.e-6; 3211cf9b237SStefano Zampini 3221cf9b237SStefano Zampini ierr = MatDuplicate(A,MAT_COPY_VALUES,&B);CHKERRQ(ierr); 3231cf9b237SStefano Zampini ierr = MatDenseGetArray(B,&array);CHKERRQ(ierr); 3241cf9b237SStefano Zampini ierr = MatGetSize(B,&n,NULL);CHKERRQ(ierr); 3251cf9b237SStefano Zampini for (i=0;i<n;i++) { 3261cf9b237SStefano Zampini PetscInt j; 3271cf9b237SStefano Zampini for (j=i+1;j<n;j++) { 3281cf9b237SStefano Zampini PetscReal thresh = chop*(PetscAbsScalar(array[i*(n+1)])+PetscAbsScalar(array[j*(n+1)])); 3291cf9b237SStefano Zampini if (PetscAbsScalar(array[i*n+j]) < thresh) array[i*n+j] = 0.; 3301cf9b237SStefano Zampini if (PetscAbsScalar(array[j*n+i]) < thresh) array[j*n+i] = 0.; 3311cf9b237SStefano Zampini } 3321cf9b237SStefano Zampini } 3331cf9b237SStefano Zampini ierr = MatDenseRestoreArray(B,&array);CHKERRQ(ierr); 3341cf9b237SStefano Zampini ierr = MatConvert(B,MATSEQAIJ,MAT_REUSE_MATRIX,&B);CHKERRQ(ierr); 3351cf9b237SStefano Zampini } 3364f1b2e48SStefano Zampini } else { 3374f1b2e48SStefano Zampini B = A; 3384f1b2e48SStefano Zampini } 3394f1b2e48SStefano Zampini ierr = MatGetRowIJ(B,0,PETSC_TRUE,PETSC_FALSE,&n,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 3404f1b2e48SStefano Zampini 3414f1b2e48SStefano Zampini /* if filter is true, then removes entries lower than PETSC_SMALL in magnitude */ 3424f1b2e48SStefano Zampini if (filter) { 3434f1b2e48SStefano Zampini PetscScalar *data; 3444f1b2e48SStefano Zampini PetscInt j,cum; 3454f1b2e48SStefano Zampini 3464f1b2e48SStefano Zampini ierr = PetscCalloc2(n+1,&xadj_filtered,xadj[n],&adjncy_filtered);CHKERRQ(ierr); 3474f1b2e48SStefano Zampini ierr = MatSeqAIJGetArray(B,&data);CHKERRQ(ierr); 3484f1b2e48SStefano Zampini cum = 0; 3494f1b2e48SStefano Zampini for (i=0;i<n;i++) { 3504f1b2e48SStefano Zampini PetscInt t; 3514f1b2e48SStefano Zampini 3524f1b2e48SStefano Zampini for (j=xadj[i];j<xadj[i+1];j++) { 3534f1b2e48SStefano Zampini if (PetscUnlikely(PetscAbsScalar(data[j]) < PETSC_SMALL)) { 3544f1b2e48SStefano Zampini continue; 3554f1b2e48SStefano Zampini } 3564f1b2e48SStefano Zampini adjncy_filtered[cum+xadj_filtered[i]++] = adjncy[j]; 3574f1b2e48SStefano Zampini } 3584f1b2e48SStefano Zampini t = xadj_filtered[i]; 3594f1b2e48SStefano Zampini xadj_filtered[i] = cum; 3604f1b2e48SStefano Zampini cum += t; 3614f1b2e48SStefano Zampini } 3624f1b2e48SStefano Zampini ierr = MatSeqAIJRestoreArray(B,&data);CHKERRQ(ierr); 3634f1b2e48SStefano Zampini } else { 3644f1b2e48SStefano Zampini xadj_filtered = NULL; 3654f1b2e48SStefano Zampini adjncy_filtered = NULL; 3664f1b2e48SStefano Zampini } 3674f1b2e48SStefano Zampini 3684f1b2e48SStefano Zampini /* compute local connected components using PCBDDCGraph */ 3694f1b2e48SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,n,0,1,&is_dummy);CHKERRQ(ierr); 3704f1b2e48SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(is_dummy,&l2gmap_dummy);CHKERRQ(ierr); 3714f1b2e48SStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 3724f1b2e48SStefano Zampini ierr = PCBDDCGraphCreate(&graph);CHKERRQ(ierr); 3734f1b2e48SStefano Zampini ierr = PCBDDCGraphInit(graph,l2gmap_dummy,n);CHKERRQ(ierr); 3744f1b2e48SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&l2gmap_dummy);CHKERRQ(ierr); 3754f1b2e48SStefano Zampini if (xadj_filtered) { 3764f1b2e48SStefano Zampini graph->xadj = xadj_filtered; 3774f1b2e48SStefano Zampini graph->adjncy = adjncy_filtered; 3784f1b2e48SStefano Zampini } else { 3794f1b2e48SStefano Zampini graph->xadj = xadj; 3804f1b2e48SStefano Zampini graph->adjncy = adjncy; 3814f1b2e48SStefano Zampini } 3824f1b2e48SStefano Zampini ierr = PCBDDCGraphSetUp(graph,1,NULL,NULL,0,NULL,NULL);CHKERRQ(ierr); 3834f1b2e48SStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(graph);CHKERRQ(ierr); 3844f1b2e48SStefano Zampini /* partial clean up */ 3854f1b2e48SStefano Zampini ierr = PetscFree2(xadj_filtered,adjncy_filtered);CHKERRQ(ierr); 3864f1b2e48SStefano Zampini ierr = MatRestoreRowIJ(B,0,PETSC_TRUE,PETSC_FALSE,&n,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 3871cf9b237SStefano Zampini if (A != B) { 3884f1b2e48SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 3894f1b2e48SStefano Zampini } 3904f1b2e48SStefano Zampini 3914f1b2e48SStefano Zampini /* get back data */ 3921cf9b237SStefano Zampini if (ncc) *ncc = graph->ncc; 3931cf9b237SStefano Zampini if (cc) { 3944f1b2e48SStefano Zampini ierr = PetscMalloc1(graph->ncc,&cc_n);CHKERRQ(ierr); 3954f1b2e48SStefano Zampini for (i=0;i<graph->ncc;i++) { 3964f1b2e48SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,graph->cptr[i+1]-graph->cptr[i],graph->queue+graph->cptr[i],PETSC_COPY_VALUES,&cc_n[i]);CHKERRQ(ierr); 3974f1b2e48SStefano Zampini } 3984f1b2e48SStefano Zampini *cc = cc_n; 3991cf9b237SStefano Zampini } 4004f1b2e48SStefano Zampini /* clean up graph */ 4014f1b2e48SStefano Zampini graph->xadj = 0; 4024f1b2e48SStefano Zampini graph->adjncy = 0; 4034f1b2e48SStefano Zampini ierr = PCBDDCGraphDestroy(&graph);CHKERRQ(ierr); 4044f1b2e48SStefano Zampini PetscFunctionReturn(0); 4054f1b2e48SStefano Zampini } 4064f1b2e48SStefano Zampini 4074f1b2e48SStefano Zampini #undef __FUNCT__ 4085408967cSStefano Zampini #define __FUNCT__ "PCBDDCBenignCheck" 4095408967cSStefano Zampini PetscErrorCode PCBDDCBenignCheck(PC pc, IS zerodiag) 4105408967cSStefano Zampini { 4115408967cSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 4125408967cSStefano Zampini PC_IS* pcis = (PC_IS*)(pc->data); 413dee84bffSStefano Zampini IS dirIS = NULL; 4144f1b2e48SStefano Zampini PetscInt i; 4155408967cSStefano Zampini PetscErrorCode ierr; 4165408967cSStefano Zampini 4175408967cSStefano Zampini PetscFunctionBegin; 418dee84bffSStefano Zampini ierr = PCBDDCGraphGetDirichletDofs(pcbddc->mat_graph,&dirIS);CHKERRQ(ierr); 4195408967cSStefano Zampini if (zerodiag) { 4205408967cSStefano Zampini Mat A; 4215408967cSStefano Zampini Vec vec3_N; 4225408967cSStefano Zampini PetscScalar *vals; 4235408967cSStefano Zampini const PetscInt *idxs; 424d12d3064SStefano Zampini PetscInt nz,*count; 4255408967cSStefano Zampini 4265408967cSStefano Zampini /* p0 */ 4275408967cSStefano Zampini ierr = VecSet(pcis->vec1_N,0.);CHKERRQ(ierr); 4285408967cSStefano Zampini ierr = PetscMalloc1(pcis->n,&vals);CHKERRQ(ierr); 4295408967cSStefano Zampini ierr = ISGetLocalSize(zerodiag,&nz);CHKERRQ(ierr); 4305408967cSStefano Zampini ierr = ISGetIndices(zerodiag,&idxs);CHKERRQ(ierr); 4314f1b2e48SStefano Zampini for (i=0;i<nz;i++) vals[i] = 1.; 4325408967cSStefano Zampini ierr = VecSetValues(pcis->vec1_N,nz,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 4335408967cSStefano Zampini ierr = VecAssemblyBegin(pcis->vec1_N);CHKERRQ(ierr); 4345408967cSStefano Zampini ierr = VecAssemblyEnd(pcis->vec1_N);CHKERRQ(ierr); 4355408967cSStefano Zampini /* v_I */ 4365408967cSStefano Zampini ierr = VecSetRandom(pcis->vec2_N,NULL);CHKERRQ(ierr); 4375408967cSStefano Zampini for (i=0;i<nz;i++) vals[i] = 0.; 4385408967cSStefano Zampini ierr = VecSetValues(pcis->vec2_N,nz,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 4395408967cSStefano Zampini ierr = ISRestoreIndices(zerodiag,&idxs);CHKERRQ(ierr); 4405408967cSStefano Zampini ierr = ISGetIndices(pcis->is_B_local,&idxs);CHKERRQ(ierr); 4415408967cSStefano Zampini for (i=0;i<pcis->n_B;i++) vals[i] = 0.; 4425408967cSStefano Zampini ierr = VecSetValues(pcis->vec2_N,pcis->n_B,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 4435408967cSStefano Zampini ierr = ISRestoreIndices(pcis->is_B_local,&idxs);CHKERRQ(ierr); 4445408967cSStefano Zampini if (dirIS) { 4455408967cSStefano Zampini PetscInt n; 4465408967cSStefano Zampini 4475408967cSStefano Zampini ierr = ISGetLocalSize(dirIS,&n);CHKERRQ(ierr); 4485408967cSStefano Zampini ierr = ISGetIndices(dirIS,&idxs);CHKERRQ(ierr); 4495408967cSStefano Zampini for (i=0;i<n;i++) vals[i] = 0.; 4505408967cSStefano Zampini ierr = VecSetValues(pcis->vec2_N,n,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 4515408967cSStefano Zampini ierr = ISRestoreIndices(dirIS,&idxs);CHKERRQ(ierr); 4525408967cSStefano Zampini } 4535408967cSStefano Zampini ierr = VecAssemblyBegin(pcis->vec2_N);CHKERRQ(ierr); 4545408967cSStefano Zampini ierr = VecAssemblyEnd(pcis->vec2_N);CHKERRQ(ierr); 4555408967cSStefano Zampini ierr = VecDuplicate(pcis->vec1_N,&vec3_N);CHKERRQ(ierr); 4565408967cSStefano Zampini ierr = VecSet(vec3_N,0.);CHKERRQ(ierr); 4575408967cSStefano Zampini ierr = MatISGetLocalMat(pc->mat,&A);CHKERRQ(ierr); 4585408967cSStefano Zampini ierr = MatMult(A,pcis->vec1_N,vec3_N);CHKERRQ(ierr); 4595408967cSStefano Zampini ierr = VecDot(vec3_N,pcis->vec2_N,&vals[0]);CHKERRQ(ierr); 4604f1b2e48SStefano Zampini if (PetscAbsScalar(vals[0]) > PETSC_SMALL) { 4615408967cSStefano Zampini SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_SUP,"Benign trick can not be applied! b(v_I,p_0) = %f (should be numerically 0.)",PetscAbsScalar(vals[0])); 4625408967cSStefano Zampini } 4635408967cSStefano Zampini ierr = PetscFree(vals);CHKERRQ(ierr); 4645408967cSStefano Zampini ierr = VecDestroy(&vec3_N);CHKERRQ(ierr); 465d12d3064SStefano Zampini 466d12d3064SStefano Zampini /* there should not be any pressure dofs lying on the interface */ 467d12d3064SStefano Zampini ierr = PetscCalloc1(pcis->n,&count);CHKERRQ(ierr); 468d12d3064SStefano Zampini ierr = ISGetIndices(pcis->is_B_local,&idxs);CHKERRQ(ierr); 469d12d3064SStefano Zampini for (i=0;i<pcis->n_B;i++) count[idxs[i]]++; 470d12d3064SStefano Zampini ierr = ISRestoreIndices(pcis->is_B_local,&idxs);CHKERRQ(ierr); 471d12d3064SStefano Zampini ierr = ISGetIndices(zerodiag,&idxs);CHKERRQ(ierr); 472d4d8cf7bSStefano Zampini for (i=0;i<nz;i++) { 473d12d3064SStefano Zampini if (count[idxs[i]]) { 474d12d3064SStefano Zampini SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_SUP,"Benign trick can not be applied! pressure dof %d is an interface dof",idxs[i]); 475d12d3064SStefano Zampini } 476d4d8cf7bSStefano Zampini } 477d12d3064SStefano Zampini ierr = ISRestoreIndices(zerodiag,&idxs);CHKERRQ(ierr); 478d12d3064SStefano Zampini ierr = PetscFree(count);CHKERRQ(ierr); 4795408967cSStefano Zampini } 480dee84bffSStefano Zampini ierr = ISDestroy(&dirIS);CHKERRQ(ierr); 4815408967cSStefano Zampini 4825408967cSStefano Zampini /* check PCBDDCBenignGetOrSetP0 */ 4835408967cSStefano Zampini ierr = VecSetRandom(pcis->vec1_global,NULL);CHKERRQ(ierr); 4844f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) pcbddc->benign_p0[i] = -PetscGlobalRank-i; 4855408967cSStefano Zampini ierr = PCBDDCBenignGetOrSetP0(pc,pcis->vec1_global,PETSC_FALSE);CHKERRQ(ierr); 4864f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) pcbddc->benign_p0[i] = 1; 4875408967cSStefano Zampini ierr = PCBDDCBenignGetOrSetP0(pc,pcis->vec1_global,PETSC_TRUE);CHKERRQ(ierr); 4884f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 4894f1b2e48SStefano Zampini if ((PetscInt)PetscRealPart(pcbddc->benign_p0[i]) != -PetscGlobalRank-i) { 4904f1b2e48SStefano Zampini SETERRQ3(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Error testing PCBDDCBenignGetOrSetP0! Found %1.4e at %d instead of %1.4e\n",pcbddc->benign_p0[i],i,-PetscGlobalRank-i);CHKERRQ(ierr); 4914f1b2e48SStefano Zampini } 4925408967cSStefano Zampini } 4935408967cSStefano Zampini PetscFunctionReturn(0); 4945408967cSStefano Zampini } 4955408967cSStefano Zampini 4965408967cSStefano Zampini #undef __FUNCT__ 497339f8db1SStefano Zampini #define __FUNCT__ "PCBDDCBenignDetectSaddlePoint" 498339f8db1SStefano Zampini PetscErrorCode PCBDDCBenignDetectSaddlePoint(PC pc, IS *zerodiaglocal) 499339f8db1SStefano Zampini { 500339f8db1SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 5014f1b2e48SStefano Zampini IS pressures,zerodiag,*zerodiag_subs; 502b0f5fe93SStefano Zampini PetscInt nz,n; 5034f1b2e48SStefano Zampini PetscBool sorted,have_null,has_null_pressures,recompute_zerodiag; 504339f8db1SStefano Zampini PetscErrorCode ierr; 505339f8db1SStefano Zampini 506339f8db1SStefano Zampini PetscFunctionBegin; 5079f47a83aSStefano Zampini ierr = PetscSFDestroy(&pcbddc->benign_sf);CHKERRQ(ierr); 5089f47a83aSStefano Zampini ierr = MatDestroy(&pcbddc->benign_B0);CHKERRQ(ierr); 509339f8db1SStefano Zampini ierr = MatDestroy(&pcbddc->benign_original_mat);CHKERRQ(ierr); 510a3df083aSStefano Zampini for (n=0;n<pcbddc->benign_n;n++) { 511a3df083aSStefano Zampini ierr = ISDestroy(&pcbddc->benign_zerodiag_subs[n]);CHKERRQ(ierr); 512a3df083aSStefano Zampini } 513a3df083aSStefano Zampini ierr = PetscFree(pcbddc->benign_zerodiag_subs);CHKERRQ(ierr); 514a3df083aSStefano Zampini pcbddc->benign_n = 0; 5154f1b2e48SStefano Zampini /* if a local info on dofs is present, assumes the last field is represented by "pressures" 5164f1b2e48SStefano Zampini otherwise, it uses only zerodiagonal dofs (ok if the pressure block is all zero; it could fail if it is not) 5174f1b2e48SStefano Zampini Checks if all the pressure dofs in each subdomain have a zero diagonal 5184f1b2e48SStefano Zampini If not, a change of basis on pressures is not needed 5194f1b2e48SStefano Zampini since the local Schur complements are SPD 5204f1b2e48SStefano Zampini */ 5214f1b2e48SStefano Zampini has_null_pressures = PETSC_TRUE; 5224f1b2e48SStefano Zampini have_null = PETSC_TRUE; 52340fa8d13SStefano Zampini if (pcbddc->n_ISForDofsLocal) { 5244f1b2e48SStefano Zampini PetscInt npl,*idxs,p = pcbddc->n_ISForDofsLocal-1; 5254f1b2e48SStefano Zampini 5264f1b2e48SStefano Zampini /* Dofs splitting for BDDC cannot have PETSC_COMM_SELF, so create a sequential IS */ 5274f1b2e48SStefano Zampini ierr = ISGetLocalSize(pcbddc->ISForDofsLocal[p],&npl);CHKERRQ(ierr); 5284f1b2e48SStefano Zampini ierr = ISGetIndices(pcbddc->ISForDofsLocal[p],(const PetscInt**)&idxs);CHKERRQ(ierr); 5294f1b2e48SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,npl,idxs,PETSC_COPY_VALUES,&pressures);CHKERRQ(ierr); 530ba14f8e3SStefano Zampini ierr = ISRestoreIndices(pcbddc->ISForDofsLocal[p],(const PetscInt**)&idxs);CHKERRQ(ierr); 53140fa8d13SStefano Zampini ierr = ISSorted(pressures,&sorted);CHKERRQ(ierr); 53240fa8d13SStefano Zampini if (!sorted) { 53340fa8d13SStefano Zampini ierr = ISSort(pressures);CHKERRQ(ierr); 53440fa8d13SStefano Zampini } 53540fa8d13SStefano Zampini } else { 53640fa8d13SStefano Zampini pressures = NULL; 53740fa8d13SStefano Zampini } 53897d764eeSStefano Zampini /* pcis has not been setup yet, so get the local size from the subdomain matrix */ 53997d764eeSStefano Zampini ierr = MatGetLocalSize(pcbddc->local_mat,&n,NULL);CHKERRQ(ierr); 54097d764eeSStefano Zampini ierr = MatFindZeroDiagonals(pcbddc->local_mat,&zerodiag);CHKERRQ(ierr); 541339f8db1SStefano Zampini ierr = ISSorted(zerodiag,&sorted);CHKERRQ(ierr); 542339f8db1SStefano Zampini if (!sorted) { 543339f8db1SStefano Zampini ierr = ISSort(zerodiag);CHKERRQ(ierr); 544339f8db1SStefano Zampini } 545339f8db1SStefano Zampini ierr = ISGetLocalSize(zerodiag,&nz);CHKERRQ(ierr); 5464f1b2e48SStefano Zampini if (!nz) { 5474f1b2e48SStefano Zampini if (n) have_null = PETSC_FALSE; 5484f1b2e48SStefano Zampini has_null_pressures = PETSC_FALSE; 54940fa8d13SStefano Zampini ierr = ISDestroy(&zerodiag);CHKERRQ(ierr); 55040fa8d13SStefano Zampini } 5514f1b2e48SStefano Zampini recompute_zerodiag = PETSC_FALSE; 5524f1b2e48SStefano Zampini /* in case disconnected subdomains info is present, split the pressures accordingly (otherwise the benign trick could fail) */ 5534f1b2e48SStefano Zampini zerodiag_subs = NULL; 5544f1b2e48SStefano Zampini pcbddc->benign_n = 0; 5554f1b2e48SStefano Zampini if (has_null_pressures) { 5564f1b2e48SStefano Zampini IS *subs; 5574f1b2e48SStefano Zampini PetscInt nsubs,i; 5584f1b2e48SStefano Zampini 5594f1b2e48SStefano Zampini subs = pcbddc->local_subs; 5604f1b2e48SStefano Zampini nsubs = pcbddc->n_local_subs; 5614f1b2e48SStefano Zampini if (nsubs > 1) { 5624f1b2e48SStefano Zampini ierr = PetscCalloc1(nsubs,&zerodiag_subs);CHKERRQ(ierr); 5634f1b2e48SStefano Zampini for (i=0;i<nsubs;i++) { 5644f1b2e48SStefano Zampini ISLocalToGlobalMapping l2g; 5654f1b2e48SStefano Zampini IS t_zerodiag_subs; 5664f1b2e48SStefano Zampini PetscInt nl; 5674f1b2e48SStefano Zampini 5684f1b2e48SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(subs[i],&l2g);CHKERRQ(ierr); 5694f1b2e48SStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(l2g,IS_GTOLM_DROP,zerodiag,&t_zerodiag_subs);CHKERRQ(ierr); 5704f1b2e48SStefano Zampini ierr = ISGetLocalSize(t_zerodiag_subs,&nl);CHKERRQ(ierr); 5714f1b2e48SStefano Zampini if (nl) { 5724f1b2e48SStefano Zampini PetscBool valid = PETSC_TRUE; 5734f1b2e48SStefano Zampini 5744f1b2e48SStefano Zampini if (pressures) { 5754f1b2e48SStefano Zampini IS t_pressure_subs; 5764f1b2e48SStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(l2g,IS_GTOLM_DROP,pressures,&t_pressure_subs);CHKERRQ(ierr); 5774f1b2e48SStefano Zampini ierr = ISEqual(t_pressure_subs,t_zerodiag_subs,&valid);CHKERRQ(ierr); 5784f1b2e48SStefano Zampini ierr = ISDestroy(&t_pressure_subs);CHKERRQ(ierr); 5794f1b2e48SStefano Zampini } 5804f1b2e48SStefano Zampini if (valid) { 5814f1b2e48SStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(l2g,t_zerodiag_subs,&zerodiag_subs[pcbddc->benign_n]);CHKERRQ(ierr); 5824f1b2e48SStefano Zampini pcbddc->benign_n++; 5834f1b2e48SStefano Zampini } else { 5844f1b2e48SStefano Zampini recompute_zerodiag = PETSC_TRUE; 5854f1b2e48SStefano Zampini } 5864f1b2e48SStefano Zampini } 5874f1b2e48SStefano Zampini ierr = ISDestroy(&t_zerodiag_subs);CHKERRQ(ierr); 5884f1b2e48SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&l2g);CHKERRQ(ierr); 5894f1b2e48SStefano Zampini } 5904f1b2e48SStefano Zampini } else { /* there's just one subdomain (or zero if they have not been detected */ 5914f1b2e48SStefano Zampini PetscBool valid = PETSC_TRUE; 5924f1b2e48SStefano Zampini if (pressures) { 5934f1b2e48SStefano Zampini ierr = ISEqual(pressures,zerodiag,&valid);CHKERRQ(ierr); 5944f1b2e48SStefano Zampini } 5954f1b2e48SStefano Zampini if (valid) { 5964f1b2e48SStefano Zampini pcbddc->benign_n = 1; 597ca92afb2SStefano Zampini ierr = PetscMalloc1(pcbddc->benign_n,&zerodiag_subs);CHKERRQ(ierr); 5984f1b2e48SStefano Zampini ierr = PetscObjectReference((PetscObject)zerodiag);CHKERRQ(ierr); 5994f1b2e48SStefano Zampini zerodiag_subs[0] = zerodiag; 6004f1b2e48SStefano Zampini } 6014f1b2e48SStefano Zampini } 6024f1b2e48SStefano Zampini } 6034f1b2e48SStefano Zampini 6044f1b2e48SStefano Zampini if (!pcbddc->benign_n) { 6054f1b2e48SStefano Zampini ierr = ISDestroy(&zerodiag);CHKERRQ(ierr); 6064f1b2e48SStefano Zampini recompute_zerodiag = PETSC_FALSE; 6074f1b2e48SStefano Zampini has_null_pressures = PETSC_FALSE; 6084f1b2e48SStefano Zampini have_null = PETSC_FALSE; 6094f1b2e48SStefano Zampini } 6104f1b2e48SStefano Zampini 6114f1b2e48SStefano Zampini /* final check for null pressures */ 6124f1b2e48SStefano Zampini if (zerodiag && pressures) { 6134f1b2e48SStefano Zampini PetscInt nz,np; 6144f1b2e48SStefano Zampini ierr = ISGetLocalSize(zerodiag,&nz);CHKERRQ(ierr); 6154f1b2e48SStefano Zampini ierr = ISGetLocalSize(pressures,&np);CHKERRQ(ierr); 6164f1b2e48SStefano Zampini if (nz != np) have_null = PETSC_FALSE; 6174f1b2e48SStefano Zampini } 6184f1b2e48SStefano Zampini 6194f1b2e48SStefano Zampini if (recompute_zerodiag) { 6204f1b2e48SStefano Zampini ierr = ISDestroy(&zerodiag);CHKERRQ(ierr); 6214f1b2e48SStefano Zampini if (pcbddc->benign_n == 1) { 6224f1b2e48SStefano Zampini ierr = PetscObjectReference((PetscObject)zerodiag_subs[0]);CHKERRQ(ierr); 6234f1b2e48SStefano Zampini zerodiag = zerodiag_subs[0]; 6244f1b2e48SStefano Zampini } else { 6254f1b2e48SStefano Zampini PetscInt i,nzn,*new_idxs; 6264f1b2e48SStefano Zampini 6274f1b2e48SStefano Zampini nzn = 0; 6284f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 6294f1b2e48SStefano Zampini PetscInt ns; 6304f1b2e48SStefano Zampini ierr = ISGetLocalSize(zerodiag_subs[i],&ns);CHKERRQ(ierr); 6314f1b2e48SStefano Zampini nzn += ns; 6324f1b2e48SStefano Zampini } 6334f1b2e48SStefano Zampini ierr = PetscMalloc1(nzn,&new_idxs);CHKERRQ(ierr); 6344f1b2e48SStefano Zampini nzn = 0; 6354f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 6364f1b2e48SStefano Zampini PetscInt ns,*idxs; 6374f1b2e48SStefano Zampini ierr = ISGetLocalSize(zerodiag_subs[i],&ns);CHKERRQ(ierr); 6384f1b2e48SStefano Zampini ierr = ISGetIndices(zerodiag_subs[i],(const PetscInt**)&idxs);CHKERRQ(ierr); 6394f1b2e48SStefano Zampini ierr = PetscMemcpy(new_idxs+nzn,idxs,ns*sizeof(PetscInt));CHKERRQ(ierr); 6404f1b2e48SStefano Zampini ierr = ISRestoreIndices(zerodiag_subs[i],(const PetscInt**)&idxs);CHKERRQ(ierr); 6414f1b2e48SStefano Zampini nzn += ns; 6424f1b2e48SStefano Zampini } 6434f1b2e48SStefano Zampini ierr = PetscSortInt(nzn,new_idxs);CHKERRQ(ierr); 6444f1b2e48SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,nzn,new_idxs,PETSC_OWN_POINTER,&zerodiag);CHKERRQ(ierr); 6454f1b2e48SStefano Zampini } 6464f1b2e48SStefano Zampini have_null = PETSC_FALSE; 6474f1b2e48SStefano Zampini } 6484f1b2e48SStefano Zampini 6494f1b2e48SStefano Zampini if (has_null_pressures) { 6504f1b2e48SStefano Zampini IS zerodiagc; 6514f1b2e48SStefano Zampini const PetscInt *idxs,*idxsc; 6524f1b2e48SStefano Zampini PetscInt i,s,*nnz; 6534f1b2e48SStefano Zampini 6544f1b2e48SStefano Zampini ierr = ISGetLocalSize(zerodiag,&nz);CHKERRQ(ierr); 655339f8db1SStefano Zampini ierr = ISComplement(zerodiag,0,n,&zerodiagc);CHKERRQ(ierr); 656339f8db1SStefano Zampini ierr = ISGetIndices(zerodiagc,&idxsc);CHKERRQ(ierr); 657339f8db1SStefano Zampini /* local change of basis for pressures */ 658339f8db1SStefano Zampini ierr = MatDestroy(&pcbddc->benign_change);CHKERRQ(ierr); 65997d764eeSStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)pcbddc->local_mat),&pcbddc->benign_change);CHKERRQ(ierr); 660339f8db1SStefano Zampini ierr = MatSetType(pcbddc->benign_change,MATAIJ);CHKERRQ(ierr); 661339f8db1SStefano Zampini ierr = MatSetSizes(pcbddc->benign_change,n,n,PETSC_DECIDE,PETSC_DECIDE);CHKERRQ(ierr); 662339f8db1SStefano Zampini ierr = PetscMalloc1(n,&nnz);CHKERRQ(ierr); 6634f1b2e48SStefano Zampini for (i=0;i<n-nz;i++) nnz[idxsc[i]] = 1; /* identity on velocities plus pressure dofs for non-singular subdomains */ 6644f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 6654f1b2e48SStefano Zampini PetscInt nzs,j; 6664f1b2e48SStefano Zampini 6674f1b2e48SStefano Zampini ierr = ISGetLocalSize(zerodiag_subs[i],&nzs);CHKERRQ(ierr); 6684f1b2e48SStefano Zampini ierr = ISGetIndices(zerodiag_subs[i],&idxs);CHKERRQ(ierr); 6694f1b2e48SStefano Zampini for (j=0;j<nzs-1;j++) nnz[idxs[j]] = 2; /* change on pressures */ 6704f1b2e48SStefano Zampini nnz[idxs[nzs-1]] = nzs; /* last local pressure dof in subdomain */ 6714f1b2e48SStefano Zampini ierr = ISRestoreIndices(zerodiag_subs[i],&idxs);CHKERRQ(ierr); 6724f1b2e48SStefano Zampini } 673339f8db1SStefano Zampini ierr = MatSeqAIJSetPreallocation(pcbddc->benign_change,0,nnz);CHKERRQ(ierr); 674339f8db1SStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 675339f8db1SStefano Zampini /* set identity on velocities */ 676339f8db1SStefano Zampini for (i=0;i<n-nz;i++) { 677339f8db1SStefano Zampini ierr = MatSetValue(pcbddc->benign_change,idxsc[i],idxsc[i],1.,INSERT_VALUES);CHKERRQ(ierr); 678339f8db1SStefano Zampini } 6794f1b2e48SStefano Zampini ierr = ISRestoreIndices(zerodiagc,&idxsc);CHKERRQ(ierr); 6804f1b2e48SStefano Zampini ierr = ISDestroy(&zerodiagc);CHKERRQ(ierr); 6819f47a83aSStefano Zampini ierr = PetscFree3(pcbddc->benign_p0_lidx,pcbddc->benign_p0_gidx,pcbddc->benign_p0);CHKERRQ(ierr); 6824f1b2e48SStefano Zampini ierr = PetscMalloc3(pcbddc->benign_n,&pcbddc->benign_p0_lidx,pcbddc->benign_n,&pcbddc->benign_p0_gidx,pcbddc->benign_n,&pcbddc->benign_p0);CHKERRQ(ierr); 683339f8db1SStefano Zampini /* set change on pressures */ 6844f1b2e48SStefano Zampini for (s=0;s<pcbddc->benign_n;s++) { 6854f1b2e48SStefano Zampini PetscScalar *array; 6864f1b2e48SStefano Zampini PetscInt nzs; 6874f1b2e48SStefano Zampini 6884f1b2e48SStefano Zampini ierr = ISGetLocalSize(zerodiag_subs[s],&nzs);CHKERRQ(ierr); 6894f1b2e48SStefano Zampini ierr = ISGetIndices(zerodiag_subs[s],&idxs);CHKERRQ(ierr); 6904f1b2e48SStefano Zampini for (i=0;i<nzs-1;i++) { 691339f8db1SStefano Zampini PetscScalar vals[2]; 692339f8db1SStefano Zampini PetscInt cols[2]; 693339f8db1SStefano Zampini 694339f8db1SStefano Zampini cols[0] = idxs[i]; 6954f1b2e48SStefano Zampini cols[1] = idxs[nzs-1]; 696339f8db1SStefano Zampini vals[0] = 1.; 697b0f5fe93SStefano Zampini vals[1] = 1.; 6984f1b2e48SStefano Zampini ierr = MatSetValues(pcbddc->benign_change,1,cols,2,cols,vals,INSERT_VALUES);CHKERRQ(ierr); 699339f8db1SStefano Zampini } 7004f1b2e48SStefano Zampini ierr = PetscMalloc1(nzs,&array);CHKERRQ(ierr); 7014f1b2e48SStefano Zampini for (i=0;i<nzs-1;i++) array[i] = -1.; 7024f1b2e48SStefano Zampini array[nzs-1] = 1.; 7034f1b2e48SStefano Zampini ierr = MatSetValues(pcbddc->benign_change,1,idxs+nzs-1,nzs,idxs,array,INSERT_VALUES);CHKERRQ(ierr); 7044f1b2e48SStefano Zampini /* store local idxs for p0 */ 7054f1b2e48SStefano Zampini pcbddc->benign_p0_lidx[s] = idxs[nzs-1]; 7064f1b2e48SStefano Zampini ierr = ISRestoreIndices(zerodiag_subs[s],&idxs);CHKERRQ(ierr); 707339f8db1SStefano Zampini ierr = PetscFree(array);CHKERRQ(ierr); 7084f1b2e48SStefano Zampini } 709339f8db1SStefano Zampini ierr = MatAssemblyBegin(pcbddc->benign_change,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 710339f8db1SStefano Zampini ierr = MatAssemblyEnd(pcbddc->benign_change,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 711a3df083aSStefano Zampini /* project if needed */ 712a3df083aSStefano Zampini if (pcbddc->benign_change_explicit) { 71397d764eeSStefano Zampini ierr = MatPtAP(pcbddc->local_mat,pcbddc->benign_change,MAT_INITIAL_MATRIX,2.0,&pcbddc->benign_original_mat);CHKERRQ(ierr); 714339f8db1SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 71597d764eeSStefano Zampini ierr = MatSeqAIJCompress(pcbddc->benign_original_mat,&pcbddc->local_mat);CHKERRQ(ierr); 71697d764eeSStefano Zampini ierr = MatDestroy(&pcbddc->benign_original_mat);CHKERRQ(ierr); 717a3df083aSStefano Zampini } 7184f1b2e48SStefano Zampini /* store global idxs for p0 */ 7194f1b2e48SStefano Zampini ierr = ISLocalToGlobalMappingApply(pc->pmat->rmap->mapping,pcbddc->benign_n,pcbddc->benign_p0_lidx,pcbddc->benign_p0_gidx);CHKERRQ(ierr); 720339f8db1SStefano Zampini } 721ca92afb2SStefano Zampini pcbddc->benign_zerodiag_subs = zerodiag_subs; 7224f1b2e48SStefano Zampini ierr = ISDestroy(&pressures);CHKERRQ(ierr); 723b0f5fe93SStefano Zampini 724b0f5fe93SStefano Zampini /* determines if the coarse solver will be singular or not */ 725b0f5fe93SStefano Zampini ierr = MPI_Allreduce(&have_null,&pcbddc->benign_null,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 726339f8db1SStefano Zampini *zerodiaglocal = zerodiag; 727339f8db1SStefano Zampini PetscFunctionReturn(0); 728339f8db1SStefano Zampini } 729339f8db1SStefano Zampini 730339f8db1SStefano Zampini #undef __FUNCT__ 731015636ebSStefano Zampini #define __FUNCT__ "PCBDDCBenignGetOrSetP0" 732015636ebSStefano Zampini PetscErrorCode PCBDDCBenignGetOrSetP0(PC pc, Vec v, PetscBool get) 733efc2fbd9SStefano Zampini { 734efc2fbd9SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 735efc2fbd9SStefano Zampini PetscErrorCode ierr; 736efc2fbd9SStefano Zampini 737efc2fbd9SStefano Zampini PetscFunctionBegin; 738efc2fbd9SStefano Zampini if (!pcbddc->benign_sf) { 739efc2fbd9SStefano Zampini ierr = PetscSFCreate(PetscObjectComm((PetscObject)pc),&pcbddc->benign_sf);CHKERRQ(ierr); 7404f1b2e48SStefano Zampini ierr = PetscSFSetGraphLayout(pcbddc->benign_sf,pc->pmat->rmap,pcbddc->benign_n,NULL,PETSC_OWN_POINTER,pcbddc->benign_p0_gidx);CHKERRQ(ierr); 741efc2fbd9SStefano Zampini } 742015636ebSStefano Zampini if (get) { /* use SF to get values */ 743efc2fbd9SStefano Zampini PetscScalar *array; 744efc2fbd9SStefano Zampini 745efc2fbd9SStefano Zampini ierr = VecGetArrayRead(v,(const PetscScalar**)&array);CHKERRQ(ierr); 7464f1b2e48SStefano Zampini ierr = PetscSFBcastBegin(pcbddc->benign_sf,MPIU_SCALAR,array,pcbddc->benign_p0);CHKERRQ(ierr); 7474f1b2e48SStefano Zampini ierr = PetscSFBcastEnd(pcbddc->benign_sf,MPIU_SCALAR,array,pcbddc->benign_p0);CHKERRQ(ierr); 748efc2fbd9SStefano Zampini ierr = VecRestoreArrayRead(v,(const PetscScalar**)&array);CHKERRQ(ierr); 7494f1b2e48SStefano Zampini } else { /* use VecSetValues (not scalable, I should try to find a better solution (defining a new MPI_OP for reduction) */ 7504f1b2e48SStefano Zampini ierr = VecSetValues(v,pcbddc->benign_n,pcbddc->benign_p0_gidx,pcbddc->benign_p0,INSERT_VALUES);CHKERRQ(ierr); 751efc2fbd9SStefano Zampini ierr = VecAssemblyBegin(v);CHKERRQ(ierr); 752efc2fbd9SStefano Zampini ierr = VecAssemblyEnd(v);CHKERRQ(ierr); 753efc2fbd9SStefano Zampini } 754efc2fbd9SStefano Zampini PetscFunctionReturn(0); 755efc2fbd9SStefano Zampini } 756efc2fbd9SStefano Zampini 757efc2fbd9SStefano Zampini #undef __FUNCT__ 758c263805aSStefano Zampini #define __FUNCT__ "PCBDDCBenignPopOrPushB0" 759c263805aSStefano Zampini PetscErrorCode PCBDDCBenignPopOrPushB0(PC pc, PetscBool pop) 760c263805aSStefano Zampini { 761c263805aSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 762c263805aSStefano Zampini PetscErrorCode ierr; 763c263805aSStefano Zampini 764c263805aSStefano Zampini PetscFunctionBegin; 765c263805aSStefano Zampini /* TODO: add error checking 766c263805aSStefano Zampini - avoid nested pop (or push) calls. 767c263805aSStefano Zampini - cannot push before pop. 7681c604dc7SStefano Zampini - cannot call this if pcbddc->local_mat is NULL 769c263805aSStefano Zampini */ 7704f1b2e48SStefano Zampini if (!pcbddc->benign_n) { 771efc2fbd9SStefano Zampini PetscFunctionReturn(0); 772efc2fbd9SStefano Zampini } 773c263805aSStefano Zampini if (pop) { 774a3df083aSStefano Zampini if (pcbddc->benign_change_explicit) { 7754f1b2e48SStefano Zampini IS is_p0; 7764f1b2e48SStefano Zampini MatReuse reuse; 777c263805aSStefano Zampini 778c263805aSStefano Zampini /* extract B_0 */ 7794f1b2e48SStefano Zampini reuse = MAT_INITIAL_MATRIX; 7804f1b2e48SStefano Zampini if (pcbddc->benign_B0) { 7814f1b2e48SStefano Zampini reuse = MAT_REUSE_MATRIX; 7824f1b2e48SStefano Zampini } 7834f1b2e48SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,pcbddc->benign_n,pcbddc->benign_p0_lidx,PETSC_COPY_VALUES,&is_p0);CHKERRQ(ierr); 7844f1b2e48SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,is_p0,NULL,reuse,&pcbddc->benign_B0);CHKERRQ(ierr); 785c263805aSStefano Zampini /* remove rows and cols from local problem */ 786c263805aSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_KEEP_NONZERO_PATTERN,PETSC_TRUE);CHKERRQ(ierr); 78797d764eeSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_FALSE);CHKERRQ(ierr); 7884f1b2e48SStefano Zampini ierr = MatZeroRowsColumnsIS(pcbddc->local_mat,is_p0,1.0,NULL,NULL);CHKERRQ(ierr); 7894f1b2e48SStefano Zampini ierr = ISDestroy(&is_p0);CHKERRQ(ierr); 790a3df083aSStefano Zampini } else { 791a3df083aSStefano Zampini Mat_IS *matis = (Mat_IS*)pc->pmat->data; 792a3df083aSStefano Zampini PetscScalar *vals; 793a3df083aSStefano Zampini PetscInt i,n,*idxs_ins; 794a3df083aSStefano Zampini 795a3df083aSStefano Zampini ierr = VecGetLocalSize(matis->y,&n);CHKERRQ(ierr); 796a3df083aSStefano Zampini ierr = PetscMalloc2(n,&idxs_ins,n,&vals);CHKERRQ(ierr); 797a3df083aSStefano Zampini if (!pcbddc->benign_B0) { 7980b5adadeSStefano Zampini PetscInt *nnz; 799a3df083aSStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)pcbddc->local_mat),&pcbddc->benign_B0);CHKERRQ(ierr); 800a3df083aSStefano Zampini ierr = MatSetType(pcbddc->benign_B0,MATAIJ);CHKERRQ(ierr); 801a3df083aSStefano Zampini ierr = MatSetSizes(pcbddc->benign_B0,pcbddc->benign_n,n,PETSC_DECIDE,PETSC_DECIDE);CHKERRQ(ierr); 802331e053bSStefano Zampini ierr = PetscMalloc1(pcbddc->benign_n,&nnz);CHKERRQ(ierr); 803331e053bSStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 804331e053bSStefano Zampini ierr = ISGetLocalSize(pcbddc->benign_zerodiag_subs[i],&nnz[i]);CHKERRQ(ierr); 805331e053bSStefano Zampini nnz[i] = n - nnz[i]; 806331e053bSStefano Zampini } 807331e053bSStefano Zampini ierr = MatSeqAIJSetPreallocation(pcbddc->benign_B0,0,nnz);CHKERRQ(ierr); 808331e053bSStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 809331e053bSStefano Zampini } 810a3df083aSStefano Zampini 811a3df083aSStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 812a3df083aSStefano Zampini PetscScalar *array; 813a3df083aSStefano Zampini PetscInt *idxs,j,nz,cum; 814a3df083aSStefano Zampini 815a3df083aSStefano Zampini ierr = VecSet(matis->x,0.);CHKERRQ(ierr); 816a3df083aSStefano Zampini ierr = ISGetLocalSize(pcbddc->benign_zerodiag_subs[i],&nz);CHKERRQ(ierr); 817a3df083aSStefano Zampini ierr = ISGetIndices(pcbddc->benign_zerodiag_subs[i],(const PetscInt**)&idxs);CHKERRQ(ierr); 818a3df083aSStefano Zampini for (j=0;j<nz;j++) vals[j] = 1.; 819a3df083aSStefano Zampini ierr = VecSetValues(matis->x,nz,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 820a3df083aSStefano Zampini ierr = VecAssemblyBegin(matis->x);CHKERRQ(ierr); 821a3df083aSStefano Zampini ierr = VecAssemblyEnd(matis->x);CHKERRQ(ierr); 822a3df083aSStefano Zampini ierr = VecSet(matis->y,0.);CHKERRQ(ierr); 823a3df083aSStefano Zampini ierr = MatMult(matis->A,matis->x,matis->y);CHKERRQ(ierr); 824a3df083aSStefano Zampini ierr = VecGetArray(matis->y,&array);CHKERRQ(ierr); 825a3df083aSStefano Zampini cum = 0; 826a3df083aSStefano Zampini for (j=0;j<n;j++) { 827a3df083aSStefano Zampini if (PetscUnlikely(PetscAbsReal(array[j]) > PETSC_SMALL)) { 828a3df083aSStefano Zampini vals[cum] = array[j]; 829a3df083aSStefano Zampini idxs_ins[cum] = j; 830a3df083aSStefano Zampini cum++; 831a3df083aSStefano Zampini } 832a3df083aSStefano Zampini } 833a3df083aSStefano Zampini ierr = MatSetValues(pcbddc->benign_B0,1,&i,cum,idxs_ins,vals,INSERT_VALUES);CHKERRQ(ierr); 834a3df083aSStefano Zampini ierr = VecRestoreArray(matis->y,&array);CHKERRQ(ierr); 835a3df083aSStefano Zampini ierr = ISRestoreIndices(pcbddc->benign_zerodiag_subs[i],(const PetscInt**)&idxs);CHKERRQ(ierr); 836a3df083aSStefano Zampini } 837a3df083aSStefano Zampini ierr = MatAssemblyBegin(pcbddc->benign_B0,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 838a3df083aSStefano Zampini ierr = MatAssemblyEnd(pcbddc->benign_B0,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 839a3df083aSStefano Zampini ierr = PetscFree2(idxs_ins,vals);CHKERRQ(ierr); 840a3df083aSStefano Zampini } 841c263805aSStefano Zampini } else { /* push */ 842a3df083aSStefano Zampini if (pcbddc->benign_change_explicit) { 8434f1b2e48SStefano Zampini PetscInt i; 8444f1b2e48SStefano Zampini 8454f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 8464f1b2e48SStefano Zampini PetscScalar *B0_vals; 8474f1b2e48SStefano Zampini PetscInt *B0_cols,B0_ncol; 8484f1b2e48SStefano Zampini 8494f1b2e48SStefano Zampini ierr = MatGetRow(pcbddc->benign_B0,i,&B0_ncol,(const PetscInt**)&B0_cols,(const PetscScalar**)&B0_vals);CHKERRQ(ierr); 8504f1b2e48SStefano Zampini ierr = MatSetValues(pcbddc->local_mat,1,pcbddc->benign_p0_lidx+i,B0_ncol,B0_cols,B0_vals,INSERT_VALUES);CHKERRQ(ierr); 8517b034428SStefano Zampini ierr = MatSetValues(pcbddc->local_mat,B0_ncol,B0_cols,1,pcbddc->benign_p0_lidx+i,B0_vals,INSERT_VALUES);CHKERRQ(ierr); 8524f1b2e48SStefano Zampini ierr = MatSetValue(pcbddc->local_mat,pcbddc->benign_p0_lidx[i],pcbddc->benign_p0_lidx[i],0.0,INSERT_VALUES);CHKERRQ(ierr); 8534f1b2e48SStefano Zampini ierr = MatRestoreRow(pcbddc->benign_B0,i,&B0_ncol,(const PetscInt**)&B0_cols,(const PetscScalar**)&B0_vals);CHKERRQ(ierr); 8544f1b2e48SStefano Zampini } 855c263805aSStefano Zampini ierr = MatAssemblyBegin(pcbddc->local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 856c263805aSStefano Zampini ierr = MatAssemblyEnd(pcbddc->local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 857a3df083aSStefano Zampini } else { 858a3df083aSStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Cannot push B0!\n"); 859a3df083aSStefano Zampini } 860c263805aSStefano Zampini } 861c263805aSStefano Zampini PetscFunctionReturn(0); 862c263805aSStefano Zampini } 863c263805aSStefano Zampini 864c263805aSStefano Zampini #undef __FUNCT__ 865b1b3d7a2SStefano Zampini #define __FUNCT__ "PCBDDCAdaptiveSelection" 86608122e43SStefano Zampini PetscErrorCode PCBDDCAdaptiveSelection(PC pc) 867b1b3d7a2SStefano Zampini { 868b1b3d7a2SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 86908122e43SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 87008122e43SStefano Zampini PetscBLASInt B_dummyint,B_neigs,B_ierr,B_lwork; 87108122e43SStefano Zampini PetscBLASInt *B_iwork,*B_ifail; 87208122e43SStefano Zampini PetscScalar *work,lwork; 87308122e43SStefano Zampini PetscScalar *St,*S,*eigv; 87408122e43SStefano Zampini PetscScalar *Sarray,*Starray; 87508122e43SStefano Zampini PetscReal *eigs,thresh; 8761b968477SStefano Zampini PetscInt i,nmax,nmin,nv,cum,mss,cum2,cumarray,maxneigs; 877f6f667cfSStefano Zampini PetscBool allocated_S_St; 87808122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 87908122e43SStefano Zampini PetscReal *rwork; 88008122e43SStefano Zampini #endif 881b1b3d7a2SStefano Zampini PetscErrorCode ierr; 882b1b3d7a2SStefano Zampini 883b1b3d7a2SStefano Zampini PetscFunctionBegin; 884df4d28bfSStefano Zampini if (!sub_schurs->schur_explicit) { 885df4d28bfSStefano Zampini SETERRQ(PetscObjectComm((PetscObject)pc),PETSC_ERR_SUP,"Adaptive selection of constraints requires MUMPS and/or MKL_CPARDISO"); 88608122e43SStefano Zampini } 88708122e43SStefano Zampini 88806a4e24aSStefano Zampini if (sub_schurs->n_subs && (!sub_schurs->is_hermitian || !sub_schurs->is_posdef)) { 88906a4e24aSStefano Zampini SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_SUP,"Adaptive selection not yet implemented for general matrix pencils (herm %d, posdef %d)\n",sub_schurs->is_hermitian,sub_schurs->is_posdef); 89006a4e24aSStefano Zampini } 89106a4e24aSStefano Zampini 892fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 893fd14bc51SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 894fd14bc51SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 895fd14bc51SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Check adaptive selection of constraints\n");CHKERRQ(ierr); 8961575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 897fd14bc51SStefano Zampini } 898fd14bc51SStefano Zampini 899e496cd5dSStefano Zampini if (pcbddc->dbg_flag) { 900e496cd5dSStefano Zampini PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d cc %d (%d,%d).\n",PetscGlobalRank,sub_schurs->n_subs,sub_schurs->is_hermitian,sub_schurs->is_posdef); 901e496cd5dSStefano Zampini } 902e496cd5dSStefano Zampini 90308122e43SStefano Zampini /* max size of subsets */ 90408122e43SStefano Zampini mss = 0; 90508122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 90608122e43SStefano Zampini PetscInt subset_size; 907862806e4SStefano Zampini 90808122e43SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 90908122e43SStefano Zampini mss = PetscMax(mss,subset_size); 91008122e43SStefano Zampini } 91108122e43SStefano Zampini 91208122e43SStefano Zampini /* min/max and threshold */ 91308122e43SStefano Zampini nmax = pcbddc->adaptive_nmax > 0 ? pcbddc->adaptive_nmax : mss; 914f6f667cfSStefano Zampini nmin = pcbddc->adaptive_nmin > 0 ? pcbddc->adaptive_nmin : 0; 91508122e43SStefano Zampini nmax = PetscMax(nmin,nmax); 916f6f667cfSStefano Zampini allocated_S_St = PETSC_FALSE; 917f6f667cfSStefano Zampini if (nmin) { 918f6f667cfSStefano Zampini allocated_S_St = PETSC_TRUE; 919f6f667cfSStefano Zampini } 92008122e43SStefano Zampini 92108122e43SStefano Zampini /* allocate lapack workspace */ 92208122e43SStefano Zampini cum = cum2 = 0; 92308122e43SStefano Zampini maxneigs = 0; 92408122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 92508122e43SStefano Zampini PetscInt n,subset_size; 926f6f667cfSStefano Zampini 92708122e43SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 92808122e43SStefano Zampini n = PetscMin(subset_size,nmax); 9299162d606SStefano Zampini cum += subset_size; 9309162d606SStefano Zampini cum2 += subset_size*n; 93108122e43SStefano Zampini maxneigs = PetscMax(maxneigs,n); 93208122e43SStefano Zampini } 93308122e43SStefano Zampini if (mss) { 9349ab7bb16SStefano Zampini if (sub_schurs->is_hermitian && sub_schurs->is_posdef) { 93508122e43SStefano Zampini PetscBLASInt B_itype = 1; 93608122e43SStefano Zampini PetscBLASInt B_N = mss; 9374c6709b3SStefano Zampini PetscReal zero = 0.0; 9384c6709b3SStefano Zampini PetscReal eps = 0.0; /* dlamch? */ 93908122e43SStefano Zampini 94008122e43SStefano Zampini B_lwork = -1; 94108122e43SStefano Zampini S = NULL; 94208122e43SStefano Zampini St = NULL; 943a58a30b4SStefano Zampini eigs = NULL; 944a58a30b4SStefano Zampini eigv = NULL; 945a58a30b4SStefano Zampini B_iwork = NULL; 946a58a30b4SStefano Zampini B_ifail = NULL; 947d1710679SStefano Zampini #if defined(PETSC_USE_COMPLEX) 948d1710679SStefano Zampini rwork = NULL; 949d1710679SStefano Zampini #endif 9508bec7fa6SStefano Zampini thresh = 1.0; 95108122e43SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 95208122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 95308122e43SStefano Zampini PetscStackCallBLAS("LAPACKsygvx",LAPACKsygvx_(&B_itype,"V","V","L",&B_N,St,&B_N,S,&B_N,&zero,&thresh,&B_dummyint,&B_dummyint,&eps,&B_neigs,eigs,eigv,&B_N,&lwork,&B_lwork,rwork,B_iwork,B_ifail,&B_ierr)); 95408122e43SStefano Zampini #else 95508122e43SStefano Zampini PetscStackCallBLAS("LAPACKsygvx",LAPACKsygvx_(&B_itype,"V","V","L",&B_N,St,&B_N,S,&B_N,&zero,&thresh,&B_dummyint,&B_dummyint,&eps,&B_neigs,eigs,eigv,&B_N,&lwork,&B_lwork,B_iwork,B_ifail,&B_ierr)); 95608122e43SStefano Zampini #endif 95708122e43SStefano Zampini if (B_ierr != 0) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to SYGVX Lapack routine %d",(int)B_ierr); 95808122e43SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 95908122e43SStefano Zampini } else { 96008122e43SStefano Zampini /* TODO */ 96108122e43SStefano Zampini } 96208122e43SStefano Zampini } else { 96308122e43SStefano Zampini lwork = 0; 96408122e43SStefano Zampini } 96508122e43SStefano Zampini 96608122e43SStefano Zampini nv = 0; 967d62866d3SStefano Zampini if (sub_schurs->is_vertices && pcbddc->use_vertices) { /* complement set of active subsets, each entry is a vertex (boundary made by active subsets, vertices and dirichlet dofs) */ 968d62866d3SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_vertices,&nv);CHKERRQ(ierr); 96908122e43SStefano Zampini } 9704c6709b3SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lwork),&B_lwork);CHKERRQ(ierr); 971f6f667cfSStefano Zampini if (allocated_S_St) { 972f6f667cfSStefano Zampini ierr = PetscMalloc2(mss*mss,&S,mss*mss,&St);CHKERRQ(ierr); 973f6f667cfSStefano Zampini } 974f6f667cfSStefano Zampini ierr = PetscMalloc5(mss*mss,&eigv,mss,&eigs,B_lwork,&work,5*mss,&B_iwork,mss,&B_ifail);CHKERRQ(ierr); 97508122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 97608122e43SStefano Zampini ierr = PetscMalloc1(7*mss,&rwork);CHKERRQ(ierr); 97708122e43SStefano Zampini #endif 9789162d606SStefano Zampini ierr = PetscMalloc5(nv+sub_schurs->n_subs,&pcbddc->adaptive_constraints_n, 9799162d606SStefano Zampini nv+sub_schurs->n_subs+1,&pcbddc->adaptive_constraints_idxs_ptr, 9809162d606SStefano Zampini nv+sub_schurs->n_subs+1,&pcbddc->adaptive_constraints_data_ptr, 98108122e43SStefano Zampini nv+cum,&pcbddc->adaptive_constraints_idxs, 9829162d606SStefano Zampini nv+cum2,&pcbddc->adaptive_constraints_data);CHKERRQ(ierr); 98308122e43SStefano Zampini ierr = PetscMemzero(pcbddc->adaptive_constraints_n,(nv+sub_schurs->n_subs)*sizeof(PetscInt));CHKERRQ(ierr); 98408122e43SStefano Zampini 98508122e43SStefano Zampini maxneigs = 0; 98608122e43SStefano Zampini cum = cum2 = cumarray = 0; 9879162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr[0] = 0; 9889162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr[0] = 0; 989d62866d3SStefano Zampini if (sub_schurs->is_vertices && pcbddc->use_vertices) { 99008122e43SStefano Zampini const PetscInt *idxs; 99108122e43SStefano Zampini 992d62866d3SStefano Zampini ierr = ISGetIndices(sub_schurs->is_vertices,&idxs);CHKERRQ(ierr); 99308122e43SStefano Zampini for (cum=0;cum<nv;cum++) { 99408122e43SStefano Zampini pcbddc->adaptive_constraints_n[cum] = 1; 99508122e43SStefano Zampini pcbddc->adaptive_constraints_idxs[cum] = idxs[cum]; 99608122e43SStefano Zampini pcbddc->adaptive_constraints_data[cum] = 1.0; 9979162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr[cum+1] = pcbddc->adaptive_constraints_idxs_ptr[cum]+1; 9989162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr[cum+1] = pcbddc->adaptive_constraints_data_ptr[cum]+1; 99908122e43SStefano Zampini } 1000d62866d3SStefano Zampini ierr = ISRestoreIndices(sub_schurs->is_vertices,&idxs);CHKERRQ(ierr); 100108122e43SStefano Zampini } 100208122e43SStefano Zampini 100308122e43SStefano Zampini if (mss) { /* multilevel */ 100408122e43SStefano Zampini ierr = MatSeqAIJGetArray(sub_schurs->sum_S_Ej_inv_all,&Sarray);CHKERRQ(ierr); 100508122e43SStefano Zampini ierr = MatSeqAIJGetArray(sub_schurs->sum_S_Ej_tilda_all,&Starray);CHKERRQ(ierr); 100608122e43SStefano Zampini } 100708122e43SStefano Zampini 1008ffd830a3SStefano Zampini thresh = pcbddc->adaptive_threshold; 100908122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 101008122e43SStefano Zampini const PetscInt *idxs; 1011f6f667cfSStefano Zampini PetscReal infty = PETSC_MAX_REAL; 1012862806e4SStefano Zampini PetscInt j,subset_size,eigs_start = 0; 101308122e43SStefano Zampini PetscBLASInt B_N; 1014aff50787SStefano Zampini PetscBool same_data = PETSC_FALSE; 101508122e43SStefano Zampini 1016862806e4SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 1017ffd830a3SStefano Zampini ierr = ISGetIndices(sub_schurs->is_subs[i],&idxs);CHKERRQ(ierr); 1018f6f667cfSStefano Zampini ierr = PetscBLASIntCast(subset_size,&B_N);CHKERRQ(ierr); 1019f6f667cfSStefano Zampini if (allocated_S_St) { /* S and S_t should be copied since we could need them later */ 10209ab7bb16SStefano Zampini if (sub_schurs->is_hermitian) { 1021aff50787SStefano Zampini PetscInt j,k; 1022aff50787SStefano Zampini if (sub_schurs->n_subs == 1) { /* zeroing memory to use PetscMemcmp later */ 1023aff50787SStefano Zampini ierr = PetscMemzero(S,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 1024aff50787SStefano Zampini ierr = PetscMemzero(St,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 102508122e43SStefano Zampini } 102608122e43SStefano Zampini for (j=0;j<subset_size;j++) { 1027aff50787SStefano Zampini for (k=j;k<subset_size;k++) { 1028aff50787SStefano Zampini S [j*subset_size+k] = Sarray [cumarray+j*subset_size+k]; 1029aff50787SStefano Zampini St[j*subset_size+k] = Starray[cumarray+j*subset_size+k]; 1030aff50787SStefano Zampini } 103108122e43SStefano Zampini } 103208122e43SStefano Zampini } else { 103308122e43SStefano Zampini ierr = PetscMemcpy(S,Sarray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 103408122e43SStefano Zampini ierr = PetscMemcpy(St,Starray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 103508122e43SStefano Zampini } 10368bec7fa6SStefano Zampini } else { 1037f6f667cfSStefano Zampini S = Sarray + cumarray; 1038f6f667cfSStefano Zampini St = Starray + cumarray; 10398bec7fa6SStefano Zampini } 1040aff50787SStefano Zampini /* see if we can save some work */ 1041aff50787SStefano Zampini if (sub_schurs->n_subs == 1) { 1042aff50787SStefano Zampini ierr = PetscMemcmp(S,St,subset_size*subset_size*sizeof(PetscScalar),&same_data);CHKERRQ(ierr); 1043aff50787SStefano Zampini } 1044aff50787SStefano Zampini 1045aff50787SStefano Zampini if (same_data) { /* there's no need of constraints here, deluxe scaling is enough */ 1046aff50787SStefano Zampini B_neigs = 0; 1047aff50787SStefano Zampini } else { 10489ab7bb16SStefano Zampini if (sub_schurs->is_hermitian && sub_schurs->is_posdef) { 104908122e43SStefano Zampini PetscBLASInt B_itype = 1; 1050f6f667cfSStefano Zampini PetscBLASInt B_IL, B_IU; 10514c6709b3SStefano Zampini PetscReal eps = -1.0; /* dlamch? */ 10529552c7c7SStefano Zampini PetscInt nmin_s; 105308122e43SStefano Zampini 1054fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 10558bec7fa6SStefano Zampini PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Computing for sub %d/%d %d %d.\n",i,sub_schurs->n_subs,subset_size,pcbddc->mat_graph->count[idxs[0]]); 1056fd14bc51SStefano Zampini } 1057d16cbb6bSStefano Zampini 105808122e43SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 1059d16cbb6bSStefano Zampini if (thresh > 1.+PETSC_SMALL) { 1060d16cbb6bSStefano Zampini 1061d16cbb6bSStefano Zampini /* ask for eigenvalues larger than thresh */ 106208122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 1063f6f667cfSStefano Zampini PetscStackCallBLAS("LAPACKsygvx",LAPACKsygvx_(&B_itype,"V","V","L",&B_N,St,&B_N,S,&B_N,&thresh,&infty,&B_IL,&B_IU,&eps,&B_neigs,eigs,eigv,&B_N,work,&B_lwork,rwork,B_iwork,B_ifail,&B_ierr)); 106408122e43SStefano Zampini #else 1065f6f667cfSStefano Zampini PetscStackCallBLAS("LAPACKsygvx",LAPACKsygvx_(&B_itype,"V","V","L",&B_N,St,&B_N,S,&B_N,&thresh,&infty,&B_IL,&B_IU,&eps,&B_neigs,eigs,eigv,&B_N,work,&B_lwork,B_iwork,B_ifail,&B_ierr)); 106608122e43SStefano Zampini #endif 1067d16cbb6bSStefano Zampini } else { 1068d16cbb6bSStefano Zampini B_IU = PetscMax(1,PetscMin(B_N,nmax)); 1069d16cbb6bSStefano Zampini B_IL = 1; 1070d16cbb6bSStefano Zampini #if defined(PETSC_USE_COMPLEX) 1071d16cbb6bSStefano Zampini PetscStackCallBLAS("LAPACKsygvx",LAPACKsygvx_(&B_itype,"V","I","L",&B_N,St,&B_N,S,&B_N,&thresh,&infty,&B_IL,&B_IU,&eps,&B_neigs,eigs,eigv,&B_N,work,&B_lwork,rwork,B_iwork,B_ifail,&B_ierr)); 1072d16cbb6bSStefano Zampini #else 1073d16cbb6bSStefano Zampini PetscStackCallBLAS("LAPACKsygvx",LAPACKsygvx_(&B_itype,"V","I","L",&B_N,St,&B_N,S,&B_N,&thresh,&infty,&B_IL,&B_IU,&eps,&B_neigs,eigs,eigv,&B_N,work,&B_lwork,B_iwork,B_ifail,&B_ierr)); 1074d16cbb6bSStefano Zampini #endif 1075d16cbb6bSStefano Zampini } 107608122e43SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 107708122e43SStefano Zampini if (B_ierr) { 107808122e43SStefano Zampini if (B_ierr < 0 ) { 107908122e43SStefano Zampini SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYGVX Lapack routine: illegal value for argument %d",-(int)B_ierr); 108008122e43SStefano Zampini } else if (B_ierr <= B_N) { 108108122e43SStefano Zampini SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYGVX Lapack routine: %d eigenvalues failed to converge",(int)B_ierr); 108208122e43SStefano Zampini } else { 10839552c7c7SStefano Zampini SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYGVX Lapack routine: leading minor of order %d is not positive definite",(int)B_ierr-B_N-1); 108408122e43SStefano Zampini } 108508122e43SStefano Zampini } 108608122e43SStefano Zampini 108708122e43SStefano Zampini if (B_neigs > nmax) { 1088fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 1089fd14bc51SStefano Zampini PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," found %d eigs, more than maximum required %d.\n",B_neigs,nmax); 1090fd14bc51SStefano Zampini } 1091f6f667cfSStefano Zampini eigs_start = B_neigs -nmax; 109208122e43SStefano Zampini B_neigs = nmax; 109308122e43SStefano Zampini } 109408122e43SStefano Zampini 10959552c7c7SStefano Zampini nmin_s = PetscMin(nmin,B_N); 10969552c7c7SStefano Zampini if (B_neigs < nmin_s) { 109708122e43SStefano Zampini PetscBLASInt B_neigs2; 109808122e43SStefano Zampini 1099f6f667cfSStefano Zampini B_IU = B_N - B_neigs; 1100f6f667cfSStefano Zampini B_IL = B_N - nmin_s + 1; 1101fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 1102fd14bc51SStefano Zampini PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," found %d eigs, less than minimum required %d. Asking for %d to %d incl (fortran like)\n",B_neigs,nmin,B_IL,B_IU); 1103fd14bc51SStefano Zampini } 11049ab7bb16SStefano Zampini if (sub_schurs->is_hermitian) { 110508122e43SStefano Zampini PetscInt j; 110608122e43SStefano Zampini for (j=0;j<subset_size;j++) { 110708122e43SStefano Zampini ierr = PetscMemcpy(S+j*(subset_size+1),Sarray+cumarray+j*(subset_size+1),(subset_size-j)*sizeof(PetscScalar));CHKERRQ(ierr); 110808122e43SStefano Zampini } 110908122e43SStefano Zampini for (j=0;j<subset_size;j++) { 111008122e43SStefano Zampini ierr = PetscMemcpy(St+j*(subset_size+1),Starray+cumarray+j*(subset_size+1),(subset_size-j)*sizeof(PetscScalar));CHKERRQ(ierr); 111108122e43SStefano Zampini } 111208122e43SStefano Zampini } else { 111308122e43SStefano Zampini ierr = PetscMemcpy(S,Sarray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 111408122e43SStefano Zampini ierr = PetscMemcpy(St,Starray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 111508122e43SStefano Zampini } 111608122e43SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 111708122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 1118f6f667cfSStefano Zampini PetscStackCallBLAS("LAPACKsygvx",LAPACKsygvx_(&B_itype,"V","I","L",&B_N,St,&B_N,S,&B_N,&thresh,&infty,&B_IL,&B_IU,&eps,&B_neigs2,eigs+B_neigs,eigv+B_neigs*subset_size,&B_N,work,&B_lwork,rwork,B_iwork,B_ifail,&B_ierr)); 111908122e43SStefano Zampini #else 1120f6f667cfSStefano Zampini PetscStackCallBLAS("LAPACKsygvx",LAPACKsygvx_(&B_itype,"V","I","L",&B_N,St,&B_N,S,&B_N,&thresh,&infty,&B_IL,&B_IU,&eps,&B_neigs2,eigs+B_neigs,eigv+B_neigs*subset_size,&B_N,work,&B_lwork,B_iwork,B_ifail,&B_ierr)); 112108122e43SStefano Zampini #endif 112208122e43SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 112308122e43SStefano Zampini B_neigs += B_neigs2; 112408122e43SStefano Zampini } 112508122e43SStefano Zampini if (B_ierr) { 112608122e43SStefano Zampini if (B_ierr < 0 ) { 112708122e43SStefano Zampini SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYGVX Lapack routine: illegal value for argument %d",-(int)B_ierr); 112808122e43SStefano Zampini } else if (B_ierr <= B_N) { 112908122e43SStefano Zampini SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYGVX Lapack routine: %d eigenvalues failed to converge",(int)B_ierr); 113008122e43SStefano Zampini } else { 11319552c7c7SStefano Zampini SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYGVX Lapack routine: leading minor of order %d is not positive definite",(int)B_ierr-B_N-1); 113208122e43SStefano Zampini } 113308122e43SStefano Zampini } 1134fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 1135ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," -> Got %d eigs\n",B_neigs);CHKERRQ(ierr); 113608122e43SStefano Zampini for (j=0;j<B_neigs;j++) { 113708122e43SStefano Zampini if (eigs[j] == 0.0) { 1138ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," Inf\n");CHKERRQ(ierr); 113908122e43SStefano Zampini } else { 1140ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," %1.6e\n",eigs[j+eigs_start]);CHKERRQ(ierr); 1141fd14bc51SStefano Zampini } 114208122e43SStefano Zampini } 114308122e43SStefano Zampini } 114408122e43SStefano Zampini } else { 114508122e43SStefano Zampini /* TODO */ 114608122e43SStefano Zampini } 1147aff50787SStefano Zampini } 11486c3e6151SStefano Zampini /* change the basis back to the original one */ 11496c3e6151SStefano Zampini if (sub_schurs->change) { 11506c3e6151SStefano Zampini Mat change,change_subs,phi,phit; 11516c3e6151SStefano Zampini IS issubs; 11526c3e6151SStefano Zampini 11536c3e6151SStefano Zampini if (pcbddc->dbg_flag > 1) { 11546c3e6151SStefano Zampini PetscInt ii; 11556c3e6151SStefano Zampini for (ii=0;ii<B_neigs;ii++) { 11566c3e6151SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," -> Eigenvector (old basis) %d/%d (%d)\n",ii,B_neigs,B_N);CHKERRQ(ierr); 11576c3e6151SStefano Zampini for (j=0;j<B_N;j++) { 11586c3e6151SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," %1.4e\n",eigv[(ii+eigs_start)*subset_size+j]);CHKERRQ(ierr); 11596c3e6151SStefano Zampini } 11606c3e6151SStefano Zampini } 11616c3e6151SStefano Zampini } 11626c3e6151SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,subset_size,cum2,1,&issubs);CHKERRQ(ierr); 11636c3e6151SStefano Zampini ierr = KSPGetOperators(sub_schurs->change,&change,NULL);CHKERRQ(ierr); 11646c3e6151SStefano Zampini ierr = MatGetSubMatrix(change,issubs,issubs,MAT_INITIAL_MATRIX,&change_subs);CHKERRQ(ierr); 11656c3e6151SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,subset_size,B_neigs,eigv+eigs_start*subset_size,&phit);CHKERRQ(ierr); 11666c3e6151SStefano Zampini ierr = MatMatMult(change_subs,phit,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&phi);CHKERRQ(ierr); 11676c3e6151SStefano Zampini ierr = MatCopy(phi,phit,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 11686c3e6151SStefano Zampini ierr = MatDestroy(&phit);CHKERRQ(ierr); 11696c3e6151SStefano Zampini ierr = MatDestroy(&phi);CHKERRQ(ierr); 11706c3e6151SStefano Zampini ierr = MatDestroy(&change_subs);CHKERRQ(ierr); 11716c3e6151SStefano Zampini ierr = ISDestroy(&issubs);CHKERRQ(ierr); 11726c3e6151SStefano Zampini } 11738bec7fa6SStefano Zampini maxneigs = PetscMax(B_neigs,maxneigs); 11748bec7fa6SStefano Zampini pcbddc->adaptive_constraints_n[i+nv] = B_neigs; 11759162d606SStefano Zampini if (B_neigs) { 11769162d606SStefano 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); 1177fd14bc51SStefano Zampini 1178fd14bc51SStefano Zampini if (pcbddc->dbg_flag > 1) { 11799552c7c7SStefano Zampini PetscInt ii; 11809552c7c7SStefano Zampini for (ii=0;ii<B_neigs;ii++) { 1181ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," -> Eigenvector %d/%d (%d)\n",ii,B_neigs,B_N);CHKERRQ(ierr); 11829552c7c7SStefano Zampini for (j=0;j<B_N;j++) { 1183ac47001eSStefano Zampini #if defined(PETSC_USE_COMPLEX) 1184ac47001eSStefano Zampini PetscReal r = PetscRealPart(pcbddc->adaptive_constraints_data[ii*subset_size+j+pcbddc->adaptive_constraints_data_ptr[cum]]); 1185ac47001eSStefano Zampini PetscReal c = PetscImaginaryPart(pcbddc->adaptive_constraints_data[ii*subset_size+j+pcbddc->adaptive_constraints_data_ptr[cum]]); 1186ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," %1.4e + %1.4e i\n",r,c);CHKERRQ(ierr); 1187ac47001eSStefano Zampini #else 1188ac47001eSStefano 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); 1189ac47001eSStefano Zampini #endif 11909552c7c7SStefano Zampini } 11919552c7c7SStefano Zampini } 1192fd14bc51SStefano Zampini } 11939162d606SStefano Zampini ierr = PetscMemcpy(pcbddc->adaptive_constraints_idxs+pcbddc->adaptive_constraints_idxs_ptr[cum],idxs,subset_size*sizeof(PetscInt));CHKERRQ(ierr); 11949162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr[cum+1] = pcbddc->adaptive_constraints_idxs_ptr[cum] + subset_size; 11959162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr[cum+1] = pcbddc->adaptive_constraints_data_ptr[cum] + subset_size*B_neigs; 11969162d606SStefano Zampini cum++; 119708122e43SStefano Zampini } 119808122e43SStefano Zampini ierr = ISRestoreIndices(sub_schurs->is_subs[i],&idxs);CHKERRQ(ierr); 119908122e43SStefano Zampini /* shift for next computation */ 120008122e43SStefano Zampini cumarray += subset_size*subset_size; 12016c3e6151SStefano Zampini cum2 += subset_size; 120208122e43SStefano Zampini } 1203fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 1204fd14bc51SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1205fd14bc51SStefano Zampini } 120608122e43SStefano Zampini 120708122e43SStefano Zampini if (mss) { 120808122e43SStefano Zampini ierr = MatSeqAIJRestoreArray(sub_schurs->sum_S_Ej_inv_all,&Sarray);CHKERRQ(ierr); 120908122e43SStefano Zampini ierr = MatSeqAIJRestoreArray(sub_schurs->sum_S_Ej_tilda_all,&Starray);CHKERRQ(ierr); 1210f6f667cfSStefano Zampini /* destroy matrices (junk) */ 1211f6f667cfSStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_inv_all);CHKERRQ(ierr); 1212f6f667cfSStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_tilda_all);CHKERRQ(ierr); 121308122e43SStefano Zampini } 1214f6f667cfSStefano Zampini if (allocated_S_St) { 1215f6f667cfSStefano Zampini ierr = PetscFree2(S,St);CHKERRQ(ierr); 1216f6f667cfSStefano Zampini } 1217f6f667cfSStefano Zampini ierr = PetscFree5(eigv,eigs,work,B_iwork,B_ifail);CHKERRQ(ierr); 121808122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 121908122e43SStefano Zampini ierr = PetscFree(rwork);CHKERRQ(ierr); 122008122e43SStefano Zampini #endif 122108122e43SStefano Zampini if (pcbddc->dbg_flag) { 12221b968477SStefano Zampini PetscInt maxneigs_r; 122308122e43SStefano Zampini ierr = MPI_Allreduce(&maxneigs,&maxneigs_r,1,MPIU_INT,MPI_MAX,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 12249b28b941SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Maximum number of constraints per cc %d\n",maxneigs_r);CHKERRQ(ierr); 122508122e43SStefano Zampini } 122608122e43SStefano Zampini PetscFunctionReturn(0); 122708122e43SStefano Zampini } 1228b1b3d7a2SStefano Zampini 1229674ae819SStefano Zampini #undef __FUNCT__ 1230c8587f34SStefano Zampini #define __FUNCT__ "PCBDDCSetUpSolvers" 1231c8587f34SStefano Zampini PetscErrorCode PCBDDCSetUpSolvers(PC pc) 1232c8587f34SStefano Zampini { 1233c8587f34SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 12348629588bSStefano Zampini PetscScalar *coarse_submat_vals; 1235c8587f34SStefano Zampini PetscErrorCode ierr; 1236c8587f34SStefano Zampini 1237c8587f34SStefano Zampini PetscFunctionBegin; 1238f4ddd8eeSStefano Zampini /* Setup local scatters R_to_B and (optionally) R_to_D */ 12395e8657edSStefano Zampini /* PCBDDCSetUpLocalWorkVectors should be called first! */ 1240c8587f34SStefano Zampini ierr = PCBDDCSetUpLocalScatters(pc);CHKERRQ(ierr); 1241c8587f34SStefano Zampini 1242684f6988SStefano Zampini /* Setup local neumann solver ksp_R */ 12430fccc4e9SStefano Zampini /* PCBDDCSetUpLocalScatters should be called first! */ 1244684f6988SStefano Zampini ierr = PCBDDCSetUpLocalSolvers(pc,PETSC_FALSE,PETSC_TRUE);CHKERRQ(ierr); 1245c8587f34SStefano Zampini 1246c8587f34SStefano Zampini /* Change global null space passed in by the user if change of basis has been requested */ 1247b9b85e73SStefano Zampini if (pcbddc->NullSpace && pcbddc->ChangeOfBasisMatrix) { 1248c8587f34SStefano Zampini ierr = PCBDDCNullSpaceAdaptGlobal(pc);CHKERRQ(ierr); 1249c8587f34SStefano Zampini } 1250c8587f34SStefano Zampini 12518629588bSStefano Zampini /* 12528629588bSStefano Zampini Setup local correction and local part of coarse basis. 12538629588bSStefano Zampini Gives back the dense local part of the coarse matrix in column major ordering 12548629588bSStefano Zampini */ 125547f4ddc3SStefano Zampini ierr = PCBDDCSetUpCorrection(pc,&coarse_submat_vals);CHKERRQ(ierr); 12568629588bSStefano Zampini 12578629588bSStefano Zampini /* Compute total number of coarse nodes and setup coarse solver */ 12588629588bSStefano Zampini ierr = PCBDDCSetUpCoarseSolver(pc,coarse_submat_vals);CHKERRQ(ierr); 12598629588bSStefano Zampini 12608629588bSStefano Zampini /* free */ 12618629588bSStefano Zampini ierr = PetscFree(coarse_submat_vals);CHKERRQ(ierr); 1262c8587f34SStefano Zampini PetscFunctionReturn(0); 1263c8587f34SStefano Zampini } 1264c8587f34SStefano Zampini 1265c8587f34SStefano Zampini #undef __FUNCT__ 1266674ae819SStefano Zampini #define __FUNCT__ "PCBDDCResetCustomization" 1267674ae819SStefano Zampini PetscErrorCode PCBDDCResetCustomization(PC pc) 1268674ae819SStefano Zampini { 1269674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 1270674ae819SStefano Zampini PetscErrorCode ierr; 1271674ae819SStefano Zampini 1272674ae819SStefano Zampini PetscFunctionBegin; 1273674ae819SStefano Zampini ierr = PCBDDCGraphResetCSR(pcbddc->mat_graph);CHKERRQ(ierr); 1274674ae819SStefano Zampini ierr = ISDestroy(&pcbddc->user_primal_vertices);CHKERRQ(ierr); 127530368db7SStefano Zampini ierr = ISDestroy(&pcbddc->user_primal_vertices_local);CHKERRQ(ierr); 1276674ae819SStefano Zampini ierr = MatNullSpaceDestroy(&pcbddc->NullSpace);CHKERRQ(ierr); 1277674ae819SStefano Zampini ierr = ISDestroy(&pcbddc->NeumannBoundaries);CHKERRQ(ierr); 1278785d1243SStefano Zampini ierr = ISDestroy(&pcbddc->NeumannBoundariesLocal);CHKERRQ(ierr); 1279674ae819SStefano Zampini ierr = ISDestroy(&pcbddc->DirichletBoundaries);CHKERRQ(ierr); 1280f4ddd8eeSStefano Zampini ierr = MatNullSpaceDestroy(&pcbddc->onearnullspace);CHKERRQ(ierr); 1281f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->onearnullvecs_state);CHKERRQ(ierr); 1282785d1243SStefano Zampini ierr = ISDestroy(&pcbddc->DirichletBoundariesLocal);CHKERRQ(ierr); 128363602bcaSStefano Zampini ierr = PCBDDCSetDofsSplitting(pc,0,NULL);CHKERRQ(ierr); 128463602bcaSStefano Zampini ierr = PCBDDCSetDofsSplittingLocal(pc,0,NULL);CHKERRQ(ierr); 1285674ae819SStefano Zampini PetscFunctionReturn(0); 1286674ae819SStefano Zampini } 1287674ae819SStefano Zampini 1288674ae819SStefano Zampini #undef __FUNCT__ 1289674ae819SStefano Zampini #define __FUNCT__ "PCBDDCResetTopography" 1290674ae819SStefano Zampini PetscErrorCode PCBDDCResetTopography(PC pc) 1291674ae819SStefano Zampini { 1292674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 12934f1b2e48SStefano Zampini PetscInt i; 1294674ae819SStefano Zampini PetscErrorCode ierr; 1295674ae819SStefano Zampini 1296674ae819SStefano Zampini PetscFunctionBegin; 1297b9b85e73SStefano Zampini ierr = MatDestroy(&pcbddc->user_ChangeOfBasisMatrix);CHKERRQ(ierr); 1298674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 1299674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 1300674ae819SStefano Zampini ierr = PCBDDCGraphReset(pcbddc->mat_graph);CHKERRQ(ierr); 13014f1b2e48SStefano Zampini for (i=0;i<pcbddc->n_local_subs;i++) { 13024f1b2e48SStefano Zampini ierr = ISDestroy(&pcbddc->local_subs[i]);CHKERRQ(ierr); 13034f1b2e48SStefano Zampini } 13044f1b2e48SStefano Zampini ierr = PetscFree(pcbddc->local_subs);CHKERRQ(ierr); 1305b96c3477SStefano Zampini ierr = PCBDDCSubSchursReset(pcbddc->sub_schurs);CHKERRQ(ierr); 1306674ae819SStefano Zampini PetscFunctionReturn(0); 1307674ae819SStefano Zampini } 1308674ae819SStefano Zampini 1309674ae819SStefano Zampini #undef __FUNCT__ 1310674ae819SStefano Zampini #define __FUNCT__ "PCBDDCResetSolvers" 1311674ae819SStefano Zampini PetscErrorCode PCBDDCResetSolvers(PC pc) 1312674ae819SStefano Zampini { 1313674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 1314674ae819SStefano Zampini PetscErrorCode ierr; 1315674ae819SStefano Zampini 1316674ae819SStefano Zampini PetscFunctionBegin; 1317674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->coarse_vec);CHKERRQ(ierr); 131858da7f69SStefano Zampini if (pcbddc->coarse_phi_B) { 1319ca92afb2SStefano Zampini PetscScalar *array; 132006656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&array);CHKERRQ(ierr); 132106656605SStefano Zampini ierr = PetscFree(array);CHKERRQ(ierr); 132258da7f69SStefano Zampini } 1323674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_B);CHKERRQ(ierr); 1324674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_D);CHKERRQ(ierr); 132515aaf578SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_B);CHKERRQ(ierr); 132615aaf578SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_D);CHKERRQ(ierr); 1327674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec1_P);CHKERRQ(ierr); 1328674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec1_C);CHKERRQ(ierr); 1329674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat2);CHKERRQ(ierr); 133006656605SStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat1);CHKERRQ(ierr); 1331674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec1_R);CHKERRQ(ierr); 1332674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec2_R);CHKERRQ(ierr); 13338ce42a96SStefano Zampini ierr = ISDestroy(&pcbddc->is_R_local);CHKERRQ(ierr); 1334674ae819SStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_B);CHKERRQ(ierr); 1335674ae819SStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_D);CHKERRQ(ierr); 1336674ae819SStefano Zampini ierr = VecScatterDestroy(&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 1337f4ddd8eeSStefano Zampini ierr = KSPDestroy(&pcbddc->ksp_D);CHKERRQ(ierr); 1338f4ddd8eeSStefano Zampini ierr = KSPDestroy(&pcbddc->ksp_R);CHKERRQ(ierr); 1339f4ddd8eeSStefano Zampini ierr = KSPDestroy(&pcbddc->coarse_ksp);CHKERRQ(ierr); 1340f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 1341727cdba6SStefano Zampini ierr = PetscFree(pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 13420e6343abSStefano Zampini ierr = PetscFree2(pcbddc->local_primal_ref_node,pcbddc->local_primal_ref_mult);CHKERRQ(ierr); 1343f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->global_primal_indices);CHKERRQ(ierr); 134470cf5478SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling);CHKERRQ(ierr); 13456e683305SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 134681d14e9dSStefano Zampini ierr = MatDestroy(&pcbddc->benign_change);CHKERRQ(ierr); 13470369aaf7SStefano Zampini ierr = VecDestroy(&pcbddc->benign_vec);CHKERRQ(ierr); 13488b9f24d4SStefano Zampini ierr = MatDestroy(&pcbddc->benign_original_mat);CHKERRQ(ierr); 13494f1b2e48SStefano Zampini ierr = MatDestroy(&pcbddc->benign_B0);CHKERRQ(ierr); 13508b9f24d4SStefano Zampini ierr = PetscSFDestroy(&pcbddc->benign_sf);CHKERRQ(ierr); 1351ca92afb2SStefano Zampini if (pcbddc->benign_zerodiag_subs) { 1352ca92afb2SStefano Zampini PetscInt i; 1353ca92afb2SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 1354ca92afb2SStefano Zampini ierr = ISDestroy(&pcbddc->benign_zerodiag_subs[i]);CHKERRQ(ierr); 1355ca92afb2SStefano Zampini } 1356ca92afb2SStefano Zampini ierr = PetscFree(pcbddc->benign_zerodiag_subs);CHKERRQ(ierr); 1357ca92afb2SStefano Zampini } 13584f1b2e48SStefano Zampini ierr = PetscFree3(pcbddc->benign_p0_lidx,pcbddc->benign_p0_gidx,pcbddc->benign_p0);CHKERRQ(ierr); 1359674ae819SStefano Zampini PetscFunctionReturn(0); 1360674ae819SStefano Zampini } 1361674ae819SStefano Zampini 1362674ae819SStefano Zampini #undef __FUNCT__ 1363f4ddd8eeSStefano Zampini #define __FUNCT__ "PCBDDCSetUpLocalWorkVectors" 1364f4ddd8eeSStefano Zampini PetscErrorCode PCBDDCSetUpLocalWorkVectors(PC pc) 13656bfb1811SStefano Zampini { 13666bfb1811SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 13676bfb1811SStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 13686bfb1811SStefano Zampini VecType impVecType; 13694f1b2e48SStefano Zampini PetscInt n_constraints,n_R,old_size; 13706bfb1811SStefano Zampini PetscErrorCode ierr; 13716bfb1811SStefano Zampini 13726bfb1811SStefano Zampini PetscFunctionBegin; 1373f4ddd8eeSStefano Zampini if (!pcbddc->ConstraintMatrix) { 1374019a44ceSStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"BDDC Constraint matrix has not been created"); 1375f4ddd8eeSStefano Zampini } 1376e7b262bdSStefano Zampini /* get sizes */ 13774f1b2e48SStefano Zampini n_constraints = pcbddc->local_primal_size - pcbddc->benign_n - pcbddc->n_vertices; 1378b371cd4fSStefano Zampini n_R = pcis->n - pcbddc->n_vertices; 13796bfb1811SStefano Zampini ierr = VecGetType(pcis->vec1_N,&impVecType);CHKERRQ(ierr); 1380e7b262bdSStefano Zampini /* local work vectors (try to avoid unneeded work)*/ 1381e7b262bdSStefano Zampini /* R nodes */ 1382e7b262bdSStefano Zampini old_size = -1; 1383e7b262bdSStefano Zampini if (pcbddc->vec1_R) { 1384e7b262bdSStefano Zampini ierr = VecGetSize(pcbddc->vec1_R,&old_size);CHKERRQ(ierr); 1385e7b262bdSStefano Zampini } 1386e7b262bdSStefano Zampini if (n_R != old_size) { 1387e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec1_R);CHKERRQ(ierr); 1388e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec2_R);CHKERRQ(ierr); 13896bfb1811SStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_R);CHKERRQ(ierr); 13906bfb1811SStefano Zampini ierr = VecSetSizes(pcbddc->vec1_R,PETSC_DECIDE,n_R);CHKERRQ(ierr); 13916bfb1811SStefano Zampini ierr = VecSetType(pcbddc->vec1_R,impVecType);CHKERRQ(ierr); 13926bfb1811SStefano Zampini ierr = VecDuplicate(pcbddc->vec1_R,&pcbddc->vec2_R);CHKERRQ(ierr); 1393e7b262bdSStefano Zampini } 1394e7b262bdSStefano Zampini /* local primal dofs */ 1395e7b262bdSStefano Zampini old_size = -1; 1396e7b262bdSStefano Zampini if (pcbddc->vec1_P) { 1397e7b262bdSStefano Zampini ierr = VecGetSize(pcbddc->vec1_P,&old_size);CHKERRQ(ierr); 1398e7b262bdSStefano Zampini } 1399e9189074SStefano Zampini if (pcbddc->local_primal_size != old_size) { 1400e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec1_P);CHKERRQ(ierr); 140183b7ccabSStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_P);CHKERRQ(ierr); 1402e9189074SStefano Zampini ierr = VecSetSizes(pcbddc->vec1_P,PETSC_DECIDE,pcbddc->local_primal_size);CHKERRQ(ierr); 14036bfb1811SStefano Zampini ierr = VecSetType(pcbddc->vec1_P,impVecType);CHKERRQ(ierr); 1404e7b262bdSStefano Zampini } 1405e7b262bdSStefano Zampini /* local explicit constraints */ 1406e7b262bdSStefano Zampini old_size = -1; 1407e7b262bdSStefano Zampini if (pcbddc->vec1_C) { 1408e7b262bdSStefano Zampini ierr = VecGetSize(pcbddc->vec1_C,&old_size);CHKERRQ(ierr); 1409e7b262bdSStefano Zampini } 1410e7b262bdSStefano Zampini if (n_constraints && n_constraints != old_size) { 1411e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec1_C);CHKERRQ(ierr); 141283b7ccabSStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_C);CHKERRQ(ierr); 141383b7ccabSStefano Zampini ierr = VecSetSizes(pcbddc->vec1_C,PETSC_DECIDE,n_constraints);CHKERRQ(ierr); 141483b7ccabSStefano Zampini ierr = VecSetType(pcbddc->vec1_C,impVecType);CHKERRQ(ierr); 141583b7ccabSStefano Zampini } 14166bfb1811SStefano Zampini PetscFunctionReturn(0); 14176bfb1811SStefano Zampini } 14186bfb1811SStefano Zampini 14196bfb1811SStefano Zampini #undef __FUNCT__ 142047f4ddc3SStefano Zampini #define __FUNCT__ "PCBDDCSetUpCorrection" 142147f4ddc3SStefano Zampini PetscErrorCode PCBDDCSetUpCorrection(PC pc, PetscScalar **coarse_submat_vals_n) 142288ebb749SStefano Zampini { 142325084f0cSStefano Zampini PetscErrorCode ierr; 142425084f0cSStefano Zampini /* pointers to pcis and pcbddc */ 142588ebb749SStefano Zampini PC_IS* pcis = (PC_IS*)pc->data; 142688ebb749SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 1427d62866d3SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 142825084f0cSStefano Zampini /* submatrices of local problem */ 142980677318SStefano Zampini Mat A_RV,A_VR,A_VV,local_auxmat2_R; 1430fc227af8SStefano Zampini /* submatrices for benign trick */ 1431fc227af8SStefano Zampini Mat B0_R = NULL, B0_V = NULL; 143206656605SStefano Zampini /* submatrices of local coarse problem */ 143306656605SStefano Zampini Mat S_VV,S_CV,S_VC,S_CC; 143425084f0cSStefano Zampini /* working matrices */ 143506656605SStefano Zampini Mat C_CR; 143625084f0cSStefano Zampini /* additional working stuff */ 143706656605SStefano Zampini PC pc_R; 14384f1b2e48SStefano Zampini Mat F; 1439a3df083aSStefano Zampini PetscBool isLU,isCHOL,isILU,need_benign_correction; 144006656605SStefano Zampini 144125084f0cSStefano Zampini PetscScalar *coarse_submat_vals; /* TODO: use a PETSc matrix */ 144206656605SStefano Zampini PetscScalar *work; 144306656605SStefano Zampini PetscInt *idx_V_B; 1444ffd830a3SStefano Zampini PetscInt lda_rhs,n,n_vertices,n_constraints,*p0_lidx_I; 144506656605SStefano Zampini PetscInt i,n_R,n_D,n_B; 1446ffd830a3SStefano Zampini 144725084f0cSStefano Zampini /* some shortcuts to scalars */ 144806656605SStefano Zampini PetscScalar one=1.0,m_one=-1.0; 144988ebb749SStefano Zampini 145088ebb749SStefano Zampini PetscFunctionBegin; 1451ffd830a3SStefano Zampini if (!pcbddc->symmetric_primal && pcbddc->benign_n) { 1452ffd830a3SStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Non-symmetric primal basis computation with benign trick not yet implemented"); 1453ffd830a3SStefano Zampini } 1454ffd830a3SStefano Zampini 1455ffd830a3SStefano Zampini /* Set Non-overlapping dimensions */ 1456b371cd4fSStefano Zampini n_vertices = pcbddc->n_vertices; 14574f1b2e48SStefano Zampini n_constraints = pcbddc->local_primal_size - pcbddc->benign_n - n_vertices; 1458b371cd4fSStefano Zampini n_B = pcis->n_B; 1459b371cd4fSStefano Zampini n_D = pcis->n - n_B; 146088ebb749SStefano Zampini n_R = pcis->n - n_vertices; 146188ebb749SStefano Zampini 146288ebb749SStefano Zampini /* vertices in boundary numbering */ 1463785e854fSJed Brown ierr = PetscMalloc1(n_vertices,&idx_V_B);CHKERRQ(ierr); 14640e6343abSStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,n_vertices,pcbddc->local_primal_ref_node,&i,idx_V_B);CHKERRQ(ierr); 146588ebb749SStefano Zampini if (i != n_vertices) { 146622d5777bSStefano Zampini SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Error in boundary numbering for BDDC vertices! %d != %d\n",n_vertices,i); 146788ebb749SStefano Zampini } 146888ebb749SStefano Zampini 146906656605SStefano Zampini /* Subdomain contribution (Non-overlapping) to coarse matrix */ 1470019a44ceSStefano Zampini ierr = PetscCalloc1(pcbddc->local_primal_size*pcbddc->local_primal_size,&coarse_submat_vals);CHKERRQ(ierr); 147106656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_vertices,n_vertices,coarse_submat_vals,&S_VV);CHKERRQ(ierr); 147206656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_VV,pcbddc->local_primal_size);CHKERRQ(ierr); 147306656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_constraints,n_vertices,coarse_submat_vals+n_vertices,&S_CV);CHKERRQ(ierr); 147406656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_CV,pcbddc->local_primal_size);CHKERRQ(ierr); 147506656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_vertices,n_constraints,coarse_submat_vals+pcbddc->local_primal_size*n_vertices,&S_VC);CHKERRQ(ierr); 147606656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_VC,pcbddc->local_primal_size);CHKERRQ(ierr); 147706656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_constraints,n_constraints,coarse_submat_vals+(pcbddc->local_primal_size+1)*n_vertices,&S_CC);CHKERRQ(ierr); 147806656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_CC,pcbddc->local_primal_size);CHKERRQ(ierr); 147906656605SStefano Zampini 148006656605SStefano Zampini /* determine if can use MatSolve routines instead of calling KSPSolve on ksp_R */ 148106656605SStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_R);CHKERRQ(ierr); 148206656605SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_R,PCLU,&isLU);CHKERRQ(ierr); 148306656605SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_R,PCILU,&isILU);CHKERRQ(ierr); 148406656605SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_R,PCCHOLESKY,&isCHOL);CHKERRQ(ierr); 1485ffd830a3SStefano Zampini lda_rhs = n_R; 1486a3df083aSStefano Zampini need_benign_correction = PETSC_FALSE; 148706656605SStefano Zampini if (isLU || isILU || isCHOL) { 148806656605SStefano Zampini ierr = PCFactorGetMatrix(pc_R,&F);CHKERRQ(ierr); 1489df4d28bfSStefano Zampini } else if (sub_schurs->reuse_solver) { 1490df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 1491d62866d3SStefano Zampini MatFactorType type; 1492d62866d3SStefano Zampini 1493df4d28bfSStefano Zampini F = reuse_solver->F; 14946816873aSStefano Zampini ierr = MatGetFactorType(F,&type);CHKERRQ(ierr); 1495d62866d3SStefano Zampini if (type == MAT_FACTOR_CHOLESKY) isCHOL = PETSC_TRUE; 1496ffd830a3SStefano Zampini ierr = MatGetSize(F,&lda_rhs,NULL);CHKERRQ(ierr); 1497df4d28bfSStefano Zampini need_benign_correction = !!reuse_solver->benign_n; 149806656605SStefano Zampini } else { 149906656605SStefano Zampini F = NULL; 150006656605SStefano Zampini } 150106656605SStefano Zampini 1502ffd830a3SStefano Zampini /* allocate workspace */ 1503ffd830a3SStefano Zampini n = 0; 1504ffd830a3SStefano Zampini if (n_constraints) { 1505ffd830a3SStefano Zampini n += lda_rhs*n_constraints; 1506ffd830a3SStefano Zampini } 1507ffd830a3SStefano Zampini if (n_vertices) { 1508ffd830a3SStefano Zampini n = PetscMax(2*lda_rhs*n_vertices,n); 1509ffd830a3SStefano Zampini n = PetscMax((lda_rhs+n_B)*n_vertices,n); 1510ffd830a3SStefano Zampini } 1511ffd830a3SStefano Zampini ierr = PetscMalloc1(n,&work);CHKERRQ(ierr); 1512ffd830a3SStefano Zampini 151388ebb749SStefano Zampini /* Precompute stuffs needed for preprocessing and application of BDDC*/ 151488ebb749SStefano Zampini if (n_constraints) { 151506656605SStefano Zampini Mat M1,M2,M3; 151680677318SStefano Zampini Mat auxmat; 151706656605SStefano Zampini IS is_aux; 151880677318SStefano Zampini PetscScalar *array,*array2; 151906656605SStefano Zampini 1520f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat1);CHKERRQ(ierr); 152180677318SStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat2);CHKERRQ(ierr); 152288ebb749SStefano Zampini 152325084f0cSStefano Zampini /* Extract constraints on R nodes: C_{CR} */ 152425084f0cSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,n_constraints,n_vertices,1,&is_aux);CHKERRQ(ierr); 15258ce42a96SStefano Zampini ierr = MatGetSubMatrix(pcbddc->ConstraintMatrix,is_aux,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&C_CR);CHKERRQ(ierr); 152680677318SStefano Zampini ierr = MatGetSubMatrix(pcbddc->ConstraintMatrix,is_aux,pcis->is_B_local,MAT_INITIAL_MATRIX,&auxmat);CHKERRQ(ierr); 152788ebb749SStefano Zampini 152880677318SStefano Zampini /* Assemble local_auxmat2_R = (- A_{RR}^{-1} C^T_{CR}) needed by BDDC setup */ 152980677318SStefano Zampini /* Assemble pcbddc->local_auxmat2 = R_to_B (- A_{RR}^{-1} C^T_{CR}) needed by BDDC application */ 1530ffd830a3SStefano Zampini ierr = PetscMemzero(work,lda_rhs*n_constraints*sizeof(PetscScalar));CHKERRQ(ierr); 153188ebb749SStefano Zampini for (i=0;i<n_constraints;i++) { 153206656605SStefano Zampini const PetscScalar *row_cmat_values; 153306656605SStefano Zampini const PetscInt *row_cmat_indices; 153406656605SStefano Zampini PetscInt size_of_constraint,j; 153588ebb749SStefano Zampini 153606656605SStefano Zampini ierr = MatGetRow(C_CR,i,&size_of_constraint,&row_cmat_indices,&row_cmat_values);CHKERRQ(ierr); 153706656605SStefano Zampini for (j=0;j<size_of_constraint;j++) { 1538ffd830a3SStefano Zampini work[row_cmat_indices[j]+i*lda_rhs] = -row_cmat_values[j]; 153906656605SStefano Zampini } 154006656605SStefano Zampini ierr = MatRestoreRow(C_CR,i,&size_of_constraint,&row_cmat_indices,&row_cmat_values);CHKERRQ(ierr); 154106656605SStefano Zampini } 1542ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_constraints,NULL,&local_auxmat2_R);CHKERRQ(ierr); 154306656605SStefano Zampini if (F) { 154406656605SStefano Zampini Mat B; 154506656605SStefano Zampini 1546ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_constraints,work,&B);CHKERRQ(ierr); 1547a3df083aSStefano Zampini if (need_benign_correction) { 1548a3df083aSStefano Zampini PetscScalar *marr; 1549df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 1550a3df083aSStefano Zampini 1551a3df083aSStefano Zampini ierr = MatDenseGetArray(B,&marr);CHKERRQ(ierr); 1552a3df083aSStefano Zampini for (i=0;i<n_constraints;i++) { 1553a3df083aSStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,marr+i*lda_rhs);CHKERRQ(ierr); 1554df4d28bfSStefano Zampini ierr = PCBDDCReuseSolversChangeInterior(reuse_solver,pcbddc->vec1_R,NULL,PETSC_FALSE);CHKERRQ(ierr); 1555a3df083aSStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 1556a3df083aSStefano Zampini } 1557a3df083aSStefano Zampini ierr = MatDenseRestoreArray(B,&marr);CHKERRQ(ierr); 1558a3df083aSStefano Zampini } 155980677318SStefano Zampini ierr = MatMatSolve(F,B,local_auxmat2_R);CHKERRQ(ierr); 1560a3df083aSStefano Zampini if (need_benign_correction) { 1561a3df083aSStefano Zampini PetscScalar *marr; 1562df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 1563a3df083aSStefano Zampini 1564a3df083aSStefano Zampini ierr = MatDenseGetArray(local_auxmat2_R,&marr);CHKERRQ(ierr); 1565a3df083aSStefano Zampini for (i=0;i<n_constraints;i++) { 1566a3df083aSStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,marr+i*lda_rhs);CHKERRQ(ierr); 1567df4d28bfSStefano Zampini ierr = PCBDDCReuseSolversChangeInterior(reuse_solver,pcbddc->vec1_R,NULL,PETSC_TRUE);CHKERRQ(ierr); 1568a3df083aSStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 1569a3df083aSStefano Zampini } 1570a3df083aSStefano Zampini ierr = MatDenseRestoreArray(local_auxmat2_R,&marr);CHKERRQ(ierr); 1571a3df083aSStefano Zampini } 157206656605SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 157306656605SStefano Zampini } else { 157480677318SStefano Zampini PetscScalar *marr; 157580677318SStefano Zampini 157680677318SStefano Zampini ierr = MatDenseGetArray(local_auxmat2_R,&marr);CHKERRQ(ierr); 157706656605SStefano Zampini for (i=0;i<n_constraints;i++) { 1578ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*lda_rhs);CHKERRQ(ierr); 1579ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,marr+i*lda_rhs);CHKERRQ(ierr); 158006656605SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 158106656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 158206656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 158306656605SStefano Zampini } 158480677318SStefano Zampini ierr = MatDenseRestoreArray(local_auxmat2_R,&marr);CHKERRQ(ierr); 158506656605SStefano Zampini } 158680677318SStefano Zampini if (!pcbddc->switch_static) { 158780677318SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,n_constraints,NULL,&pcbddc->local_auxmat2);CHKERRQ(ierr); 158880677318SStefano Zampini ierr = MatDenseGetArray(pcbddc->local_auxmat2,&array);CHKERRQ(ierr); 158980677318SStefano Zampini ierr = MatDenseGetArray(local_auxmat2_R,&array2);CHKERRQ(ierr); 159080677318SStefano Zampini for (i=0;i<n_constraints;i++) { 1591ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,array2+i*lda_rhs);CHKERRQ(ierr); 159280677318SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,array+i*n_B);CHKERRQ(ierr); 159380677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 159480677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 159580677318SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 159680677318SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 159780677318SStefano Zampini } 159880677318SStefano Zampini ierr = MatDenseRestoreArray(local_auxmat2_R,&array2);CHKERRQ(ierr); 159980677318SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->local_auxmat2,&array);CHKERRQ(ierr); 160080677318SStefano Zampini ierr = MatMatMult(auxmat,pcbddc->local_auxmat2,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&M3);CHKERRQ(ierr); 160180677318SStefano Zampini } else { 1602ffd830a3SStefano Zampini if (lda_rhs != n_R) { 1603ffd830a3SStefano Zampini IS dummy; 1604ffd830a3SStefano Zampini 1605ffd830a3SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,n_R,0,1,&dummy);CHKERRQ(ierr); 1606ffd830a3SStefano Zampini ierr = MatGetSubMatrix(local_auxmat2_R,dummy,dummy,MAT_INITIAL_MATRIX,&pcbddc->local_auxmat2);CHKERRQ(ierr); 1607ffd830a3SStefano Zampini ierr = ISDestroy(&dummy);CHKERRQ(ierr); 1608ffd830a3SStefano Zampini } else { 160980677318SStefano Zampini ierr = PetscObjectReference((PetscObject)local_auxmat2_R);CHKERRQ(ierr); 161080677318SStefano Zampini pcbddc->local_auxmat2 = local_auxmat2_R; 1611ffd830a3SStefano Zampini } 161225084f0cSStefano Zampini ierr = MatMatMult(C_CR,pcbddc->local_auxmat2,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&M3);CHKERRQ(ierr); 161380677318SStefano Zampini } 161480677318SStefano Zampini ierr = ISDestroy(&is_aux);CHKERRQ(ierr); 161580677318SStefano Zampini /* Assemble explicitly S_CC = ( C_{CR} A_{RR}^{-1} C^T_{CR} )^{-1} */ 161680677318SStefano Zampini ierr = MatScale(M3,m_one);CHKERRQ(ierr); 161706656605SStefano Zampini ierr = MatDuplicate(M3,MAT_DO_NOT_COPY_VALUES,&M1);CHKERRQ(ierr); 161806656605SStefano Zampini ierr = MatDuplicate(M3,MAT_DO_NOT_COPY_VALUES,&M2);CHKERRQ(ierr); 161980677318SStefano Zampini if (isCHOL) { 162080677318SStefano Zampini ierr = MatCholeskyFactor(M3,NULL,NULL);CHKERRQ(ierr); 162180677318SStefano Zampini } else { 162225084f0cSStefano Zampini ierr = MatLUFactor(M3,NULL,NULL,NULL);CHKERRQ(ierr); 162380677318SStefano Zampini } 162480677318SStefano Zampini ierr = VecSet(pcbddc->vec1_C,one);CHKERRQ(ierr); 162506656605SStefano Zampini ierr = MatDiagonalSet(M2,pcbddc->vec1_C,INSERT_VALUES);CHKERRQ(ierr); 162625084f0cSStefano Zampini ierr = MatMatSolve(M3,M2,M1);CHKERRQ(ierr); 162725084f0cSStefano Zampini ierr = MatDestroy(&M2);CHKERRQ(ierr); 162825084f0cSStefano Zampini ierr = MatDestroy(&M3);CHKERRQ(ierr); 162980677318SStefano Zampini /* Assemble local_auxmat1 = S_CC*C_{CB} needed by BDDC application in KSP and in preproc */ 163080677318SStefano Zampini ierr = MatMatMult(M1,auxmat,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&pcbddc->local_auxmat1);CHKERRQ(ierr); 163180677318SStefano Zampini ierr = MatDestroy(&auxmat);CHKERRQ(ierr); 163206656605SStefano Zampini ierr = MatCopy(M1,S_CC,SAME_NONZERO_PATTERN);CHKERRQ(ierr); /* S_CC can have a different LDA, MatMatSolve doesn't support it */ 163306656605SStefano Zampini ierr = MatDestroy(&M1);CHKERRQ(ierr); 1634f4ddd8eeSStefano Zampini } 1635fc227af8SStefano Zampini 1636fc227af8SStefano Zampini /* Extract B0_R */ 16374f1b2e48SStefano Zampini if (pcbddc->benign_n) { 1638d16cbb6bSStefano Zampini IS dummy; 1639d16cbb6bSStefano Zampini PetscReal norm; 1640d16cbb6bSStefano Zampini 16414f1b2e48SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcbddc->benign_n,0,1,&dummy);CHKERRQ(ierr); 16424f1b2e48SStefano Zampini ierr = MatGetSubMatrix(pcbddc->benign_B0,dummy,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&B0_R);CHKERRQ(ierr); 1643d16cbb6bSStefano Zampini ierr = MatNorm(B0_R,NORM_INFINITY,&norm);CHKERRQ(ierr); 1644fc227af8SStefano Zampini if (norm < PETSC_SMALL) { 1645d16cbb6bSStefano Zampini ierr = MatDestroy(&B0_R);CHKERRQ(ierr); 1646fc227af8SStefano Zampini } 1647d16cbb6bSStefano Zampini ierr = ISDestroy(&dummy);CHKERRQ(ierr); 1648d16cbb6bSStefano Zampini } 1649d16cbb6bSStefano Zampini 1650fc227af8SStefano Zampini /* Get submatrices from subdomain matrix */ 165188ebb749SStefano Zampini if (n_vertices) { 165206656605SStefano Zampini IS is_aux; 16533a50541eSStefano Zampini 1654df4d28bfSStefano Zampini if (sub_schurs->reuse_solver) { /* is_R_local is not sorted, ISComplement doesn't like it */ 16556816873aSStefano Zampini IS tis; 16566816873aSStefano Zampini 16576816873aSStefano Zampini ierr = ISDuplicate(pcbddc->is_R_local,&tis);CHKERRQ(ierr); 16586816873aSStefano Zampini ierr = ISSort(tis);CHKERRQ(ierr); 16596816873aSStefano Zampini ierr = ISComplement(tis,0,pcis->n,&is_aux);CHKERRQ(ierr); 16606816873aSStefano Zampini ierr = ISDestroy(&tis);CHKERRQ(ierr); 16616816873aSStefano Zampini } else { 16623a50541eSStefano Zampini ierr = ISComplement(pcbddc->is_R_local,0,pcis->n,&is_aux);CHKERRQ(ierr); 16636816873aSStefano Zampini } 16649577ea80SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,pcbddc->is_R_local,is_aux,MAT_INITIAL_MATRIX,&A_RV);CHKERRQ(ierr); 16659577ea80SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,is_aux,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&A_VR);CHKERRQ(ierr); 166604708bb6SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,is_aux,is_aux,MAT_INITIAL_MATRIX,&A_VV);CHKERRQ(ierr); 16674f1b2e48SStefano Zampini if (pcbddc->benign_n) { 1668019a44ceSStefano Zampini IS dummy; 1669019a44ceSStefano Zampini 16704f1b2e48SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcbddc->benign_n,0,1,&dummy);CHKERRQ(ierr); 16714f1b2e48SStefano Zampini ierr = MatGetSubMatrix(pcbddc->benign_B0,dummy,is_aux,MAT_INITIAL_MATRIX,&B0_V);CHKERRQ(ierr); 1672019a44ceSStefano Zampini ierr = ISDestroy(&dummy);CHKERRQ(ierr); 1673019a44ceSStefano Zampini } 167425084f0cSStefano Zampini ierr = ISDestroy(&is_aux);CHKERRQ(ierr); 167588ebb749SStefano Zampini } 167688ebb749SStefano Zampini 167788ebb749SStefano Zampini /* Matrix of coarse basis functions (local) */ 1678f4ddd8eeSStefano Zampini if (pcbddc->coarse_phi_B) { 167906656605SStefano Zampini PetscInt on_B,on_primal,on_D=n_D; 168006656605SStefano Zampini if (pcbddc->coarse_phi_D) { 168106656605SStefano Zampini ierr = MatGetSize(pcbddc->coarse_phi_D,&on_D,NULL);CHKERRQ(ierr); 168206656605SStefano Zampini } 1683f4ddd8eeSStefano Zampini ierr = MatGetSize(pcbddc->coarse_phi_B,&on_B,&on_primal);CHKERRQ(ierr); 168406656605SStefano Zampini if (on_B != n_B || on_primal != pcbddc->local_primal_size || on_D != n_D) { 168506656605SStefano Zampini PetscScalar *marray; 168606656605SStefano Zampini 168706656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&marray);CHKERRQ(ierr); 168806656605SStefano Zampini ierr = PetscFree(marray);CHKERRQ(ierr); 1689f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_B);CHKERRQ(ierr); 1690f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_B);CHKERRQ(ierr); 1691f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_D);CHKERRQ(ierr); 1692f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_D);CHKERRQ(ierr); 1693f4ddd8eeSStefano Zampini } 1694f4ddd8eeSStefano Zampini } 169506656605SStefano Zampini 1696f4ddd8eeSStefano Zampini if (!pcbddc->coarse_phi_B) { 169706656605SStefano Zampini PetscScalar *marray; 169888ebb749SStefano Zampini 169906656605SStefano Zampini n = n_B*pcbddc->local_primal_size; 17008eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 170106656605SStefano Zampini n += n_D*pcbddc->local_primal_size; 170288ebb749SStefano Zampini } 17033301b35fSStefano Zampini if (!pcbddc->symmetric_primal) { 170406656605SStefano Zampini n *= 2; 170588ebb749SStefano Zampini } 170606656605SStefano Zampini ierr = PetscCalloc1(n,&marray);CHKERRQ(ierr); 170706656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,pcbddc->local_primal_size,marray,&pcbddc->coarse_phi_B);CHKERRQ(ierr); 170806656605SStefano Zampini n = n_B*pcbddc->local_primal_size; 17098eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 171006656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_D,pcbddc->local_primal_size,marray+n,&pcbddc->coarse_phi_D);CHKERRQ(ierr); 171106656605SStefano Zampini n += n_D*pcbddc->local_primal_size; 171288ebb749SStefano Zampini } 17133301b35fSStefano Zampini if (!pcbddc->symmetric_primal) { 171406656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,pcbddc->local_primal_size,marray+n,&pcbddc->coarse_psi_B);CHKERRQ(ierr); 17158eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 171606656605SStefano Zampini n = n_B*pcbddc->local_primal_size; 171706656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_D,pcbddc->local_primal_size,marray+n,&pcbddc->coarse_psi_D);CHKERRQ(ierr); 171888ebb749SStefano Zampini } 171988ebb749SStefano Zampini } else { 1720c0553b1fSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->coarse_phi_B);CHKERRQ(ierr); 1721c0553b1fSStefano Zampini pcbddc->coarse_psi_B = pcbddc->coarse_phi_B; 17221b968477SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 1723c0553b1fSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->coarse_phi_D);CHKERRQ(ierr); 1724c0553b1fSStefano Zampini pcbddc->coarse_psi_D = pcbddc->coarse_phi_D; 1725c0553b1fSStefano Zampini } 172688ebb749SStefano Zampini } 172706656605SStefano Zampini } 1728019a44ceSStefano Zampini 172906656605SStefano Zampini /* We are now ready to evaluate coarse basis functions and subdomain contribution to coarse problem */ 17304f1b2e48SStefano Zampini p0_lidx_I = NULL; 17314f1b2e48SStefano Zampini if (pcbddc->benign_n && (pcbddc->switch_static || pcbddc->dbg_flag)) { 1732d12edf2fSStefano Zampini const PetscInt *idxs; 1733d12edf2fSStefano Zampini 1734d12edf2fSStefano Zampini ierr = ISGetIndices(pcis->is_I_local,&idxs);CHKERRQ(ierr); 17354f1b2e48SStefano Zampini ierr = PetscMalloc1(pcbddc->benign_n,&p0_lidx_I);CHKERRQ(ierr); 17364f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 17374f1b2e48SStefano Zampini ierr = PetscFindInt(pcbddc->benign_p0_lidx[i],pcis->n-pcis->n_B,idxs,&p0_lidx_I[i]);CHKERRQ(ierr); 17384f1b2e48SStefano Zampini } 1739d12edf2fSStefano Zampini ierr = ISRestoreIndices(pcis->is_I_local,&idxs);CHKERRQ(ierr); 1740d12edf2fSStefano Zampini } 1741d16cbb6bSStefano Zampini 174206656605SStefano Zampini /* vertices */ 174306656605SStefano Zampini if (n_vertices) { 174416f15bc4SStefano Zampini 1745ffd830a3SStefano Zampini ierr = MatConvert(A_VV,MATDENSE,MAT_REUSE_MATRIX,&A_VV);CHKERRQ(ierr); 174604708bb6SStefano Zampini 174716f15bc4SStefano Zampini if (n_R) { 174814393ed6SStefano Zampini Mat A_RRmA_RV,A_RV_bcorr=NULL,S_VVt; /* S_VVt with LDA=N */ 174906656605SStefano Zampini PetscBLASInt B_N,B_one = 1; 175016f15bc4SStefano Zampini PetscScalar *x,*y; 175104708bb6SStefano Zampini PetscBool isseqaij; 175206656605SStefano Zampini 175321eccb56SStefano Zampini ierr = MatScale(A_RV,m_one);CHKERRQ(ierr); 175414393ed6SStefano Zampini if (need_benign_correction) { 175514393ed6SStefano Zampini ISLocalToGlobalMapping RtoN; 175614393ed6SStefano Zampini IS is_p0; 175714393ed6SStefano Zampini PetscInt *idxs_p0,n; 175814393ed6SStefano Zampini 175914393ed6SStefano Zampini ierr = PetscMalloc1(pcbddc->benign_n,&idxs_p0);CHKERRQ(ierr); 176014393ed6SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(pcbddc->is_R_local,&RtoN);CHKERRQ(ierr); 176114393ed6SStefano Zampini ierr = ISGlobalToLocalMappingApply(RtoN,IS_GTOLM_DROP,pcbddc->benign_n,pcbddc->benign_p0_lidx,&n,idxs_p0);CHKERRQ(ierr); 176214393ed6SStefano Zampini if (n != pcbddc->benign_n) { 176314393ed6SStefano Zampini SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Error in R numbering for benign p0! %d != %d\n",n,pcbddc->benign_n); 176414393ed6SStefano Zampini } 176514393ed6SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&RtoN);CHKERRQ(ierr); 176614393ed6SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,n,idxs_p0,PETSC_OWN_POINTER,&is_p0);CHKERRQ(ierr); 176714393ed6SStefano Zampini ierr = MatGetSubMatrix(A_RV,is_p0,NULL,MAT_INITIAL_MATRIX,&A_RV_bcorr);CHKERRQ(ierr); 176814393ed6SStefano Zampini ierr = ISDestroy(&is_p0);CHKERRQ(ierr); 176914393ed6SStefano Zampini } 177014393ed6SStefano Zampini 1771ffd830a3SStefano Zampini if (lda_rhs == n_R) { 1772ffd830a3SStefano Zampini ierr = MatConvert(A_RV,MATDENSE,MAT_REUSE_MATRIX,&A_RV);CHKERRQ(ierr); 1773ffd830a3SStefano Zampini } else { 1774ca92afb2SStefano Zampini PetscScalar *av,*array; 1775ca92afb2SStefano Zampini const PetscInt *xadj,*adjncy; 1776ca92afb2SStefano Zampini PetscInt n; 1777ca92afb2SStefano Zampini PetscBool flg_row; 1778ffd830a3SStefano Zampini 1779ca92afb2SStefano Zampini array = work+lda_rhs*n_vertices; 1780ca92afb2SStefano Zampini ierr = PetscMemzero(array,lda_rhs*n_vertices*sizeof(PetscScalar));CHKERRQ(ierr); 1781ca92afb2SStefano Zampini ierr = MatConvert(A_RV,MATSEQAIJ,MAT_REUSE_MATRIX,&A_RV);CHKERRQ(ierr); 1782ca92afb2SStefano Zampini ierr = MatGetRowIJ(A_RV,0,PETSC_FALSE,PETSC_FALSE,&n,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 1783ca92afb2SStefano Zampini ierr = MatSeqAIJGetArray(A_RV,&av);CHKERRQ(ierr); 1784ca92afb2SStefano Zampini for (i=0;i<n;i++) { 1785ca92afb2SStefano Zampini PetscInt j; 1786ca92afb2SStefano Zampini for (j=xadj[i];j<xadj[i+1];j++) array[lda_rhs*adjncy[j]+i] = av[j]; 1787ffd830a3SStefano Zampini } 1788ca92afb2SStefano Zampini ierr = MatRestoreRowIJ(A_RV,0,PETSC_FALSE,PETSC_FALSE,&n,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 1789ca92afb2SStefano Zampini ierr = MatDestroy(&A_RV);CHKERRQ(ierr); 1790ca92afb2SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_vertices,array,&A_RV);CHKERRQ(ierr); 1791ffd830a3SStefano Zampini } 1792ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_vertices,work,&A_RRmA_RV);CHKERRQ(ierr); 179314393ed6SStefano Zampini if (F) { 1794a3df083aSStefano Zampini if (need_benign_correction) { 1795df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 1796a3df083aSStefano Zampini PetscScalar *marr; 1797a3df083aSStefano Zampini 1798a3df083aSStefano Zampini ierr = MatDenseGetArray(A_RV,&marr);CHKERRQ(ierr); 179914393ed6SStefano Zampini /* need \Phi^T A_RV = (I+L)A_RV, L given by 180014393ed6SStefano Zampini 180114393ed6SStefano Zampini | 0 0 0 | (V) 180214393ed6SStefano Zampini L = | 0 0 -1 | (P-p0) 180314393ed6SStefano Zampini | 0 0 -1 | (p0) 180414393ed6SStefano Zampini 180514393ed6SStefano Zampini */ 1806df4d28bfSStefano Zampini for (i=0;i<reuse_solver->benign_n;i++) { 180714393ed6SStefano Zampini const PetscScalar *vals; 180814393ed6SStefano Zampini const PetscInt *idxs,*idxs_zero; 180914393ed6SStefano Zampini PetscInt n,j,nz; 181014393ed6SStefano Zampini 1811df4d28bfSStefano Zampini ierr = ISGetLocalSize(reuse_solver->benign_zerodiag_subs[i],&nz);CHKERRQ(ierr); 1812df4d28bfSStefano Zampini ierr = ISGetIndices(reuse_solver->benign_zerodiag_subs[i],&idxs_zero);CHKERRQ(ierr); 181314393ed6SStefano Zampini ierr = MatGetRow(A_RV_bcorr,i,&n,&idxs,&vals);CHKERRQ(ierr); 181414393ed6SStefano Zampini for (j=0;j<n;j++) { 181514393ed6SStefano Zampini PetscScalar val = vals[j]; 181614393ed6SStefano Zampini PetscInt k,col = idxs[j]; 181714393ed6SStefano Zampini for (k=0;k<nz;k++) marr[idxs_zero[k]+lda_rhs*col] -= val; 181814393ed6SStefano Zampini } 181914393ed6SStefano Zampini ierr = MatRestoreRow(A_RV_bcorr,i,&n,&idxs,&vals);CHKERRQ(ierr); 1820df4d28bfSStefano Zampini ierr = ISRestoreIndices(reuse_solver->benign_zerodiag_subs[i],&idxs_zero);CHKERRQ(ierr); 182114393ed6SStefano Zampini } 182214393ed6SStefano Zampini /* need to correct the rhs */ 1823a3df083aSStefano Zampini for (i=0;i<n_vertices;i++) { 1824a3df083aSStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,marr+i*lda_rhs);CHKERRQ(ierr); 1825df4d28bfSStefano Zampini ierr = PCBDDCReuseSolversChangeInterior(reuse_solver,pcbddc->vec1_R,NULL,PETSC_FALSE);CHKERRQ(ierr); 1826a3df083aSStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 1827a3df083aSStefano Zampini } 1828a3df083aSStefano Zampini ierr = MatDenseRestoreArray(A_RV,&marr);CHKERRQ(ierr); 1829a3df083aSStefano Zampini } 183006656605SStefano Zampini ierr = MatMatSolve(F,A_RV,A_RRmA_RV);CHKERRQ(ierr); 183114393ed6SStefano Zampini /* need to correct the solution */ 1832a3df083aSStefano Zampini if (need_benign_correction) { 1833df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 1834a3df083aSStefano Zampini PetscScalar *marr; 1835a3df083aSStefano Zampini 1836a3df083aSStefano Zampini ierr = MatDenseGetArray(A_RRmA_RV,&marr);CHKERRQ(ierr); 1837a3df083aSStefano Zampini for (i=0;i<n_vertices;i++) { 1838a3df083aSStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,marr+i*lda_rhs);CHKERRQ(ierr); 1839df4d28bfSStefano Zampini ierr = PCBDDCReuseSolversChangeInterior(reuse_solver,pcbddc->vec1_R,NULL,PETSC_TRUE);CHKERRQ(ierr); 1840a3df083aSStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 1841a3df083aSStefano Zampini } 1842a3df083aSStefano Zampini ierr = MatDenseRestoreArray(A_RRmA_RV,&marr);CHKERRQ(ierr); 1843a3df083aSStefano Zampini } 184406656605SStefano Zampini } else { 184506656605SStefano Zampini ierr = MatDenseGetArray(A_RV,&y);CHKERRQ(ierr); 184606656605SStefano Zampini for (i=0;i<n_vertices;i++) { 1847ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,y+i*lda_rhs);CHKERRQ(ierr); 1848ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,work+i*lda_rhs);CHKERRQ(ierr); 184906656605SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 185006656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 185106656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 185206656605SStefano Zampini } 185306656605SStefano Zampini ierr = MatDenseRestoreArray(A_RV,&y);CHKERRQ(ierr); 185406656605SStefano Zampini } 185580677318SStefano Zampini ierr = MatDestroy(&A_RV);CHKERRQ(ierr); 1856ffd830a3SStefano Zampini /* S_VV and S_CV */ 185706656605SStefano Zampini if (n_constraints) { 185806656605SStefano Zampini Mat B; 185980677318SStefano Zampini 1860ffd830a3SStefano Zampini ierr = PetscMemzero(work+lda_rhs*n_vertices,n_B*n_vertices*sizeof(PetscScalar));CHKERRQ(ierr); 186180677318SStefano Zampini for (i=0;i<n_vertices;i++) { 1862ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*lda_rhs);CHKERRQ(ierr); 1863ffd830a3SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,work+lda_rhs*n_vertices+i*n_B);CHKERRQ(ierr); 186480677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 186580677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 186680677318SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 186780677318SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 186880677318SStefano Zampini } 1869ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,n_vertices,work+lda_rhs*n_vertices,&B);CHKERRQ(ierr); 187080677318SStefano Zampini ierr = MatMatMult(pcbddc->local_auxmat1,B,MAT_REUSE_MATRIX,PETSC_DEFAULT,&S_CV);CHKERRQ(ierr); 187180677318SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 1872ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_vertices,work+lda_rhs*n_vertices,&B);CHKERRQ(ierr); 187380677318SStefano Zampini ierr = MatMatMult(local_auxmat2_R,S_CV,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B);CHKERRQ(ierr); 187406656605SStefano Zampini ierr = MatScale(S_CV,m_one);CHKERRQ(ierr); 1875ffd830a3SStefano Zampini ierr = PetscBLASIntCast(lda_rhs*n_vertices,&B_N);CHKERRQ(ierr); 1876ffd830a3SStefano Zampini PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&B_N,&one,work+lda_rhs*n_vertices,&B_one,work,&B_one)); 187706656605SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 187806656605SStefano Zampini } 187904708bb6SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)A_VR,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 188004708bb6SStefano Zampini if (!isseqaij) { /* MatMatMult with SEQ(S)BAIJ below will raise an error */ 188104708bb6SStefano Zampini ierr = MatConvert(A_VR,MATSEQAIJ,MAT_REUSE_MATRIX,&A_VR);CHKERRQ(ierr); 188204708bb6SStefano Zampini } 1883ffd830a3SStefano Zampini if (lda_rhs != n_R) { 1884ffd830a3SStefano Zampini ierr = MatDestroy(&A_RRmA_RV);CHKERRQ(ierr); 1885ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_vertices,work,&A_RRmA_RV);CHKERRQ(ierr); 1886ffd830a3SStefano Zampini ierr = MatSeqDenseSetLDA(A_RRmA_RV,lda_rhs);CHKERRQ(ierr); 1887ffd830a3SStefano Zampini } 188806656605SStefano Zampini ierr = MatMatMult(A_VR,A_RRmA_RV,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&S_VVt);CHKERRQ(ierr); 188914393ed6SStefano Zampini /* need A_VR * \Phi * A_RRmA_RV = A_VR * (I+L)^T * A_RRmA_RV, L given as before */ 189014393ed6SStefano Zampini if (need_benign_correction) { 1891df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 189214393ed6SStefano Zampini PetscScalar *marr,*sums; 189314393ed6SStefano Zampini 189414393ed6SStefano Zampini ierr = PetscMalloc1(n_vertices,&sums);CHKERRQ(ierr); 189514393ed6SStefano Zampini ierr = MatDenseGetArray(S_VVt,&marr); 1896df4d28bfSStefano Zampini for (i=0;i<reuse_solver->benign_n;i++) { 189714393ed6SStefano Zampini const PetscScalar *vals; 189814393ed6SStefano Zampini const PetscInt *idxs,*idxs_zero; 189914393ed6SStefano Zampini PetscInt n,j,nz; 190014393ed6SStefano Zampini 1901df4d28bfSStefano Zampini ierr = ISGetLocalSize(reuse_solver->benign_zerodiag_subs[i],&nz);CHKERRQ(ierr); 1902df4d28bfSStefano Zampini ierr = ISGetIndices(reuse_solver->benign_zerodiag_subs[i],&idxs_zero);CHKERRQ(ierr); 190314393ed6SStefano Zampini for (j=0;j<n_vertices;j++) { 190414393ed6SStefano Zampini PetscInt k; 190514393ed6SStefano Zampini sums[j] = 0.; 190614393ed6SStefano Zampini for (k=0;k<nz;k++) sums[j] += work[idxs_zero[k]+j*lda_rhs]; 190714393ed6SStefano Zampini } 190814393ed6SStefano Zampini ierr = MatGetRow(A_RV_bcorr,i,&n,&idxs,&vals);CHKERRQ(ierr); 190914393ed6SStefano Zampini for (j=0;j<n;j++) { 191014393ed6SStefano Zampini PetscScalar val = vals[j]; 191114393ed6SStefano Zampini PetscInt k; 191214393ed6SStefano Zampini for (k=0;k<n_vertices;k++) { 191314393ed6SStefano Zampini marr[idxs[j]+k*n_vertices] += val*sums[k]; 191414393ed6SStefano Zampini } 191514393ed6SStefano Zampini } 191614393ed6SStefano Zampini ierr = MatRestoreRow(A_RV_bcorr,i,&n,&idxs,&vals);CHKERRQ(ierr); 1917df4d28bfSStefano Zampini ierr = ISRestoreIndices(reuse_solver->benign_zerodiag_subs[i],&idxs_zero);CHKERRQ(ierr); 191814393ed6SStefano Zampini } 191914393ed6SStefano Zampini ierr = PetscFree(sums);CHKERRQ(ierr); 192014393ed6SStefano Zampini ierr = MatDenseRestoreArray(S_VVt,&marr); 192114393ed6SStefano Zampini ierr = MatDestroy(&A_RV_bcorr);CHKERRQ(ierr); 192214393ed6SStefano Zampini } 192380677318SStefano Zampini ierr = MatDestroy(&A_RRmA_RV);CHKERRQ(ierr); 192406656605SStefano Zampini ierr = PetscBLASIntCast(n_vertices*n_vertices,&B_N);CHKERRQ(ierr); 192506656605SStefano Zampini ierr = MatDenseGetArray(A_VV,&x);CHKERRQ(ierr); 192606656605SStefano Zampini ierr = MatDenseGetArray(S_VVt,&y);CHKERRQ(ierr); 192706656605SStefano Zampini PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&B_N,&one,x,&B_one,y,&B_one)); 192806656605SStefano Zampini ierr = MatDenseRestoreArray(A_VV,&x);CHKERRQ(ierr); 192906656605SStefano Zampini ierr = MatDenseRestoreArray(S_VVt,&y);CHKERRQ(ierr); 193006656605SStefano Zampini ierr = MatCopy(S_VVt,S_VV,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 1931d16cbb6bSStefano Zampini ierr = MatDestroy(&S_VVt);CHKERRQ(ierr); 1932019a44ceSStefano Zampini } else { 1933d16cbb6bSStefano Zampini ierr = MatCopy(A_VV,S_VV,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 1934d16cbb6bSStefano Zampini } 19354f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 1936019a44ceSStefano Zampini const PetscScalar *vals; 1937019a44ceSStefano Zampini const PetscInt *idxs; 19384f1b2e48SStefano Zampini PetscInt n,j,primal_idx; 1939019a44ceSStefano Zampini 19404f1b2e48SStefano Zampini ierr = MatGetRow(B0_V,i,&n,&idxs,&vals);CHKERRQ(ierr); 19414f1b2e48SStefano Zampini primal_idx = pcbddc->local_primal_size - pcbddc->benign_n + i; 1942d16cbb6bSStefano Zampini for (j=0;j<n;j++) { 19434f1b2e48SStefano Zampini coarse_submat_vals[primal_idx*pcbddc->local_primal_size+idxs[j]] = vals[j]; 19444f1b2e48SStefano Zampini coarse_submat_vals[idxs[j]*pcbddc->local_primal_size+primal_idx] = vals[j]; 1945019a44ceSStefano Zampini } 19464f1b2e48SStefano Zampini ierr = MatRestoreRow(B0_V,i,&n,&idxs,&vals);CHKERRQ(ierr); 194716f15bc4SStefano Zampini } 194821eccb56SStefano Zampini ierr = MatDestroy(&A_VV);CHKERRQ(ierr); 1949d16cbb6bSStefano Zampini 195006656605SStefano Zampini /* coarse basis functions */ 195106656605SStefano Zampini for (i=0;i<n_vertices;i++) { 195216f15bc4SStefano Zampini PetscScalar *y; 195316f15bc4SStefano Zampini 1954ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+lda_rhs*i);CHKERRQ(ierr); 195506656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 195606656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*i);CHKERRQ(ierr); 195706656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 195806656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 195906656605SStefano Zampini y[n_B*i+idx_V_B[i]] = 1.0; 196006656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 196106656605SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 196206656605SStefano Zampini 196306656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 19644f1b2e48SStefano Zampini PetscInt j; 19654f1b2e48SStefano Zampini 196606656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 196706656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*i);CHKERRQ(ierr); 196806656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 196906656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 197006656605SStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 19714f1b2e48SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) y[n_D*i+p0_lidx_I[j]] = 0.0; 197206656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 197306656605SStefano Zampini } 197406656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 197506656605SStefano Zampini } 197604708bb6SStefano Zampini /* if n_R == 0 the object is not destroyed */ 197704708bb6SStefano Zampini ierr = MatDestroy(&A_RV);CHKERRQ(ierr); 197806656605SStefano Zampini } 197906656605SStefano Zampini 198006656605SStefano Zampini if (n_constraints) { 198106656605SStefano Zampini Mat B; 198206656605SStefano Zampini 1983ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_constraints,work,&B);CHKERRQ(ierr); 198406656605SStefano Zampini ierr = MatScale(S_CC,m_one);CHKERRQ(ierr); 198580677318SStefano Zampini ierr = MatMatMult(local_auxmat2_R,S_CC,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B);CHKERRQ(ierr); 198606656605SStefano Zampini ierr = MatScale(S_CC,m_one);CHKERRQ(ierr); 198706656605SStefano Zampini if (n_vertices) { 198880677318SStefano Zampini if (isCHOL) { /* if we can solve the interior problem with cholesky, we should also be fine with transposing here */ 198980677318SStefano Zampini ierr = MatTranspose(S_CV,MAT_REUSE_MATRIX,&S_VC);CHKERRQ(ierr); 199080677318SStefano Zampini } else { 199180677318SStefano Zampini Mat S_VCt; 199280677318SStefano Zampini 1993ffd830a3SStefano Zampini if (lda_rhs != n_R) { 1994ffd830a3SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 1995ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_vertices,work,&B);CHKERRQ(ierr); 1996ffd830a3SStefano Zampini ierr = MatSeqDenseSetLDA(B,lda_rhs);CHKERRQ(ierr); 1997ffd830a3SStefano Zampini } 199880677318SStefano Zampini ierr = MatMatMult(A_VR,B,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&S_VCt);CHKERRQ(ierr); 199980677318SStefano Zampini ierr = MatCopy(S_VCt,S_VC,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 200080677318SStefano Zampini ierr = MatDestroy(&S_VCt);CHKERRQ(ierr); 200180677318SStefano Zampini } 200206656605SStefano Zampini } 200306656605SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 200406656605SStefano Zampini /* coarse basis functions */ 200506656605SStefano Zampini for (i=0;i<n_constraints;i++) { 200606656605SStefano Zampini PetscScalar *y; 200706656605SStefano Zampini 2008ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+lda_rhs*i);CHKERRQ(ierr); 200906656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 201006656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*(i+n_vertices));CHKERRQ(ierr); 201106656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 201206656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 201306656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 201406656605SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 201506656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 20164f1b2e48SStefano Zampini PetscInt j; 20174f1b2e48SStefano Zampini 201806656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 201906656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*(i+n_vertices));CHKERRQ(ierr); 202006656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 202106656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 202206656605SStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 20234f1b2e48SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) y[n_D*i+p0_lidx_I[j]] = 0.0; 202406656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 202506656605SStefano Zampini } 2026fc227af8SStefano Zampini if (B0_R) { 2027fc227af8SStefano Zampini Vec benign_proj; 2028fc227af8SStefano Zampini const PetscScalar *array; 2029fc227af8SStefano Zampini PetscInt j; 2030fc227af8SStefano Zampini 2031fc227af8SStefano Zampini ierr = MatCreateVecs(B0_R,NULL,&benign_proj);CHKERRQ(ierr); 2032fc227af8SStefano Zampini ierr = MatMult(B0_R,pcbddc->vec1_R,benign_proj);CHKERRQ(ierr); 2033fc227af8SStefano Zampini ierr = VecGetArrayRead(benign_proj,&array);CHKERRQ(ierr); 2034fc227af8SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) { 2035fc227af8SStefano Zampini PetscInt primal_idx = pcbddc->local_primal_size - pcbddc->benign_n + j; 2036fc227af8SStefano Zampini coarse_submat_vals[primal_idx*pcbddc->local_primal_size+i+n_vertices] = array[j]; 2037fc227af8SStefano Zampini coarse_submat_vals[(i+n_vertices)*pcbddc->local_primal_size+primal_idx] = array[j]; 2038fc227af8SStefano Zampini } 2039fc227af8SStefano Zampini ierr = VecRestoreArrayRead(benign_proj,&array);CHKERRQ(ierr); 2040fc227af8SStefano Zampini ierr = VecDestroy(&benign_proj);CHKERRQ(ierr); 2041fc227af8SStefano Zampini } 204206656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 204306656605SStefano Zampini } 204406656605SStefano Zampini } 204580677318SStefano Zampini if (n_constraints) { 204680677318SStefano Zampini ierr = MatDestroy(&local_auxmat2_R);CHKERRQ(ierr); 204780677318SStefano Zampini } 20484f1b2e48SStefano Zampini ierr = PetscFree(p0_lidx_I);CHKERRQ(ierr); 2049019a44ceSStefano Zampini ierr = MatDestroy(&B0_V);CHKERRQ(ierr); 2050fc227af8SStefano Zampini ierr = MatDestroy(&B0_R);CHKERRQ(ierr); 2051019a44ceSStefano Zampini 205206656605SStefano Zampini /* compute other basis functions for non-symmetric problems */ 20533301b35fSStefano Zampini if (!pcbddc->symmetric_primal) { 2054ffd830a3SStefano Zampini Mat B_V=NULL,B_C=NULL; 2055ffd830a3SStefano Zampini PetscScalar *marray; 205606656605SStefano Zampini 205706656605SStefano Zampini if (n_constraints) { 2058ffd830a3SStefano Zampini Mat S_CCT,C_CRT; 205906656605SStefano Zampini 2060ffd830a3SStefano Zampini ierr = MatTranspose(C_CR,MAT_INITIAL_MATRIX,&C_CRT);CHKERRQ(ierr); 206106656605SStefano Zampini ierr = MatTranspose(S_CC,MAT_INITIAL_MATRIX,&S_CCT);CHKERRQ(ierr); 2062ffd830a3SStefano Zampini ierr = MatMatMult(C_CRT,S_CCT,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&B_C);CHKERRQ(ierr); 206316f15bc4SStefano Zampini ierr = MatDestroy(&S_CCT);CHKERRQ(ierr); 206406656605SStefano Zampini if (n_vertices) { 2065ffd830a3SStefano Zampini Mat S_VCT; 206606656605SStefano Zampini 206706656605SStefano Zampini ierr = MatTranspose(S_VC,MAT_INITIAL_MATRIX,&S_VCT);CHKERRQ(ierr); 2068ffd830a3SStefano Zampini ierr = MatMatMult(C_CRT,S_VCT,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&B_V);CHKERRQ(ierr); 206916f15bc4SStefano Zampini ierr = MatDestroy(&S_VCT);CHKERRQ(ierr); 207006656605SStefano Zampini } 2071ffd830a3SStefano Zampini ierr = MatDestroy(&C_CRT);CHKERRQ(ierr); 207206656605SStefano Zampini } 207316f15bc4SStefano Zampini if (n_vertices && n_R) { 2074ffd830a3SStefano Zampini PetscScalar *av,*marray; 2075ffd830a3SStefano Zampini const PetscInt *xadj,*adjncy; 2076ffd830a3SStefano Zampini PetscInt n; 2077ffd830a3SStefano Zampini PetscBool flg_row; 207806656605SStefano Zampini 2079ffd830a3SStefano Zampini /* B_V = B_V - A_VR^T */ 2080ffd830a3SStefano Zampini ierr = MatConvert(A_VR,MATSEQAIJ,MAT_REUSE_MATRIX,&A_VR);CHKERRQ(ierr); 2081ffd830a3SStefano Zampini ierr = MatGetRowIJ(A_VR,0,PETSC_FALSE,PETSC_FALSE,&n,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 2082ffd830a3SStefano Zampini ierr = MatSeqAIJGetArray(A_VR,&av);CHKERRQ(ierr); 2083ffd830a3SStefano Zampini ierr = MatDenseGetArray(B_V,&marray);CHKERRQ(ierr); 2084ffd830a3SStefano Zampini for (i=0;i<n;i++) { 2085ffd830a3SStefano Zampini PetscInt j; 2086ffd830a3SStefano Zampini for (j=xadj[i];j<xadj[i+1];j++) marray[i*n_R + adjncy[j]] -= av[j]; 2087ffd830a3SStefano Zampini } 2088ffd830a3SStefano Zampini ierr = MatDenseRestoreArray(B_V,&marray);CHKERRQ(ierr); 2089ffd830a3SStefano Zampini ierr = MatRestoreRowIJ(A_VR,0,PETSC_FALSE,PETSC_FALSE,&n,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 2090ffd830a3SStefano Zampini ierr = MatDestroy(&A_VR);CHKERRQ(ierr); 209106656605SStefano Zampini } 209206656605SStefano Zampini 2093ffd830a3SStefano Zampini /* currently there's no support for MatTransposeMatSolve(F,B,X) */ 2094ffd830a3SStefano Zampini ierr = MatDenseGetArray(B_V,&marray);CHKERRQ(ierr); 2095ffd830a3SStefano Zampini for (i=0;i<n_vertices;i++) { 2096ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,marray+i*n_R);CHKERRQ(ierr); 2097ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,work+i*n_R);CHKERRQ(ierr); 209806656605SStefano Zampini ierr = KSPSolveTranspose(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 209906656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 210006656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 210106656605SStefano Zampini } 2102ffd830a3SStefano Zampini ierr = MatDenseRestoreArray(B_V,&marray);CHKERRQ(ierr); 2103ffd830a3SStefano Zampini ierr = MatDenseGetArray(B_C,&marray);CHKERRQ(ierr); 2104ffd830a3SStefano Zampini for (i=n_vertices;i<n_constraints+n_vertices;i++) { 2105ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,marray+(i-n_vertices)*n_R);CHKERRQ(ierr); 2106ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,work+i*n_R);CHKERRQ(ierr); 2107ffd830a3SStefano Zampini ierr = KSPSolveTranspose(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 2108ffd830a3SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 2109ffd830a3SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 211006656605SStefano Zampini } 2111ffd830a3SStefano Zampini ierr = MatDenseRestoreArray(B_C,&marray);CHKERRQ(ierr); 211206656605SStefano Zampini /* coarse basis functions */ 211306656605SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 211406656605SStefano Zampini PetscScalar *y; 211506656605SStefano Zampini 2116ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*n_R);CHKERRQ(ierr); 211706656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_psi_B,&y);CHKERRQ(ierr); 211806656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*i);CHKERRQ(ierr); 211906656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 212006656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 212106656605SStefano Zampini if (i<n_vertices) { 212206656605SStefano Zampini y[n_B*i+idx_V_B[i]] = 1.0; 212306656605SStefano Zampini } 212406656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_psi_B,&y);CHKERRQ(ierr); 212506656605SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 212606656605SStefano Zampini 212706656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 212806656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_psi_D,&y);CHKERRQ(ierr); 212906656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*i);CHKERRQ(ierr); 213006656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 213106656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 213206656605SStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 213306656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_psi_D,&y);CHKERRQ(ierr); 213406656605SStefano Zampini } 213506656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 213606656605SStefano Zampini } 2137ffd830a3SStefano Zampini ierr = MatDestroy(&B_V);CHKERRQ(ierr); 2138ffd830a3SStefano Zampini ierr = MatDestroy(&B_C);CHKERRQ(ierr); 213906656605SStefano Zampini } 2140d62866d3SStefano Zampini /* free memory */ 214188ebb749SStefano Zampini ierr = PetscFree(idx_V_B);CHKERRQ(ierr); 214206656605SStefano Zampini ierr = MatDestroy(&S_VV);CHKERRQ(ierr); 214306656605SStefano Zampini ierr = MatDestroy(&S_CV);CHKERRQ(ierr); 214406656605SStefano Zampini ierr = MatDestroy(&S_VC);CHKERRQ(ierr); 214506656605SStefano Zampini ierr = MatDestroy(&S_CC);CHKERRQ(ierr); 2146d62866d3SStefano Zampini ierr = PetscFree(work);CHKERRQ(ierr); 2147d62866d3SStefano Zampini if (n_vertices) { 2148d62866d3SStefano Zampini ierr = MatDestroy(&A_VR);CHKERRQ(ierr); 2149d62866d3SStefano Zampini } 2150d62866d3SStefano Zampini if (n_constraints) { 2151d62866d3SStefano Zampini ierr = MatDestroy(&C_CR);CHKERRQ(ierr); 2152d62866d3SStefano Zampini } 215388ebb749SStefano Zampini /* Checking coarse_sub_mat and coarse basis functios */ 215488ebb749SStefano Zampini /* Symmetric case : It should be \Phi^{(j)^T} A^{(j)} \Phi^{(j)}=coarse_sub_mat */ 215588ebb749SStefano Zampini /* Non-symmetric case : It should be \Psi^{(j)^T} A^{(j)} \Phi^{(j)}=coarse_sub_mat */ 2156d12edf2fSStefano Zampini if (pcbddc->dbg_flag) { 215788ebb749SStefano Zampini Mat coarse_sub_mat; 215825084f0cSStefano Zampini Mat AUXMAT,TM1,TM2,TM3,TM4; 215988ebb749SStefano Zampini Mat coarse_phi_D,coarse_phi_B; 216088ebb749SStefano Zampini Mat coarse_psi_D,coarse_psi_B; 216188ebb749SStefano Zampini Mat A_II,A_BB,A_IB,A_BI; 21628bec7fa6SStefano Zampini Mat C_B,CPHI; 21638bec7fa6SStefano Zampini IS is_dummy; 21648bec7fa6SStefano Zampini Vec mones; 216588ebb749SStefano Zampini MatType checkmattype=MATSEQAIJ; 216688ebb749SStefano Zampini PetscReal real_value; 216788ebb749SStefano Zampini 2168a3df083aSStefano Zampini if (pcbddc->benign_n && !pcbddc->benign_change_explicit) { 2169a3df083aSStefano Zampini Mat A; 2170a3df083aSStefano Zampini ierr = PCBDDCBenignProject(pc,NULL,NULL,&A);CHKERRQ(ierr); 2171a3df083aSStefano Zampini ierr = MatGetSubMatrix(A,pcis->is_I_local,pcis->is_I_local,MAT_INITIAL_MATRIX,&A_II);CHKERRQ(ierr); 2172a3df083aSStefano Zampini ierr = MatGetSubMatrix(A,pcis->is_I_local,pcis->is_B_local,MAT_INITIAL_MATRIX,&A_IB);CHKERRQ(ierr); 2173a3df083aSStefano Zampini ierr = MatGetSubMatrix(A,pcis->is_B_local,pcis->is_I_local,MAT_INITIAL_MATRIX,&A_BI);CHKERRQ(ierr); 2174a3df083aSStefano Zampini ierr = MatGetSubMatrix(A,pcis->is_B_local,pcis->is_B_local,MAT_INITIAL_MATRIX,&A_BB);CHKERRQ(ierr); 2175a3df083aSStefano Zampini ierr = MatDestroy(&A);CHKERRQ(ierr); 2176a3df083aSStefano Zampini } else { 217788ebb749SStefano Zampini ierr = MatConvert(pcis->A_II,checkmattype,MAT_INITIAL_MATRIX,&A_II);CHKERRQ(ierr); 217888ebb749SStefano Zampini ierr = MatConvert(pcis->A_IB,checkmattype,MAT_INITIAL_MATRIX,&A_IB);CHKERRQ(ierr); 217988ebb749SStefano Zampini ierr = MatConvert(pcis->A_BI,checkmattype,MAT_INITIAL_MATRIX,&A_BI);CHKERRQ(ierr); 218088ebb749SStefano Zampini ierr = MatConvert(pcis->A_BB,checkmattype,MAT_INITIAL_MATRIX,&A_BB);CHKERRQ(ierr); 2181a3df083aSStefano Zampini } 218288ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_phi_D,checkmattype,MAT_INITIAL_MATRIX,&coarse_phi_D);CHKERRQ(ierr); 218388ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_phi_B,checkmattype,MAT_INITIAL_MATRIX,&coarse_phi_B);CHKERRQ(ierr); 2184ffd830a3SStefano Zampini if (!pcbddc->symmetric_primal) { 218588ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_psi_D,checkmattype,MAT_INITIAL_MATRIX,&coarse_psi_D);CHKERRQ(ierr); 218688ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_psi_B,checkmattype,MAT_INITIAL_MATRIX,&coarse_psi_B);CHKERRQ(ierr); 218788ebb749SStefano Zampini } 218888ebb749SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,pcbddc->local_primal_size,pcbddc->local_primal_size,coarse_submat_vals,&coarse_sub_mat);CHKERRQ(ierr); 218988ebb749SStefano Zampini 219025084f0cSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 21913301b35fSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Check coarse sub mat computation (symmetric %d)\n",pcbddc->symmetric_primal);CHKERRQ(ierr); 219225084f0cSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 2193ffd830a3SStefano Zampini if (!pcbddc->symmetric_primal) { 219488ebb749SStefano Zampini ierr = MatMatMult(A_II,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 219588ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM1);CHKERRQ(ierr); 219688ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 219788ebb749SStefano Zampini ierr = MatMatMult(A_BB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 219888ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM2);CHKERRQ(ierr); 219988ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 220088ebb749SStefano Zampini ierr = MatMatMult(A_IB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 220188ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM3);CHKERRQ(ierr); 220288ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 220388ebb749SStefano Zampini ierr = MatMatMult(A_BI,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 220488ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM4);CHKERRQ(ierr); 220588ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 220688ebb749SStefano Zampini } else { 220788ebb749SStefano Zampini ierr = MatPtAP(A_II,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&TM1);CHKERRQ(ierr); 220888ebb749SStefano Zampini ierr = MatPtAP(A_BB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&TM2);CHKERRQ(ierr); 220988ebb749SStefano Zampini ierr = MatMatMult(A_IB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 221088ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_phi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM3);CHKERRQ(ierr); 221188ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 221288ebb749SStefano Zampini ierr = MatMatMult(A_BI,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 221388ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_phi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM4);CHKERRQ(ierr); 221488ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 221588ebb749SStefano Zampini } 221688ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM2,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 221788ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM3,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 221888ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM4,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 221988ebb749SStefano Zampini ierr = MatConvert(TM1,MATSEQDENSE,MAT_REUSE_MATRIX,&TM1);CHKERRQ(ierr); 22204f1b2e48SStefano Zampini if (pcbddc->benign_n) { 2221fc227af8SStefano Zampini Mat B0_B,B0_BPHI; 2222d12edf2fSStefano Zampini PetscScalar *data,*data2; 22234f1b2e48SStefano Zampini PetscInt j; 2224d12edf2fSStefano Zampini 22254f1b2e48SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcbddc->benign_n,0,1,&is_dummy);CHKERRQ(ierr); 2226fc227af8SStefano Zampini ierr = MatGetSubMatrix(pcbddc->benign_B0,is_dummy,pcis->is_B_local,MAT_INITIAL_MATRIX,&B0_B);CHKERRQ(ierr); 2227d12edf2fSStefano Zampini ierr = MatMatMult(B0_B,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&B0_BPHI);CHKERRQ(ierr); 2228d12edf2fSStefano Zampini ierr = MatConvert(B0_BPHI,MATSEQDENSE,MAT_REUSE_MATRIX,&B0_BPHI);CHKERRQ(ierr); 2229d12edf2fSStefano Zampini ierr = MatDenseGetArray(TM1,&data);CHKERRQ(ierr); 2230d12edf2fSStefano Zampini ierr = MatDenseGetArray(B0_BPHI,&data2);CHKERRQ(ierr); 22314f1b2e48SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) { 22324f1b2e48SStefano Zampini PetscInt primal_idx = pcbddc->local_primal_size - pcbddc->benign_n + j; 2233d12edf2fSStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 22344f1b2e48SStefano Zampini data[primal_idx*pcbddc->local_primal_size+i] += data2[i*pcbddc->benign_n+j]; 22354f1b2e48SStefano Zampini data[i*pcbddc->local_primal_size+primal_idx] += data2[i*pcbddc->benign_n+j]; 22364f1b2e48SStefano Zampini } 2237d12edf2fSStefano Zampini } 2238d12edf2fSStefano Zampini ierr = MatDenseRestoreArray(TM1,&data);CHKERRQ(ierr); 2239d12edf2fSStefano Zampini ierr = MatDenseRestoreArray(B0_BPHI,&data2);CHKERRQ(ierr); 2240d12edf2fSStefano Zampini ierr = MatDestroy(&B0_B);CHKERRQ(ierr); 2241d12edf2fSStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 2242d12edf2fSStefano Zampini ierr = MatDestroy(&B0_BPHI);CHKERRQ(ierr); 2243d12edf2fSStefano Zampini } 2244d12edf2fSStefano Zampini #if 0 2245d12edf2fSStefano Zampini { 2246d12edf2fSStefano Zampini PetscViewer viewer; 2247d12edf2fSStefano Zampini char filename[256]; 2248ffd830a3SStefano Zampini sprintf(filename,"details_local_coarse_mat%d_level%d.m",PetscGlobalRank,pcbddc->current_level); 2249d12edf2fSStefano Zampini ierr = PetscViewerASCIIOpen(PETSC_COMM_SELF,filename,&viewer);CHKERRQ(ierr); 2250d12edf2fSStefano Zampini ierr = PetscViewerSetFormat(viewer,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 2251ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)coarse_sub_mat,"computed");CHKERRQ(ierr); 2252ffd830a3SStefano Zampini ierr = MatView(coarse_sub_mat,viewer);CHKERRQ(ierr); 2253ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)TM1,"projected");CHKERRQ(ierr); 2254d12edf2fSStefano Zampini ierr = MatView(TM1,viewer);CHKERRQ(ierr); 2255ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)pcbddc->coarse_phi_B,"phi_B");CHKERRQ(ierr); 2256ffd830a3SStefano Zampini ierr = MatView(pcbddc->coarse_phi_B,viewer);CHKERRQ(ierr); 2257ffd830a3SStefano Zampini if (pcbddc->coarse_phi_D) { 2258ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)pcbddc->coarse_phi_D,"phi_D");CHKERRQ(ierr); 2259ffd830a3SStefano Zampini ierr = MatView(pcbddc->coarse_phi_D,viewer);CHKERRQ(ierr); 2260ffd830a3SStefano Zampini } 2261ffd830a3SStefano Zampini if (pcbddc->coarse_psi_B) { 2262ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)pcbddc->coarse_psi_B,"psi_B");CHKERRQ(ierr); 2263ffd830a3SStefano Zampini ierr = MatView(pcbddc->coarse_psi_B,viewer);CHKERRQ(ierr); 2264ffd830a3SStefano Zampini } 2265ffd830a3SStefano Zampini if (pcbddc->coarse_phi_B) { 2266ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)pcbddc->coarse_psi_D,"psi_D");CHKERRQ(ierr); 2267ffd830a3SStefano Zampini ierr = MatView(pcbddc->coarse_psi_D,viewer);CHKERRQ(ierr); 2268ffd830a3SStefano Zampini } 2269d12edf2fSStefano Zampini ierr = PetscViewerDestroy(&viewer);CHKERRQ(ierr); 2270d12edf2fSStefano Zampini } 2271d12edf2fSStefano Zampini #endif 227281d9aea3SBarry Smith ierr = MatAXPY(TM1,m_one,coarse_sub_mat,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 22738bec7fa6SStefano Zampini ierr = MatNorm(TM1,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 22741575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 227506656605SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d matrix error % 1.14e\n",PetscGlobalRank,real_value);CHKERRQ(ierr); 22768bec7fa6SStefano Zampini 22778bec7fa6SStefano Zampini /* check constraints */ 22784f1b2e48SStefano Zampini if (!pcbddc->benign_n) { /* TODO: add benign case */ 22798bec7fa6SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcbddc->local_primal_size,0,1,&is_dummy);CHKERRQ(ierr); 22808bec7fa6SStefano Zampini ierr = MatGetSubMatrix(pcbddc->ConstraintMatrix,is_dummy,pcis->is_B_local,MAT_INITIAL_MATRIX,&C_B); 22818bec7fa6SStefano Zampini ierr = MatMatMult(C_B,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&CPHI);CHKERRQ(ierr); 22828bec7fa6SStefano Zampini ierr = MatCreateVecs(CPHI,&mones,NULL);CHKERRQ(ierr); 22838bec7fa6SStefano Zampini ierr = VecSet(mones,-1.0);CHKERRQ(ierr); 22848bec7fa6SStefano Zampini ierr = MatDiagonalSet(CPHI,mones,ADD_VALUES);CHKERRQ(ierr); 22858bec7fa6SStefano Zampini ierr = MatNorm(CPHI,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 2286bdae7319SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d phi constraints error % 1.14e\n",PetscGlobalRank,real_value);CHKERRQ(ierr); 2287ffd830a3SStefano Zampini if (!pcbddc->symmetric_primal) { 2288bdae7319SStefano Zampini ierr = MatMatMult(C_B,coarse_psi_B,MAT_REUSE_MATRIX,1.0,&CPHI);CHKERRQ(ierr); 2289bdae7319SStefano Zampini ierr = VecSet(mones,-1.0);CHKERRQ(ierr); 2290bdae7319SStefano Zampini ierr = MatDiagonalSet(CPHI,mones,ADD_VALUES);CHKERRQ(ierr); 2291bdae7319SStefano Zampini ierr = MatNorm(CPHI,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 2292bdae7319SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d psi constraints error % 1.14e\n",PetscGlobalRank,real_value);CHKERRQ(ierr); 229388ebb749SStefano Zampini } 22948bec7fa6SStefano Zampini ierr = MatDestroy(&C_B);CHKERRQ(ierr); 22958bec7fa6SStefano Zampini ierr = MatDestroy(&CPHI);CHKERRQ(ierr); 22968bec7fa6SStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 22978bec7fa6SStefano Zampini ierr = VecDestroy(&mones);CHKERRQ(ierr); 2298d12edf2fSStefano Zampini } 229925084f0cSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 230088ebb749SStefano Zampini ierr = MatDestroy(&A_II);CHKERRQ(ierr); 230188ebb749SStefano Zampini ierr = MatDestroy(&A_BB);CHKERRQ(ierr); 230288ebb749SStefano Zampini ierr = MatDestroy(&A_IB);CHKERRQ(ierr); 230388ebb749SStefano Zampini ierr = MatDestroy(&A_BI);CHKERRQ(ierr); 230488ebb749SStefano Zampini ierr = MatDestroy(&TM1);CHKERRQ(ierr); 230588ebb749SStefano Zampini ierr = MatDestroy(&TM2);CHKERRQ(ierr); 230688ebb749SStefano Zampini ierr = MatDestroy(&TM3);CHKERRQ(ierr); 230788ebb749SStefano Zampini ierr = MatDestroy(&TM4);CHKERRQ(ierr); 230888ebb749SStefano Zampini ierr = MatDestroy(&coarse_phi_D);CHKERRQ(ierr); 230988ebb749SStefano Zampini ierr = MatDestroy(&coarse_phi_B);CHKERRQ(ierr); 2310ffd830a3SStefano Zampini if (!pcbddc->symmetric_primal) { 231188ebb749SStefano Zampini ierr = MatDestroy(&coarse_psi_D);CHKERRQ(ierr); 231288ebb749SStefano Zampini ierr = MatDestroy(&coarse_psi_B);CHKERRQ(ierr); 231388ebb749SStefano Zampini } 231488ebb749SStefano Zampini ierr = MatDestroy(&coarse_sub_mat);CHKERRQ(ierr); 231588ebb749SStefano Zampini } 23168629588bSStefano Zampini /* get back data */ 23178629588bSStefano Zampini *coarse_submat_vals_n = coarse_submat_vals; 231888ebb749SStefano Zampini PetscFunctionReturn(0); 231988ebb749SStefano Zampini } 232088ebb749SStefano Zampini 232188ebb749SStefano Zampini #undef __FUNCT__ 2322d65f70fdSStefano Zampini #define __FUNCT__ "MatGetSubMatrixUnsorted" 2323d648f858SStefano Zampini PetscErrorCode MatGetSubMatrixUnsorted(Mat A, IS isrow, IS iscol, Mat* B) 2324aa0d41d4SStefano Zampini { 2325d65f70fdSStefano Zampini Mat *work_mat; 2326d65f70fdSStefano Zampini IS isrow_s,iscol_s; 2327d65f70fdSStefano Zampini PetscBool rsorted,csorted; 2328d65f70fdSStefano Zampini PetscInt rsize,*idxs_perm_r,csize,*idxs_perm_c; 2329aa0d41d4SStefano Zampini PetscErrorCode ierr; 2330aa0d41d4SStefano Zampini 2331aa0d41d4SStefano Zampini PetscFunctionBegin; 2332d65f70fdSStefano Zampini ierr = ISSorted(isrow,&rsorted);CHKERRQ(ierr); 2333d65f70fdSStefano Zampini ierr = ISSorted(iscol,&csorted);CHKERRQ(ierr); 2334d65f70fdSStefano Zampini ierr = ISGetLocalSize(isrow,&rsize);CHKERRQ(ierr); 2335d65f70fdSStefano Zampini ierr = ISGetLocalSize(iscol,&csize);CHKERRQ(ierr); 2336aa0d41d4SStefano Zampini 2337d65f70fdSStefano Zampini if (!rsorted) { 2338906d46d4SStefano Zampini const PetscInt *idxs; 2339906d46d4SStefano Zampini PetscInt *idxs_sorted,i; 2340aa0d41d4SStefano Zampini 2341d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_perm_r);CHKERRQ(ierr); 2342d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_sorted);CHKERRQ(ierr); 2343d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 2344d65f70fdSStefano Zampini idxs_perm_r[i] = i; 2345aa0d41d4SStefano Zampini } 2346d65f70fdSStefano Zampini ierr = ISGetIndices(isrow,&idxs);CHKERRQ(ierr); 2347d65f70fdSStefano Zampini ierr = PetscSortIntWithPermutation(rsize,idxs,idxs_perm_r);CHKERRQ(ierr); 2348d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 2349d65f70fdSStefano Zampini idxs_sorted[i] = idxs[idxs_perm_r[i]]; 2350aa0d41d4SStefano Zampini } 2351d65f70fdSStefano Zampini ierr = ISRestoreIndices(isrow,&idxs);CHKERRQ(ierr); 2352d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,rsize,idxs_sorted,PETSC_OWN_POINTER,&isrow_s);CHKERRQ(ierr); 2353d65f70fdSStefano Zampini } else { 2354d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)isrow);CHKERRQ(ierr); 2355d65f70fdSStefano Zampini isrow_s = isrow; 2356aa0d41d4SStefano Zampini } 2357906d46d4SStefano Zampini 2358d65f70fdSStefano Zampini if (!csorted) { 2359d65f70fdSStefano Zampini if (isrow == iscol) { 2360d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)isrow_s);CHKERRQ(ierr); 2361d65f70fdSStefano Zampini iscol_s = isrow_s; 2362d65f70fdSStefano Zampini } else { 2363d65f70fdSStefano Zampini const PetscInt *idxs; 2364d65f70fdSStefano Zampini PetscInt *idxs_sorted,i; 2365906d46d4SStefano Zampini 2366d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_perm_c);CHKERRQ(ierr); 2367d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_sorted);CHKERRQ(ierr); 2368d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 2369d65f70fdSStefano Zampini idxs_perm_c[i] = i; 2370d65f70fdSStefano Zampini } 2371d65f70fdSStefano Zampini ierr = ISGetIndices(iscol,&idxs);CHKERRQ(ierr); 2372d65f70fdSStefano Zampini ierr = PetscSortIntWithPermutation(csize,idxs,idxs_perm_c);CHKERRQ(ierr); 2373d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 2374d65f70fdSStefano Zampini idxs_sorted[i] = idxs[idxs_perm_c[i]]; 2375d65f70fdSStefano Zampini } 2376d65f70fdSStefano Zampini ierr = ISRestoreIndices(iscol,&idxs);CHKERRQ(ierr); 2377d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,csize,idxs_sorted,PETSC_OWN_POINTER,&iscol_s);CHKERRQ(ierr); 2378d65f70fdSStefano Zampini } 2379d65f70fdSStefano Zampini } else { 2380d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)iscol);CHKERRQ(ierr); 2381d65f70fdSStefano Zampini iscol_s = iscol; 2382d65f70fdSStefano Zampini } 2383d65f70fdSStefano Zampini 2384d648f858SStefano Zampini ierr = MatGetSubMatrices(A,1,&isrow_s,&iscol_s,MAT_INITIAL_MATRIX,&work_mat);CHKERRQ(ierr); 2385d65f70fdSStefano Zampini 2386d65f70fdSStefano Zampini if (!rsorted || !csorted) { 2387906d46d4SStefano Zampini Mat new_mat; 2388d65f70fdSStefano Zampini IS is_perm_r,is_perm_c; 2389906d46d4SStefano Zampini 2390d65f70fdSStefano Zampini if (!rsorted) { 2391d65f70fdSStefano Zampini PetscInt *idxs_r,i; 2392d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_r);CHKERRQ(ierr); 2393d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 2394d65f70fdSStefano Zampini idxs_r[idxs_perm_r[i]] = i; 2395906d46d4SStefano Zampini } 2396d65f70fdSStefano Zampini ierr = PetscFree(idxs_perm_r);CHKERRQ(ierr); 2397d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,rsize,idxs_r,PETSC_OWN_POINTER,&is_perm_r);CHKERRQ(ierr); 2398d65f70fdSStefano Zampini } else { 2399d65f70fdSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,rsize,0,1,&is_perm_r);CHKERRQ(ierr); 2400906d46d4SStefano Zampini } 2401d65f70fdSStefano Zampini ierr = ISSetPermutation(is_perm_r);CHKERRQ(ierr); 2402d65f70fdSStefano Zampini 2403d65f70fdSStefano Zampini if (!csorted) { 2404d65f70fdSStefano Zampini if (isrow_s == iscol_s) { 2405d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)is_perm_r);CHKERRQ(ierr); 2406d65f70fdSStefano Zampini is_perm_c = is_perm_r; 2407d65f70fdSStefano Zampini } else { 2408d65f70fdSStefano Zampini PetscInt *idxs_c,i; 2409d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_c);CHKERRQ(ierr); 2410d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 2411d65f70fdSStefano Zampini idxs_c[idxs_perm_c[i]] = i; 2412d65f70fdSStefano Zampini } 2413d65f70fdSStefano Zampini ierr = PetscFree(idxs_perm_c);CHKERRQ(ierr); 2414d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,csize,idxs_c,PETSC_OWN_POINTER,&is_perm_c);CHKERRQ(ierr); 2415d65f70fdSStefano Zampini } 2416d65f70fdSStefano Zampini } else { 2417d65f70fdSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,csize,0,1,&is_perm_c);CHKERRQ(ierr); 2418d65f70fdSStefano Zampini } 2419d65f70fdSStefano Zampini ierr = ISSetPermutation(is_perm_c);CHKERRQ(ierr); 2420d65f70fdSStefano Zampini 2421d65f70fdSStefano Zampini ierr = MatPermute(work_mat[0],is_perm_r,is_perm_c,&new_mat);CHKERRQ(ierr); 2422d65f70fdSStefano Zampini ierr = MatDestroy(&work_mat[0]);CHKERRQ(ierr); 2423d65f70fdSStefano Zampini work_mat[0] = new_mat; 2424d65f70fdSStefano Zampini ierr = ISDestroy(&is_perm_r);CHKERRQ(ierr); 2425d65f70fdSStefano Zampini ierr = ISDestroy(&is_perm_c);CHKERRQ(ierr); 2426d65f70fdSStefano Zampini } 2427d65f70fdSStefano Zampini 2428d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)work_mat[0]);CHKERRQ(ierr); 2429d65f70fdSStefano Zampini *B = work_mat[0]; 2430d65f70fdSStefano Zampini ierr = MatDestroyMatrices(1,&work_mat);CHKERRQ(ierr); 2431d65f70fdSStefano Zampini ierr = ISDestroy(&isrow_s);CHKERRQ(ierr); 2432d65f70fdSStefano Zampini ierr = ISDestroy(&iscol_s);CHKERRQ(ierr); 2433d65f70fdSStefano Zampini PetscFunctionReturn(0); 2434d65f70fdSStefano Zampini } 2435d65f70fdSStefano Zampini 2436d65f70fdSStefano Zampini #undef __FUNCT__ 24375e8657edSStefano Zampini #define __FUNCT__ "PCBDDCComputeLocalMatrix" 24385e8657edSStefano Zampini PetscErrorCode PCBDDCComputeLocalMatrix(PC pc, Mat ChangeOfBasisMatrix) 2439aa0d41d4SStefano Zampini { 2440aa0d41d4SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 24415e8657edSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 2442d65f70fdSStefano Zampini Mat new_mat; 24435e8657edSStefano Zampini IS is_local,is_global; 2444d65f70fdSStefano Zampini PetscInt local_size; 2445d65f70fdSStefano Zampini PetscBool isseqaij; 2446aa0d41d4SStefano Zampini PetscErrorCode ierr; 2447aa0d41d4SStefano Zampini 2448aa0d41d4SStefano Zampini PetscFunctionBegin; 2449aa0d41d4SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 24505e8657edSStefano Zampini ierr = MatGetSize(matis->A,&local_size,NULL);CHKERRQ(ierr); 24515e8657edSStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)matis->A),local_size,0,1,&is_local);CHKERRQ(ierr); 2452b087196eSStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(pc->pmat->rmap->mapping,is_local,&is_global);CHKERRQ(ierr); 2453aa0d41d4SStefano Zampini ierr = ISDestroy(&is_local);CHKERRQ(ierr); 2454d648f858SStefano Zampini ierr = MatGetSubMatrixUnsorted(ChangeOfBasisMatrix,is_global,is_global,&new_mat);CHKERRQ(ierr); 2455aa0d41d4SStefano Zampini ierr = ISDestroy(&is_global);CHKERRQ(ierr); 2456906d46d4SStefano Zampini 2457906d46d4SStefano Zampini /* check */ 2458906d46d4SStefano Zampini if (pcbddc->dbg_flag) { 2459906d46d4SStefano Zampini Vec x,x_change; 2460906d46d4SStefano Zampini PetscReal error; 2461906d46d4SStefano Zampini 24625e8657edSStefano Zampini ierr = MatCreateVecs(ChangeOfBasisMatrix,&x,&x_change);CHKERRQ(ierr); 2463906d46d4SStefano Zampini ierr = VecSetRandom(x,NULL);CHKERRQ(ierr); 24645e8657edSStefano Zampini ierr = MatMult(ChangeOfBasisMatrix,x,x_change);CHKERRQ(ierr); 2465e176bc59SStefano Zampini ierr = VecScatterBegin(matis->cctx,x,matis->x,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2466e176bc59SStefano Zampini ierr = VecScatterEnd(matis->cctx,x,matis->x,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2467d65f70fdSStefano Zampini ierr = MatMult(new_mat,matis->x,matis->y);CHKERRQ(ierr); 2468e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,matis->y,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2469e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,matis->y,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2470906d46d4SStefano Zampini ierr = VecAXPY(x,-1.0,x_change);CHKERRQ(ierr); 2471906d46d4SStefano Zampini ierr = VecNorm(x,NORM_INFINITY,&error);CHKERRQ(ierr); 2472906d46d4SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 2473906d46d4SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Error global vs local change on N: %1.6e\n",error);CHKERRQ(ierr); 2474906d46d4SStefano Zampini ierr = VecDestroy(&x);CHKERRQ(ierr); 2475906d46d4SStefano Zampini ierr = VecDestroy(&x_change);CHKERRQ(ierr); 2476906d46d4SStefano Zampini } 2477906d46d4SStefano Zampini 247822d5777bSStefano Zampini /* TODO: HOW TO WORK WITH BAIJ and SBAIJ and SEQDENSE? */ 24799b28b3ffSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)matis->A,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 248022d5777bSStefano Zampini if (isseqaij) { 24811cf9b237SStefano Zampini Mat M; 24821cf9b237SStefano Zampini 24831cf9b237SStefano Zampini ierr = MatPtAP(matis->A,new_mat,MAT_INITIAL_MATRIX,2.0,&M);CHKERRQ(ierr); 24841cf9b237SStefano Zampini ierr = MatSeqAIJCompress(M,&pcbddc->local_mat);CHKERRQ(ierr); 24851cf9b237SStefano Zampini ierr = MatDestroy(&M);CHKERRQ(ierr); 2486aa0d41d4SStefano Zampini } else { 24871cf9b237SStefano Zampini Mat work_mat,M; 24881cf9b237SStefano Zampini 2489aa0d41d4SStefano Zampini ierr = MatConvert(matis->A,MATSEQAIJ,MAT_INITIAL_MATRIX,&work_mat);CHKERRQ(ierr); 24901cf9b237SStefano Zampini ierr = MatPtAP(work_mat,new_mat,MAT_INITIAL_MATRIX,2.0,&M);CHKERRQ(ierr); 24911cf9b237SStefano Zampini ierr = MatSeqAIJCompress(M,&pcbddc->local_mat);CHKERRQ(ierr); 24921cf9b237SStefano Zampini ierr = MatDestroy(&M);CHKERRQ(ierr); 2493aa0d41d4SStefano Zampini ierr = MatDestroy(&work_mat);CHKERRQ(ierr); 2494aa0d41d4SStefano Zampini } 24953301b35fSStefano Zampini if (matis->A->symmetric_set) { 24963301b35fSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_SYMMETRIC,matis->A->symmetric);CHKERRQ(ierr); 2497e496cd5dSStefano Zampini #if !defined(PETSC_USE_COMPLEX) 24983301b35fSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_HERMITIAN,matis->A->symmetric);CHKERRQ(ierr); 2499e496cd5dSStefano Zampini #endif 25003301b35fSStefano Zampini } 250145a1bb75SStefano Zampini /* 250245a1bb75SStefano Zampini ierr = PetscViewerSetFormat(PETSC_VIEWER_STDOUT_SELF,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 2503d65f70fdSStefano Zampini ierr = MatView(new_mat,(PetscViewer)0);CHKERRQ(ierr); 250445a1bb75SStefano Zampini */ 2505d65f70fdSStefano Zampini ierr = MatDestroy(&new_mat);CHKERRQ(ierr); 2506aa0d41d4SStefano Zampini PetscFunctionReturn(0); 2507aa0d41d4SStefano Zampini } 2508aa0d41d4SStefano Zampini 2509aa0d41d4SStefano Zampini #undef __FUNCT__ 2510a64d13efSStefano Zampini #define __FUNCT__ "PCBDDCSetUpLocalScatters" 25118ce42a96SStefano Zampini PetscErrorCode PCBDDCSetUpLocalScatters(PC pc) 2512a64d13efSStefano Zampini { 2513a64d13efSStefano Zampini PC_IS* pcis = (PC_IS*)(pc->data); 2514a64d13efSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 2515d62866d3SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 251653892102SStefano Zampini PetscInt *idx_R_local=NULL; 25173a50541eSStefano Zampini PetscInt n_vertices,i,j,n_R,n_D,n_B; 25183a50541eSStefano Zampini PetscInt vbs,bs; 25196816873aSStefano Zampini PetscBT bitmask=NULL; 2520a64d13efSStefano Zampini PetscErrorCode ierr; 2521a64d13efSStefano Zampini 2522a64d13efSStefano Zampini PetscFunctionBegin; 2523b23d619eSStefano Zampini /* 2524b23d619eSStefano Zampini No need to setup local scatters if 2525b23d619eSStefano Zampini - primal space is unchanged 2526b23d619eSStefano Zampini AND 2527b23d619eSStefano Zampini - we actually have locally some primal dofs (could not be true in multilevel or for isolated subdomains) 2528b23d619eSStefano Zampini AND 2529b23d619eSStefano Zampini - we are not in debugging mode (this is needed since there are Synchronized prints at the end of the subroutine 2530b23d619eSStefano Zampini */ 2531b23d619eSStefano Zampini if (!pcbddc->new_primal_space_local && pcbddc->local_primal_size && !pcbddc->dbg_flag) { 2532f4ddd8eeSStefano Zampini PetscFunctionReturn(0); 2533f4ddd8eeSStefano Zampini } 2534f4ddd8eeSStefano Zampini /* destroy old objects */ 2535f4ddd8eeSStefano Zampini ierr = ISDestroy(&pcbddc->is_R_local);CHKERRQ(ierr); 2536f4ddd8eeSStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_B);CHKERRQ(ierr); 2537f4ddd8eeSStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_D);CHKERRQ(ierr); 2538a64d13efSStefano Zampini /* Set Non-overlapping dimensions */ 2539b371cd4fSStefano Zampini n_B = pcis->n_B; 2540b371cd4fSStefano Zampini n_D = pcis->n - n_B; 2541b371cd4fSStefano Zampini n_vertices = pcbddc->n_vertices; 25423a50541eSStefano Zampini 2543a64d13efSStefano Zampini /* Dohrmann's notation: dofs splitted in R (Remaining: all dofs but the vertices) and V (Vertices) */ 25446816873aSStefano Zampini 254553892102SStefano Zampini /* create auxiliary bitmask and allocate workspace */ 2546df4d28bfSStefano Zampini if (!sub_schurs->reuse_solver) { 2547854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n-n_vertices,&idx_R_local);CHKERRQ(ierr); 2548a64d13efSStefano Zampini ierr = PetscBTCreate(pcis->n,&bitmask);CHKERRQ(ierr); 2549a64d13efSStefano Zampini for (i=0;i<n_vertices;i++) { 25500e6343abSStefano Zampini ierr = PetscBTSet(bitmask,pcbddc->local_primal_ref_node[i]);CHKERRQ(ierr); 2551a64d13efSStefano Zampini } 2552a64d13efSStefano Zampini 2553a64d13efSStefano Zampini for (i=0, n_R=0; i<pcis->n; i++) { 25544641a718SStefano Zampini if (!PetscBTLookup(bitmask,i)) { 25556816873aSStefano Zampini idx_R_local[n_R++] = i; 2556a64d13efSStefano Zampini } 2557a64d13efSStefano Zampini } 2558df4d28bfSStefano Zampini } else { /* A different ordering (already computed) is present if we are reusing the Schur solver */ 2559df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 25606816873aSStefano Zampini 2561df4d28bfSStefano Zampini ierr = ISGetIndices(reuse_solver->is_R,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 2562df4d28bfSStefano Zampini ierr = ISGetLocalSize(reuse_solver->is_R,&n_R);CHKERRQ(ierr); 25636816873aSStefano Zampini } 25643a50541eSStefano Zampini 25653a50541eSStefano Zampini /* Block code */ 25663a50541eSStefano Zampini vbs = 1; 25673a50541eSStefano Zampini ierr = MatGetBlockSize(pcbddc->local_mat,&bs);CHKERRQ(ierr); 25683a50541eSStefano Zampini if (bs>1 && !(n_vertices%bs)) { 25693a50541eSStefano Zampini PetscBool is_blocked = PETSC_TRUE; 25703a50541eSStefano Zampini PetscInt *vary; 2571df4d28bfSStefano Zampini if (!sub_schurs->reuse_solver) { 2572785e854fSJed Brown ierr = PetscMalloc1(pcis->n/bs,&vary);CHKERRQ(ierr); 25733a50541eSStefano Zampini ierr = PetscMemzero(vary,pcis->n/bs*sizeof(PetscInt));CHKERRQ(ierr); 2574d3df7717SStefano Zampini /* Verify that the vertex indices correspond to each element in a block (code taken from sbaij2.c) */ 2575d3df7717SStefano 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 */ 25760e6343abSStefano Zampini for (i=0; i<n_vertices; i++) vary[pcbddc->local_primal_ref_node[i]/bs]++; 2577d3df7717SStefano Zampini for (i=0; i<pcis->n/bs; i++) { 25783a50541eSStefano Zampini if (vary[i]!=0 && vary[i]!=bs) { 25793a50541eSStefano Zampini is_blocked = PETSC_FALSE; 25803a50541eSStefano Zampini break; 25813a50541eSStefano Zampini } 25823a50541eSStefano Zampini } 2583d3df7717SStefano Zampini ierr = PetscFree(vary);CHKERRQ(ierr); 2584d3df7717SStefano Zampini } else { 2585d3df7717SStefano Zampini /* Verify directly the R set */ 2586d3df7717SStefano Zampini for (i=0; i<n_R/bs; i++) { 2587d3df7717SStefano Zampini PetscInt j,node=idx_R_local[bs*i]; 2588d3df7717SStefano Zampini for (j=1; j<bs; j++) { 2589d3df7717SStefano Zampini if (node != idx_R_local[bs*i+j]-j) { 2590d3df7717SStefano Zampini is_blocked = PETSC_FALSE; 2591d3df7717SStefano Zampini break; 2592d3df7717SStefano Zampini } 2593d3df7717SStefano Zampini } 2594d3df7717SStefano Zampini } 2595d3df7717SStefano Zampini } 25963a50541eSStefano Zampini if (is_blocked) { /* build compressed IS for R nodes (complement of vertices) */ 25973a50541eSStefano Zampini vbs = bs; 25983a50541eSStefano Zampini for (i=0;i<n_R/vbs;i++) { 25993a50541eSStefano Zampini idx_R_local[i] = idx_R_local[vbs*i]/vbs; 26003a50541eSStefano Zampini } 26013a50541eSStefano Zampini } 26023a50541eSStefano Zampini } 26033a50541eSStefano Zampini ierr = ISCreateBlock(PETSC_COMM_SELF,vbs,n_R/vbs,idx_R_local,PETSC_COPY_VALUES,&pcbddc->is_R_local);CHKERRQ(ierr); 2604df4d28bfSStefano Zampini if (sub_schurs->reuse_solver) { 2605df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 260653892102SStefano Zampini 2607df4d28bfSStefano Zampini ierr = ISRestoreIndices(reuse_solver->is_R,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 2608df4d28bfSStefano Zampini ierr = ISDestroy(&reuse_solver->is_R);CHKERRQ(ierr); 260953892102SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->is_R_local);CHKERRQ(ierr); 2610df4d28bfSStefano Zampini reuse_solver->is_R = pcbddc->is_R_local; 261153892102SStefano Zampini } else { 26123a50541eSStefano Zampini ierr = PetscFree(idx_R_local);CHKERRQ(ierr); 261353892102SStefano Zampini } 2614a64d13efSStefano Zampini 2615a64d13efSStefano Zampini /* print some info if requested */ 2616a64d13efSStefano Zampini if (pcbddc->dbg_flag) { 2617a64d13efSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 2618a64d13efSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 26191575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 2620a64d13efSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d local dimensions\n",PetscGlobalRank);CHKERRQ(ierr); 2621a64d13efSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"local_size = %d, dirichlet_size = %d, boundary_size = %d\n",pcis->n,n_D,n_B);CHKERRQ(ierr); 26224f1b2e48SStefano 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); 2623a64d13efSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 2624a64d13efSStefano Zampini } 2625a64d13efSStefano Zampini 2626a64d13efSStefano Zampini /* VecScatters pcbddc->R_to_B and (optionally) pcbddc->R_to_D */ 2627df4d28bfSStefano Zampini if (!sub_schurs->reuse_solver) { 26286816873aSStefano Zampini IS is_aux1,is_aux2; 26296816873aSStefano Zampini PetscInt *aux_array1,*aux_array2,*is_indices,*idx_R_local; 26306816873aSStefano Zampini 26313a50541eSStefano Zampini ierr = ISGetIndices(pcbddc->is_R_local,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 2632854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n_B-n_vertices,&aux_array1);CHKERRQ(ierr); 2633854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n_B-n_vertices,&aux_array2);CHKERRQ(ierr); 2634a64d13efSStefano Zampini ierr = ISGetIndices(pcis->is_I_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 26354641a718SStefano Zampini for (i=0; i<n_D; i++) { 26364641a718SStefano Zampini ierr = PetscBTSet(bitmask,is_indices[i]);CHKERRQ(ierr); 26374641a718SStefano Zampini } 2638a64d13efSStefano Zampini ierr = ISRestoreIndices(pcis->is_I_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 2639a64d13efSStefano Zampini for (i=0, j=0; i<n_R; i++) { 26404641a718SStefano Zampini if (!PetscBTLookup(bitmask,idx_R_local[i])) { 26414641a718SStefano Zampini aux_array1[j++] = i; 2642a64d13efSStefano Zampini } 2643a64d13efSStefano Zampini } 2644a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array1,PETSC_OWN_POINTER,&is_aux1);CHKERRQ(ierr); 2645a64d13efSStefano Zampini ierr = ISGetIndices(pcis->is_B_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 2646a64d13efSStefano Zampini for (i=0, j=0; i<n_B; i++) { 26474641a718SStefano Zampini if (!PetscBTLookup(bitmask,is_indices[i])) { 26484641a718SStefano Zampini aux_array2[j++] = i; 2649a64d13efSStefano Zampini } 2650a64d13efSStefano Zampini } 2651a64d13efSStefano Zampini ierr = ISRestoreIndices(pcis->is_B_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 2652a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array2,PETSC_OWN_POINTER,&is_aux2);CHKERRQ(ierr); 2653a64d13efSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,is_aux1,pcis->vec1_B,is_aux2,&pcbddc->R_to_B);CHKERRQ(ierr); 2654a64d13efSStefano Zampini ierr = ISDestroy(&is_aux1);CHKERRQ(ierr); 2655a64d13efSStefano Zampini ierr = ISDestroy(&is_aux2);CHKERRQ(ierr); 2656a64d13efSStefano Zampini 26578eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 2658785e854fSJed Brown ierr = PetscMalloc1(n_D,&aux_array1);CHKERRQ(ierr); 2659a64d13efSStefano Zampini for (i=0, j=0; i<n_R; i++) { 26604641a718SStefano Zampini if (PetscBTLookup(bitmask,idx_R_local[i])) { 26614641a718SStefano Zampini aux_array1[j++] = i; 2662a64d13efSStefano Zampini } 2663a64d13efSStefano Zampini } 2664a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array1,PETSC_OWN_POINTER,&is_aux1);CHKERRQ(ierr); 2665a64d13efSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,is_aux1,pcis->vec1_D,(IS)0,&pcbddc->R_to_D);CHKERRQ(ierr); 2666a64d13efSStefano Zampini ierr = ISDestroy(&is_aux1);CHKERRQ(ierr); 2667a64d13efSStefano Zampini } 26684641a718SStefano Zampini ierr = PetscBTDestroy(&bitmask);CHKERRQ(ierr); 26693a50541eSStefano Zampini ierr = ISRestoreIndices(pcbddc->is_R_local,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 2670d62866d3SStefano Zampini } else { 2671df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 26726816873aSStefano Zampini IS tis; 26736816873aSStefano Zampini PetscInt schur_size; 26746816873aSStefano Zampini 2675df4d28bfSStefano Zampini ierr = ISGetLocalSize(reuse_solver->is_B,&schur_size);CHKERRQ(ierr); 26766816873aSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,schur_size,n_D,1,&tis);CHKERRQ(ierr); 2677df4d28bfSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,tis,pcis->vec1_B,reuse_solver->is_B,&pcbddc->R_to_B);CHKERRQ(ierr); 26786816873aSStefano Zampini ierr = ISDestroy(&tis);CHKERRQ(ierr); 26796816873aSStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 26806816873aSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,n_D,0,1,&tis);CHKERRQ(ierr); 26816816873aSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,tis,pcis->vec1_D,(IS)0,&pcbddc->R_to_D);CHKERRQ(ierr); 26826816873aSStefano Zampini ierr = ISDestroy(&tis);CHKERRQ(ierr); 2683d62866d3SStefano Zampini } 2684d62866d3SStefano Zampini } 2685a64d13efSStefano Zampini PetscFunctionReturn(0); 2686a64d13efSStefano Zampini } 2687a64d13efSStefano Zampini 2688304d26faSStefano Zampini 2689304d26faSStefano Zampini #undef __FUNCT__ 2690304d26faSStefano Zampini #define __FUNCT__ "PCBDDCSetUpLocalSolvers" 2691684f6988SStefano Zampini PetscErrorCode PCBDDCSetUpLocalSolvers(PC pc, PetscBool dirichlet, PetscBool neumann) 2692304d26faSStefano Zampini { 2693304d26faSStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 2694304d26faSStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 2695304d26faSStefano Zampini PC pc_temp; 2696304d26faSStefano Zampini Mat A_RR; 2697f4ddd8eeSStefano Zampini MatReuse reuse; 2698304d26faSStefano Zampini PetscScalar m_one = -1.0; 2699304d26faSStefano Zampini PetscReal value; 270004708bb6SStefano Zampini PetscInt n_D,n_R; 27019577ea80SStefano Zampini PetscBool use_exact,use_exact_reduced,issbaij; 2702304d26faSStefano Zampini PetscErrorCode ierr; 2703e604994aSStefano Zampini /* prefixes stuff */ 2704312be037SStefano Zampini char dir_prefix[256],neu_prefix[256],str_level[16]; 2705e604994aSStefano Zampini size_t len; 2706304d26faSStefano Zampini 2707304d26faSStefano Zampini PetscFunctionBegin; 2708304d26faSStefano Zampini 2709e604994aSStefano Zampini /* compute prefixes */ 2710e604994aSStefano Zampini ierr = PetscStrcpy(dir_prefix,"");CHKERRQ(ierr); 2711e604994aSStefano Zampini ierr = PetscStrcpy(neu_prefix,"");CHKERRQ(ierr); 2712e604994aSStefano Zampini if (!pcbddc->current_level) { 2713e604994aSStefano Zampini ierr = PetscStrcpy(dir_prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 2714e604994aSStefano Zampini ierr = PetscStrcpy(neu_prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 2715e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,"pc_bddc_dirichlet_");CHKERRQ(ierr); 2716e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,"pc_bddc_neumann_");CHKERRQ(ierr); 2717e604994aSStefano Zampini } else { 2718e604994aSStefano Zampini ierr = PetscStrcpy(str_level,"");CHKERRQ(ierr); 2719312be037SStefano Zampini sprintf(str_level,"l%d_",(int)(pcbddc->current_level)); 2720e604994aSStefano Zampini ierr = PetscStrlen(((PetscObject)pc)->prefix,&len);CHKERRQ(ierr); 2721e604994aSStefano Zampini len -= 15; /* remove "pc_bddc_coarse_" */ 2722312be037SStefano Zampini if (pcbddc->current_level>1) len -= 3; /* remove "lX_" with X level number */ 2723312be037SStefano Zampini if (pcbddc->current_level>10) len -= 1; /* remove another char from level number */ 272434d6797cSStefano Zampini ierr = PetscStrncpy(dir_prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 272534d6797cSStefano Zampini ierr = PetscStrncpy(neu_prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 2726e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,"pc_bddc_dirichlet_");CHKERRQ(ierr); 2727e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,"pc_bddc_neumann_");CHKERRQ(ierr); 2728e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,str_level);CHKERRQ(ierr); 2729e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,str_level);CHKERRQ(ierr); 2730e604994aSStefano Zampini } 2731e604994aSStefano Zampini 2732304d26faSStefano Zampini /* DIRICHLET PROBLEM */ 2733684f6988SStefano Zampini if (dirichlet) { 2734d5574798SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 2735450f8f5eSStefano Zampini if (pcbddc->benign_n && !pcbddc->benign_change_explicit) { 2736450f8f5eSStefano Zampini if (!sub_schurs->reuse_solver) { 2737450f8f5eSStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Not yet implemented\n"); 2738450f8f5eSStefano Zampini } 2739450f8f5eSStefano Zampini if (pcbddc->dbg_flag) { 2740a3df083aSStefano Zampini Mat A_IIn; 2741a3df083aSStefano Zampini 2742a3df083aSStefano Zampini ierr = PCBDDCBenignProject(pc,pcis->is_I_local,pcis->is_I_local,&A_IIn);CHKERRQ(ierr); 2743a3df083aSStefano Zampini ierr = MatDestroy(&pcis->A_II);CHKERRQ(ierr); 2744a3df083aSStefano Zampini pcis->A_II = A_IIn; 2745a3df083aSStefano Zampini } 2746450f8f5eSStefano Zampini } 27473301b35fSStefano Zampini if (pcbddc->local_mat->symmetric_set) { 27483301b35fSStefano Zampini ierr = MatSetOption(pcis->A_II,MAT_SYMMETRIC,pcbddc->local_mat->symmetric_set);CHKERRQ(ierr); 2749964fefecSStefano Zampini } 2750ac78edfcSStefano Zampini /* Matrix for Dirichlet problem is pcis->A_II */ 2751964fefecSStefano Zampini n_D = pcis->n - pcis->n_B; 2752304d26faSStefano Zampini if (!pcbddc->ksp_D) { /* create object if not yet build */ 2753304d26faSStefano Zampini ierr = KSPCreate(PETSC_COMM_SELF,&pcbddc->ksp_D);CHKERRQ(ierr); 2754304d26faSStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->ksp_D,(PetscObject)pc,1);CHKERRQ(ierr); 2755304d26faSStefano Zampini /* default */ 2756304d26faSStefano Zampini ierr = KSPSetType(pcbddc->ksp_D,KSPPREONLY);CHKERRQ(ierr); 2757e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->ksp_D,dir_prefix);CHKERRQ(ierr); 27589577ea80SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcis->A_II,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 2759304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_D,&pc_temp);CHKERRQ(ierr); 27609577ea80SStefano Zampini if (issbaij) { 27619577ea80SStefano Zampini ierr = PCSetType(pc_temp,PCCHOLESKY);CHKERRQ(ierr); 27629577ea80SStefano Zampini } else { 2763304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCLU);CHKERRQ(ierr); 27649577ea80SStefano Zampini } 2765304d26faSStefano Zampini /* Allow user's customization */ 2766304d26faSStefano Zampini ierr = KSPSetFromOptions(pcbddc->ksp_D);CHKERRQ(ierr); 2767304d26faSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 2768304d26faSStefano Zampini } 2769d1e9a80fSBarry Smith ierr = KSPSetOperators(pcbddc->ksp_D,pcis->A_II,pcis->A_II);CHKERRQ(ierr); 2770df4d28bfSStefano Zampini if (sub_schurs->reuse_solver) { 2771df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 2772d62866d3SStefano Zampini 2773df4d28bfSStefano Zampini ierr = KSPSetPC(pcbddc->ksp_D,reuse_solver->interior_solver);CHKERRQ(ierr); 2774d5574798SStefano Zampini } 2775304d26faSStefano Zampini /* umfpack interface has a bug when matrix dimension is zero. TODO solve from umfpack interface */ 2776304d26faSStefano Zampini if (!n_D) { 2777304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_D,&pc_temp);CHKERRQ(ierr); 2778304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCNONE);CHKERRQ(ierr); 2779304d26faSStefano Zampini } 2780304d26faSStefano Zampini /* Set Up KSP for Dirichlet problem of BDDC */ 2781304d26faSStefano Zampini ierr = KSPSetUp(pcbddc->ksp_D);CHKERRQ(ierr); 2782304d26faSStefano Zampini /* set ksp_D into pcis data */ 2783304d26faSStefano Zampini ierr = KSPDestroy(&pcis->ksp_D);CHKERRQ(ierr); 2784304d26faSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->ksp_D);CHKERRQ(ierr); 2785304d26faSStefano Zampini pcis->ksp_D = pcbddc->ksp_D; 2786684f6988SStefano Zampini } 2787304d26faSStefano Zampini 2788304d26faSStefano Zampini /* NEUMANN PROBLEM */ 2789684f6988SStefano Zampini A_RR = 0; 2790684f6988SStefano Zampini if (neumann) { 2791d62866d3SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 279204708bb6SStefano Zampini PetscInt ibs,mbs; 279304708bb6SStefano Zampini PetscBool issbaij; 279404708bb6SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 2795f4ddd8eeSStefano Zampini /* Matrix for Neumann problem is A_RR -> we need to create/reuse it at this point */ 27968ce42a96SStefano Zampini ierr = ISGetSize(pcbddc->is_R_local,&n_R);CHKERRQ(ierr); 2797f4ddd8eeSStefano Zampini if (pcbddc->ksp_R) { /* already created ksp */ 2798f4ddd8eeSStefano Zampini PetscInt nn_R; 279981d9aea3SBarry Smith ierr = KSPGetOperators(pcbddc->ksp_R,NULL,&A_RR);CHKERRQ(ierr); 2800f4ddd8eeSStefano Zampini ierr = PetscObjectReference((PetscObject)A_RR);CHKERRQ(ierr); 2801f4ddd8eeSStefano Zampini ierr = MatGetSize(A_RR,&nn_R,NULL);CHKERRQ(ierr); 2802f4ddd8eeSStefano Zampini if (nn_R != n_R) { /* old ksp is not reusable, so reset it */ 2803f4ddd8eeSStefano Zampini ierr = KSPReset(pcbddc->ksp_R);CHKERRQ(ierr); 2804f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 2805f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 2806f4ddd8eeSStefano Zampini } else { /* same sizes, but nonzero pattern depend on primal vertices so it can be changed */ 2807727cdba6SStefano Zampini if (pcbddc->new_primal_space_local) { /* we are not sure the matrix will have the same nonzero pattern */ 2808f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 2809f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 2810f4ddd8eeSStefano Zampini } else { /* safe to reuse the matrix */ 2811f4ddd8eeSStefano Zampini reuse = MAT_REUSE_MATRIX; 2812f4ddd8eeSStefano Zampini } 2813f4ddd8eeSStefano Zampini } 2814f4ddd8eeSStefano Zampini /* last check */ 2815d1e9a80fSBarry Smith if (pc->flag == DIFFERENT_NONZERO_PATTERN) { 2816f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 2817f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 2818f4ddd8eeSStefano Zampini } 2819f4ddd8eeSStefano Zampini } else { /* first time, so we need to create the matrix */ 2820f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 2821f4ddd8eeSStefano Zampini } 2822f4ddd8eeSStefano Zampini /* extract A_RR */ 2823af732b37SStefano Zampini ierr = MatGetBlockSize(pcbddc->local_mat,&mbs);CHKERRQ(ierr); 2824af732b37SStefano Zampini ierr = ISGetBlockSize(pcbddc->is_R_local,&ibs);CHKERRQ(ierr); 282504708bb6SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcbddc->local_mat,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 282604708bb6SStefano Zampini if (ibs != mbs) { /* need to convert to SEQAIJ to extract any submatrix with is_R_local */ 282704708bb6SStefano Zampini if (matis->A == pcbddc->local_mat) { 282804708bb6SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 282904708bb6SStefano Zampini ierr = MatConvert(matis->A,MATSEQAIJ,MAT_INITIAL_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 2830af732b37SStefano Zampini } else { 283104708bb6SStefano Zampini ierr = MatConvert(pcbddc->local_mat,MATSEQAIJ,MAT_REUSE_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 28326816873aSStefano Zampini } 283304708bb6SStefano Zampini } else if (issbaij) { /* need to convert to BAIJ to get offdiagonal blocks */ 283404708bb6SStefano Zampini if (matis->A == pcbddc->local_mat) { 283504708bb6SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 283604708bb6SStefano Zampini ierr = MatConvert(matis->A,MATSEQBAIJ,MAT_INITIAL_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 283704708bb6SStefano Zampini } else { 283804708bb6SStefano Zampini ierr = MatConvert(pcbddc->local_mat,MATSEQBAIJ,MAT_REUSE_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 283904708bb6SStefano Zampini } 284004708bb6SStefano Zampini } 284116e386b8SStefano Zampini if (pcbddc->benign_n && !pcbddc->benign_change_explicit && pcbddc->dbg_flag) { 284216e386b8SStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 284316e386b8SStefano Zampini ierr = PCBDDCBenignProject(pc,pcbddc->is_R_local,pcbddc->is_R_local,&A_RR);CHKERRQ(ierr); 284416e386b8SStefano Zampini } else { 2845f4ddd8eeSStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,pcbddc->is_R_local,pcbddc->is_R_local,reuse,&A_RR);CHKERRQ(ierr); 284616e386b8SStefano Zampini } 28473301b35fSStefano Zampini if (pcbddc->local_mat->symmetric_set) { 28483301b35fSStefano Zampini ierr = MatSetOption(A_RR,MAT_SYMMETRIC,pcbddc->local_mat->symmetric_set);CHKERRQ(ierr); 28496816873aSStefano Zampini } 2850f4ddd8eeSStefano Zampini if (!pcbddc->ksp_R) { /* create object if not present */ 2851304d26faSStefano Zampini ierr = KSPCreate(PETSC_COMM_SELF,&pcbddc->ksp_R);CHKERRQ(ierr); 2852304d26faSStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->ksp_R,(PetscObject)pc,1);CHKERRQ(ierr); 2853304d26faSStefano Zampini /* default */ 2854304d26faSStefano Zampini ierr = KSPSetType(pcbddc->ksp_R,KSPPREONLY);CHKERRQ(ierr); 2855e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->ksp_R,neu_prefix);CHKERRQ(ierr); 2856304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_temp);CHKERRQ(ierr); 28579577ea80SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)A_RR,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 28589577ea80SStefano Zampini if (issbaij) { 28599577ea80SStefano Zampini ierr = PCSetType(pc_temp,PCCHOLESKY);CHKERRQ(ierr); 28609577ea80SStefano Zampini } else { 2861304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCLU);CHKERRQ(ierr); 28629577ea80SStefano Zampini } 2863304d26faSStefano Zampini /* Allow user's customization */ 2864304d26faSStefano Zampini ierr = KSPSetFromOptions(pcbddc->ksp_R);CHKERRQ(ierr); 2865304d26faSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 2866304d26faSStefano Zampini } 2867304d26faSStefano Zampini /* umfpack interface has a bug when matrix dimension is zero. TODO solve from umfpack interface */ 2868304d26faSStefano Zampini if (!n_R) { 2869304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_temp);CHKERRQ(ierr); 2870304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCNONE);CHKERRQ(ierr); 2871304d26faSStefano Zampini } 2872df4d28bfSStefano Zampini /* Reuse solver if it is present */ 2873df4d28bfSStefano Zampini if (sub_schurs->reuse_solver) { 2874df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 2875d62866d3SStefano Zampini 2876df4d28bfSStefano Zampini ierr = KSPSetPC(pcbddc->ksp_R,reuse_solver->correction_solver);CHKERRQ(ierr); 2877d62866d3SStefano Zampini } 2878ffd830a3SStefano Zampini ierr = KSPSetOperators(pcbddc->ksp_R,A_RR,A_RR);CHKERRQ(ierr); 2879304d26faSStefano Zampini /* Set Up KSP for Neumann problem of BDDC */ 2880304d26faSStefano Zampini ierr = KSPSetUp(pcbddc->ksp_R);CHKERRQ(ierr); 2881684f6988SStefano Zampini } 28826816873aSStefano Zampini /* free Neumann problem's matrix */ 28836816873aSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 2884304d26faSStefano Zampini 2885304d26faSStefano Zampini /* check Dirichlet and Neumann solvers and adapt them if a nullspace correction is needed */ 28860fccc4e9SStefano Zampini if (pcbddc->NullSpace || pcbddc->dbg_flag) { 2887684f6988SStefano Zampini if (pcbddc->dbg_flag) { 2888684f6988SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 28891575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 2890684f6988SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 2891684f6988SStefano Zampini } 2892684f6988SStefano Zampini if (dirichlet) { /* Dirichlet */ 28930fccc4e9SStefano Zampini ierr = VecSetRandom(pcis->vec1_D,NULL);CHKERRQ(ierr); 28940fccc4e9SStefano Zampini ierr = MatMult(pcis->A_II,pcis->vec1_D,pcis->vec2_D);CHKERRQ(ierr); 28950fccc4e9SStefano Zampini ierr = KSPSolve(pcbddc->ksp_D,pcis->vec2_D,pcis->vec2_D);CHKERRQ(ierr); 28960fccc4e9SStefano Zampini ierr = VecAXPY(pcis->vec1_D,m_one,pcis->vec2_D);CHKERRQ(ierr); 28970fccc4e9SStefano Zampini ierr = VecNorm(pcis->vec1_D,NORM_INFINITY,&value);CHKERRQ(ierr); 2898304d26faSStefano Zampini /* need to be adapted? */ 2899b8ffe317SStefano Zampini use_exact = (PetscAbsReal(value) > 1.e-4 ? PETSC_FALSE : PETSC_TRUE); 2900b8ffe317SStefano Zampini ierr = MPI_Allreduce(&use_exact,&use_exact_reduced,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 2901b8ffe317SStefano Zampini ierr = PCBDDCSetUseExactDirichlet(pc,use_exact_reduced);CHKERRQ(ierr); 2902304d26faSStefano Zampini /* print info */ 2903304d26faSStefano Zampini if (pcbddc->dbg_flag) { 2904e604994aSStefano 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); 2905304d26faSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 2906304d26faSStefano Zampini } 2907b8ffe317SStefano Zampini if (pcbddc->NullSpace && !use_exact_reduced && !pcbddc->switch_static) { 2908298c0119SStefano Zampini ierr = PCBDDCNullSpaceAssembleCorrection(pc,PETSC_TRUE,pcis->is_I_local);CHKERRQ(ierr); 2909304d26faSStefano Zampini } 2910684f6988SStefano Zampini } 2911684f6988SStefano Zampini if (neumann) { /* Neumann */ 29126816873aSStefano Zampini ierr = KSPGetOperators(pcbddc->ksp_R,&A_RR,NULL);CHKERRQ(ierr); 29130fccc4e9SStefano Zampini ierr = VecSetRandom(pcbddc->vec1_R,NULL);CHKERRQ(ierr); 29140fccc4e9SStefano Zampini ierr = MatMult(A_RR,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 29150fccc4e9SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec2_R,pcbddc->vec2_R);CHKERRQ(ierr); 29160fccc4e9SStefano Zampini ierr = VecAXPY(pcbddc->vec1_R,m_one,pcbddc->vec2_R);CHKERRQ(ierr); 29170fccc4e9SStefano Zampini ierr = VecNorm(pcbddc->vec1_R,NORM_INFINITY,&value);CHKERRQ(ierr); 2918304d26faSStefano Zampini /* need to be adapted? */ 2919b8ffe317SStefano Zampini use_exact = (PetscAbsReal(value) > 1.e-4 ? PETSC_FALSE : PETSC_TRUE); 2920b8ffe317SStefano Zampini ierr = MPI_Allreduce(&use_exact,&use_exact_reduced,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 2921304d26faSStefano Zampini /* print info */ 2922304d26faSStefano Zampini if (pcbddc->dbg_flag) { 2923e604994aSStefano 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); 2924304d26faSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 2925304d26faSStefano Zampini } 2926b8ffe317SStefano Zampini if (pcbddc->NullSpace && !use_exact_reduced) { /* is it the right logic? */ 2927298c0119SStefano Zampini ierr = PCBDDCNullSpaceAssembleCorrection(pc,PETSC_FALSE,pcbddc->is_R_local);CHKERRQ(ierr); 2928304d26faSStefano Zampini } 29290fccc4e9SStefano Zampini } 2930684f6988SStefano Zampini } 2931304d26faSStefano Zampini PetscFunctionReturn(0); 2932304d26faSStefano Zampini } 2933304d26faSStefano Zampini 2934304d26faSStefano Zampini #undef __FUNCT__ 2935ba15a52eSStefano Zampini #define __FUNCT__ "PCBDDCSolveSubstructureCorrection" 293680677318SStefano Zampini static PetscErrorCode PCBDDCSolveSubstructureCorrection(PC pc, Vec inout_B, Vec inout_D, PetscBool applytranspose) 2937674ae819SStefano Zampini { 2938674ae819SStefano Zampini PetscErrorCode ierr; 2939674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 2940be83ff47SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 2941674ae819SStefano Zampini 2942674ae819SStefano Zampini PetscFunctionBegin; 2943df4d28bfSStefano Zampini if (!sub_schurs->reuse_solver) { 294480677318SStefano Zampini ierr = VecSet(pcbddc->vec1_R,0.);CHKERRQ(ierr); 294520c7b377SStefano Zampini } 294680677318SStefano Zampini if (!pcbddc->switch_static) { 294780677318SStefano Zampini if (applytranspose && pcbddc->local_auxmat1) { 294880677318SStefano Zampini ierr = MatMultTranspose(pcbddc->local_auxmat2,inout_B,pcbddc->vec1_C);CHKERRQ(ierr); 294980677318SStefano Zampini ierr = MatMultTransposeAdd(pcbddc->local_auxmat1,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr); 295020c7b377SStefano Zampini } 2951df4d28bfSStefano Zampini if (!sub_schurs->reuse_solver) { 295280677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 295380677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 295420c7b377SStefano Zampini } else { 2955df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 2956be83ff47SStefano Zampini 2957df4d28bfSStefano Zampini ierr = VecScatterBegin(reuse_solver->correction_scatter_B,inout_B,reuse_solver->rhs_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2958df4d28bfSStefano Zampini ierr = VecScatterEnd(reuse_solver->correction_scatter_B,inout_B,reuse_solver->rhs_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 295920c7b377SStefano Zampini } 2960be83ff47SStefano Zampini } else { 296180677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 296280677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 296380677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,inout_D,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 296480677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,inout_D,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 296580677318SStefano Zampini if (applytranspose && pcbddc->local_auxmat1) { 296680677318SStefano Zampini ierr = MatMultTranspose(pcbddc->local_auxmat2,pcbddc->vec1_R,pcbddc->vec1_C);CHKERRQ(ierr); 296780677318SStefano Zampini ierr = MatMultTransposeAdd(pcbddc->local_auxmat1,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr); 296880677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 296980677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2970674ae819SStefano Zampini } 2971674ae819SStefano Zampini } 2972df4d28bfSStefano Zampini if (!sub_schurs->reuse_solver) { 297380677318SStefano Zampini if (applytranspose) { 297480677318SStefano Zampini ierr = KSPSolveTranspose(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 297580677318SStefano Zampini } else { 297680677318SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 297780677318SStefano Zampini } 2978be83ff47SStefano Zampini } else { 2979df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 2980be83ff47SStefano Zampini 2981be83ff47SStefano Zampini if (applytranspose) { 2982df4d28bfSStefano Zampini ierr = MatFactorSolveSchurComplementTranspose(reuse_solver->F,reuse_solver->rhs_B,reuse_solver->sol_B);CHKERRQ(ierr); 2983be83ff47SStefano Zampini } else { 2984df4d28bfSStefano Zampini ierr = MatFactorSolveSchurComplement(reuse_solver->F,reuse_solver->rhs_B,reuse_solver->sol_B);CHKERRQ(ierr); 2985be83ff47SStefano Zampini } 2986be83ff47SStefano Zampini } 298780677318SStefano Zampini ierr = VecSet(inout_B,0.);CHKERRQ(ierr); 298880677318SStefano Zampini if (!pcbddc->switch_static) { 2989df4d28bfSStefano Zampini if (!sub_schurs->reuse_solver) { 299080677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 299180677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2992be83ff47SStefano Zampini } else { 2993df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 2994be83ff47SStefano Zampini 2995df4d28bfSStefano Zampini ierr = VecScatterBegin(reuse_solver->correction_scatter_B,reuse_solver->sol_B,inout_B,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2996df4d28bfSStefano Zampini ierr = VecScatterEnd(reuse_solver->correction_scatter_B,reuse_solver->sol_B,inout_B,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2997be83ff47SStefano Zampini } 299880677318SStefano Zampini if (!applytranspose && pcbddc->local_auxmat1) { 299980677318SStefano Zampini ierr = MatMult(pcbddc->local_auxmat1,inout_B,pcbddc->vec1_C);CHKERRQ(ierr); 300080677318SStefano Zampini ierr = MatMultAdd(pcbddc->local_auxmat2,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr); 300180677318SStefano Zampini } 300280677318SStefano Zampini } else { 300380677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 300480677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 300580677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 300680677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 300780677318SStefano Zampini if (!applytranspose && pcbddc->local_auxmat1) { 300880677318SStefano Zampini ierr = MatMult(pcbddc->local_auxmat1,inout_B,pcbddc->vec1_C);CHKERRQ(ierr); 300980677318SStefano Zampini ierr = MatMultAdd(pcbddc->local_auxmat2,pcbddc->vec1_C,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 301080677318SStefano Zampini } 301180677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 301280677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 301380677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 301480677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 3015674ae819SStefano Zampini } 3016674ae819SStefano Zampini PetscFunctionReturn(0); 3017674ae819SStefano Zampini } 3018674ae819SStefano Zampini 3019dc359a40SStefano Zampini /* parameter apply transpose determines if the interface preconditioner should be applied transposed or not */ 3020674ae819SStefano Zampini #undef __FUNCT__ 3021674ae819SStefano Zampini #define __FUNCT__ "PCBDDCApplyInterfacePreconditioner" 3022dc359a40SStefano Zampini PetscErrorCode PCBDDCApplyInterfacePreconditioner(PC pc, PetscBool applytranspose) 3023674ae819SStefano Zampini { 3024674ae819SStefano Zampini PetscErrorCode ierr; 3025674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 3026674ae819SStefano Zampini PC_IS* pcis = (PC_IS*) (pc->data); 3027674ae819SStefano Zampini const PetscScalar zero = 0.0; 3028674ae819SStefano Zampini 3029674ae819SStefano Zampini PetscFunctionBegin; 3030dc359a40SStefano Zampini /* Application of PSI^T or PHI^T (depending on applytranspose, see comment above) */ 30314fee134fSStefano Zampini if (!pcbddc->benign_apply_coarse_only) { 3032dc359a40SStefano Zampini if (applytranspose) { 3033674ae819SStefano Zampini ierr = MatMultTranspose(pcbddc->coarse_phi_B,pcis->vec1_B,pcbddc->vec1_P);CHKERRQ(ierr); 30348eeda7d8SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultTransposeAdd(pcbddc->coarse_phi_D,pcis->vec1_D,pcbddc->vec1_P,pcbddc->vec1_P);CHKERRQ(ierr); } 3035dc359a40SStefano Zampini } else { 3036674ae819SStefano Zampini ierr = MatMultTranspose(pcbddc->coarse_psi_B,pcis->vec1_B,pcbddc->vec1_P);CHKERRQ(ierr); 3037674ae819SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultTransposeAdd(pcbddc->coarse_psi_D,pcis->vec1_D,pcbddc->vec1_P,pcbddc->vec1_P);CHKERRQ(ierr); } 303815aaf578SStefano Zampini } 30394fee134fSStefano Zampini } else { 30404fee134fSStefano Zampini ierr = VecSet(pcbddc->vec1_P,zero);CHKERRQ(ierr); 30414fee134fSStefano Zampini } 3042efc2fbd9SStefano Zampini 3043efc2fbd9SStefano Zampini /* add p0 to the last value of vec1_P holding the coarse dof relative to p0 */ 30444f1b2e48SStefano Zampini if (pcbddc->benign_n) { 3045efc2fbd9SStefano Zampini PetscScalar *array; 30464f1b2e48SStefano Zampini PetscInt j; 3047efc2fbd9SStefano Zampini 3048efc2fbd9SStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 30494f1b2e48SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) array[pcbddc->local_primal_size-pcbddc->benign_n+j] += pcbddc->benign_p0[j]; 3050efc2fbd9SStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 3051efc2fbd9SStefano Zampini } 3052efc2fbd9SStefano Zampini 305312edc857SStefano Zampini /* start communications from local primal nodes to rhs of coarse solver */ 305412edc857SStefano Zampini ierr = VecSet(pcbddc->coarse_vec,zero);CHKERRQ(ierr); 305512edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataBegin(pc,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 305612edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataEnd(pc,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 305712edc857SStefano Zampini 30589f00e9b4SStefano Zampini /* Coarse solution -> rhs and sol updated inside PCBDDCScattarCoarseDataBegin/End */ 305912edc857SStefano Zampini /* TODO remove null space when doing multilevel */ 306012edc857SStefano Zampini if (pcbddc->coarse_ksp) { 306151694757SStefano Zampini Mat coarse_mat; 3062964fefecSStefano Zampini Vec rhs,sol; 306351694757SStefano Zampini MatNullSpace nullsp; 3064964fefecSStefano Zampini 3065964fefecSStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&rhs);CHKERRQ(ierr); 3066964fefecSStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&sol);CHKERRQ(ierr); 306751694757SStefano Zampini ierr = KSPGetOperators(pcbddc->coarse_ksp,&coarse_mat,NULL);CHKERRQ(ierr); 306851694757SStefano Zampini ierr = MatGetNullSpace(coarse_mat,&nullsp);CHKERRQ(ierr); 306951694757SStefano Zampini if (nullsp) { 307051694757SStefano Zampini ierr = MatNullSpaceRemove(nullsp,rhs);CHKERRQ(ierr); 307151694757SStefano Zampini } 307212edc857SStefano Zampini if (applytranspose) { 3073964fefecSStefano Zampini ierr = KSPSolveTranspose(pcbddc->coarse_ksp,rhs,sol);CHKERRQ(ierr); 307412edc857SStefano Zampini } else { 3075964fefecSStefano Zampini ierr = KSPSolve(pcbddc->coarse_ksp,rhs,sol);CHKERRQ(ierr); 307612edc857SStefano Zampini } 307751694757SStefano Zampini if (nullsp) { 307851694757SStefano Zampini ierr = MatNullSpaceRemove(nullsp,sol);CHKERRQ(ierr); 307951694757SStefano Zampini } 308012edc857SStefano Zampini } 3081674ae819SStefano Zampini 3082674ae819SStefano Zampini /* Local solution on R nodes */ 30834fee134fSStefano Zampini if (pcis->n && !pcbddc->benign_apply_coarse_only) { 308480677318SStefano Zampini ierr = PCBDDCSolveSubstructureCorrection(pc,pcis->vec1_B,pcis->vec1_D,applytranspose);CHKERRQ(ierr); 30859f00e9b4SStefano Zampini } 30869f00e9b4SStefano Zampini /* communications from coarse sol to local primal nodes */ 30879f00e9b4SStefano Zampini ierr = PCBDDCScatterCoarseDataBegin(pc,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 308812edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataEnd(pc,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 3089674ae819SStefano Zampini 30904fee134fSStefano Zampini /* Sum contributions from the two levels */ 30914fee134fSStefano Zampini if (!pcbddc->benign_apply_coarse_only) { 3092dc359a40SStefano Zampini if (applytranspose) { 3093dc359a40SStefano Zampini ierr = MatMultAdd(pcbddc->coarse_psi_B,pcbddc->vec1_P,pcis->vec1_B,pcis->vec1_B);CHKERRQ(ierr); 3094dc359a40SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultAdd(pcbddc->coarse_psi_D,pcbddc->vec1_P,pcis->vec1_D,pcis->vec1_D);CHKERRQ(ierr); } 3095dc359a40SStefano Zampini } else { 3096674ae819SStefano Zampini ierr = MatMultAdd(pcbddc->coarse_phi_B,pcbddc->vec1_P,pcis->vec1_B,pcis->vec1_B);CHKERRQ(ierr); 30978eeda7d8SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultAdd(pcbddc->coarse_phi_D,pcbddc->vec1_P,pcis->vec1_D,pcis->vec1_D);CHKERRQ(ierr); } 3098dc359a40SStefano Zampini } 3099efc2fbd9SStefano Zampini /* store p0 */ 31004f1b2e48SStefano Zampini if (pcbddc->benign_n) { 3101efc2fbd9SStefano Zampini PetscScalar *array; 31024f1b2e48SStefano Zampini PetscInt j; 3103efc2fbd9SStefano Zampini 3104efc2fbd9SStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 31054f1b2e48SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) pcbddc->benign_p0[j] = array[pcbddc->local_primal_size-pcbddc->benign_n+j]; 3106efc2fbd9SStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 3107efc2fbd9SStefano Zampini } 31084fee134fSStefano Zampini } else { /* expand the coarse solution */ 31094fee134fSStefano Zampini if (applytranspose) { 31104fee134fSStefano Zampini ierr = MatMult(pcbddc->coarse_psi_B,pcbddc->vec1_P,pcis->vec1_B);CHKERRQ(ierr); 31114fee134fSStefano Zampini } else { 31124fee134fSStefano Zampini ierr = MatMult(pcbddc->coarse_phi_B,pcbddc->vec1_P,pcis->vec1_B);CHKERRQ(ierr); 31134fee134fSStefano Zampini } 31144fee134fSStefano Zampini } 3115674ae819SStefano Zampini PetscFunctionReturn(0); 3116674ae819SStefano Zampini } 3117674ae819SStefano Zampini 3118674ae819SStefano Zampini #undef __FUNCT__ 3119674ae819SStefano Zampini #define __FUNCT__ "PCBDDCScatterCoarseDataBegin" 312012edc857SStefano Zampini PetscErrorCode PCBDDCScatterCoarseDataBegin(PC pc,InsertMode imode, ScatterMode smode) 3121674ae819SStefano Zampini { 3122674ae819SStefano Zampini PetscErrorCode ierr; 3123674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 312458da7f69SStefano Zampini PetscScalar *array; 312512edc857SStefano Zampini Vec from,to; 3126674ae819SStefano Zampini 3127674ae819SStefano Zampini PetscFunctionBegin; 312812edc857SStefano Zampini if (smode == SCATTER_REVERSE) { /* from global to local -> get data from coarse solution */ 312912edc857SStefano Zampini from = pcbddc->coarse_vec; 313012edc857SStefano Zampini to = pcbddc->vec1_P; 313112edc857SStefano Zampini if (pcbddc->coarse_ksp) { /* get array from coarse processes */ 313212edc857SStefano Zampini Vec tvec; 313358da7f69SStefano Zampini 313458da7f69SStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 313558da7f69SStefano Zampini ierr = VecResetArray(tvec);CHKERRQ(ierr); 313612edc857SStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 313758da7f69SStefano Zampini ierr = VecGetArray(tvec,&array);CHKERRQ(ierr); 313858da7f69SStefano Zampini ierr = VecPlaceArray(from,array);CHKERRQ(ierr); 313958da7f69SStefano Zampini ierr = VecRestoreArray(tvec,&array);CHKERRQ(ierr); 314012edc857SStefano Zampini } 314112edc857SStefano Zampini } else { /* from local to global -> put data in coarse right hand side */ 314212edc857SStefano Zampini from = pcbddc->vec1_P; 314312edc857SStefano Zampini to = pcbddc->coarse_vec; 314412edc857SStefano Zampini } 314512edc857SStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,from,to,imode,smode);CHKERRQ(ierr); 3146674ae819SStefano Zampini PetscFunctionReturn(0); 3147674ae819SStefano Zampini } 3148674ae819SStefano Zampini 3149674ae819SStefano Zampini #undef __FUNCT__ 3150674ae819SStefano Zampini #define __FUNCT__ "PCBDDCScatterCoarseDataEnd" 315112edc857SStefano Zampini PetscErrorCode PCBDDCScatterCoarseDataEnd(PC pc, InsertMode imode, ScatterMode smode) 3152674ae819SStefano Zampini { 3153674ae819SStefano Zampini PetscErrorCode ierr; 3154674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 315558da7f69SStefano Zampini PetscScalar *array; 315612edc857SStefano Zampini Vec from,to; 3157674ae819SStefano Zampini 3158674ae819SStefano Zampini PetscFunctionBegin; 315912edc857SStefano Zampini if (smode == SCATTER_REVERSE) { /* from global to local -> get data from coarse solution */ 316012edc857SStefano Zampini from = pcbddc->coarse_vec; 316112edc857SStefano Zampini to = pcbddc->vec1_P; 316212edc857SStefano Zampini } else { /* from local to global -> put data in coarse right hand side */ 316312edc857SStefano Zampini from = pcbddc->vec1_P; 316412edc857SStefano Zampini to = pcbddc->coarse_vec; 316512edc857SStefano Zampini } 316612edc857SStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,from,to,imode,smode);CHKERRQ(ierr); 316712edc857SStefano Zampini if (smode == SCATTER_FORWARD) { 316812edc857SStefano Zampini if (pcbddc->coarse_ksp) { /* get array from coarse processes */ 316912edc857SStefano Zampini Vec tvec; 317058da7f69SStefano Zampini 317112edc857SStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 317258da7f69SStefano Zampini ierr = VecGetArray(to,&array);CHKERRQ(ierr); 317358da7f69SStefano Zampini ierr = VecPlaceArray(tvec,array);CHKERRQ(ierr); 317458da7f69SStefano Zampini ierr = VecRestoreArray(to,&array);CHKERRQ(ierr); 317558da7f69SStefano Zampini } 317658da7f69SStefano Zampini } else { 317758da7f69SStefano Zampini if (pcbddc->coarse_ksp) { /* restore array of pcbddc->coarse_vec */ 317858da7f69SStefano Zampini ierr = VecResetArray(from);CHKERRQ(ierr); 317912edc857SStefano Zampini } 318012edc857SStefano Zampini } 3181674ae819SStefano Zampini PetscFunctionReturn(0); 3182674ae819SStefano Zampini } 3183674ae819SStefano Zampini 3184984c4197SStefano Zampini /* uncomment for testing purposes */ 3185984c4197SStefano Zampini /* #define PETSC_MISSING_LAPACK_GESVD 1 */ 3186674ae819SStefano Zampini #undef __FUNCT__ 3187674ae819SStefano Zampini #define __FUNCT__ "PCBDDCConstraintsSetUp" 3188674ae819SStefano Zampini PetscErrorCode PCBDDCConstraintsSetUp(PC pc) 3189674ae819SStefano Zampini { 3190674ae819SStefano Zampini PetscErrorCode ierr; 3191674ae819SStefano Zampini PC_IS* pcis = (PC_IS*)(pc->data); 3192674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 3193674ae819SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 3194984c4197SStefano Zampini /* one and zero */ 3195984c4197SStefano Zampini PetscScalar one=1.0,zero=0.0; 3196984c4197SStefano Zampini /* space to store constraints and their local indices */ 31979162d606SStefano Zampini PetscScalar *constraints_data; 31989162d606SStefano Zampini PetscInt *constraints_idxs,*constraints_idxs_B; 31999162d606SStefano Zampini PetscInt *constraints_idxs_ptr,*constraints_data_ptr; 32009162d606SStefano Zampini PetscInt *constraints_n; 3201984c4197SStefano Zampini /* iterators */ 3202b3d85658SStefano Zampini PetscInt i,j,k,total_counts,total_counts_cc,cum; 3203984c4197SStefano Zampini /* BLAS integers */ 3204e310c8b4SStefano Zampini PetscBLASInt lwork,lierr; 3205e310c8b4SStefano Zampini PetscBLASInt Blas_N,Blas_M,Blas_K,Blas_one=1; 3206c4303822SStefano Zampini PetscBLASInt Blas_LDA,Blas_LDB,Blas_LDC; 3207727cdba6SStefano Zampini /* reuse */ 32080e6343abSStefano Zampini PetscInt olocal_primal_size,olocal_primal_size_cc; 32090e6343abSStefano Zampini PetscInt *olocal_primal_ref_node,*olocal_primal_ref_mult; 3210984c4197SStefano Zampini /* change of basis */ 3211b3d85658SStefano Zampini PetscBool qr_needed; 32129162d606SStefano Zampini PetscBT change_basis,qr_needed_idx; 3213984c4197SStefano Zampini /* auxiliary stuff */ 321464efe560SStefano Zampini PetscInt *nnz,*is_indices; 32158a0068c3SStefano Zampini PetscInt ncc; 3216984c4197SStefano Zampini /* some quantities */ 321745a1bb75SStefano Zampini PetscInt n_vertices,total_primal_vertices,valid_constraints; 3218a58a30b4SStefano Zampini PetscInt size_of_constraint,max_size_of_constraint=0,max_constraints,temp_constraints; 3219984c4197SStefano Zampini 3220674ae819SStefano Zampini PetscFunctionBegin; 32218e61c736SStefano Zampini /* Destroy Mat objects computed previously */ 32228e61c736SStefano Zampini ierr = MatDestroy(&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 32238e61c736SStefano Zampini ierr = MatDestroy(&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 3224088faed8SStefano Zampini /* save info on constraints from previous setup (if any) */ 3225088faed8SStefano Zampini olocal_primal_size = pcbddc->local_primal_size; 32260e6343abSStefano Zampini olocal_primal_size_cc = pcbddc->local_primal_size_cc; 32270e6343abSStefano Zampini ierr = PetscMalloc2(olocal_primal_size_cc,&olocal_primal_ref_node,olocal_primal_size_cc,&olocal_primal_ref_mult);CHKERRQ(ierr); 32280e6343abSStefano Zampini ierr = PetscMemcpy(olocal_primal_ref_node,pcbddc->local_primal_ref_node,olocal_primal_size_cc*sizeof(PetscInt));CHKERRQ(ierr); 32290e6343abSStefano Zampini ierr = PetscMemcpy(olocal_primal_ref_mult,pcbddc->local_primal_ref_mult,olocal_primal_size_cc*sizeof(PetscInt));CHKERRQ(ierr); 32300e6343abSStefano Zampini ierr = PetscFree2(pcbddc->local_primal_ref_node,pcbddc->local_primal_ref_mult);CHKERRQ(ierr); 3231088faed8SStefano Zampini ierr = PetscFree(pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 3232cf5a6209SStefano Zampini 3233cf5a6209SStefano Zampini /* print some info */ 3234cf5a6209SStefano Zampini if (pcbddc->dbg_flag) { 3235cf5a6209SStefano Zampini IS vertices; 3236cf5a6209SStefano Zampini PetscInt nv,nedges,nfaces; 3237cf5a6209SStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,&nfaces,NULL,&nedges,NULL,&vertices);CHKERRQ(ierr); 3238cf5a6209SStefano Zampini ierr = ISGetSize(vertices,&nv);CHKERRQ(ierr); 3239cf5a6209SStefano Zampini ierr = ISDestroy(&vertices);CHKERRQ(ierr); 32401575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 3241cf5a6209SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"--------------------------------------------------------------\n");CHKERRQ(ierr); 3242cf5a6209SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate vertices (%d)\n",PetscGlobalRank,nv,pcbddc->use_vertices);CHKERRQ(ierr); 3243fd14bc51SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate edges (%d)\n",PetscGlobalRank,nedges,pcbddc->use_edges);CHKERRQ(ierr); 3244fd14bc51SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate faces (%d)\n",PetscGlobalRank,nfaces,pcbddc->use_faces);CHKERRQ(ierr); 3245cf5a6209SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 32461575c14dSBarry Smith ierr = PetscViewerASCIIPopSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 3247cf5a6209SStefano Zampini } 3248cf5a6209SStefano Zampini 3249cf5a6209SStefano Zampini if (!pcbddc->adaptive_selection) { 32509162d606SStefano Zampini IS ISForVertices,*ISForFaces,*ISForEdges; 3251cf5a6209SStefano Zampini MatNullSpace nearnullsp; 3252cf5a6209SStefano Zampini const Vec *nearnullvecs; 3253cf5a6209SStefano Zampini Vec *localnearnullsp; 3254cf5a6209SStefano Zampini PetscScalar *array; 3255cf5a6209SStefano Zampini PetscInt n_ISForFaces,n_ISForEdges,nnsp_size; 3256cf5a6209SStefano Zampini PetscBool nnsp_has_cnst; 3257674ae819SStefano Zampini /* LAPACK working arrays for SVD or POD */ 3258b3d85658SStefano Zampini PetscBool skip_lapack,boolforchange; 3259674ae819SStefano Zampini PetscScalar *work; 3260674ae819SStefano Zampini PetscReal *singular_vals; 3261674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX) 3262674ae819SStefano Zampini PetscReal *rwork; 3263674ae819SStefano Zampini #endif 3264674ae819SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 3265674ae819SStefano Zampini PetscScalar *temp_basis,*correlation_mat; 3266674ae819SStefano Zampini #else 3267964fefecSStefano Zampini PetscBLASInt dummy_int=1; 3268964fefecSStefano Zampini PetscScalar dummy_scalar=1.; 3269674ae819SStefano Zampini #endif 3270674ae819SStefano Zampini 3271674ae819SStefano Zampini /* Get index sets for faces, edges and vertices from graph */ 3272d06fc5fdSStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,&n_ISForFaces,&ISForFaces,&n_ISForEdges,&ISForEdges,&ISForVertices);CHKERRQ(ierr); 3273d06fc5fdSStefano Zampini /* free unneeded index sets */ 3274d06fc5fdSStefano Zampini if (!pcbddc->use_vertices) { 3275d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForVertices);CHKERRQ(ierr); 3276674ae819SStefano Zampini } 3277d06fc5fdSStefano Zampini if (!pcbddc->use_edges) { 3278d06fc5fdSStefano Zampini for (i=0;i<n_ISForEdges;i++) { 3279d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForEdges[i]);CHKERRQ(ierr); 3280d06fc5fdSStefano Zampini } 3281d06fc5fdSStefano Zampini ierr = PetscFree(ISForEdges);CHKERRQ(ierr); 3282d06fc5fdSStefano Zampini n_ISForEdges = 0; 3283d06fc5fdSStefano Zampini } 3284d06fc5fdSStefano Zampini if (!pcbddc->use_faces) { 3285d06fc5fdSStefano Zampini for (i=0;i<n_ISForFaces;i++) { 3286d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForFaces[i]);CHKERRQ(ierr); 3287d06fc5fdSStefano Zampini } 3288d06fc5fdSStefano Zampini ierr = PetscFree(ISForFaces);CHKERRQ(ierr); 3289d06fc5fdSStefano Zampini n_ISForFaces = 0; 3290d06fc5fdSStefano Zampini } 329170022509SStefano Zampini 329270022509SStefano Zampini #if defined(PETSC_USE_DEBUG) 329370022509SStefano Zampini /* HACK: when solving singular problems not using vertices, a change of basis is mandatory. 329470022509SStefano Zampini Also use_change_of_basis should be consistent among processors */ 329570022509SStefano Zampini if (pcbddc->NullSpace) { 329670022509SStefano Zampini PetscBool tbool[2],gbool[2]; 329770022509SStefano Zampini 329870022509SStefano Zampini if (!ISForVertices && !pcbddc->user_ChangeOfBasisMatrix) { 3299b8ffe317SStefano Zampini pcbddc->use_change_of_basis = PETSC_TRUE; 3300d06fc5fdSStefano Zampini if (!ISForEdges) { 3301d06fc5fdSStefano Zampini pcbddc->use_change_on_faces = PETSC_TRUE; 3302d06fc5fdSStefano Zampini } 3303b8ffe317SStefano Zampini } 3304d06fc5fdSStefano Zampini tbool[0] = pcbddc->use_change_of_basis; 3305d06fc5fdSStefano Zampini tbool[1] = pcbddc->use_change_on_faces; 3306d06fc5fdSStefano Zampini ierr = MPI_Allreduce(tbool,gbool,2,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 3307d06fc5fdSStefano Zampini pcbddc->use_change_of_basis = gbool[0]; 3308d06fc5fdSStefano Zampini pcbddc->use_change_on_faces = gbool[1]; 330998a51de6SStefano Zampini } 331070022509SStefano Zampini #endif 331108122e43SStefano Zampini 3312674ae819SStefano Zampini /* check if near null space is attached to global mat */ 3313674ae819SStefano Zampini ierr = MatGetNearNullSpace(pc->pmat,&nearnullsp);CHKERRQ(ierr); 3314674ae819SStefano Zampini if (nearnullsp) { 3315674ae819SStefano Zampini ierr = MatNullSpaceGetVecs(nearnullsp,&nnsp_has_cnst,&nnsp_size,&nearnullvecs);CHKERRQ(ierr); 3316f4ddd8eeSStefano Zampini /* remove any stored info */ 3317f4ddd8eeSStefano Zampini ierr = MatNullSpaceDestroy(&pcbddc->onearnullspace);CHKERRQ(ierr); 3318f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->onearnullvecs_state);CHKERRQ(ierr); 3319f4ddd8eeSStefano Zampini /* store information for BDDC solver reuse */ 3320f4ddd8eeSStefano Zampini ierr = PetscObjectReference((PetscObject)nearnullsp);CHKERRQ(ierr); 3321f4ddd8eeSStefano Zampini pcbddc->onearnullspace = nearnullsp; 3322473ba861SJed Brown ierr = PetscMalloc1(nnsp_size,&pcbddc->onearnullvecs_state);CHKERRQ(ierr); 3323f4ddd8eeSStefano Zampini for (i=0;i<nnsp_size;i++) { 3324f4ddd8eeSStefano Zampini ierr = PetscObjectStateGet((PetscObject)nearnullvecs[i],&pcbddc->onearnullvecs_state[i]);CHKERRQ(ierr); 3325f4ddd8eeSStefano Zampini } 3326984c4197SStefano Zampini } else { /* if near null space is not provided BDDC uses constants by default */ 3327984c4197SStefano Zampini nnsp_size = 0; 3328674ae819SStefano Zampini nnsp_has_cnst = PETSC_TRUE; 3329674ae819SStefano Zampini } 3330984c4197SStefano Zampini /* get max number of constraints on a single cc */ 3331984c4197SStefano Zampini max_constraints = nnsp_size; 3332984c4197SStefano Zampini if (nnsp_has_cnst) max_constraints++; 3333984c4197SStefano Zampini 3334674ae819SStefano Zampini /* 3335674ae819SStefano Zampini Evaluate maximum storage size needed by the procedure 33369162d606SStefano Zampini - Indices for connected component i stored at "constraints_idxs + constraints_idxs_ptr[i]" 33379162d606SStefano Zampini - Values for constraints on connected component i stored at "constraints_data + constraints_data_ptr[i]" 33389162d606SStefano Zampini There can be multiple constraints per connected component 3339674ae819SStefano Zampini */ 3340674ae819SStefano Zampini n_vertices = 0; 3341674ae819SStefano Zampini if (ISForVertices) { 3342674ae819SStefano Zampini ierr = ISGetSize(ISForVertices,&n_vertices);CHKERRQ(ierr); 3343674ae819SStefano Zampini } 33449162d606SStefano Zampini ncc = n_vertices+n_ISForFaces+n_ISForEdges; 33459162d606SStefano Zampini ierr = PetscMalloc3(ncc+1,&constraints_idxs_ptr,ncc+1,&constraints_data_ptr,ncc,&constraints_n);CHKERRQ(ierr); 33469162d606SStefano Zampini 33479162d606SStefano Zampini total_counts = n_ISForFaces+n_ISForEdges; 33489162d606SStefano Zampini total_counts *= max_constraints; 3349674ae819SStefano Zampini total_counts += n_vertices; 33504641a718SStefano Zampini ierr = PetscBTCreate(total_counts,&change_basis);CHKERRQ(ierr); 33519162d606SStefano Zampini 3352674ae819SStefano Zampini total_counts = 0; 3353674ae819SStefano Zampini max_size_of_constraint = 0; 3354674ae819SStefano Zampini for (i=0;i<n_ISForEdges+n_ISForFaces;i++) { 33559162d606SStefano Zampini IS used_is; 3356674ae819SStefano Zampini if (i<n_ISForEdges) { 33579162d606SStefano Zampini used_is = ISForEdges[i]; 3358674ae819SStefano Zampini } else { 33599162d606SStefano Zampini used_is = ISForFaces[i-n_ISForEdges]; 3360674ae819SStefano Zampini } 33619162d606SStefano Zampini ierr = ISGetSize(used_is,&j);CHKERRQ(ierr); 3362674ae819SStefano Zampini total_counts += j; 3363674ae819SStefano Zampini max_size_of_constraint = PetscMax(j,max_size_of_constraint); 3364674ae819SStefano Zampini } 33659162d606SStefano 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); 33669162d606SStefano Zampini 3367984c4197SStefano Zampini /* get local part of global near null space vectors */ 3368785e854fSJed Brown ierr = PetscMalloc1(nnsp_size,&localnearnullsp);CHKERRQ(ierr); 3369984c4197SStefano Zampini for (k=0;k<nnsp_size;k++) { 3370984c4197SStefano Zampini ierr = VecDuplicate(pcis->vec1_N,&localnearnullsp[k]);CHKERRQ(ierr); 3371e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,nearnullvecs[k],localnearnullsp[k],INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 3372e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,nearnullvecs[k],localnearnullsp[k],INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 3373984c4197SStefano Zampini } 3374674ae819SStefano Zampini 3375242a89d7SStefano Zampini /* whether or not to skip lapack calls */ 3376242a89d7SStefano Zampini skip_lapack = PETSC_TRUE; 3377a773dcb8SStefano Zampini if (n_ISForFaces+n_ISForEdges && max_constraints > 1 && !pcbddc->use_nnsp_true) skip_lapack = PETSC_FALSE; 3378242a89d7SStefano Zampini 3379984c4197SStefano Zampini /* First we issue queries to allocate optimal workspace for LAPACKgesvd (or LAPACKsyev if SVD is missing) */ 3380a773dcb8SStefano Zampini if (!skip_lapack) { 3381674ae819SStefano Zampini PetscScalar temp_work; 3382911cabfeSStefano Zampini 3383674ae819SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 3384984c4197SStefano Zampini /* Proper Orthogonal Decomposition (POD) using the snapshot method */ 3385785e854fSJed Brown ierr = PetscMalloc1(max_constraints*max_constraints,&correlation_mat);CHKERRQ(ierr); 3386785e854fSJed Brown ierr = PetscMalloc1(max_constraints,&singular_vals);CHKERRQ(ierr); 3387785e854fSJed Brown ierr = PetscMalloc1(max_size_of_constraint*max_constraints,&temp_basis);CHKERRQ(ierr); 3388674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX) 3389785e854fSJed Brown ierr = PetscMalloc1(3*max_constraints,&rwork);CHKERRQ(ierr); 3390674ae819SStefano Zampini #endif 3391674ae819SStefano Zampini /* now we evaluate the optimal workspace using query with lwork=-1 */ 3392c8244a33SStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_N);CHKERRQ(ierr); 3393c8244a33SStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_LDA);CHKERRQ(ierr); 3394674ae819SStefano Zampini lwork = -1; 3395674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 3396674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 3397c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,&temp_work,&lwork,&lierr)); 3398674ae819SStefano Zampini #else 3399c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,&temp_work,&lwork,rwork,&lierr)); 3400674ae819SStefano Zampini #endif 3401674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 3402984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to SYEV Lapack routine %d",(int)lierr); 3403674ae819SStefano Zampini #else /* on missing GESVD */ 3404674ae819SStefano Zampini /* SVD */ 3405674ae819SStefano Zampini PetscInt max_n,min_n; 3406674ae819SStefano Zampini max_n = max_size_of_constraint; 3407984c4197SStefano Zampini min_n = max_constraints; 3408984c4197SStefano Zampini if (max_size_of_constraint < max_constraints) { 3409674ae819SStefano Zampini min_n = max_size_of_constraint; 3410984c4197SStefano Zampini max_n = max_constraints; 3411674ae819SStefano Zampini } 3412785e854fSJed Brown ierr = PetscMalloc1(min_n,&singular_vals);CHKERRQ(ierr); 3413674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX) 3414785e854fSJed Brown ierr = PetscMalloc1(5*min_n,&rwork);CHKERRQ(ierr); 3415674ae819SStefano Zampini #endif 3416674ae819SStefano Zampini /* now we evaluate the optimal workspace using query with lwork=-1 */ 3417674ae819SStefano Zampini lwork = -1; 3418e310c8b4SStefano Zampini ierr = PetscBLASIntCast(max_n,&Blas_M);CHKERRQ(ierr); 3419e310c8b4SStefano Zampini ierr = PetscBLASIntCast(min_n,&Blas_N);CHKERRQ(ierr); 3420b7d8b9f8SStefano Zampini ierr = PetscBLASIntCast(max_n,&Blas_LDA);CHKERRQ(ierr); 3421674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 3422674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 34239162d606SStefano 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)); 3424674ae819SStefano Zampini #else 34259162d606SStefano 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)); 3426674ae819SStefano Zampini #endif 3427674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 3428984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to GESVD Lapack routine %d",(int)lierr); 3429984c4197SStefano Zampini #endif /* on missing GESVD */ 3430674ae819SStefano Zampini /* Allocate optimal workspace */ 3431674ae819SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(temp_work),&lwork);CHKERRQ(ierr); 3432854ce69bSBarry Smith ierr = PetscMalloc1(lwork,&work);CHKERRQ(ierr); 3433674ae819SStefano Zampini } 3434674ae819SStefano Zampini /* Now we can loop on constraining sets */ 3435674ae819SStefano Zampini total_counts = 0; 34369162d606SStefano Zampini constraints_idxs_ptr[0] = 0; 34379162d606SStefano Zampini constraints_data_ptr[0] = 0; 3438674ae819SStefano Zampini /* vertices */ 34399162d606SStefano Zampini if (n_vertices) { 3440674ae819SStefano Zampini ierr = ISGetIndices(ISForVertices,(const PetscInt**)&is_indices);CHKERRQ(ierr); 34419162d606SStefano Zampini ierr = PetscMemcpy(constraints_idxs,is_indices,n_vertices*sizeof(PetscInt));CHKERRQ(ierr); 3442674ae819SStefano Zampini for (i=0;i<n_vertices;i++) { 34439162d606SStefano Zampini constraints_n[total_counts] = 1; 34449162d606SStefano Zampini constraints_data[total_counts] = 1.0; 34459162d606SStefano Zampini constraints_idxs_ptr[total_counts+1] = constraints_idxs_ptr[total_counts]+1; 34469162d606SStefano Zampini constraints_data_ptr[total_counts+1] = constraints_data_ptr[total_counts]+1; 3447674ae819SStefano Zampini total_counts++; 3448674ae819SStefano Zampini } 3449674ae819SStefano Zampini ierr = ISRestoreIndices(ISForVertices,(const PetscInt**)&is_indices);CHKERRQ(ierr); 3450674ae819SStefano Zampini n_vertices = total_counts; 3451674ae819SStefano Zampini } 3452984c4197SStefano Zampini 3453674ae819SStefano Zampini /* edges and faces */ 34549162d606SStefano Zampini total_counts_cc = total_counts; 3455911cabfeSStefano Zampini for (ncc=0;ncc<n_ISForEdges+n_ISForFaces;ncc++) { 34569162d606SStefano Zampini IS used_is; 34579162d606SStefano Zampini PetscBool idxs_copied = PETSC_FALSE; 34589162d606SStefano Zampini 3459911cabfeSStefano Zampini if (ncc<n_ISForEdges) { 34609162d606SStefano Zampini used_is = ISForEdges[ncc]; 3461984c4197SStefano Zampini boolforchange = pcbddc->use_change_of_basis; /* change or not the basis on the edge */ 3462674ae819SStefano Zampini } else { 34639162d606SStefano Zampini used_is = ISForFaces[ncc-n_ISForEdges]; 3464984c4197SStefano Zampini boolforchange = (PetscBool)(pcbddc->use_change_of_basis && pcbddc->use_change_on_faces); /* change or not the basis on the face */ 3465674ae819SStefano Zampini } 3466674ae819SStefano Zampini temp_constraints = 0; /* zero the number of constraints I have on this conn comp */ 34679162d606SStefano Zampini 34689162d606SStefano Zampini ierr = ISGetSize(used_is,&size_of_constraint);CHKERRQ(ierr); 34699162d606SStefano Zampini ierr = ISGetIndices(used_is,(const PetscInt**)&is_indices);CHKERRQ(ierr); 3470984c4197SStefano Zampini /* change of basis should not be performed on local periodic nodes */ 3471984c4197SStefano Zampini if (pcbddc->mat_graph->mirrors && pcbddc->mat_graph->mirrors[is_indices[0]]) boolforchange = PETSC_FALSE; 3472674ae819SStefano Zampini if (nnsp_has_cnst) { 34735b08dc53SStefano Zampini PetscScalar quad_value; 34749162d606SStefano Zampini 34759162d606SStefano Zampini ierr = PetscMemcpy(constraints_idxs + constraints_idxs_ptr[total_counts_cc],is_indices,size_of_constraint*sizeof(PetscInt));CHKERRQ(ierr); 34769162d606SStefano Zampini idxs_copied = PETSC_TRUE; 34779162d606SStefano Zampini 3478a773dcb8SStefano Zampini if (!pcbddc->use_nnsp_true) { 3479674ae819SStefano Zampini quad_value = (PetscScalar)(1.0/PetscSqrtReal((PetscReal)size_of_constraint)); 3480a773dcb8SStefano Zampini } else { 3481a773dcb8SStefano Zampini quad_value = 1.0; 3482a773dcb8SStefano Zampini } 3483674ae819SStefano Zampini for (j=0;j<size_of_constraint;j++) { 34849162d606SStefano Zampini constraints_data[constraints_data_ptr[total_counts_cc]+j] = quad_value; 3485674ae819SStefano Zampini } 34869162d606SStefano Zampini temp_constraints++; 3487674ae819SStefano Zampini total_counts++; 3488674ae819SStefano Zampini } 3489674ae819SStefano Zampini for (k=0;k<nnsp_size;k++) { 3490984c4197SStefano Zampini PetscReal real_value; 34919162d606SStefano Zampini PetscScalar *ptr_to_data; 34929162d606SStefano Zampini 3493984c4197SStefano Zampini ierr = VecGetArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 34949162d606SStefano Zampini ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]+temp_constraints*size_of_constraint]; 3495674ae819SStefano Zampini for (j=0;j<size_of_constraint;j++) { 34969162d606SStefano Zampini ptr_to_data[j] = array[is_indices[j]]; 3497674ae819SStefano Zampini } 3498984c4197SStefano Zampini ierr = VecRestoreArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 3499984c4197SStefano Zampini /* check if array is null on the connected component */ 3500e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 35019162d606SStefano Zampini PetscStackCallBLAS("BLASasum",real_value = BLASasum_(&Blas_N,ptr_to_data,&Blas_one)); 35025b08dc53SStefano Zampini if (real_value > 0.0) { /* keep indices and values */ 3503674ae819SStefano Zampini temp_constraints++; 3504674ae819SStefano Zampini total_counts++; 35059162d606SStefano Zampini if (!idxs_copied) { 35069162d606SStefano Zampini ierr = PetscMemcpy(constraints_idxs + constraints_idxs_ptr[total_counts_cc],is_indices,size_of_constraint*sizeof(PetscInt));CHKERRQ(ierr); 35079162d606SStefano Zampini idxs_copied = PETSC_TRUE; 3508674ae819SStefano Zampini } 3509674ae819SStefano Zampini } 35109162d606SStefano Zampini } 35119162d606SStefano Zampini ierr = ISRestoreIndices(used_is,(const PetscInt**)&is_indices);CHKERRQ(ierr); 351245a1bb75SStefano Zampini valid_constraints = temp_constraints; 3513eb97c9d2SStefano Zampini if (!pcbddc->use_nnsp_true && temp_constraints) { 3514a773dcb8SStefano Zampini if (temp_constraints == 1) { /* just normalize the constraint */ 35159162d606SStefano Zampini PetscScalar norm,*ptr_to_data; 35169162d606SStefano Zampini 35179162d606SStefano Zampini ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]]; 3518a773dcb8SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 35199162d606SStefano Zampini PetscStackCallBLAS("BLASdot",norm = BLASdot_(&Blas_N,ptr_to_data,&Blas_one,ptr_to_data,&Blas_one)); 3520a773dcb8SStefano Zampini norm = 1.0/PetscSqrtReal(PetscRealPart(norm)); 35219162d606SStefano Zampini PetscStackCallBLAS("BLASscal",BLASscal_(&Blas_N,&norm,ptr_to_data,&Blas_one)); 3522a773dcb8SStefano Zampini } else { /* perform SVD */ 3523984c4197SStefano Zampini PetscReal tol = 1.0e-8; /* tolerance for retaining eigenmodes */ 35249162d606SStefano Zampini PetscScalar *ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]]; 3525674ae819SStefano Zampini 3526674ae819SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 3527984c4197SStefano Zampini /* SVD: Y = U*S*V^H -> U (eigenvectors of Y*Y^H) = Y*V*(S)^\dag 3528984c4197SStefano Zampini POD: Y^H*Y = V*D*V^H, D = S^H*S -> U = Y*V*D^(-1/2) 3529984c4197SStefano Zampini -> When PETSC_USE_COMPLEX and PETSC_MISSING_LAPACK_GESVD are defined 3530984c4197SStefano Zampini the constraints basis will differ (by a complex factor with absolute value equal to 1) 3531984c4197SStefano Zampini from that computed using LAPACKgesvd 3532984c4197SStefano Zampini -> This is due to a different computation of eigenvectors in LAPACKheev 3533984c4197SStefano Zampini -> The quality of the POD-computed basis will be the same */ 3534984c4197SStefano Zampini ierr = PetscMemzero(correlation_mat,temp_constraints*temp_constraints*sizeof(PetscScalar));CHKERRQ(ierr); 3535674ae819SStefano Zampini /* Store upper triangular part of correlation matrix */ 3536e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 3537984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 3538674ae819SStefano Zampini for (j=0;j<temp_constraints;j++) { 3539674ae819SStefano Zampini for (k=0;k<j+1;k++) { 35409162d606SStefano 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)); 3541674ae819SStefano Zampini } 3542674ae819SStefano Zampini } 3543e310c8b4SStefano Zampini /* compute eigenvalues and eigenvectors of correlation matrix */ 3544e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 3545e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_LDA);CHKERRQ(ierr); 3546674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 3547c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,work,&lwork,&lierr)); 3548674ae819SStefano Zampini #else 3549c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,work,&lwork,rwork,&lierr)); 3550674ae819SStefano Zampini #endif 3551674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 3552984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYEV Lapack routine %d",(int)lierr); 3553984c4197SStefano Zampini /* retain eigenvalues greater than tol: note that LAPACKsyev gives eigs in ascending order */ 3554674ae819SStefano Zampini j = 0; 3555984c4197SStefano Zampini while (j < temp_constraints && singular_vals[j] < tol) j++; 3556674ae819SStefano Zampini total_counts = total_counts-j; 355745a1bb75SStefano Zampini valid_constraints = temp_constraints-j; 3558e310c8b4SStefano Zampini /* scale and copy POD basis into used quadrature memory */ 3559c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 3560c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 3561c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_K);CHKERRQ(ierr); 3562c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 3563c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_LDB);CHKERRQ(ierr); 3564c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDC);CHKERRQ(ierr); 3565674ae819SStefano Zampini if (j<temp_constraints) { 3566984c4197SStefano Zampini PetscInt ii; 3567984c4197SStefano Zampini for (k=j;k<temp_constraints;k++) singular_vals[k] = 1.0/PetscSqrtReal(singular_vals[k]); 3568674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 35699162d606SStefano 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)); 3570674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 3571984c4197SStefano Zampini for (k=0;k<temp_constraints-j;k++) { 3572674ae819SStefano Zampini for (ii=0;ii<size_of_constraint;ii++) { 35739162d606SStefano 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]; 3574674ae819SStefano Zampini } 3575674ae819SStefano Zampini } 3576674ae819SStefano Zampini } 3577674ae819SStefano Zampini #else /* on missing GESVD */ 3578e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 3579e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 3580b7d8b9f8SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 3581674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 3582674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 35839162d606SStefano 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)); 3584674ae819SStefano Zampini #else 35859162d606SStefano 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)); 3586674ae819SStefano Zampini #endif 3587984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GESVD Lapack routine %d",(int)lierr); 3588674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 3589984c4197SStefano Zampini /* retain eigenvalues greater than tol: note that LAPACKgesvd gives eigs in descending order */ 3590e310c8b4SStefano Zampini k = temp_constraints; 3591e310c8b4SStefano Zampini if (k > size_of_constraint) k = size_of_constraint; 3592674ae819SStefano Zampini j = 0; 3593e310c8b4SStefano Zampini while (j < k && singular_vals[k-j-1] < tol) j++; 359445a1bb75SStefano Zampini valid_constraints = k-j; 3595911cabfeSStefano Zampini total_counts = total_counts-temp_constraints+valid_constraints; 3596984c4197SStefano Zampini #endif /* on missing GESVD */ 3597674ae819SStefano Zampini } 3598a773dcb8SStefano Zampini } 35999162d606SStefano Zampini /* update pointers information */ 36009162d606SStefano Zampini if (valid_constraints) { 36019162d606SStefano Zampini constraints_n[total_counts_cc] = valid_constraints; 36029162d606SStefano Zampini constraints_idxs_ptr[total_counts_cc+1] = constraints_idxs_ptr[total_counts_cc]+size_of_constraint; 36039162d606SStefano Zampini constraints_data_ptr[total_counts_cc+1] = constraints_data_ptr[total_counts_cc]+size_of_constraint*valid_constraints; 36049162d606SStefano Zampini /* set change_of_basis flag */ 360545a1bb75SStefano Zampini if (boolforchange) { 3606b3d85658SStefano Zampini PetscBTSet(change_basis,total_counts_cc); 36079162d606SStefano Zampini } 3608b3d85658SStefano Zampini total_counts_cc++; 360945a1bb75SStefano Zampini } 361045a1bb75SStefano Zampini } 3611984c4197SStefano Zampini /* free workspace */ 36128f1c130eSStefano Zampini if (!skip_lapack) { 3613984c4197SStefano Zampini ierr = PetscFree(work);CHKERRQ(ierr); 3614984c4197SStefano Zampini #if defined(PETSC_USE_COMPLEX) 3615984c4197SStefano Zampini ierr = PetscFree(rwork);CHKERRQ(ierr); 3616984c4197SStefano Zampini #endif 3617984c4197SStefano Zampini ierr = PetscFree(singular_vals);CHKERRQ(ierr); 3618984c4197SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 3619984c4197SStefano Zampini ierr = PetscFree(correlation_mat);CHKERRQ(ierr); 3620984c4197SStefano Zampini ierr = PetscFree(temp_basis);CHKERRQ(ierr); 3621984c4197SStefano Zampini #endif 3622984c4197SStefano Zampini } 3623984c4197SStefano Zampini for (k=0;k<nnsp_size;k++) { 3624984c4197SStefano Zampini ierr = VecDestroy(&localnearnullsp[k]);CHKERRQ(ierr); 3625984c4197SStefano Zampini } 3626984c4197SStefano Zampini ierr = PetscFree(localnearnullsp);CHKERRQ(ierr); 3627cf5a6209SStefano Zampini /* free index sets of faces, edges and vertices */ 3628cf5a6209SStefano Zampini for (i=0;i<n_ISForFaces;i++) { 3629cf5a6209SStefano Zampini ierr = ISDestroy(&ISForFaces[i]);CHKERRQ(ierr); 3630cf5a6209SStefano Zampini } 3631cf5a6209SStefano Zampini if (n_ISForFaces) { 3632cf5a6209SStefano Zampini ierr = PetscFree(ISForFaces);CHKERRQ(ierr); 3633cf5a6209SStefano Zampini } 3634cf5a6209SStefano Zampini for (i=0;i<n_ISForEdges;i++) { 3635cf5a6209SStefano Zampini ierr = ISDestroy(&ISForEdges[i]);CHKERRQ(ierr); 3636cf5a6209SStefano Zampini } 3637cf5a6209SStefano Zampini if (n_ISForEdges) { 3638cf5a6209SStefano Zampini ierr = PetscFree(ISForEdges);CHKERRQ(ierr); 3639cf5a6209SStefano Zampini } 3640cf5a6209SStefano Zampini ierr = ISDestroy(&ISForVertices);CHKERRQ(ierr); 364108122e43SStefano Zampini } else { 364208122e43SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 3643984c4197SStefano Zampini 364408122e43SStefano Zampini total_counts = 0; 364508122e43SStefano Zampini n_vertices = 0; 3646d62866d3SStefano Zampini if (sub_schurs->is_vertices && pcbddc->use_vertices) { 3647d62866d3SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_vertices,&n_vertices);CHKERRQ(ierr); 364808122e43SStefano Zampini } 364908122e43SStefano Zampini max_constraints = 0; 36509162d606SStefano Zampini total_counts_cc = 0; 365108122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs+n_vertices;i++) { 365208122e43SStefano Zampini total_counts += pcbddc->adaptive_constraints_n[i]; 36539162d606SStefano Zampini if (pcbddc->adaptive_constraints_n[i]) total_counts_cc++; 365408122e43SStefano Zampini max_constraints = PetscMax(max_constraints,pcbddc->adaptive_constraints_n[i]); 365508122e43SStefano Zampini } 36569162d606SStefano Zampini constraints_idxs_ptr = pcbddc->adaptive_constraints_idxs_ptr; 36579162d606SStefano Zampini constraints_data_ptr = pcbddc->adaptive_constraints_data_ptr; 36589162d606SStefano Zampini constraints_idxs = pcbddc->adaptive_constraints_idxs; 36599162d606SStefano Zampini constraints_data = pcbddc->adaptive_constraints_data; 366074d5cdf7SStefano Zampini /* constraints_n differs from pcbddc->adaptive_constraints_n */ 36619162d606SStefano Zampini ierr = PetscMalloc1(total_counts_cc,&constraints_n);CHKERRQ(ierr); 36629162d606SStefano Zampini total_counts_cc = 0; 36639162d606SStefano Zampini for (i=0;i<sub_schurs->n_subs+n_vertices;i++) { 36649162d606SStefano Zampini if (pcbddc->adaptive_constraints_n[i]) { 36659162d606SStefano Zampini constraints_n[total_counts_cc++] = pcbddc->adaptive_constraints_n[i]; 366608122e43SStefano Zampini } 366708122e43SStefano Zampini } 36689162d606SStefano Zampini #if 0 36699162d606SStefano Zampini printf("Found %d totals (%d)\n",total_counts_cc,total_counts); 36709162d606SStefano Zampini for (i=0;i<total_counts_cc;i++) { 36719162d606SStefano Zampini printf("const %d, start %d",i,constraints_idxs_ptr[i]); 36729162d606SStefano Zampini printf(" end %d:\n",constraints_idxs_ptr[i+1]); 36739162d606SStefano Zampini for (j=constraints_idxs_ptr[i];j<constraints_idxs_ptr[i+1];j++) { 36749162d606SStefano Zampini printf(" %d",constraints_idxs[j]); 36759162d606SStefano Zampini } 36769162d606SStefano Zampini printf("\n"); 36779162d606SStefano Zampini printf("number of cc: %d\n",constraints_n[i]); 36789162d606SStefano Zampini } 36791b968477SStefano Zampini for (i=0;i<n_vertices;i++) { 36808bec7fa6SStefano Zampini PetscPrintf(PETSC_COMM_SELF,"[%d] vertex %d, n %d\n",PetscGlobalRank,i,pcbddc->adaptive_constraints_n[i]); 36811b968477SStefano Zampini } 36821b968477SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 36838bec7fa6SStefano 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]); 36841b968477SStefano Zampini } 368508122e43SStefano Zampini #endif 368608122e43SStefano Zampini 36878bec7fa6SStefano Zampini max_size_of_constraint = 0; 36889162d606SStefano 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]); 36899162d606SStefano Zampini ierr = PetscMalloc1(constraints_idxs_ptr[total_counts_cc],&constraints_idxs_B);CHKERRQ(ierr); 369008122e43SStefano Zampini /* Change of basis */ 3691b3d85658SStefano Zampini ierr = PetscBTCreate(total_counts_cc,&change_basis);CHKERRQ(ierr); 369208122e43SStefano Zampini if (pcbddc->use_change_of_basis) { 369308122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 369408122e43SStefano Zampini if (PetscBTLookup(sub_schurs->is_edge,i) || pcbddc->use_change_on_faces) { 3695b3d85658SStefano Zampini ierr = PetscBTSet(change_basis,i+n_vertices);CHKERRQ(ierr); 369608122e43SStefano Zampini } 369708122e43SStefano Zampini } 369808122e43SStefano Zampini } 369908122e43SStefano Zampini } 3700984c4197SStefano Zampini pcbddc->local_primal_size = total_counts; 37014f1b2e48SStefano Zampini ierr = PetscMalloc1(pcbddc->local_primal_size+pcbddc->benign_n,&pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 370208122e43SStefano Zampini 37039162d606SStefano Zampini /* map constraints_idxs in boundary numbering */ 37049162d606SStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,constraints_idxs_ptr[total_counts_cc],constraints_idxs,&i,constraints_idxs_B);CHKERRQ(ierr); 37059162d606SStefano Zampini if (i != constraints_idxs_ptr[total_counts_cc]) { 37069162d606SStefano 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); 370708122e43SStefano Zampini } 3708674ae819SStefano Zampini 3709674ae819SStefano Zampini /* Create constraint matrix */ 3710674ae819SStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 371116f15bc4SStefano Zampini ierr = MatSetType(pcbddc->ConstraintMatrix,MATAIJ);CHKERRQ(ierr); 3712984c4197SStefano Zampini ierr = MatSetSizes(pcbddc->ConstraintMatrix,pcbddc->local_primal_size,pcis->n,pcbddc->local_primal_size,pcis->n);CHKERRQ(ierr); 3713984c4197SStefano Zampini 3714984c4197SStefano Zampini /* find primal_dofs: subdomain corners plus dofs selected as primal after change of basis */ 3715a717540cSStefano Zampini /* determine if a QR strategy is needed for change of basis */ 3716a717540cSStefano Zampini qr_needed = PETSC_FALSE; 371774d5cdf7SStefano Zampini ierr = PetscBTCreate(total_counts_cc,&qr_needed_idx);CHKERRQ(ierr); 3718984c4197SStefano Zampini total_primal_vertices=0; 3719b3d85658SStefano Zampini pcbddc->local_primal_size_cc = 0; 37209162d606SStefano Zampini for (i=0;i<total_counts_cc;i++) { 37219162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 3722984c4197SStefano Zampini if (size_of_constraint == 1) { 37239162d606SStefano Zampini pcbddc->primal_indices_local_idxs[total_primal_vertices++] = constraints_idxs[constraints_idxs_ptr[i]]; 3724b3d85658SStefano Zampini pcbddc->local_primal_size_cc += 1; 372564efe560SStefano Zampini } else if (PetscBTLookup(change_basis,i)) { 37269162d606SStefano Zampini for (k=0;k<constraints_n[i];k++) { 37279162d606SStefano Zampini pcbddc->primal_indices_local_idxs[total_primal_vertices++] = constraints_idxs[constraints_idxs_ptr[i]+k]; 3728a717540cSStefano Zampini } 3729b3d85658SStefano Zampini pcbddc->local_primal_size_cc += constraints_n[i]; 373074d5cdf7SStefano Zampini if (constraints_n[i] > 1 || pcbddc->use_qr_single || pcbddc->faster_deluxe) { 3731a717540cSStefano Zampini PetscBTSet(qr_needed_idx,i); 3732a717540cSStefano Zampini qr_needed = PETSC_TRUE; 3733a717540cSStefano Zampini } 3734fa434743SStefano Zampini } else { 3735b3d85658SStefano Zampini pcbddc->local_primal_size_cc += 1; 3736fa434743SStefano Zampini } 3737a717540cSStefano Zampini } 3738b371cd4fSStefano Zampini /* note that the local variable n_vertices used below stores the number of pointwise constraints */ 3739b371cd4fSStefano Zampini pcbddc->n_vertices = total_primal_vertices; 3740674ae819SStefano Zampini /* permute indices in order to have a sorted set of vertices */ 374170022509SStefano Zampini ierr = PetscSortInt(total_primal_vertices,pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 3742b3d85658SStefano Zampini 37434f1b2e48SStefano 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); 37440e6343abSStefano Zampini ierr = PetscMemcpy(pcbddc->local_primal_ref_node,pcbddc->primal_indices_local_idxs,total_primal_vertices*sizeof(PetscInt));CHKERRQ(ierr); 37450e6343abSStefano Zampini for (i=0;i<total_primal_vertices;i++) pcbddc->local_primal_ref_mult[i] = 1; 3746984c4197SStefano Zampini 3747984c4197SStefano Zampini /* nonzero structure of constraint matrix */ 374874d5cdf7SStefano Zampini /* and get reference dof for local constraints */ 3749785e854fSJed Brown ierr = PetscMalloc1(pcbddc->local_primal_size,&nnz);CHKERRQ(ierr); 3750984c4197SStefano Zampini for (i=0;i<total_primal_vertices;i++) nnz[i] = 1; 375174d5cdf7SStefano Zampini 3752984c4197SStefano Zampini j = total_primal_vertices; 375374d5cdf7SStefano Zampini total_counts = total_primal_vertices; 3754b3d85658SStefano Zampini cum = total_primal_vertices; 37559162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 37564641a718SStefano Zampini if (!PetscBTLookup(change_basis,i)) { 3757b3d85658SStefano Zampini pcbddc->local_primal_ref_node[cum] = constraints_idxs[constraints_idxs_ptr[i]]; 3758b3d85658SStefano Zampini pcbddc->local_primal_ref_mult[cum] = constraints_n[i]; 3759b3d85658SStefano Zampini cum++; 37609162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 376174d5cdf7SStefano Zampini for (k=0;k<constraints_n[i];k++) { 376274d5cdf7SStefano Zampini pcbddc->primal_indices_local_idxs[total_counts++] = constraints_idxs[constraints_idxs_ptr[i]+k]; 376374d5cdf7SStefano Zampini nnz[j+k] = size_of_constraint; 376474d5cdf7SStefano Zampini } 37659162d606SStefano Zampini j += constraints_n[i]; 3766674ae819SStefano Zampini } 3767674ae819SStefano Zampini } 3768674ae819SStefano Zampini ierr = MatSeqAIJSetPreallocation(pcbddc->ConstraintMatrix,0,nnz);CHKERRQ(ierr); 3769674ae819SStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 3770088faed8SStefano Zampini 3771674ae819SStefano Zampini /* set values in constraint matrix */ 3772984c4197SStefano Zampini for (i=0;i<total_primal_vertices;i++) { 37730e6343abSStefano Zampini ierr = MatSetValue(pcbddc->ConstraintMatrix,i,pcbddc->local_primal_ref_node[i],1.0,INSERT_VALUES);CHKERRQ(ierr); 3774674ae819SStefano Zampini } 3775984c4197SStefano Zampini total_counts = total_primal_vertices; 37769162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 37774641a718SStefano Zampini if (!PetscBTLookup(change_basis,i)) { 37789162d606SStefano Zampini PetscInt *cols; 37799162d606SStefano Zampini 37809162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 37819162d606SStefano Zampini cols = constraints_idxs+constraints_idxs_ptr[i]; 37829162d606SStefano Zampini for (k=0;k<constraints_n[i];k++) { 37839162d606SStefano Zampini PetscInt row = total_counts+k; 37849162d606SStefano Zampini PetscScalar *vals; 37859162d606SStefano Zampini 37869162d606SStefano Zampini vals = constraints_data+constraints_data_ptr[i]+k*size_of_constraint; 37879162d606SStefano Zampini ierr = MatSetValues(pcbddc->ConstraintMatrix,1,&row,size_of_constraint,cols,vals,INSERT_VALUES);CHKERRQ(ierr); 37889162d606SStefano Zampini } 37899162d606SStefano Zampini total_counts += constraints_n[i]; 3790674ae819SStefano Zampini } 3791674ae819SStefano Zampini } 3792674ae819SStefano Zampini /* assembling */ 3793674ae819SStefano Zampini ierr = MatAssemblyBegin(pcbddc->ConstraintMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3794674ae819SStefano Zampini ierr = MatAssemblyEnd(pcbddc->ConstraintMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3795088faed8SStefano Zampini 3796984c4197SStefano Zampini /* 379745a1bb75SStefano Zampini ierr = PetscViewerSetFormat(PETSC_VIEWER_STDOUT_SELF,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 3798984c4197SStefano Zampini ierr = MatView(pcbddc->ConstraintMatrix,(PetscViewer)0);CHKERRQ(ierr); 3799984c4197SStefano Zampini */ 3800674ae819SStefano Zampini /* Create matrix for change of basis. We don't need it in case pcbddc->use_change_of_basis is FALSE */ 3801674ae819SStefano Zampini if (pcbddc->use_change_of_basis) { 3802026de310SStefano Zampini /* dual and primal dofs on a single cc */ 3803984c4197SStefano Zampini PetscInt dual_dofs,primal_dofs; 3804984c4197SStefano Zampini /* working stuff for GEQRF */ 380581d9aea3SBarry Smith PetscScalar *qr_basis,*qr_tau = NULL,*qr_work,lqr_work_t; 3806984c4197SStefano Zampini PetscBLASInt lqr_work; 3807984c4197SStefano Zampini /* working stuff for UNGQR */ 3808984c4197SStefano Zampini PetscScalar *gqr_work,lgqr_work_t; 3809984c4197SStefano Zampini PetscBLASInt lgqr_work; 3810984c4197SStefano Zampini /* working stuff for TRTRS */ 3811984c4197SStefano Zampini PetscScalar *trs_rhs; 38123f08241aSStefano Zampini PetscBLASInt Blas_NRHS; 3813984c4197SStefano Zampini /* pointers for values insertion into change of basis matrix */ 3814984c4197SStefano Zampini PetscInt *start_rows,*start_cols; 3815984c4197SStefano Zampini PetscScalar *start_vals; 3816984c4197SStefano Zampini /* working stuff for values insertion */ 38174641a718SStefano Zampini PetscBT is_primal; 381864efe560SStefano Zampini PetscInt *aux_primal_numbering_B; 3819906d46d4SStefano Zampini /* matrix sizes */ 3820906d46d4SStefano Zampini PetscInt global_size,local_size; 3821906d46d4SStefano Zampini /* temporary change of basis */ 3822906d46d4SStefano Zampini Mat localChangeOfBasisMatrix; 3823cf5a6209SStefano Zampini /* extra space for debugging */ 3824cf5a6209SStefano Zampini PetscScalar *dbg_work; 3825984c4197SStefano Zampini 3826906d46d4SStefano Zampini /* local temporary change of basis acts on local interfaces -> dimension is n_B x n_B */ 3827906d46d4SStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&localChangeOfBasisMatrix);CHKERRQ(ierr); 382816f15bc4SStefano Zampini ierr = MatSetType(localChangeOfBasisMatrix,MATAIJ);CHKERRQ(ierr); 3829bbb9e6c6SStefano Zampini ierr = MatSetSizes(localChangeOfBasisMatrix,pcis->n,pcis->n,pcis->n,pcis->n);CHKERRQ(ierr); 3830906d46d4SStefano Zampini /* nonzeros for local mat */ 3831bbb9e6c6SStefano Zampini ierr = PetscMalloc1(pcis->n,&nnz);CHKERRQ(ierr); 3832bbb9e6c6SStefano Zampini for (i=0;i<pcis->n;i++) nnz[i]=1; 38339162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 3834a717540cSStefano Zampini if (PetscBTLookup(change_basis,i)) { 38359162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 3836a717540cSStefano Zampini if (PetscBTLookup(qr_needed_idx,i)) { 38379162d606SStefano Zampini for (j=0;j<size_of_constraint;j++) nnz[constraints_idxs[constraints_idxs_ptr[i]+j]] = size_of_constraint; 3838a717540cSStefano Zampini } else { 38399162d606SStefano Zampini nnz[constraints_idxs[constraints_idxs_ptr[i]]] = size_of_constraint; 38409162d606SStefano Zampini for (j=1;j<size_of_constraint;j++) nnz[constraints_idxs[constraints_idxs_ptr[i]+j]] = 2; 3841a717540cSStefano Zampini } 3842a717540cSStefano Zampini } 3843a717540cSStefano Zampini } 3844906d46d4SStefano Zampini ierr = MatSeqAIJSetPreallocation(localChangeOfBasisMatrix,0,nnz);CHKERRQ(ierr); 3845bbb9e6c6SStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 3846a717540cSStefano Zampini /* Set initial identity in the matrix */ 3847bbb9e6c6SStefano Zampini for (i=0;i<pcis->n;i++) { 3848906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,i,i,1.0,INSERT_VALUES);CHKERRQ(ierr); 3849a717540cSStefano Zampini } 3850a717540cSStefano Zampini 3851a717540cSStefano Zampini if (pcbddc->dbg_flag) { 3852a717540cSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"--------------------------------------------------------------\n");CHKERRQ(ierr); 3853a717540cSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Checking change of basis computation for subdomain %04d\n",PetscGlobalRank);CHKERRQ(ierr); 3854a717540cSStefano Zampini } 3855a717540cSStefano Zampini 3856a717540cSStefano Zampini 3857a717540cSStefano Zampini /* Now we loop on the constraints which need a change of basis */ 3858a717540cSStefano Zampini /* 3859a717540cSStefano Zampini Change of basis matrix is evaluated similarly to the FIRST APPROACH in 3860a717540cSStefano Zampini Klawonn and Widlund, Dual-primal FETI-DP methods for linear elasticity, (see Sect 6.2.1) 3861a717540cSStefano Zampini 3862a6b551f4SStefano Zampini Basic blocks of change of basis matrix T computed by 3863a717540cSStefano Zampini 3864a6b551f4SStefano 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) 3865a6b551f4SStefano Zampini 3866a6b551f4SStefano Zampini | 1 0 ... 0 s_1/S | 3867a6b551f4SStefano Zampini | 0 1 ... 0 s_2/S | 3868a717540cSStefano Zampini | ... | 3869a6b551f4SStefano Zampini | 0 ... 1 s_{n-1}/S | 3870a6b551f4SStefano Zampini | -s_1/s_n ... -s_{n-1}/s_n s_n/S | 3871a717540cSStefano Zampini 3872a6b551f4SStefano Zampini with S = \sum_{i=1}^n s_i^2 3873a6b551f4SStefano Zampini NOTE: in the above example, the primal dof is the last one of the edge in LOCAL ordering 3874a6b551f4SStefano Zampini in the current implementation, the primal dof is the first one of the edge in GLOBAL ordering 3875a6b551f4SStefano Zampini 3876a6b551f4SStefano Zampini - QR decomposition of constraints otherwise 3877a717540cSStefano Zampini */ 3878a717540cSStefano Zampini if (qr_needed) { 3879984c4197SStefano Zampini /* space to store Q */ 3880854ce69bSBarry Smith ierr = PetscMalloc1(max_size_of_constraint*max_size_of_constraint,&qr_basis);CHKERRQ(ierr); 3881984c4197SStefano Zampini /* first we issue queries for optimal work */ 38823f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_M);CHKERRQ(ierr); 38833f08241aSStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_N);CHKERRQ(ierr); 38843f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 3885984c4197SStefano Zampini lqr_work = -1; 38863f08241aSStefano Zampini PetscStackCallBLAS("LAPACKgeqrf",LAPACKgeqrf_(&Blas_M,&Blas_N,qr_basis,&Blas_LDA,qr_tau,&lqr_work_t,&lqr_work,&lierr)); 3887984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to GEQRF Lapack routine %d",(int)lierr); 3888984c4197SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lqr_work_t),&lqr_work);CHKERRQ(ierr); 3889785e854fSJed Brown ierr = PetscMalloc1((PetscInt)PetscRealPart(lqr_work_t),&qr_work);CHKERRQ(ierr); 3890984c4197SStefano Zampini lgqr_work = -1; 38913f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_M);CHKERRQ(ierr); 38923f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_N);CHKERRQ(ierr); 38933f08241aSStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_K);CHKERRQ(ierr); 38943f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 38953f08241aSStefano Zampini if (Blas_K>Blas_M) Blas_K=Blas_M; /* adjust just for computing optimal work */ 38963f08241aSStefano Zampini PetscStackCallBLAS("LAPACKungqr",LAPACKungqr_(&Blas_M,&Blas_N,&Blas_K,qr_basis,&Blas_LDA,qr_tau,&lgqr_work_t,&lgqr_work,&lierr)); 3897984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to UNGQR Lapack routine %d",(int)lierr); 3898984c4197SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lgqr_work_t),&lgqr_work);CHKERRQ(ierr); 3899785e854fSJed Brown ierr = PetscMalloc1((PetscInt)PetscRealPart(lgqr_work_t),&gqr_work);CHKERRQ(ierr); 3900984c4197SStefano Zampini /* array to store scaling factors for reflectors */ 3901785e854fSJed Brown ierr = PetscMalloc1(max_constraints,&qr_tau);CHKERRQ(ierr); 3902984c4197SStefano Zampini /* array to store rhs and solution of triangular solver */ 3903785e854fSJed Brown ierr = PetscMalloc1(max_constraints*max_constraints,&trs_rhs);CHKERRQ(ierr); 3904a717540cSStefano Zampini /* allocating workspace for check */ 3905a717540cSStefano Zampini if (pcbddc->dbg_flag) { 3906cf5a6209SStefano Zampini ierr = PetscMalloc1(max_size_of_constraint*(max_constraints+max_size_of_constraint),&dbg_work);CHKERRQ(ierr); 3907a717540cSStefano Zampini } 3908a717540cSStefano Zampini } 3909984c4197SStefano Zampini /* array to store whether a node is primal or not */ 39104641a718SStefano Zampini ierr = PetscBTCreate(pcis->n_B,&is_primal);CHKERRQ(ierr); 3911473ba861SJed Brown ierr = PetscMalloc1(total_primal_vertices,&aux_primal_numbering_B);CHKERRQ(ierr); 39120e6343abSStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,total_primal_vertices,pcbddc->local_primal_ref_node,&i,aux_primal_numbering_B);CHKERRQ(ierr); 391339e2fb2aSStefano Zampini if (i != total_primal_vertices) { 391439e2fb2aSStefano Zampini SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_SUP,"Error in boundary numbering for BDDC vertices! %d != %d\n",total_primal_vertices,i); 39154641a718SStefano Zampini } 391639e2fb2aSStefano Zampini for (i=0;i<total_primal_vertices;i++) { 391739e2fb2aSStefano Zampini ierr = PetscBTSet(is_primal,aux_primal_numbering_B[i]);CHKERRQ(ierr); 391839e2fb2aSStefano Zampini } 391939e2fb2aSStefano Zampini ierr = PetscFree(aux_primal_numbering_B);CHKERRQ(ierr); 3920984c4197SStefano Zampini 3921a717540cSStefano Zampini /* loop on constraints and see whether or not they need a change of basis and compute it */ 39229162d606SStefano Zampini for (total_counts=n_vertices;total_counts<total_counts_cc;total_counts++) { 39239162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[total_counts+1]-constraints_idxs_ptr[total_counts]; 39244641a718SStefano Zampini if (PetscBTLookup(change_basis,total_counts)) { 3925984c4197SStefano Zampini /* get constraint info */ 39269162d606SStefano Zampini primal_dofs = constraints_n[total_counts]; 3927984c4197SStefano Zampini dual_dofs = size_of_constraint-primal_dofs; 3928984c4197SStefano Zampini 3929984c4197SStefano Zampini if (pcbddc->dbg_flag) { 39309162d606SStefano 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); 3931674ae819SStefano Zampini } 3932984c4197SStefano Zampini 3933fa434743SStefano Zampini if (PetscBTLookup(qr_needed_idx,total_counts)) { /* QR */ 3934a717540cSStefano Zampini 3935a717540cSStefano Zampini /* copy quadrature constraints for change of basis check */ 3936a717540cSStefano Zampini if (pcbddc->dbg_flag) { 39379162d606SStefano Zampini ierr = PetscMemcpy(dbg_work,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 3938a717540cSStefano Zampini } 3939984c4197SStefano Zampini /* copy temporary constraints into larger work vector (in order to store all columns of Q) */ 39409162d606SStefano Zampini ierr = PetscMemcpy(qr_basis,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 3941984c4197SStefano Zampini 3942984c4197SStefano Zampini /* compute QR decomposition of constraints */ 39433f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 39443f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 39453f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 3946674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 39473f08241aSStefano Zampini PetscStackCallBLAS("LAPACKgeqrf",LAPACKgeqrf_(&Blas_M,&Blas_N,qr_basis,&Blas_LDA,qr_tau,qr_work,&lqr_work,&lierr)); 3948984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GEQRF Lapack routine %d",(int)lierr); 3949674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 3950984c4197SStefano Zampini 3951984c4197SStefano Zampini /* explictly compute R^-T */ 3952984c4197SStefano Zampini ierr = PetscMemzero(trs_rhs,primal_dofs*primal_dofs*sizeof(*trs_rhs));CHKERRQ(ierr); 3953984c4197SStefano Zampini for (j=0;j<primal_dofs;j++) trs_rhs[j*(primal_dofs+1)] = 1.0; 39543f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 39553f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_NRHS);CHKERRQ(ierr); 39563f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 39573f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDB);CHKERRQ(ierr); 3958984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 39593f08241aSStefano Zampini PetscStackCallBLAS("LAPACKtrtrs",LAPACKtrtrs_("U","T","N",&Blas_N,&Blas_NRHS,qr_basis,&Blas_LDA,trs_rhs,&Blas_LDB,&lierr)); 3960984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in TRTRS Lapack routine %d",(int)lierr); 3961984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 3962984c4197SStefano Zampini 3963a717540cSStefano Zampini /* explicitly compute all columns of Q (Q = [Q1 | Q2] ) overwriting QR factorization in qr_basis */ 39643f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 39653f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 39663f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_K);CHKERRQ(ierr); 39673f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 3968984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 39693f08241aSStefano Zampini PetscStackCallBLAS("LAPACKungqr",LAPACKungqr_(&Blas_M,&Blas_N,&Blas_K,qr_basis,&Blas_LDA,qr_tau,gqr_work,&lgqr_work,&lierr)); 3970984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in UNGQR Lapack routine %d",(int)lierr); 3971984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 3972984c4197SStefano Zampini 3973984c4197SStefano Zampini /* first primal_dofs columns of Q need to be re-scaled in order to be unitary w.r.t constraints 3974984c4197SStefano Zampini i.e. C_{pxn}*Q_{nxn} should be equal to [I_pxp | 0_pxd] (see check below) 3975984c4197SStefano Zampini where n=size_of_constraint, p=primal_dofs, d=dual_dofs (n=p+d), I and 0 identity and null matrix resp. */ 39763f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 39773f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 39783f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_K);CHKERRQ(ierr); 39793f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 39803f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDB);CHKERRQ(ierr); 39813f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDC);CHKERRQ(ierr); 3982984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 39839162d606SStefano 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)); 3984984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 39859162d606SStefano Zampini ierr = PetscMemcpy(qr_basis,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 3986984c4197SStefano Zampini 3987984c4197SStefano Zampini /* insert values in change of basis matrix respecting global ordering of new primal dofs */ 39889162d606SStefano Zampini start_rows = &constraints_idxs[constraints_idxs_ptr[total_counts]]; 3989984c4197SStefano Zampini /* insert cols for primal dofs */ 3990984c4197SStefano Zampini for (j=0;j<primal_dofs;j++) { 3991984c4197SStefano Zampini start_vals = &qr_basis[j*size_of_constraint]; 39929162d606SStefano Zampini start_cols = &constraints_idxs[constraints_idxs_ptr[total_counts]+j]; 3993906d46d4SStefano Zampini ierr = MatSetValues(localChangeOfBasisMatrix,size_of_constraint,start_rows,1,start_cols,start_vals,INSERT_VALUES);CHKERRQ(ierr); 3994984c4197SStefano Zampini } 3995984c4197SStefano Zampini /* insert cols for dual dofs */ 3996984c4197SStefano Zampini for (j=0,k=0;j<dual_dofs;k++) { 39979162d606SStefano Zampini if (!PetscBTLookup(is_primal,constraints_idxs_B[constraints_idxs_ptr[total_counts]+k])) { 3998984c4197SStefano Zampini start_vals = &qr_basis[(primal_dofs+j)*size_of_constraint]; 39999162d606SStefano Zampini start_cols = &constraints_idxs[constraints_idxs_ptr[total_counts]+k]; 4000906d46d4SStefano Zampini ierr = MatSetValues(localChangeOfBasisMatrix,size_of_constraint,start_rows,1,start_cols,start_vals,INSERT_VALUES);CHKERRQ(ierr); 4001984c4197SStefano Zampini j++; 4002674ae819SStefano Zampini } 4003674ae819SStefano Zampini } 4004984c4197SStefano Zampini 4005984c4197SStefano Zampini /* check change of basis */ 4006984c4197SStefano Zampini if (pcbddc->dbg_flag) { 4007984c4197SStefano Zampini PetscInt ii,jj; 4008984c4197SStefano Zampini PetscBool valid_qr=PETSC_TRUE; 4009c4303822SStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_M);CHKERRQ(ierr); 4010c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 4011c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_K);CHKERRQ(ierr); 4012c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 4013c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDB);CHKERRQ(ierr); 4014c4303822SStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDC);CHKERRQ(ierr); 4015984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 4016cf5a6209SStefano 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)); 4017984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 4018984c4197SStefano Zampini for (jj=0;jj<size_of_constraint;jj++) { 4019984c4197SStefano Zampini for (ii=0;ii<primal_dofs;ii++) { 4020cf5a6209SStefano Zampini if (ii != jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]) > 1.e-12) valid_qr = PETSC_FALSE; 4021cf5a6209SStefano 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; 4022674ae819SStefano Zampini } 4023674ae819SStefano Zampini } 4024984c4197SStefano Zampini if (!valid_qr) { 402522d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> wrong change of basis!\n");CHKERRQ(ierr); 4026984c4197SStefano Zampini for (jj=0;jj<size_of_constraint;jj++) { 4027984c4197SStefano Zampini for (ii=0;ii<primal_dofs;ii++) { 4028cf5a6209SStefano Zampini if (ii != jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]) > 1.e-12) { 4029cf5a6209SStefano 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])); 4030674ae819SStefano Zampini } 4031cf5a6209SStefano Zampini if (ii == jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]-1.0) > 1.e-12) { 4032cf5a6209SStefano 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])); 4033984c4197SStefano Zampini } 4034984c4197SStefano Zampini } 4035984c4197SStefano Zampini } 4036674ae819SStefano Zampini } else { 403722d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> right change of basis!\n");CHKERRQ(ierr); 4038674ae819SStefano Zampini } 4039674ae819SStefano Zampini } 4040a717540cSStefano Zampini } else { /* simple transformation block */ 4041a717540cSStefano Zampini PetscInt row,col; 4042a6b551f4SStefano Zampini PetscScalar val,norm; 4043a6b551f4SStefano Zampini 4044a6b551f4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 40459162d606SStefano 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)); 4046a717540cSStefano Zampini for (j=0;j<size_of_constraint;j++) { 40479162d606SStefano Zampini PetscInt row_B = constraints_idxs_B[constraints_idxs_ptr[total_counts]+j]; 40489162d606SStefano Zampini row = constraints_idxs[constraints_idxs_ptr[total_counts]+j]; 4049bbb9e6c6SStefano Zampini if (!PetscBTLookup(is_primal,row_B)) { 40509162d606SStefano Zampini col = constraints_idxs[constraints_idxs_ptr[total_counts]]; 4051906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,row,1.0,INSERT_VALUES);CHKERRQ(ierr); 40529162d606SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,col,constraints_data[constraints_data_ptr[total_counts]+j]/norm,INSERT_VALUES);CHKERRQ(ierr); 4053a717540cSStefano Zampini } else { 4054a717540cSStefano Zampini for (k=0;k<size_of_constraint;k++) { 40559162d606SStefano Zampini col = constraints_idxs[constraints_idxs_ptr[total_counts]+k]; 4056a717540cSStefano Zampini if (row != col) { 40579162d606SStefano Zampini val = -constraints_data[constraints_data_ptr[total_counts]+k]/constraints_data[constraints_data_ptr[total_counts]]; 4058a717540cSStefano Zampini } else { 40599162d606SStefano Zampini val = constraints_data[constraints_data_ptr[total_counts]]/norm; 4060a717540cSStefano Zampini } 4061906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,col,val,INSERT_VALUES);CHKERRQ(ierr); 4062a717540cSStefano Zampini } 4063a717540cSStefano Zampini } 4064a717540cSStefano Zampini } 406598a51de6SStefano Zampini if (pcbddc->dbg_flag) { 406622d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> using standard change of basis\n");CHKERRQ(ierr); 4067a717540cSStefano Zampini } 4068674ae819SStefano Zampini } 4069984c4197SStefano Zampini } else { 4070984c4197SStefano Zampini if (pcbddc->dbg_flag) { 40719162d606SStefano 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); 4072674ae819SStefano Zampini } 4073674ae819SStefano Zampini } 4074674ae819SStefano Zampini } 4075a717540cSStefano Zampini 4076a717540cSStefano Zampini /* free workspace */ 4077a717540cSStefano Zampini if (qr_needed) { 4078984c4197SStefano Zampini if (pcbddc->dbg_flag) { 4079cf5a6209SStefano Zampini ierr = PetscFree(dbg_work);CHKERRQ(ierr); 4080984c4197SStefano Zampini } 4081984c4197SStefano Zampini ierr = PetscFree(trs_rhs);CHKERRQ(ierr); 4082984c4197SStefano Zampini ierr = PetscFree(qr_tau);CHKERRQ(ierr); 4083984c4197SStefano Zampini ierr = PetscFree(qr_work);CHKERRQ(ierr); 4084984c4197SStefano Zampini ierr = PetscFree(gqr_work);CHKERRQ(ierr); 4085984c4197SStefano Zampini ierr = PetscFree(qr_basis);CHKERRQ(ierr); 4086674ae819SStefano Zampini } 4087a717540cSStefano Zampini ierr = PetscBTDestroy(&is_primal);CHKERRQ(ierr); 4088906d46d4SStefano Zampini ierr = MatAssemblyBegin(localChangeOfBasisMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4089906d46d4SStefano Zampini ierr = MatAssemblyEnd(localChangeOfBasisMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4090906d46d4SStefano Zampini 4091906d46d4SStefano Zampini /* assembling of global change of variable */ 409288c03ad3SStefano Zampini if (!pcbddc->fake_change) { 4093bbb9e6c6SStefano Zampini Mat tmat; 409416f15bc4SStefano Zampini PetscInt bs; 409516f15bc4SStefano Zampini 4096906d46d4SStefano Zampini ierr = VecGetSize(pcis->vec1_global,&global_size);CHKERRQ(ierr); 4097906d46d4SStefano Zampini ierr = VecGetLocalSize(pcis->vec1_global,&local_size);CHKERRQ(ierr); 4098bbb9e6c6SStefano Zampini ierr = MatDuplicate(pc->pmat,MAT_DO_NOT_COPY_VALUES,&tmat);CHKERRQ(ierr); 4099bbb9e6c6SStefano Zampini ierr = MatISSetLocalMat(tmat,localChangeOfBasisMatrix);CHKERRQ(ierr); 4100bbb9e6c6SStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)pc),&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 4101bbb9e6c6SStefano Zampini ierr = MatSetType(pcbddc->ChangeOfBasisMatrix,MATAIJ);CHKERRQ(ierr); 410216f15bc4SStefano Zampini ierr = MatGetBlockSize(pc->pmat,&bs);CHKERRQ(ierr); 410316f15bc4SStefano Zampini ierr = MatSetBlockSize(pcbddc->ChangeOfBasisMatrix,bs);CHKERRQ(ierr); 4104906d46d4SStefano Zampini ierr = MatSetSizes(pcbddc->ChangeOfBasisMatrix,local_size,local_size,global_size,global_size);CHKERRQ(ierr); 4105bbb9e6c6SStefano Zampini ierr = MatISSetMPIXAIJPreallocation_Private(tmat,pcbddc->ChangeOfBasisMatrix,PETSC_TRUE);CHKERRQ(ierr); 4106bbb9e6c6SStefano Zampini ierr = MatISGetMPIXAIJ(tmat,MAT_REUSE_MATRIX,&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 4107bbb9e6c6SStefano Zampini ierr = MatDestroy(&tmat);CHKERRQ(ierr); 4108bbb9e6c6SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 4109bbb9e6c6SStefano Zampini ierr = VecSet(pcis->vec1_N,1.0);CHKERRQ(ierr); 4110e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4111e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4112bbb9e6c6SStefano Zampini ierr = VecReciprocal(pcis->vec1_global);CHKERRQ(ierr); 4113bbb9e6c6SStefano Zampini ierr = MatDiagonalScale(pcbddc->ChangeOfBasisMatrix,pcis->vec1_global,NULL);CHKERRQ(ierr); 411488c03ad3SStefano Zampini 4115906d46d4SStefano Zampini /* check */ 4116906d46d4SStefano Zampini if (pcbddc->dbg_flag) { 4117906d46d4SStefano Zampini PetscReal error; 4118906d46d4SStefano Zampini Vec x,x_change; 4119906d46d4SStefano Zampini 4120906d46d4SStefano Zampini ierr = VecDuplicate(pcis->vec1_global,&x);CHKERRQ(ierr); 4121906d46d4SStefano Zampini ierr = VecDuplicate(pcis->vec1_global,&x_change);CHKERRQ(ierr); 4122906d46d4SStefano Zampini ierr = VecSetRandom(x,NULL);CHKERRQ(ierr); 4123906d46d4SStefano Zampini ierr = VecCopy(x,pcis->vec1_global);CHKERRQ(ierr); 4124e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,x,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4125e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,x,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4126bbb9e6c6SStefano Zampini ierr = MatMult(localChangeOfBasisMatrix,pcis->vec1_N,pcis->vec2_N);CHKERRQ(ierr); 4127e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec2_N,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4128e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec2_N,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4129906d46d4SStefano Zampini ierr = MatMult(pcbddc->ChangeOfBasisMatrix,pcis->vec1_global,x_change);CHKERRQ(ierr); 4130906d46d4SStefano Zampini ierr = VecAXPY(x,-1.0,x_change);CHKERRQ(ierr); 4131906d46d4SStefano Zampini ierr = VecNorm(x,NORM_INFINITY,&error);CHKERRQ(ierr); 4132906d46d4SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4133bbb9e6c6SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Error global vs local change: %1.6e\n",error);CHKERRQ(ierr); 4134906d46d4SStefano Zampini ierr = VecDestroy(&x);CHKERRQ(ierr); 4135906d46d4SStefano Zampini ierr = VecDestroy(&x_change);CHKERRQ(ierr); 4136906d46d4SStefano Zampini } 413788c03ad3SStefano Zampini } 4138b96c3477SStefano Zampini 4139b96c3477SStefano Zampini /* adapt sub_schurs computed (if any) */ 4140b96c3477SStefano Zampini if (pcbddc->use_deluxe_scaling) { 4141b96c3477SStefano Zampini PCBDDCSubSchurs sub_schurs=pcbddc->sub_schurs; 4142*bf3a8328SStefano Zampini 4143*bf3a8328SStefano Zampini if (pcbddc->fake_change && !pcbddc->deluxe_zerorows) { 4144*bf3a8328SStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"This should not happen");CHKERRQ(ierr); 4145*bf3a8328SStefano Zampini } 4146*bf3a8328SStefano Zampini if (pcbddc->use_change_of_basis && pcbddc->adaptive_userdefined) { 4147*bf3a8328SStefano Zampini SETERRQ(PetscObjectComm((PetscObject)pc),PETSC_ERR_SUP,"Cannot mix automatic change of basis, adaptive selection and user-defined constraints");CHKERRQ(ierr); 4148*bf3a8328SStefano Zampini } 4149b96c3477SStefano Zampini if (sub_schurs->S_Ej_all) { 4150ac632422SStefano Zampini Mat S_new,tmat; 4151*bf3a8328SStefano Zampini IS is_all_N,is_V_Sall = NULL; 4152bbb9e6c6SStefano Zampini 4153bbb9e6c6SStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(pcis->BtoNmap,sub_schurs->is_Ej_all,&is_all_N);CHKERRQ(ierr); 41546816873aSStefano Zampini ierr = MatGetSubMatrix(localChangeOfBasisMatrix,is_all_N,is_all_N,MAT_INITIAL_MATRIX,&tmat);CHKERRQ(ierr); 4155*bf3a8328SStefano Zampini if (pcbddc->deluxe_zerorows) { 4156*bf3a8328SStefano Zampini ISLocalToGlobalMapping NtoSall; 4157*bf3a8328SStefano Zampini IS is_V; 4158b087196eSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,pcbddc->n_vertices,pcbddc->local_primal_ref_node,PETSC_COPY_VALUES,&is_V);CHKERRQ(ierr); 4159b087196eSStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(is_all_N,&NtoSall);CHKERRQ(ierr); 4160b087196eSStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(NtoSall,IS_GTOLM_DROP,is_V,&is_V_Sall);CHKERRQ(ierr); 4161b087196eSStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&NtoSall);CHKERRQ(ierr); 4162b087196eSStefano Zampini ierr = ISDestroy(&is_V);CHKERRQ(ierr); 4163*bf3a8328SStefano Zampini } 4164*bf3a8328SStefano Zampini ierr = ISDestroy(&is_all_N);CHKERRQ(ierr); 4165ac632422SStefano Zampini ierr = MatPtAP(sub_schurs->S_Ej_all,tmat,MAT_INITIAL_MATRIX,1.0,&S_new);CHKERRQ(ierr); 4166b96c3477SStefano Zampini ierr = MatDestroy(&sub_schurs->S_Ej_all);CHKERRQ(ierr); 4167ac632422SStefano Zampini ierr = PetscObjectReference((PetscObject)S_new);CHKERRQ(ierr); 4168*bf3a8328SStefano Zampini if (pcbddc->deluxe_zerorows) { 4169*bf3a8328SStefano Zampini const PetscScalar *array; 4170*bf3a8328SStefano Zampini const PetscInt *idxs_V,*idxs_all; 4171*bf3a8328SStefano Zampini PetscInt i,n_V; 4172*bf3a8328SStefano Zampini 4173b087196eSStefano Zampini ierr = MatZeroRowsColumnsIS(S_new,is_V_Sall,1.,NULL,NULL);CHKERRQ(ierr); 4174b087196eSStefano Zampini ierr = ISGetLocalSize(is_V_Sall,&n_V);CHKERRQ(ierr); 4175b087196eSStefano Zampini ierr = ISGetIndices(is_V_Sall,&idxs_V);CHKERRQ(ierr); 4176b087196eSStefano Zampini ierr = ISGetIndices(sub_schurs->is_Ej_all,&idxs_all);CHKERRQ(ierr); 4177b087196eSStefano Zampini ierr = VecGetArrayRead(pcis->D,&array);CHKERRQ(ierr); 4178b087196eSStefano Zampini for (i=0;i<n_V;i++) { 4179b087196eSStefano Zampini PetscScalar val; 4180b087196eSStefano Zampini PetscInt idx; 4181b087196eSStefano Zampini 4182b087196eSStefano Zampini idx = idxs_V[i]; 4183b087196eSStefano Zampini val = array[idxs_all[idxs_V[i]]]; 4184b087196eSStefano Zampini ierr = MatSetValue(S_new,idx,idx,val,INSERT_VALUES);CHKERRQ(ierr); 4185b087196eSStefano Zampini } 4186b087196eSStefano Zampini ierr = MatAssemblyBegin(S_new,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4187b087196eSStefano Zampini ierr = MatAssemblyEnd(S_new,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4188*bf3a8328SStefano Zampini ierr = VecRestoreArrayRead(pcis->D,&array);CHKERRQ(ierr); 4189*bf3a8328SStefano Zampini ierr = ISRestoreIndices(sub_schurs->is_Ej_all,&idxs_all);CHKERRQ(ierr); 4190*bf3a8328SStefano Zampini ierr = ISRestoreIndices(is_V_Sall,&idxs_V);CHKERRQ(ierr); 4191*bf3a8328SStefano Zampini } 4192ac632422SStefano Zampini sub_schurs->S_Ej_all = S_new; 4193ac632422SStefano Zampini ierr = MatDestroy(&S_new);CHKERRQ(ierr); 4194ac632422SStefano Zampini if (sub_schurs->sum_S_Ej_all) { 4195ac632422SStefano Zampini ierr = MatPtAP(sub_schurs->sum_S_Ej_all,tmat,MAT_INITIAL_MATRIX,1.0,&S_new);CHKERRQ(ierr); 4196b96c3477SStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_all);CHKERRQ(ierr); 4197ac632422SStefano Zampini ierr = PetscObjectReference((PetscObject)S_new);CHKERRQ(ierr); 4198*bf3a8328SStefano Zampini if (pcbddc->deluxe_zerorows) { 4199b087196eSStefano Zampini ierr = MatZeroRowsColumnsIS(S_new,is_V_Sall,1.,NULL,NULL);CHKERRQ(ierr); 4200*bf3a8328SStefano Zampini } 4201ac632422SStefano Zampini sub_schurs->sum_S_Ej_all = S_new; 4202ac632422SStefano Zampini ierr = MatDestroy(&S_new);CHKERRQ(ierr); 4203ac632422SStefano Zampini } 4204*bf3a8328SStefano Zampini if (sub_schurs->sum_S_Ej_inv_all && pcbddc->fake_change && pcbddc->deluxe_zerorows) { 42056c3e6151SStefano Zampini PetscScalar infty = PETSC_MAX_REAL; 42066c3e6151SStefano Zampini ierr = MatPtAP(sub_schurs->sum_S_Ej_inv_all,tmat,MAT_INITIAL_MATRIX,1.0,&S_new);CHKERRQ(ierr); 42076c3e6151SStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_inv_all);CHKERRQ(ierr); 42086c3e6151SStefano Zampini ierr = PetscObjectReference((PetscObject)S_new);CHKERRQ(ierr); 42096c3e6151SStefano Zampini ierr = MatZeroRowsColumnsIS(S_new,is_V_Sall,1.,NULL,NULL);CHKERRQ(ierr); 42106c3e6151SStefano Zampini sub_schurs->sum_S_Ej_inv_all = S_new; 42116c3e6151SStefano Zampini ierr = MatDestroy(&S_new);CHKERRQ(ierr); 42126c3e6151SStefano Zampini ierr = MatPtAP(sub_schurs->sum_S_Ej_tilda_all,tmat,MAT_INITIAL_MATRIX,1.0,&S_new);CHKERRQ(ierr); 42136c3e6151SStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_tilda_all);CHKERRQ(ierr); 42146c3e6151SStefano Zampini ierr = PetscObjectReference((PetscObject)S_new);CHKERRQ(ierr); 42156c3e6151SStefano Zampini ierr = MatZeroRowsColumnsIS(S_new,is_V_Sall,infty,NULL,NULL);CHKERRQ(ierr); 42166c3e6151SStefano Zampini sub_schurs->sum_S_Ej_tilda_all = S_new; 42176c3e6151SStefano Zampini ierr = MatDestroy(&S_new);CHKERRQ(ierr); 42186c3e6151SStefano Zampini } 4219*bf3a8328SStefano Zampini 422088c03ad3SStefano Zampini if (!pcbddc->fake_change) { 4221b087196eSStefano Zampini ierr = ISDestroy(&is_V_Sall);CHKERRQ(ierr); 422288c03ad3SStefano Zampini } else { 422316e386b8SStefano Zampini if (!sub_schurs->change_primal) { 422488c03ad3SStefano Zampini sub_schurs->change_primal = is_V_Sall; 422516e386b8SStefano Zampini } else { 422616e386b8SStefano Zampini #if defined(PETSC_USE_DEBUG) 422716e386b8SStefano Zampini IS ist; 422816e386b8SStefano Zampini PetscInt n; 422916e386b8SStefano Zampini 423016e386b8SStefano Zampini ierr = ISDifference(sub_schurs->change_primal,is_V_Sall,&ist);CHKERRQ(ierr); 423116e386b8SStefano Zampini ierr = ISGetLocalSize(ist,&n);CHKERRQ(ierr); 423216e386b8SStefano Zampini if (n) { 423316e386b8SStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Not yet implemented"); 423416e386b8SStefano Zampini } 423516e386b8SStefano Zampini ierr = ISDestroy(&ist);CHKERRQ(ierr); 423616e386b8SStefano Zampini #endif 423716e386b8SStefano Zampini ierr = ISDestroy(&is_V_Sall);CHKERRQ(ierr); 423816e386b8SStefano Zampini } 423916e386b8SStefano Zampini if (!sub_schurs->change) { 424088c03ad3SStefano Zampini ierr = KSPCreate(PETSC_COMM_SELF,&sub_schurs->change);CHKERRQ(ierr); 424188c03ad3SStefano Zampini ierr = KSPSetOperators(sub_schurs->change,tmat,tmat);CHKERRQ(ierr); 424216e386b8SStefano Zampini #if defined(PETSC_USE_DEBUG) 424316e386b8SStefano Zampini } else { 424416e386b8SStefano Zampini Mat At; 424516e386b8SStefano Zampini PetscReal norm; 424616e386b8SStefano Zampini 424716e386b8SStefano Zampini ierr = KSPGetOperators(sub_schurs->change,&At,NULL);CHKERRQ(ierr); 424816e386b8SStefano Zampini ierr = MatAXPY(tmat,-1.,At,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 424916e386b8SStefano Zampini ierr = MatNorm(tmat,NORM_INFINITY,&norm);CHKERRQ(ierr); 425016e386b8SStefano Zampini if (norm > PETSC_SMALL) { 425116e386b8SStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Not yet implemented"); 425216e386b8SStefano Zampini } 425316e386b8SStefano Zampini #endif 425416e386b8SStefano Zampini } 425588c03ad3SStefano Zampini } 425688c03ad3SStefano Zampini ierr = MatDestroy(&tmat);CHKERRQ(ierr); 4257b96c3477SStefano Zampini } 4258b96c3477SStefano Zampini } 4259906d46d4SStefano Zampini ierr = MatDestroy(&localChangeOfBasisMatrix);CHKERRQ(ierr); 4260906d46d4SStefano Zampini } else if (pcbddc->user_ChangeOfBasisMatrix) { 4261b9b85e73SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->user_ChangeOfBasisMatrix);CHKERRQ(ierr); 4262b9b85e73SStefano Zampini pcbddc->ChangeOfBasisMatrix = pcbddc->user_ChangeOfBasisMatrix; 4263b9b85e73SStefano Zampini } 4264906d46d4SStefano Zampini 4265906d46d4SStefano Zampini /* set up change of basis context */ 4266906d46d4SStefano Zampini if (pcbddc->ChangeOfBasisMatrix) { 4267906d46d4SStefano Zampini PCBDDCChange_ctx change_ctx; 4268906d46d4SStefano Zampini 4269906d46d4SStefano Zampini if (!pcbddc->new_global_mat) { 4270906d46d4SStefano Zampini PetscInt global_size,local_size; 4271906d46d4SStefano Zampini 4272906d46d4SStefano Zampini ierr = VecGetSize(pcis->vec1_global,&global_size);CHKERRQ(ierr); 4273906d46d4SStefano Zampini ierr = VecGetLocalSize(pcis->vec1_global,&local_size);CHKERRQ(ierr); 4274906d46d4SStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)pc),&pcbddc->new_global_mat);CHKERRQ(ierr); 4275906d46d4SStefano Zampini ierr = MatSetSizes(pcbddc->new_global_mat,local_size,local_size,global_size,global_size);CHKERRQ(ierr); 4276906d46d4SStefano Zampini ierr = MatSetType(pcbddc->new_global_mat,MATSHELL);CHKERRQ(ierr); 4277906d46d4SStefano Zampini ierr = MatShellSetOperation(pcbddc->new_global_mat,MATOP_MULT,(void (*)(void))PCBDDCMatMult_Private);CHKERRQ(ierr); 4278906d46d4SStefano Zampini ierr = MatShellSetOperation(pcbddc->new_global_mat,MATOP_MULT_TRANSPOSE,(void (*)(void))PCBDDCMatMultTranspose_Private);CHKERRQ(ierr); 4279906d46d4SStefano Zampini ierr = PetscNew(&change_ctx);CHKERRQ(ierr); 4280906d46d4SStefano Zampini ierr = MatShellSetContext(pcbddc->new_global_mat,change_ctx);CHKERRQ(ierr); 4281906d46d4SStefano Zampini } else { 4282906d46d4SStefano Zampini ierr = MatShellGetContext(pcbddc->new_global_mat,&change_ctx);CHKERRQ(ierr); 4283906d46d4SStefano Zampini ierr = MatDestroy(&change_ctx->global_change);CHKERRQ(ierr); 4284906d46d4SStefano Zampini ierr = VecDestroyVecs(2,&change_ctx->work);CHKERRQ(ierr); 4285906d46d4SStefano Zampini } 4286906d46d4SStefano Zampini if (!pcbddc->user_ChangeOfBasisMatrix) { 4287906d46d4SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 4288906d46d4SStefano Zampini change_ctx->global_change = pcbddc->ChangeOfBasisMatrix; 4289906d46d4SStefano Zampini } else { 4290906d46d4SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->user_ChangeOfBasisMatrix);CHKERRQ(ierr); 4291906d46d4SStefano Zampini change_ctx->global_change = pcbddc->user_ChangeOfBasisMatrix; 4292906d46d4SStefano Zampini } 4293906d46d4SStefano Zampini ierr = VecDuplicateVecs(pcis->vec1_global,2,&change_ctx->work);CHKERRQ(ierr); 4294906d46d4SStefano Zampini ierr = MatSetUp(pcbddc->new_global_mat);CHKERRQ(ierr); 4295906d46d4SStefano Zampini ierr = MatAssemblyBegin(pcbddc->new_global_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4296906d46d4SStefano Zampini ierr = MatAssemblyEnd(pcbddc->new_global_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4297a8661815SStefano Zampini } else { 4298a8661815SStefano Zampini ierr = MatDestroy(&pcbddc->new_global_mat);CHKERRQ(ierr); 4299b9b85e73SStefano Zampini } 4300a717540cSStefano Zampini 43014f1b2e48SStefano Zampini /* add pressure dofs to set of primal nodes for numbering purposes */ 43024f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 43034f1b2e48SStefano Zampini pcbddc->local_primal_ref_node[pcbddc->local_primal_size_cc] = pcbddc->benign_p0_lidx[i]; 43044f1b2e48SStefano Zampini pcbddc->primal_indices_local_idxs[pcbddc->local_primal_size] = pcbddc->benign_p0_lidx[i]; 4305019a44ceSStefano Zampini pcbddc->local_primal_ref_mult[pcbddc->local_primal_size_cc] = 1; 4306019a44ceSStefano Zampini pcbddc->local_primal_size_cc++; 4307019a44ceSStefano Zampini pcbddc->local_primal_size++; 4308019a44ceSStefano Zampini } 4309019a44ceSStefano Zampini 4310019a44ceSStefano Zampini /* check if a new primal space has been introduced (also take into account benign trick) */ 4311727cdba6SStefano Zampini pcbddc->new_primal_space_local = PETSC_TRUE; 4312727cdba6SStefano Zampini if (olocal_primal_size == pcbddc->local_primal_size) { 43139f47a83aSStefano 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); 4314c1c8e736SStefano Zampini pcbddc->new_primal_space_local = (PetscBool)(!pcbddc->new_primal_space_local); 43150e6343abSStefano Zampini if (!pcbddc->new_primal_space_local) { 43169f47a83aSStefano 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); 4317727cdba6SStefano Zampini pcbddc->new_primal_space_local = (PetscBool)(!pcbddc->new_primal_space_local); 4318727cdba6SStefano Zampini } 43190e6343abSStefano Zampini } 43200e6343abSStefano Zampini ierr = PetscFree2(olocal_primal_ref_node,olocal_primal_ref_mult);CHKERRQ(ierr); 4321727cdba6SStefano Zampini /* new_primal_space will be used for numbering of coarse dofs, so it should be the same across all subdomains */ 4322727cdba6SStefano Zampini ierr = MPI_Allreduce(&pcbddc->new_primal_space_local,&pcbddc->new_primal_space,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 4323727cdba6SStefano Zampini 4324a717540cSStefano Zampini /* flush dbg viewer */ 4325b8ffe317SStefano Zampini if (pcbddc->dbg_flag) { 4326b8ffe317SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4327b8ffe317SStefano Zampini } 4328a717540cSStefano Zampini 4329e310c8b4SStefano Zampini /* free workspace */ 4330a717540cSStefano Zampini ierr = PetscBTDestroy(&qr_needed_idx);CHKERRQ(ierr); 43314641a718SStefano Zampini ierr = PetscBTDestroy(&change_basis);CHKERRQ(ierr); 433208122e43SStefano Zampini if (!pcbddc->adaptive_selection) { 43339162d606SStefano Zampini ierr = PetscFree3(constraints_idxs_ptr,constraints_data_ptr,constraints_n);CHKERRQ(ierr); 43349162d606SStefano Zampini ierr = PetscFree3(constraints_data,constraints_idxs,constraints_idxs_B);CHKERRQ(ierr); 433508122e43SStefano Zampini } else { 43369162d606SStefano Zampini ierr = PetscFree5(pcbddc->adaptive_constraints_n, 43379162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr, 43389162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr, 433908122e43SStefano Zampini pcbddc->adaptive_constraints_idxs, 434008122e43SStefano Zampini pcbddc->adaptive_constraints_data);CHKERRQ(ierr); 43419162d606SStefano Zampini ierr = PetscFree(constraints_n);CHKERRQ(ierr); 43429162d606SStefano Zampini ierr = PetscFree(constraints_idxs_B);CHKERRQ(ierr); 434308122e43SStefano Zampini } 4344674ae819SStefano Zampini PetscFunctionReturn(0); 4345674ae819SStefano Zampini } 4346674ae819SStefano Zampini 4347674ae819SStefano Zampini #undef __FUNCT__ 4348674ae819SStefano Zampini #define __FUNCT__ "PCBDDCAnalyzeInterface" 4349674ae819SStefano Zampini PetscErrorCode PCBDDCAnalyzeInterface(PC pc) 4350674ae819SStefano Zampini { 4351674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 4352674ae819SStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 4353674ae819SStefano Zampini Mat_IS *matis = (Mat_IS*)pc->pmat->data; 43547fb0e2dbSStefano Zampini PetscInt ierr,i,vertex_size,N; 4355674ae819SStefano Zampini PetscViewer viewer=pcbddc->dbg_viewer; 4356674ae819SStefano Zampini 4357674ae819SStefano Zampini PetscFunctionBegin; 43588e61c736SStefano Zampini /* Reset previously computed graph */ 43598e61c736SStefano Zampini ierr = PCBDDCGraphReset(pcbddc->mat_graph);CHKERRQ(ierr); 4360674ae819SStefano Zampini /* Init local Graph struct */ 43617fb0e2dbSStefano Zampini ierr = MatGetSize(pc->pmat,&N,NULL);CHKERRQ(ierr); 43623bbff08aSStefano Zampini ierr = PCBDDCGraphInit(pcbddc->mat_graph,pcis->mapping,N);CHKERRQ(ierr); 4363674ae819SStefano Zampini 4364575ad6abSStefano Zampini /* Check validity of the csr graph passed in by the user */ 43655099eff2SStefano Zampini if (pcbddc->mat_graph->nvtxs_csr && pcbddc->mat_graph->nvtxs_csr != pcbddc->mat_graph->nvtxs) { 43665099eff2SStefano Zampini SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONG,"Invalid size of local CSR graph! Found %d, expected %d\n",pcbddc->mat_graph->nvtxs_csr,pcbddc->mat_graph->nvtxs); 4367575ad6abSStefano Zampini } 43689577ea80SStefano Zampini 4369674ae819SStefano Zampini /* Set default CSR adjacency of local dofs if not provided by the user with PCBDDCSetLocalAdjacencyGraph */ 4370d4d8cf7bSStefano Zampini if ( (!pcbddc->mat_graph->xadj || !pcbddc->mat_graph->adjncy) && pcbddc->use_local_adj) { 43714d379d7bSStefano Zampini PetscInt *xadj,*adjncy; 43724d379d7bSStefano Zampini PetscInt nvtxs; 4373e496cd5dSStefano Zampini PetscBool flg_row=PETSC_FALSE; 4374674ae819SStefano Zampini 43752fffb893SStefano Zampini ierr = MatGetRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 43762fffb893SStefano Zampini if (flg_row) { 43774d379d7bSStefano Zampini ierr = PCBDDCSetLocalAdjacencyGraph(pc,nvtxs,xadj,adjncy,PETSC_COPY_VALUES);CHKERRQ(ierr); 4378b96c3477SStefano Zampini pcbddc->computed_rowadj = PETSC_TRUE; 43792fffb893SStefano Zampini } 43802fffb893SStefano Zampini ierr = MatRestoreRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 4381674ae819SStefano Zampini } 43829b28b941SStefano Zampini if (pcbddc->dbg_flag) { 43839b28b941SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4384674ae819SStefano Zampini } 4385674ae819SStefano Zampini 438663602bcaSStefano Zampini /* Set default dofs' splitting if no information has been provided by the user with PCBDDCSetDofsSplitting or PCBDDCSetDofsSplittingLocal */ 4387674ae819SStefano Zampini vertex_size = 1; 438863602bcaSStefano Zampini if (pcbddc->user_provided_isfordofs) { 438963602bcaSStefano Zampini if (pcbddc->n_ISForDofs) { /* need to convert from global to local and remove references to global dofs splitting */ 439095ecbf38SStefano Zampini ierr = PetscMalloc1(pcbddc->n_ISForDofs,&pcbddc->ISForDofsLocal);CHKERRQ(ierr); 439163602bcaSStefano Zampini for (i=0;i<pcbddc->n_ISForDofs;i++) { 4392e176bc59SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->rctx,pcis->vec1_global,pcis->vec1_N,pcbddc->ISForDofs[i],&pcbddc->ISForDofsLocal[i]);CHKERRQ(ierr); 439363602bcaSStefano Zampini ierr = ISDestroy(&pcbddc->ISForDofs[i]);CHKERRQ(ierr); 4394674ae819SStefano Zampini } 439563602bcaSStefano Zampini pcbddc->n_ISForDofsLocal = pcbddc->n_ISForDofs; 439663602bcaSStefano Zampini pcbddc->n_ISForDofs = 0; 439763602bcaSStefano Zampini ierr = PetscFree(pcbddc->ISForDofs);CHKERRQ(ierr); 4398674ae819SStefano Zampini } 439963602bcaSStefano Zampini /* mat block size as vertex size (used for elasticity with rigid body modes as nearnullspace) */ 4400674ae819SStefano Zampini ierr = MatGetBlockSize(matis->A,&vertex_size);CHKERRQ(ierr); 440163602bcaSStefano Zampini } else { 440263602bcaSStefano Zampini if (!pcbddc->n_ISForDofsLocal) { /* field split not present, create it in local ordering */ 440363602bcaSStefano Zampini ierr = MatGetBlockSize(pc->pmat,&pcbddc->n_ISForDofsLocal);CHKERRQ(ierr); 4404854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->n_ISForDofsLocal,&pcbddc->ISForDofsLocal);CHKERRQ(ierr); 440563602bcaSStefano Zampini for (i=0;i<pcbddc->n_ISForDofsLocal;i++) { 440663602bcaSStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)pc),pcis->n/pcbddc->n_ISForDofsLocal,i,pcbddc->n_ISForDofsLocal,&pcbddc->ISForDofsLocal[i]);CHKERRQ(ierr); 440763602bcaSStefano Zampini } 440863602bcaSStefano Zampini } 4409674ae819SStefano Zampini } 4410674ae819SStefano Zampini 4411674ae819SStefano Zampini /* Setup of Graph */ 4412785d1243SStefano Zampini if (!pcbddc->DirichletBoundariesLocal && pcbddc->DirichletBoundaries) { /* need to convert from global to local */ 4413e176bc59SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->rctx,pcis->vec1_global,pcis->vec1_N,pcbddc->DirichletBoundaries,&pcbddc->DirichletBoundariesLocal);CHKERRQ(ierr); 4414785d1243SStefano Zampini } 4415785d1243SStefano Zampini if (!pcbddc->NeumannBoundariesLocal && pcbddc->NeumannBoundaries) { /* need to convert from global to local */ 4416e176bc59SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->rctx,pcis->vec1_global,pcis->vec1_N,pcbddc->NeumannBoundaries,&pcbddc->NeumannBoundariesLocal);CHKERRQ(ierr); 4417785d1243SStefano Zampini } 441830368db7SStefano Zampini if (!pcbddc->user_primal_vertices_local && pcbddc->user_primal_vertices) { /* need to convert from global to local */ 441930368db7SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->rctx,pcis->vec1_global,pcis->vec1_N,pcbddc->user_primal_vertices,&pcbddc->user_primal_vertices_local);CHKERRQ(ierr); 442030368db7SStefano Zampini } 442130368db7SStefano Zampini ierr = PCBDDCGraphSetUp(pcbddc->mat_graph,vertex_size,pcbddc->NeumannBoundariesLocal,pcbddc->DirichletBoundariesLocal,pcbddc->n_ISForDofsLocal,pcbddc->ISForDofsLocal,pcbddc->user_primal_vertices_local);CHKERRQ(ierr); 4422674ae819SStefano Zampini 44234f1b2e48SStefano Zampini /* attach info on disconnected subdomains if present */ 44244f1b2e48SStefano Zampini if (pcbddc->n_local_subs) { 44254f1b2e48SStefano Zampini PetscInt *local_subs; 44264f1b2e48SStefano Zampini 44274f1b2e48SStefano Zampini ierr = PetscMalloc1(N,&local_subs);CHKERRQ(ierr); 44284f1b2e48SStefano Zampini for (i=0;i<pcbddc->n_local_subs;i++) { 44294f1b2e48SStefano Zampini const PetscInt *idxs; 44304f1b2e48SStefano Zampini PetscInt nl,j; 44314f1b2e48SStefano Zampini 44324f1b2e48SStefano Zampini ierr = ISGetLocalSize(pcbddc->local_subs[i],&nl);CHKERRQ(ierr); 44334f1b2e48SStefano Zampini ierr = ISGetIndices(pcbddc->local_subs[i],&idxs);CHKERRQ(ierr); 44344f1b2e48SStefano Zampini for (j=0;j<nl;j++) { 44354f1b2e48SStefano Zampini local_subs[idxs[j]] = i; 44364f1b2e48SStefano Zampini } 44374f1b2e48SStefano Zampini ierr = ISRestoreIndices(pcbddc->local_subs[i],&idxs);CHKERRQ(ierr); 44384f1b2e48SStefano Zampini } 44394f1b2e48SStefano Zampini pcbddc->mat_graph->n_local_subs = pcbddc->n_local_subs; 44404f1b2e48SStefano Zampini pcbddc->mat_graph->local_subs = local_subs; 44414f1b2e48SStefano Zampini } 44424f1b2e48SStefano Zampini 4443674ae819SStefano Zampini /* Graph's connected components analysis */ 4444674ae819SStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(pcbddc->mat_graph);CHKERRQ(ierr); 4445674ae819SStefano Zampini 4446674ae819SStefano Zampini /* print some info to stdout */ 4447674ae819SStefano Zampini if (pcbddc->dbg_flag) { 4448302440fdSBarry Smith ierr = PCBDDCGraphASCIIView(pcbddc->mat_graph,pcbddc->dbg_flag,viewer);CHKERRQ(ierr); 4449674ae819SStefano Zampini } 4450fb180af4SStefano Zampini 4451fb180af4SStefano Zampini /* mark topography has done */ 4452fb180af4SStefano Zampini pcbddc->recompute_topography = PETSC_FALSE; 4453674ae819SStefano Zampini PetscFunctionReturn(0); 4454674ae819SStefano Zampini } 4455674ae819SStefano Zampini 4456dc456d91SStefano Zampini /* given an index sets possibly with holes, renumbers the indexes removing the holes */ 4457674ae819SStefano Zampini #undef __FUNCT__ 4458674ae819SStefano Zampini #define __FUNCT__ "PCBDDCSubsetNumbering" 4459dc456d91SStefano Zampini PetscErrorCode PCBDDCSubsetNumbering(IS subset, IS subset_mult, PetscInt *N_n, IS *subset_n) 4460674ae819SStefano Zampini { 4461dc456d91SStefano Zampini PetscSF sf; 4462dc456d91SStefano Zampini PetscLayout map; 4463dc456d91SStefano Zampini const PetscInt *idxs; 4464dc456d91SStefano Zampini PetscInt *leaf_data,*root_data,*gidxs; 4465dc456d91SStefano Zampini PetscInt N,n,i,lbounds[2],gbounds[2],Nl; 4466dc456d91SStefano Zampini PetscInt n_n,nlocals,start,first_index; 4467dc456d91SStefano Zampini PetscMPIInt commsize; 4468674ae819SStefano Zampini PetscBool first_found; 4469674ae819SStefano Zampini PetscErrorCode ierr; 4470674ae819SStefano Zampini 4471674ae819SStefano Zampini PetscFunctionBegin; 4472dc456d91SStefano Zampini ierr = ISGetLocalSize(subset,&n);CHKERRQ(ierr); 4473dc456d91SStefano Zampini if (subset_mult) { 4474dc456d91SStefano Zampini PetscCheckSameComm(subset,1,subset_mult,2); 4475dc456d91SStefano Zampini ierr = ISGetLocalSize(subset,&i);CHKERRQ(ierr); 4476dc456d91SStefano Zampini if (i != n) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Local subset and multiplicity sizes don't match! %d != %d",n,i); 4477674ae819SStefano Zampini } 4478dc456d91SStefano Zampini /* create workspace layout for computing global indices of subset */ 4479dc456d91SStefano Zampini ierr = ISGetIndices(subset,&idxs);CHKERRQ(ierr); 4480dc456d91SStefano Zampini lbounds[0] = lbounds[1] = 0; 4481dc456d91SStefano Zampini for (i=0;i<n;i++) { 4482dc456d91SStefano Zampini if (idxs[i] < lbounds[0]) lbounds[0] = idxs[i]; 4483dc456d91SStefano Zampini else if (idxs[i] > lbounds[1]) lbounds[1] = idxs[i]; 4484674ae819SStefano Zampini } 4485dc456d91SStefano Zampini lbounds[0] = -lbounds[0]; 4486dc456d91SStefano Zampini ierr = MPI_Allreduce(lbounds,gbounds,2,MPIU_INT,MPI_MAX,PetscObjectComm((PetscObject)subset));CHKERRQ(ierr); 4487dc456d91SStefano Zampini gbounds[0] = -gbounds[0]; 4488dc456d91SStefano Zampini N = gbounds[1] - gbounds[0] + 1; 4489dc456d91SStefano Zampini ierr = PetscLayoutCreate(PetscObjectComm((PetscObject)subset),&map);CHKERRQ(ierr); 4490dc456d91SStefano Zampini ierr = PetscLayoutSetBlockSize(map,1);CHKERRQ(ierr); 4491dc456d91SStefano Zampini ierr = PetscLayoutSetSize(map,N);CHKERRQ(ierr); 4492dc456d91SStefano Zampini ierr = PetscLayoutSetUp(map);CHKERRQ(ierr); 4493dc456d91SStefano Zampini ierr = PetscLayoutGetLocalSize(map,&Nl);CHKERRQ(ierr); 4494dc456d91SStefano Zampini 4495dc456d91SStefano Zampini /* create sf : leaf_data == multiplicity of indexes, root data == global index in layout */ 4496dc456d91SStefano Zampini ierr = PetscMalloc2(n,&leaf_data,Nl,&root_data);CHKERRQ(ierr); 4497dc456d91SStefano Zampini if (subset_mult) { 4498dc456d91SStefano Zampini const PetscInt* idxs_mult; 4499dc456d91SStefano Zampini 4500dc456d91SStefano Zampini ierr = ISGetIndices(subset_mult,&idxs_mult);CHKERRQ(ierr); 4501dc456d91SStefano Zampini ierr = PetscMemcpy(leaf_data,idxs_mult,n*sizeof(PetscInt));CHKERRQ(ierr); 4502dc456d91SStefano Zampini ierr = ISRestoreIndices(subset_mult,&idxs_mult);CHKERRQ(ierr); 4503674ae819SStefano Zampini } else { 4504dc456d91SStefano Zampini for (i=0;i<n;i++) leaf_data[i] = 1; 4505674ae819SStefano Zampini } 4506dc456d91SStefano Zampini /* local size of new subset */ 4507dc456d91SStefano Zampini n_n = 0; 4508dc456d91SStefano Zampini for (i=0;i<n;i++) n_n += leaf_data[i]; 4509dc456d91SStefano Zampini 4510dc456d91SStefano Zampini /* global indexes in layout */ 4511dc456d91SStefano Zampini ierr = PetscMalloc1(n_n,&gidxs);CHKERRQ(ierr); /* allocating possibly extra space in gidxs which will be used later */ 4512dc456d91SStefano Zampini for (i=0;i<n;i++) gidxs[i] = idxs[i] - gbounds[0]; 4513dc456d91SStefano Zampini ierr = ISRestoreIndices(subset,&idxs);CHKERRQ(ierr); 4514dc456d91SStefano Zampini ierr = PetscSFCreate(PetscObjectComm((PetscObject)subset),&sf);CHKERRQ(ierr); 4515dc456d91SStefano Zampini ierr = PetscSFSetGraphLayout(sf,map,n,NULL,PETSC_COPY_VALUES,gidxs);CHKERRQ(ierr); 4516dc456d91SStefano Zampini ierr = PetscLayoutDestroy(&map);CHKERRQ(ierr); 4517dc456d91SStefano Zampini 4518dc456d91SStefano Zampini /* reduce from leaves to roots */ 4519dc456d91SStefano Zampini ierr = PetscMemzero(root_data,Nl*sizeof(PetscInt));CHKERRQ(ierr); 452064a8e5bcSStefano Zampini ierr = PetscSFReduceBegin(sf,MPIU_INT,leaf_data,root_data,MPI_MAX);CHKERRQ(ierr); 452164a8e5bcSStefano Zampini ierr = PetscSFReduceEnd(sf,MPIU_INT,leaf_data,root_data,MPI_MAX);CHKERRQ(ierr); 4522dc456d91SStefano Zampini 4523dc456d91SStefano Zampini /* count indexes in local part of layout */ 4524674ae819SStefano Zampini nlocals = 0; 4525674ae819SStefano Zampini first_index = -1; 4526674ae819SStefano Zampini first_found = PETSC_FALSE; 4527dc456d91SStefano Zampini for (i=0;i<Nl;i++) { 4528dc456d91SStefano Zampini if (!first_found && root_data[i]) { 4529674ae819SStefano Zampini first_found = PETSC_TRUE; 4530674ae819SStefano Zampini first_index = i; 4531674ae819SStefano Zampini } 4532dc456d91SStefano Zampini nlocals += root_data[i]; 4533674ae819SStefano Zampini } 4534dc456d91SStefano Zampini 4535dc456d91SStefano Zampini /* cumulative of number of indexes and size of subset without holes */ 45365fa240b1SStefano Zampini #if defined(PETSC_HAVE_MPI_EXSCAN) 4537dc456d91SStefano Zampini start = 0; 453864a8e5bcSStefano Zampini ierr = MPI_Exscan(&nlocals,&start,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)subset));CHKERRQ(ierr); 45395fa240b1SStefano Zampini #else 454064a8e5bcSStefano Zampini ierr = MPI_Scan(&nlocals,&start,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)subset));CHKERRQ(ierr); 45415fa240b1SStefano Zampini start = start-nlocals; 45425fa240b1SStefano Zampini #endif 45435fa240b1SStefano Zampini 4544dc456d91SStefano Zampini if (N_n) { /* compute total size of new subset if requested */ 4545dc456d91SStefano Zampini *N_n = start + nlocals; 4546dc456d91SStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)subset),&commsize);CHKERRQ(ierr); 4547dc456d91SStefano Zampini ierr = MPI_Bcast(N_n,1,MPIU_INT,commsize-1,PetscObjectComm((PetscObject)subset));CHKERRQ(ierr); 4548674ae819SStefano Zampini } 45495fa240b1SStefano Zampini 45505fa240b1SStefano Zampini /* adapt root data with cumulative */ 4551674ae819SStefano Zampini if (first_found) { 4552dc456d91SStefano Zampini PetscInt old_index; 4553dc456d91SStefano Zampini 4554dc456d91SStefano Zampini root_data[first_index] += start; 4555674ae819SStefano Zampini old_index = first_index; 4556dc456d91SStefano Zampini for (i=first_index+1;i<Nl;i++) { 4557dc456d91SStefano Zampini if (root_data[i]) { 4558dc456d91SStefano Zampini root_data[i] += root_data[old_index]; 4559674ae819SStefano Zampini old_index = i; 4560674ae819SStefano Zampini } 4561674ae819SStefano Zampini } 4562674ae819SStefano Zampini } 4563dc456d91SStefano Zampini 4564dc456d91SStefano Zampini /* from roots to leaves */ 4565dc456d91SStefano Zampini ierr = PetscSFBcastBegin(sf,MPIU_INT,root_data,leaf_data);CHKERRQ(ierr); 4566dc456d91SStefano Zampini ierr = PetscSFBcastEnd(sf,MPIU_INT,root_data,leaf_data);CHKERRQ(ierr); 4567dc456d91SStefano Zampini ierr = PetscSFDestroy(&sf);CHKERRQ(ierr); 4568dc456d91SStefano Zampini 4569dc456d91SStefano Zampini /* create new IS with global indexes without holes */ 4570dc456d91SStefano Zampini if (subset_mult) { 4571dc456d91SStefano Zampini const PetscInt* idxs_mult; 4572dc456d91SStefano Zampini PetscInt cum; 4573dc456d91SStefano Zampini 4574dc456d91SStefano Zampini cum = 0; 4575dc456d91SStefano Zampini ierr = ISGetIndices(subset_mult,&idxs_mult);CHKERRQ(ierr); 4576dc456d91SStefano Zampini for (i=0;i<n;i++) { 4577dc456d91SStefano Zampini PetscInt j; 4578dc456d91SStefano Zampini for (j=0;j<idxs_mult[i];j++) gidxs[cum++] = leaf_data[i] - idxs_mult[i] + j; 4579674ae819SStefano Zampini } 4580dc456d91SStefano Zampini ierr = ISRestoreIndices(subset_mult,&idxs_mult);CHKERRQ(ierr); 4581674ae819SStefano Zampini } else { 4582dc456d91SStefano Zampini for (i=0;i<n;i++) { 4583dc456d91SStefano Zampini gidxs[i] = leaf_data[i]-1; 4584674ae819SStefano Zampini } 4585674ae819SStefano Zampini } 4586dc456d91SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)subset),n_n,gidxs,PETSC_OWN_POINTER,subset_n);CHKERRQ(ierr); 4587dc456d91SStefano Zampini ierr = PetscFree2(leaf_data,root_data);CHKERRQ(ierr); 4588674ae819SStefano Zampini PetscFunctionReturn(0); 4589674ae819SStefano Zampini } 45909a7d3425SStefano Zampini 45919a7d3425SStefano Zampini #undef __FUNCT__ 45929a7d3425SStefano Zampini #define __FUNCT__ "PCBDDCOrthonormalizeVecs" 45939a7d3425SStefano Zampini PetscErrorCode PCBDDCOrthonormalizeVecs(PetscInt n, Vec vecs[]) 45949a7d3425SStefano Zampini { 45959a7d3425SStefano Zampini PetscInt i,j; 45969a7d3425SStefano Zampini PetscScalar *alphas; 45979a7d3425SStefano Zampini PetscErrorCode ierr; 45989a7d3425SStefano Zampini 45999a7d3425SStefano Zampini PetscFunctionBegin; 46009a7d3425SStefano Zampini /* this implements stabilized Gram-Schmidt */ 4601785e854fSJed Brown ierr = PetscMalloc1(n,&alphas);CHKERRQ(ierr); 46029a7d3425SStefano Zampini for (i=0;i<n;i++) { 46039a7d3425SStefano Zampini ierr = VecNormalize(vecs[i],NULL);CHKERRQ(ierr); 46049a7d3425SStefano Zampini if (i<n) { ierr = VecMDot(vecs[i],n-i-1,&vecs[i+1],&alphas[i+1]);CHKERRQ(ierr); } 46059a7d3425SStefano Zampini for (j=i+1;j<n;j++) { ierr = VecAXPY(vecs[j],PetscConj(-alphas[j]),vecs[i]);CHKERRQ(ierr); } 46069a7d3425SStefano Zampini } 46079a7d3425SStefano Zampini ierr = PetscFree(alphas);CHKERRQ(ierr); 46089a7d3425SStefano Zampini PetscFunctionReturn(0); 46099a7d3425SStefano Zampini } 46109a7d3425SStefano Zampini 4611e7931f94SStefano Zampini #undef __FUNCT__ 461270cf5478SStefano Zampini #define __FUNCT__ "MatISGetSubassemblingPattern" 4613b0c7d250SStefano Zampini PetscErrorCode MatISGetSubassemblingPattern(Mat mat, PetscInt n_subdomains, PetscInt redprocs, IS* is_sends) 4614e7931f94SStefano Zampini { 461552e5ac9dSStefano Zampini IS ranks_send_to; 4616e7931f94SStefano Zampini PetscInt n_neighs,*neighs,*n_shared,**shared; 4617e7931f94SStefano Zampini PetscMPIInt size,rank,color; 461852e5ac9dSStefano Zampini PetscInt *xadj,*adjncy; 461952e5ac9dSStefano Zampini PetscInt *adjncy_wgt,*v_wgt,*ranks_send_to_idx; 46203837a79fSStefano Zampini PetscInt i,local_size,threshold=0; 46212b510759SStefano Zampini PetscBool use_vwgt=PETSC_FALSE,use_square=PETSC_FALSE; 4622e7931f94SStefano Zampini PetscSubcomm subcomm; 462352e5ac9dSStefano Zampini PetscErrorCode ierr; 4624a57a6d2fSStefano Zampini 4625e7931f94SStefano Zampini PetscFunctionBegin; 46262b510759SStefano Zampini ierr = PetscOptionsGetBool(NULL,"-matis_partitioning_use_square",&use_square,NULL);CHKERRQ(ierr); 46272b510759SStefano Zampini ierr = PetscOptionsGetBool(NULL,"-matis_partitioning_use_vwgt",&use_vwgt,NULL);CHKERRQ(ierr); 46282b510759SStefano Zampini ierr = PetscOptionsGetInt(NULL,"-matis_partitioning_threshold",&threshold,NULL);CHKERRQ(ierr); 4629e7931f94SStefano Zampini 4630e7931f94SStefano Zampini /* Get info on mapping */ 46313bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetSize(mat->rmap->mapping,&local_size);CHKERRQ(ierr); 46323bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetInfo(mat->rmap->mapping,&n_neighs,&neighs,&n_shared,&shared);CHKERRQ(ierr); 4633e7931f94SStefano Zampini 4634e7931f94SStefano Zampini /* build local CSR graph of subdomains' connectivity */ 4635785e854fSJed Brown ierr = PetscMalloc1(2,&xadj);CHKERRQ(ierr); 4636e7931f94SStefano Zampini xadj[0] = 0; 4637e7931f94SStefano Zampini xadj[1] = PetscMax(n_neighs-1,0); 4638785e854fSJed Brown ierr = PetscMalloc1(xadj[1],&adjncy);CHKERRQ(ierr); 4639785e854fSJed Brown ierr = PetscMalloc1(xadj[1],&adjncy_wgt);CHKERRQ(ierr); 4640e7931f94SStefano Zampini 46412b510759SStefano Zampini if (threshold) { 4642d023bfaeSStefano Zampini PetscInt xadj_count = 0; 46432b510759SStefano Zampini for (i=1;i<n_neighs;i++) { 4644d023bfaeSStefano Zampini if (n_shared[i] > threshold) { 4645d023bfaeSStefano Zampini adjncy[xadj_count] = neighs[i]; 4646d023bfaeSStefano Zampini adjncy_wgt[xadj_count] = n_shared[i]; 4647d023bfaeSStefano Zampini xadj_count++; 4648e7931f94SStefano Zampini } 4649e7931f94SStefano Zampini } 4650d023bfaeSStefano Zampini xadj[1] = xadj_count; 4651c8587f34SStefano Zampini } else { 4652e7931f94SStefano Zampini if (xadj[1]) { 4653e7931f94SStefano Zampini ierr = PetscMemcpy(adjncy,&neighs[1],xadj[1]*sizeof(*adjncy));CHKERRQ(ierr); 4654e7931f94SStefano Zampini ierr = PetscMemcpy(adjncy_wgt,&n_shared[1],xadj[1]*sizeof(*adjncy_wgt));CHKERRQ(ierr); 4655c8587f34SStefano Zampini } 4656e7931f94SStefano Zampini } 46573bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingRestoreInfo(mat->rmap->mapping,&n_neighs,&neighs,&n_shared,&shared);CHKERRQ(ierr); 4658e7931f94SStefano Zampini if (use_square) { 4659e7931f94SStefano Zampini for (i=0;i<xadj[1];i++) { 4660e7931f94SStefano Zampini adjncy_wgt[i] = adjncy_wgt[i]*adjncy_wgt[i]; 4661e7931f94SStefano Zampini } 4662e7931f94SStefano Zampini } 4663e7931f94SStefano Zampini ierr = PetscSortIntWithArray(xadj[1],adjncy,adjncy_wgt);CHKERRQ(ierr); 4664e7931f94SStefano Zampini 46653837a79fSStefano Zampini ierr = PetscMalloc1(1,&ranks_send_to_idx);CHKERRQ(ierr); 4666e7931f94SStefano Zampini 4667e7931f94SStefano Zampini /* 4668e7931f94SStefano Zampini Restrict work on active processes only. 4669e7931f94SStefano Zampini */ 4670e7931f94SStefano Zampini ierr = PetscSubcommCreate(PetscObjectComm((PetscObject)mat),&subcomm);CHKERRQ(ierr); 4671e7931f94SStefano Zampini ierr = PetscSubcommSetNumber(subcomm,2);CHKERRQ(ierr); /* 2 groups, active process and not active processes */ 4672e7931f94SStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)mat),&rank);CHKERRQ(ierr); 46732b510759SStefano Zampini ierr = PetscMPIIntCast(!local_size,&color);CHKERRQ(ierr); 4674d3531aaaSJed Brown ierr = PetscSubcommSetTypeGeneral(subcomm,color,rank);CHKERRQ(ierr); 4675e7931f94SStefano Zampini if (color) { 4676e7931f94SStefano Zampini ierr = PetscFree(xadj);CHKERRQ(ierr); 4677e7931f94SStefano Zampini ierr = PetscFree(adjncy);CHKERRQ(ierr); 4678e7931f94SStefano Zampini ierr = PetscFree(adjncy_wgt);CHKERRQ(ierr); 4679c8587f34SStefano Zampini } else { 468052e5ac9dSStefano Zampini Mat subdomain_adj; 468152e5ac9dSStefano Zampini IS new_ranks,new_ranks_contig; 468252e5ac9dSStefano Zampini MatPartitioning partitioner; 468352e5ac9dSStefano Zampini PetscInt prank,rstart=0,rend=0; 468452e5ac9dSStefano Zampini PetscInt *is_indices,*oldranks; 4685b0c7d250SStefano Zampini PetscBool aggregate; 4686b0c7d250SStefano Zampini 4687306c2d5bSBarry Smith ierr = MPI_Comm_size(PetscSubcommChild(subcomm),&size);CHKERRQ(ierr); 4688785e854fSJed Brown ierr = PetscMalloc1(size,&oldranks);CHKERRQ(ierr); 4689e7931f94SStefano Zampini prank = rank; 4690306c2d5bSBarry Smith ierr = MPI_Allgather(&prank,1,MPIU_INT,oldranks,1,MPIU_INT,PetscSubcommChild(subcomm));CHKERRQ(ierr); 46918002ef2cSStefano Zampini /* 4692e7931f94SStefano Zampini for (i=0;i<size;i++) { 4693e7931f94SStefano Zampini PetscPrintf(subcomm->comm,"oldranks[%d] = %d\n",i,oldranks[i]); 4694c8587f34SStefano Zampini } 46958002ef2cSStefano Zampini */ 4696e7931f94SStefano Zampini for (i=0;i<xadj[1];i++) { 4697e7931f94SStefano Zampini ierr = PetscFindInt(adjncy[i],size,oldranks,&adjncy[i]);CHKERRQ(ierr); 4698c8587f34SStefano Zampini } 4699e7931f94SStefano Zampini ierr = PetscSortIntWithArray(xadj[1],adjncy,adjncy_wgt);CHKERRQ(ierr); 4700b0c7d250SStefano Zampini aggregate = ((redprocs > 0 && redprocs < size) ? PETSC_TRUE : PETSC_FALSE); 4701b0c7d250SStefano Zampini if (aggregate) { 4702b0c7d250SStefano Zampini PetscInt lrows,row,ncols,*cols; 4703b0c7d250SStefano Zampini PetscMPIInt nrank; 4704b0c7d250SStefano Zampini PetscScalar *vals; 4705b0c7d250SStefano Zampini 4706b0c7d250SStefano Zampini ierr = MPI_Comm_rank(PetscSubcommChild(subcomm),&nrank);CHKERRQ(ierr); 4707b0c7d250SStefano Zampini lrows = 0; 4708b0c7d250SStefano Zampini if (nrank<redprocs) { 4709b0c7d250SStefano Zampini lrows = size/redprocs; 4710b0c7d250SStefano Zampini if (nrank<size%redprocs) lrows++; 4711b0c7d250SStefano Zampini } 47125fa240b1SStefano Zampini ierr = MatCreateAIJ(PetscSubcommChild(subcomm),lrows,lrows,size,size,50,NULL,50,NULL,&subdomain_adj);CHKERRQ(ierr); 4713b0c7d250SStefano Zampini ierr = MatGetOwnershipRange(subdomain_adj,&rstart,&rend);CHKERRQ(ierr); 4714b0c7d250SStefano Zampini ierr = MatSetOption(subdomain_adj,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_FALSE);CHKERRQ(ierr); 4715b0c7d250SStefano Zampini ierr = MatSetOption(subdomain_adj,MAT_NEW_NONZERO_ALLOCATION_ERR,PETSC_FALSE);CHKERRQ(ierr); 4716b0c7d250SStefano Zampini row = nrank; 4717b0c7d250SStefano Zampini ncols = xadj[1]-xadj[0]; 4718b0c7d250SStefano Zampini cols = adjncy; 4719b0c7d250SStefano Zampini ierr = PetscMalloc1(ncols,&vals);CHKERRQ(ierr); 4720b0c7d250SStefano Zampini for (i=0;i<ncols;i++) vals[i] = adjncy_wgt[i]; 4721b0c7d250SStefano Zampini ierr = MatSetValues(subdomain_adj,1,&row,ncols,cols,vals,INSERT_VALUES);CHKERRQ(ierr); 4722b0c7d250SStefano Zampini ierr = MatAssemblyBegin(subdomain_adj,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4723b0c7d250SStefano Zampini ierr = MatAssemblyEnd(subdomain_adj,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 472452e5ac9dSStefano Zampini ierr = PetscFree(xadj);CHKERRQ(ierr); 472552e5ac9dSStefano Zampini ierr = PetscFree(adjncy);CHKERRQ(ierr); 472652e5ac9dSStefano Zampini ierr = PetscFree(adjncy_wgt);CHKERRQ(ierr); 4727b0c7d250SStefano Zampini ierr = PetscFree(vals);CHKERRQ(ierr); 4728b0c7d250SStefano Zampini } else { 4729306c2d5bSBarry Smith ierr = MatCreateMPIAdj(PetscSubcommChild(subcomm),1,(PetscInt)size,xadj,adjncy,adjncy_wgt,&subdomain_adj);CHKERRQ(ierr); 4730b0c7d250SStefano Zampini } 473122b6e8a2SStefano Zampini /* ierr = MatView(subdomain_adj,0);CHKERRQ(ierr); */ 4732e7931f94SStefano Zampini 4733e7931f94SStefano Zampini /* Partition */ 4734306c2d5bSBarry Smith ierr = MatPartitioningCreate(PetscSubcommChild(subcomm),&partitioner);CHKERRQ(ierr); 4735e7931f94SStefano Zampini ierr = MatPartitioningSetAdjacency(partitioner,subdomain_adj);CHKERRQ(ierr); 4736e7931f94SStefano Zampini if (use_vwgt) { 47373837a79fSStefano Zampini ierr = PetscMalloc1(1,&v_wgt);CHKERRQ(ierr); 4738e7931f94SStefano Zampini v_wgt[0] = local_size; 4739e7931f94SStefano Zampini ierr = MatPartitioningSetVertexWeights(partitioner,v_wgt);CHKERRQ(ierr); 4740c8587f34SStefano Zampini } 474128143c3dSStefano Zampini n_subdomains = PetscMin((PetscInt)size,n_subdomains); 474228143c3dSStefano Zampini ierr = MatPartitioningSetNParts(partitioner,n_subdomains);CHKERRQ(ierr); 4743e7931f94SStefano Zampini ierr = MatPartitioningSetFromOptions(partitioner);CHKERRQ(ierr); 4744e7931f94SStefano Zampini ierr = MatPartitioningApply(partitioner,&new_ranks);CHKERRQ(ierr); 474522b6e8a2SStefano Zampini /* ierr = MatPartitioningView(partitioner,0);CHKERRQ(ierr); */ 4746e7931f94SStefano Zampini 474752e5ac9dSStefano Zampini /* renumber new_ranks to avoid "holes" in new set of processors */ 474852e5ac9dSStefano Zampini ierr = PCBDDCSubsetNumbering(new_ranks,NULL,NULL,&new_ranks_contig);CHKERRQ(ierr); 474952e5ac9dSStefano Zampini ierr = ISDestroy(&new_ranks);CHKERRQ(ierr); 475052e5ac9dSStefano Zampini ierr = ISGetIndices(new_ranks_contig,(const PetscInt**)&is_indices);CHKERRQ(ierr); 4751b0c7d250SStefano Zampini if (!redprocs) { 4752b0c7d250SStefano Zampini ranks_send_to_idx[0] = oldranks[is_indices[0]]; 475328143c3dSStefano Zampini } else { 4754b0c7d250SStefano Zampini PetscInt idxs[1]; 4755b0c7d250SStefano Zampini PetscMPIInt tag; 4756b0c7d250SStefano Zampini MPI_Request *reqs; 4757b0c7d250SStefano Zampini 4758b0c7d250SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)subdomain_adj,&tag);CHKERRQ(ierr); 4759b0c7d250SStefano Zampini ierr = PetscMalloc1(rend-rstart,&reqs);CHKERRQ(ierr); 4760b0c7d250SStefano Zampini for (i=rstart;i<rend;i++) { 4761b0c7d250SStefano Zampini ierr = MPI_Isend(is_indices+i-rstart,1,MPIU_INT,i,tag,PetscSubcommChild(subcomm),&reqs[i-rstart]);CHKERRQ(ierr); 476228143c3dSStefano Zampini } 4763b0c7d250SStefano Zampini ierr = MPI_Recv(idxs,1,MPIU_INT,MPI_ANY_SOURCE,tag,PetscSubcommChild(subcomm),MPI_STATUS_IGNORE);CHKERRQ(ierr); 4764b0c7d250SStefano Zampini ierr = MPI_Waitall(rend-rstart,reqs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 4765b0c7d250SStefano Zampini ierr = PetscFree(reqs);CHKERRQ(ierr); 4766b0c7d250SStefano Zampini ranks_send_to_idx[0] = oldranks[idxs[0]]; 4767e7931f94SStefano Zampini } 476852e5ac9dSStefano Zampini ierr = ISRestoreIndices(new_ranks_contig,(const PetscInt**)&is_indices);CHKERRQ(ierr); 4769e7931f94SStefano Zampini /* clean up */ 4770e7931f94SStefano Zampini ierr = PetscFree(oldranks);CHKERRQ(ierr); 477152e5ac9dSStefano Zampini ierr = ISDestroy(&new_ranks_contig);CHKERRQ(ierr); 4772e7931f94SStefano Zampini ierr = MatDestroy(&subdomain_adj);CHKERRQ(ierr); 4773e7931f94SStefano Zampini ierr = MatPartitioningDestroy(&partitioner);CHKERRQ(ierr); 4774e7931f94SStefano Zampini } 4775e7931f94SStefano Zampini ierr = PetscSubcommDestroy(&subcomm);CHKERRQ(ierr); 4776e7931f94SStefano Zampini 4777e7931f94SStefano Zampini /* assemble parallel IS for sends */ 4778e7931f94SStefano Zampini i = 1; 4779e7931f94SStefano Zampini if (color) i=0; 4780e7931f94SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)mat),i,ranks_send_to_idx,PETSC_OWN_POINTER,&ranks_send_to);CHKERRQ(ierr); 4781e7931f94SStefano Zampini /* get back IS */ 4782e7931f94SStefano Zampini *is_sends = ranks_send_to; 4783e7931f94SStefano Zampini PetscFunctionReturn(0); 4784e7931f94SStefano Zampini } 4785e7931f94SStefano Zampini 4786e7931f94SStefano Zampini typedef enum {MATDENSE_PRIVATE=0,MATAIJ_PRIVATE,MATBAIJ_PRIVATE,MATSBAIJ_PRIVATE}MatTypePrivate; 4787e7931f94SStefano Zampini 4788e7931f94SStefano Zampini #undef __FUNCT__ 4789e7931f94SStefano Zampini #define __FUNCT__ "MatISSubassemble" 479053a05cb3SStefano 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[]) 4791e7931f94SStefano Zampini { 479270cf5478SStefano Zampini Mat local_mat; 4793e7931f94SStefano Zampini IS is_sends_internal; 47949d30be91SStefano Zampini PetscInt rows,cols,new_local_rows; 479528143c3dSStefano Zampini PetscInt i,bs,buf_size_idxs,buf_size_idxs_is,buf_size_vals; 47969d30be91SStefano Zampini PetscBool ismatis,isdense,newisdense,destroy_mat; 4797e7931f94SStefano Zampini ISLocalToGlobalMapping l2gmap; 4798e7931f94SStefano Zampini PetscInt* l2gmap_indices; 4799e7931f94SStefano Zampini const PetscInt* is_indices; 4800e7931f94SStefano Zampini MatType new_local_type; 4801e7931f94SStefano Zampini /* buffers */ 4802e7931f94SStefano Zampini PetscInt *ptr_idxs,*send_buffer_idxs,*recv_buffer_idxs; 480328143c3dSStefano Zampini PetscInt *ptr_idxs_is,*send_buffer_idxs_is,*recv_buffer_idxs_is; 48049d30be91SStefano Zampini PetscInt *recv_buffer_idxs_local; 4805e7931f94SStefano Zampini PetscScalar *ptr_vals,*send_buffer_vals,*recv_buffer_vals; 4806e7931f94SStefano Zampini /* MPI */ 480728143c3dSStefano Zampini MPI_Comm comm,comm_n; 480828143c3dSStefano Zampini PetscSubcomm subcomm; 4809e7931f94SStefano Zampini PetscMPIInt n_sends,n_recvs,commsize; 481028143c3dSStefano Zampini PetscMPIInt *iflags,*ilengths_idxs,*ilengths_vals,*ilengths_idxs_is; 481128143c3dSStefano Zampini PetscMPIInt *onodes,*onodes_is,*olengths_idxs,*olengths_idxs_is,*olengths_vals; 481228143c3dSStefano Zampini PetscMPIInt len,tag_idxs,tag_idxs_is,tag_vals,source_dest; 481328143c3dSStefano Zampini MPI_Request *send_req_idxs,*send_req_idxs_is,*send_req_vals; 481428143c3dSStefano Zampini MPI_Request *recv_req_idxs,*recv_req_idxs_is,*recv_req_vals; 4815e7931f94SStefano Zampini PetscErrorCode ierr; 4816e7931f94SStefano Zampini 4817e7931f94SStefano Zampini PetscFunctionBegin; 481828143c3dSStefano Zampini /* TODO: add missing checks */ 481928143c3dSStefano Zampini PetscValidLogicalCollectiveInt(mat,n_subdomains,3); 482028143c3dSStefano Zampini PetscValidLogicalCollectiveBool(mat,restrict_comm,4); 482128143c3dSStefano Zampini PetscValidLogicalCollectiveEnum(mat,reuse,5); 482228143c3dSStefano Zampini PetscValidLogicalCollectiveInt(mat,nis,7); 4823e7931f94SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)mat,MATIS,&ismatis);CHKERRQ(ierr); 482428143c3dSStefano Zampini if (!ismatis) SETERRQ1(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot use %s on a matrix object which is not of type MATIS",__FUNCT__); 4825e7931f94SStefano Zampini ierr = MatISGetLocalMat(mat,&local_mat);CHKERRQ(ierr); 4826e7931f94SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)local_mat,MATSEQDENSE,&isdense);CHKERRQ(ierr); 4827e7931f94SStefano Zampini if (!isdense) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Currently cannot subassemble MATIS when local matrix type is not of type SEQDENSE"); 4828e7931f94SStefano Zampini ierr = MatGetSize(local_mat,&rows,&cols);CHKERRQ(ierr); 4829e7931f94SStefano Zampini if (rows != cols) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Local MATIS matrices should be square"); 483028143c3dSStefano Zampini if (reuse == MAT_REUSE_MATRIX && *mat_n) { 483170cf5478SStefano Zampini PetscInt mrows,mcols,mnrows,mncols; 483270cf5478SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)*mat_n,MATIS,&ismatis);CHKERRQ(ierr); 483328143c3dSStefano Zampini if (!ismatis) SETERRQ(PetscObjectComm((PetscObject)*mat_n),PETSC_ERR_SUP,"Cannot reuse a matrix which is not of type MATIS"); 483470cf5478SStefano Zampini ierr = MatGetSize(mat,&mrows,&mcols);CHKERRQ(ierr); 483570cf5478SStefano Zampini ierr = MatGetSize(*mat_n,&mnrows,&mncols);CHKERRQ(ierr); 483670cf5478SStefano Zampini if (mrows != mnrows) SETERRQ2(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot reuse matrix! Wrong number of rows %D != %D",mrows,mnrows); 483770cf5478SStefano Zampini if (mcols != mncols) SETERRQ2(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot reuse matrix! Wrong number of cols %D != %D",mcols,mncols); 483870cf5478SStefano Zampini } 4839e7931f94SStefano Zampini ierr = MatGetBlockSize(local_mat,&bs);CHKERRQ(ierr); 4840e7931f94SStefano Zampini PetscValidLogicalCollectiveInt(mat,bs,0); 4841e7931f94SStefano Zampini /* prepare IS for sending if not provided */ 4842e7931f94SStefano Zampini if (!is_sends) { 484328143c3dSStefano Zampini if (!n_subdomains) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"You should specify either an IS or a target number of subdomains"); 4844b0c7d250SStefano Zampini ierr = MatISGetSubassemblingPattern(mat,n_subdomains,0,&is_sends_internal);CHKERRQ(ierr); 4845c8587f34SStefano Zampini } else { 4846e7931f94SStefano Zampini ierr = PetscObjectReference((PetscObject)is_sends);CHKERRQ(ierr); 4847e7931f94SStefano Zampini is_sends_internal = is_sends; 4848c8587f34SStefano Zampini } 4849e7931f94SStefano Zampini 4850e7931f94SStefano Zampini /* get comm */ 4851a316fed8SStefano Zampini ierr = PetscObjectGetComm((PetscObject)mat,&comm);CHKERRQ(ierr); 4852e7931f94SStefano Zampini 4853e7931f94SStefano Zampini /* compute number of sends */ 4854e7931f94SStefano Zampini ierr = ISGetLocalSize(is_sends_internal,&i);CHKERRQ(ierr); 4855e7931f94SStefano Zampini ierr = PetscMPIIntCast(i,&n_sends);CHKERRQ(ierr); 4856e7931f94SStefano Zampini 4857e7931f94SStefano Zampini /* compute number of receives */ 4858e7931f94SStefano Zampini ierr = MPI_Comm_size(comm,&commsize);CHKERRQ(ierr); 4859785e854fSJed Brown ierr = PetscMalloc1(commsize,&iflags);CHKERRQ(ierr); 4860e7931f94SStefano Zampini ierr = PetscMemzero(iflags,commsize*sizeof(*iflags));CHKERRQ(ierr); 4861e7931f94SStefano Zampini ierr = ISGetIndices(is_sends_internal,&is_indices);CHKERRQ(ierr); 4862e7931f94SStefano Zampini for (i=0;i<n_sends;i++) iflags[is_indices[i]] = 1; 4863e7931f94SStefano Zampini ierr = PetscGatherNumberOfMessages(comm,iflags,NULL,&n_recvs);CHKERRQ(ierr); 4864e7931f94SStefano Zampini ierr = PetscFree(iflags);CHKERRQ(ierr); 4865e7931f94SStefano Zampini 486628143c3dSStefano Zampini /* restrict comm if requested */ 486728143c3dSStefano Zampini subcomm = 0; 486828143c3dSStefano Zampini destroy_mat = PETSC_FALSE; 486928143c3dSStefano Zampini if (restrict_comm) { 4870779c1cceSStefano Zampini PetscMPIInt color,subcommsize; 4871779c1cceSStefano Zampini 487228143c3dSStefano Zampini color = 0; 487353a05cb3SStefano Zampini if (restrict_full) { 487453a05cb3SStefano Zampini if (!n_recvs) color = 1; /* processes not receiving anything will not partecipate in new comm (full restriction) */ 487553a05cb3SStefano Zampini } else { 487653a05cb3SStefano Zampini if (!n_recvs && n_sends) color = 1; /* just those processes that are sending but not receiving anything will not partecipate in new comm */ 487753a05cb3SStefano Zampini } 487828143c3dSStefano Zampini ierr = MPI_Allreduce(&color,&subcommsize,1,MPI_INT,MPI_SUM,comm);CHKERRQ(ierr); 487928143c3dSStefano Zampini subcommsize = commsize - subcommsize; 488028143c3dSStefano Zampini /* check if reuse has been requested */ 488128143c3dSStefano Zampini if (reuse == MAT_REUSE_MATRIX) { 488228143c3dSStefano Zampini if (*mat_n) { 488328143c3dSStefano Zampini PetscMPIInt subcommsize2; 488428143c3dSStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)*mat_n),&subcommsize2);CHKERRQ(ierr); 488528143c3dSStefano Zampini if (subcommsize != subcommsize2) SETERRQ2(PetscObjectComm((PetscObject)*mat_n),PETSC_ERR_PLIB,"Cannot reuse matrix! wrong subcomm size %d != %d",subcommsize,subcommsize2); 488628143c3dSStefano Zampini comm_n = PetscObjectComm((PetscObject)*mat_n); 488728143c3dSStefano Zampini } else { 488828143c3dSStefano Zampini comm_n = PETSC_COMM_SELF; 488928143c3dSStefano Zampini } 489028143c3dSStefano Zampini } else { /* MAT_INITIAL_MATRIX */ 4891779c1cceSStefano Zampini PetscMPIInt rank; 4892779c1cceSStefano Zampini 4893779c1cceSStefano Zampini ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 489428143c3dSStefano Zampini ierr = PetscSubcommCreate(comm,&subcomm);CHKERRQ(ierr); 489528143c3dSStefano Zampini ierr = PetscSubcommSetNumber(subcomm,2);CHKERRQ(ierr); 489628143c3dSStefano Zampini ierr = PetscSubcommSetTypeGeneral(subcomm,color,rank);CHKERRQ(ierr); 4897306c2d5bSBarry Smith comm_n = PetscSubcommChild(subcomm); 489828143c3dSStefano Zampini } 489928143c3dSStefano Zampini /* flag to destroy *mat_n if not significative */ 490028143c3dSStefano Zampini if (color) destroy_mat = PETSC_TRUE; 490128143c3dSStefano Zampini } else { 490228143c3dSStefano Zampini comm_n = comm; 490328143c3dSStefano Zampini } 490428143c3dSStefano Zampini 4905e7931f94SStefano Zampini /* prepare send/receive buffers */ 4906785e854fSJed Brown ierr = PetscMalloc1(commsize,&ilengths_idxs);CHKERRQ(ierr); 4907e7931f94SStefano Zampini ierr = PetscMemzero(ilengths_idxs,commsize*sizeof(*ilengths_idxs));CHKERRQ(ierr); 4908785e854fSJed Brown ierr = PetscMalloc1(commsize,&ilengths_vals);CHKERRQ(ierr); 4909e7931f94SStefano Zampini ierr = PetscMemzero(ilengths_vals,commsize*sizeof(*ilengths_vals));CHKERRQ(ierr); 491028143c3dSStefano Zampini if (nis) { 4911854ce69bSBarry Smith ierr = PetscCalloc1(commsize,&ilengths_idxs_is);CHKERRQ(ierr); 491228143c3dSStefano Zampini } 4913e7931f94SStefano Zampini 491428143c3dSStefano Zampini /* Get data from local matrices */ 4915e7931f94SStefano Zampini if (!isdense) { 4916a26c9d0eSStefano Zampini SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Subassembling of AIJ local matrices not yet implemented"); 4917e7931f94SStefano Zampini /* TODO: See below some guidelines on how to prepare the local buffers */ 4918e7931f94SStefano Zampini /* 4919e7931f94SStefano Zampini send_buffer_vals should contain the raw values of the local matrix 4920e7931f94SStefano Zampini send_buffer_idxs should contain: 4921e7931f94SStefano Zampini - MatType_PRIVATE type 4922e7931f94SStefano Zampini - PetscInt size_of_l2gmap 4923e7931f94SStefano Zampini - PetscInt global_row_indices[size_of_l2gmap] 4924e7931f94SStefano Zampini - PetscInt all_other_info_which_is_needed_to_compute_preallocation_and_set_values 4925e7931f94SStefano Zampini */ 4926e7931f94SStefano Zampini } else { 4927e7931f94SStefano Zampini ierr = MatDenseGetArray(local_mat,&send_buffer_vals);CHKERRQ(ierr); 49283bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetSize(mat->rmap->mapping,&i);CHKERRQ(ierr); 4929854ce69bSBarry Smith ierr = PetscMalloc1(i+2,&send_buffer_idxs);CHKERRQ(ierr); 4930e7931f94SStefano Zampini send_buffer_idxs[0] = (PetscInt)MATDENSE_PRIVATE; 4931e7931f94SStefano Zampini send_buffer_idxs[1] = i; 49323bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetIndices(mat->rmap->mapping,(const PetscInt**)&ptr_idxs);CHKERRQ(ierr); 4933e7931f94SStefano Zampini ierr = PetscMemcpy(&send_buffer_idxs[2],ptr_idxs,i*sizeof(PetscInt));CHKERRQ(ierr); 49343bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingRestoreIndices(mat->rmap->mapping,(const PetscInt**)&ptr_idxs);CHKERRQ(ierr); 4935e7931f94SStefano Zampini ierr = PetscMPIIntCast(i,&len);CHKERRQ(ierr); 4936e7931f94SStefano Zampini for (i=0;i<n_sends;i++) { 4937e7931f94SStefano Zampini ilengths_vals[is_indices[i]] = len*len; 4938e7931f94SStefano Zampini ilengths_idxs[is_indices[i]] = len+2; 4939c8587f34SStefano Zampini } 4940c8587f34SStefano Zampini } 4941e7931f94SStefano Zampini ierr = PetscGatherMessageLengths2(comm,n_sends,n_recvs,ilengths_idxs,ilengths_vals,&onodes,&olengths_idxs,&olengths_vals);CHKERRQ(ierr); 494228143c3dSStefano Zampini /* additional is (if any) */ 494328143c3dSStefano Zampini if (nis) { 494428143c3dSStefano Zampini PetscMPIInt psum; 494528143c3dSStefano Zampini PetscInt j; 494628143c3dSStefano Zampini for (j=0,psum=0;j<nis;j++) { 494728143c3dSStefano Zampini PetscInt plen; 494828143c3dSStefano Zampini ierr = ISGetLocalSize(isarray[j],&plen);CHKERRQ(ierr); 494928143c3dSStefano Zampini ierr = PetscMPIIntCast(plen,&len);CHKERRQ(ierr); 495028143c3dSStefano Zampini psum += len+1; /* indices + lenght */ 495128143c3dSStefano Zampini } 4952854ce69bSBarry Smith ierr = PetscMalloc1(psum,&send_buffer_idxs_is);CHKERRQ(ierr); 495328143c3dSStefano Zampini for (j=0,psum=0;j<nis;j++) { 495428143c3dSStefano Zampini PetscInt plen; 495528143c3dSStefano Zampini const PetscInt *is_array_idxs; 495628143c3dSStefano Zampini ierr = ISGetLocalSize(isarray[j],&plen);CHKERRQ(ierr); 495728143c3dSStefano Zampini send_buffer_idxs_is[psum] = plen; 495828143c3dSStefano Zampini ierr = ISGetIndices(isarray[j],&is_array_idxs);CHKERRQ(ierr); 495928143c3dSStefano Zampini ierr = PetscMemcpy(&send_buffer_idxs_is[psum+1],is_array_idxs,plen*sizeof(PetscInt));CHKERRQ(ierr); 496028143c3dSStefano Zampini ierr = ISRestoreIndices(isarray[j],&is_array_idxs);CHKERRQ(ierr); 496128143c3dSStefano Zampini psum += plen+1; /* indices + lenght */ 496228143c3dSStefano Zampini } 496328143c3dSStefano Zampini for (i=0;i<n_sends;i++) { 496428143c3dSStefano Zampini ilengths_idxs_is[is_indices[i]] = psum; 496528143c3dSStefano Zampini } 496628143c3dSStefano Zampini ierr = PetscGatherMessageLengths(comm,n_sends,n_recvs,ilengths_idxs_is,&onodes_is,&olengths_idxs_is);CHKERRQ(ierr); 496728143c3dSStefano Zampini } 496828143c3dSStefano Zampini 4969e7931f94SStefano Zampini buf_size_idxs = 0; 4970e7931f94SStefano Zampini buf_size_vals = 0; 497128143c3dSStefano Zampini buf_size_idxs_is = 0; 4972e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 4973e7931f94SStefano Zampini buf_size_idxs += (PetscInt)olengths_idxs[i]; 4974e7931f94SStefano Zampini buf_size_vals += (PetscInt)olengths_vals[i]; 497528143c3dSStefano Zampini if (nis) buf_size_idxs_is += (PetscInt)olengths_idxs_is[i]; 4976e7931f94SStefano Zampini } 4977785e854fSJed Brown ierr = PetscMalloc1(buf_size_idxs,&recv_buffer_idxs);CHKERRQ(ierr); 4978785e854fSJed Brown ierr = PetscMalloc1(buf_size_vals,&recv_buffer_vals);CHKERRQ(ierr); 497995ecbf38SStefano Zampini ierr = PetscMalloc1(buf_size_idxs_is,&recv_buffer_idxs_is);CHKERRQ(ierr); 4980e7931f94SStefano Zampini 4981e7931f94SStefano Zampini /* get new tags for clean communications */ 4982e7931f94SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_idxs);CHKERRQ(ierr); 4983e7931f94SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_vals);CHKERRQ(ierr); 498428143c3dSStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_idxs_is);CHKERRQ(ierr); 4985e7931f94SStefano Zampini 4986e7931f94SStefano Zampini /* allocate for requests */ 4987785e854fSJed Brown ierr = PetscMalloc1(n_sends,&send_req_idxs);CHKERRQ(ierr); 4988785e854fSJed Brown ierr = PetscMalloc1(n_sends,&send_req_vals);CHKERRQ(ierr); 498995ecbf38SStefano Zampini ierr = PetscMalloc1(n_sends,&send_req_idxs_is);CHKERRQ(ierr); 4990785e854fSJed Brown ierr = PetscMalloc1(n_recvs,&recv_req_idxs);CHKERRQ(ierr); 4991785e854fSJed Brown ierr = PetscMalloc1(n_recvs,&recv_req_vals);CHKERRQ(ierr); 499295ecbf38SStefano Zampini ierr = PetscMalloc1(n_recvs,&recv_req_idxs_is);CHKERRQ(ierr); 4993e7931f94SStefano Zampini 4994e7931f94SStefano Zampini /* communications */ 4995e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 4996e7931f94SStefano Zampini ptr_vals = recv_buffer_vals; 499728143c3dSStefano Zampini ptr_idxs_is = recv_buffer_idxs_is; 4998e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 4999e7931f94SStefano Zampini source_dest = onodes[i]; 5000e7931f94SStefano Zampini ierr = MPI_Irecv(ptr_idxs,olengths_idxs[i],MPIU_INT,source_dest,tag_idxs,comm,&recv_req_idxs[i]);CHKERRQ(ierr); 5001e7931f94SStefano Zampini ierr = MPI_Irecv(ptr_vals,olengths_vals[i],MPIU_SCALAR,source_dest,tag_vals,comm,&recv_req_vals[i]);CHKERRQ(ierr); 5002e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 5003e7931f94SStefano Zampini ptr_vals += olengths_vals[i]; 500428143c3dSStefano Zampini if (nis) { 500528143c3dSStefano 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); 500628143c3dSStefano Zampini ptr_idxs_is += olengths_idxs_is[i]; 500728143c3dSStefano Zampini } 5008e7931f94SStefano Zampini } 5009e7931f94SStefano Zampini for (i=0;i<n_sends;i++) { 5010e7931f94SStefano Zampini ierr = PetscMPIIntCast(is_indices[i],&source_dest);CHKERRQ(ierr); 5011e7931f94SStefano Zampini ierr = MPI_Isend(send_buffer_idxs,ilengths_idxs[source_dest],MPIU_INT,source_dest,tag_idxs,comm,&send_req_idxs[i]);CHKERRQ(ierr); 5012e7931f94SStefano Zampini ierr = MPI_Isend(send_buffer_vals,ilengths_vals[source_dest],MPIU_SCALAR,source_dest,tag_vals,comm,&send_req_vals[i]);CHKERRQ(ierr); 501328143c3dSStefano Zampini if (nis) { 501428143c3dSStefano 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); 501528143c3dSStefano Zampini } 5016e7931f94SStefano Zampini } 5017e7931f94SStefano Zampini ierr = ISRestoreIndices(is_sends_internal,&is_indices);CHKERRQ(ierr); 5018e7931f94SStefano Zampini ierr = ISDestroy(&is_sends_internal);CHKERRQ(ierr); 5019e7931f94SStefano Zampini 5020e7931f94SStefano Zampini /* assemble new l2g map */ 5021e7931f94SStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_idxs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 5022e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 50239d30be91SStefano Zampini new_local_rows = 0; 5024e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 50259d30be91SStefano Zampini new_local_rows += *(ptr_idxs+1); /* second element is the local size of the l2gmap */ 5026e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 5027e7931f94SStefano Zampini } 50289d30be91SStefano Zampini ierr = PetscMalloc1(new_local_rows,&l2gmap_indices);CHKERRQ(ierr); 5029e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 50309d30be91SStefano Zampini new_local_rows = 0; 5031e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 50329d30be91SStefano Zampini ierr = PetscMemcpy(&l2gmap_indices[new_local_rows],ptr_idxs+2,(*(ptr_idxs+1))*sizeof(PetscInt));CHKERRQ(ierr); 50339d30be91SStefano Zampini new_local_rows += *(ptr_idxs+1); /* second element is the local size of the l2gmap */ 5034e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 5035e7931f94SStefano Zampini } 50369d30be91SStefano Zampini ierr = PetscSortRemoveDupsInt(&new_local_rows,l2gmap_indices);CHKERRQ(ierr); 50379d30be91SStefano Zampini ierr = ISLocalToGlobalMappingCreate(comm_n,1,new_local_rows,l2gmap_indices,PETSC_COPY_VALUES,&l2gmap);CHKERRQ(ierr); 5038e7931f94SStefano Zampini ierr = PetscFree(l2gmap_indices);CHKERRQ(ierr); 5039e7931f94SStefano Zampini 5040e7931f94SStefano Zampini /* infer new local matrix type from received local matrices type */ 5041e7931f94SStefano Zampini /* currently if all local matrices are of type X, then the resulting matrix will be of type X, except for the dense case */ 5042e7931f94SStefano 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) */ 5043e7931f94SStefano Zampini if (n_recvs) { 504428143c3dSStefano Zampini MatTypePrivate new_local_type_private = (MatTypePrivate)send_buffer_idxs[0]; 5045e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 5046e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 5047e7931f94SStefano Zampini if ((PetscInt)new_local_type_private != *ptr_idxs) { 5048e7931f94SStefano Zampini new_local_type_private = MATAIJ_PRIVATE; 5049e7931f94SStefano Zampini break; 5050e7931f94SStefano Zampini } 5051e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 5052e7931f94SStefano Zampini } 5053e7931f94SStefano Zampini switch (new_local_type_private) { 505428143c3dSStefano Zampini case MATDENSE_PRIVATE: 505528143c3dSStefano Zampini if (n_recvs>1) { /* subassembling of dense matrices does not give a dense matrix! */ 5056e7931f94SStefano Zampini new_local_type = MATSEQAIJ; 5057e7931f94SStefano Zampini bs = 1; 505828143c3dSStefano Zampini } else { /* if I receive only 1 dense matrix */ 505928143c3dSStefano Zampini new_local_type = MATSEQDENSE; 506028143c3dSStefano Zampini bs = 1; 506128143c3dSStefano Zampini } 5062e7931f94SStefano Zampini break; 5063e7931f94SStefano Zampini case MATAIJ_PRIVATE: 5064e7931f94SStefano Zampini new_local_type = MATSEQAIJ; 5065e7931f94SStefano Zampini bs = 1; 5066e7931f94SStefano Zampini break; 5067e7931f94SStefano Zampini case MATBAIJ_PRIVATE: 5068e7931f94SStefano Zampini new_local_type = MATSEQBAIJ; 5069e7931f94SStefano Zampini break; 5070e7931f94SStefano Zampini case MATSBAIJ_PRIVATE: 5071e7931f94SStefano Zampini new_local_type = MATSEQSBAIJ; 5072e7931f94SStefano Zampini break; 5073e7931f94SStefano Zampini default: 50749d30be91SStefano Zampini SETERRQ2(comm,PETSC_ERR_SUP,"Unsupported private type %d in %s",new_local_type_private,__FUNCT__); 5075e7931f94SStefano Zampini break; 5076e7931f94SStefano Zampini } 507728143c3dSStefano Zampini } else { /* by default, new_local_type is seqdense */ 507828143c3dSStefano Zampini new_local_type = MATSEQDENSE; 507928143c3dSStefano Zampini bs = 1; 5080e7931f94SStefano Zampini } 5081e7931f94SStefano Zampini 508270cf5478SStefano Zampini /* create MATIS object if needed */ 508370cf5478SStefano Zampini if (reuse == MAT_INITIAL_MATRIX) { 5084e7931f94SStefano Zampini ierr = MatGetSize(mat,&rows,&cols);CHKERRQ(ierr); 5085e176bc59SStefano Zampini ierr = MatCreateIS(comm_n,bs,PETSC_DECIDE,PETSC_DECIDE,rows,cols,l2gmap,NULL,mat_n);CHKERRQ(ierr); 508670cf5478SStefano Zampini } else { 508770cf5478SStefano Zampini /* it also destroys the local matrices */ 508870cf5478SStefano Zampini ierr = MatSetLocalToGlobalMapping(*mat_n,l2gmap,l2gmap);CHKERRQ(ierr); 508970cf5478SStefano Zampini } 509070cf5478SStefano Zampini ierr = MatISGetLocalMat(*mat_n,&local_mat);CHKERRQ(ierr); 5091e7931f94SStefano Zampini ierr = MatSetType(local_mat,new_local_type);CHKERRQ(ierr); 50929d30be91SStefano Zampini 50939d30be91SStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_vals,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 50949d30be91SStefano Zampini 50959d30be91SStefano Zampini /* Global to local map of received indices */ 50969d30be91SStefano Zampini ierr = PetscMalloc1(buf_size_idxs,&recv_buffer_idxs_local);CHKERRQ(ierr); /* needed for values insertion */ 50979d30be91SStefano Zampini ierr = ISGlobalToLocalMappingApply(l2gmap,IS_GTOLM_MASK,buf_size_idxs,recv_buffer_idxs,&i,recv_buffer_idxs_local);CHKERRQ(ierr); 50989d30be91SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&l2gmap);CHKERRQ(ierr); 50999d30be91SStefano Zampini 51009d30be91SStefano Zampini /* restore attributes -> type of incoming data and its size */ 51019d30be91SStefano Zampini buf_size_idxs = 0; 51029d30be91SStefano Zampini for (i=0;i<n_recvs;i++) { 51039d30be91SStefano Zampini recv_buffer_idxs_local[buf_size_idxs] = recv_buffer_idxs[buf_size_idxs]; 51049d30be91SStefano Zampini recv_buffer_idxs_local[buf_size_idxs+1] = recv_buffer_idxs[buf_size_idxs+1]; 51059d30be91SStefano Zampini buf_size_idxs += (PetscInt)olengths_idxs[i]; 51069d30be91SStefano Zampini } 51079d30be91SStefano Zampini ierr = PetscFree(recv_buffer_idxs);CHKERRQ(ierr); 51089d30be91SStefano Zampini 51099d30be91SStefano Zampini /* set preallocation */ 51109d30be91SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)local_mat,MATSEQDENSE,&newisdense);CHKERRQ(ierr); 51119d30be91SStefano Zampini if (!newisdense) { 51129d30be91SStefano Zampini PetscInt *new_local_nnz=0; 51139d30be91SStefano Zampini 51149d30be91SStefano Zampini ptr_vals = recv_buffer_vals; 51159d30be91SStefano Zampini ptr_idxs = recv_buffer_idxs_local; 51169d30be91SStefano Zampini if (n_recvs) { 51179d30be91SStefano Zampini ierr = PetscCalloc1(new_local_rows,&new_local_nnz);CHKERRQ(ierr); 51189d30be91SStefano Zampini } 51199d30be91SStefano Zampini for (i=0;i<n_recvs;i++) { 51209d30be91SStefano Zampini PetscInt j; 51219d30be91SStefano Zampini if (*ptr_idxs == (PetscInt)MATDENSE_PRIVATE) { /* preallocation provided for dense case only */ 51229d30be91SStefano Zampini for (j=0;j<*(ptr_idxs+1);j++) { 51239d30be91SStefano Zampini new_local_nnz[*(ptr_idxs+2+j)] += *(ptr_idxs+1); 51249d30be91SStefano Zampini } 51259d30be91SStefano Zampini } else { 51269d30be91SStefano Zampini /* TODO */ 51279d30be91SStefano Zampini } 51289d30be91SStefano Zampini ptr_idxs += olengths_idxs[i]; 51299d30be91SStefano Zampini } 51309d30be91SStefano Zampini if (new_local_nnz) { 51319d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] = PetscMin(new_local_nnz[i],new_local_rows); 51329d30be91SStefano Zampini ierr = MatSeqAIJSetPreallocation(local_mat,0,new_local_nnz);CHKERRQ(ierr); 51339d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] /= bs; 51349d30be91SStefano Zampini ierr = MatSeqBAIJSetPreallocation(local_mat,bs,0,new_local_nnz);CHKERRQ(ierr); 51359d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] = PetscMax(new_local_nnz[i]-i,0); 51369d30be91SStefano Zampini ierr = MatSeqSBAIJSetPreallocation(local_mat,bs,0,new_local_nnz);CHKERRQ(ierr); 51379d30be91SStefano Zampini } else { 51389d30be91SStefano Zampini ierr = MatSetUp(local_mat);CHKERRQ(ierr); 51399d30be91SStefano Zampini } 51409d30be91SStefano Zampini ierr = PetscFree(new_local_nnz);CHKERRQ(ierr); 51419d30be91SStefano Zampini } else { 51429d30be91SStefano Zampini ierr = MatSetUp(local_mat);CHKERRQ(ierr); 51439d30be91SStefano Zampini } 5144e7931f94SStefano Zampini 5145e7931f94SStefano Zampini /* set values */ 5146e7931f94SStefano Zampini ptr_vals = recv_buffer_vals; 51479d30be91SStefano Zampini ptr_idxs = recv_buffer_idxs_local; 5148e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 5149e7931f94SStefano Zampini if (*ptr_idxs == (PetscInt)MATDENSE_PRIVATE) { /* values insertion provided for dense case only */ 5150e7931f94SStefano Zampini ierr = MatSetOption(local_mat,MAT_ROW_ORIENTED,PETSC_FALSE);CHKERRQ(ierr); 51519d30be91SStefano Zampini ierr = MatSetValues(local_mat,*(ptr_idxs+1),ptr_idxs+2,*(ptr_idxs+1),ptr_idxs+2,ptr_vals,ADD_VALUES);CHKERRQ(ierr); 5152e7931f94SStefano Zampini ierr = MatAssemblyBegin(local_mat,MAT_FLUSH_ASSEMBLY);CHKERRQ(ierr); 5153e7931f94SStefano Zampini ierr = MatAssemblyEnd(local_mat,MAT_FLUSH_ASSEMBLY);CHKERRQ(ierr); 5154e7931f94SStefano Zampini ierr = MatSetOption(local_mat,MAT_ROW_ORIENTED,PETSC_TRUE);CHKERRQ(ierr); 515528143c3dSStefano Zampini } else { 515628143c3dSStefano Zampini /* TODO */ 5157e7931f94SStefano Zampini } 5158e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 5159e7931f94SStefano Zampini ptr_vals += olengths_vals[i]; 5160e7931f94SStefano Zampini } 5161e7931f94SStefano Zampini ierr = MatAssemblyBegin(local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5162e7931f94SStefano Zampini ierr = MatAssemblyEnd(local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 516370cf5478SStefano Zampini ierr = MatAssemblyBegin(*mat_n,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 516470cf5478SStefano Zampini ierr = MatAssemblyEnd(*mat_n,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 51659d30be91SStefano Zampini ierr = PetscFree(recv_buffer_vals);CHKERRQ(ierr); 51669d30be91SStefano Zampini ierr = PetscFree(recv_buffer_idxs_local);CHKERRQ(ierr); 5167e7931f94SStefano Zampini 5168dfd14d43SStefano Zampini #if 0 516928143c3dSStefano Zampini if (!restrict_comm) { /* check */ 5170e7931f94SStefano Zampini Vec lvec,rvec; 5171e7931f94SStefano Zampini PetscReal infty_error; 5172e7931f94SStefano Zampini 51732a7a6963SBarry Smith ierr = MatCreateVecs(mat,&rvec,&lvec);CHKERRQ(ierr); 5174e7931f94SStefano Zampini ierr = VecSetRandom(rvec,NULL);CHKERRQ(ierr); 5175e7931f94SStefano Zampini ierr = MatMult(mat,rvec,lvec);CHKERRQ(ierr); 5176e7931f94SStefano Zampini ierr = VecScale(lvec,-1.0);CHKERRQ(ierr); 517770cf5478SStefano Zampini ierr = MatMultAdd(*mat_n,rvec,lvec,lvec);CHKERRQ(ierr); 5178e7931f94SStefano Zampini ierr = VecNorm(lvec,NORM_INFINITY,&infty_error);CHKERRQ(ierr); 5179e7931f94SStefano Zampini ierr = PetscPrintf(PetscObjectComm((PetscObject)mat),"Infinity error subassembling %1.6e\n",infty_error); 5180e7931f94SStefano Zampini ierr = VecDestroy(&rvec);CHKERRQ(ierr); 5181e7931f94SStefano Zampini ierr = VecDestroy(&lvec);CHKERRQ(ierr); 5182e7931f94SStefano Zampini } 518328143c3dSStefano Zampini #endif 5184e7931f94SStefano Zampini 518528143c3dSStefano Zampini /* assemble new additional is (if any) */ 518628143c3dSStefano Zampini if (nis) { 518728143c3dSStefano Zampini PetscInt **temp_idxs,*count_is,j,psum; 518828143c3dSStefano Zampini 518928143c3dSStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_idxs_is,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 5190854ce69bSBarry Smith ierr = PetscCalloc1(nis,&count_is);CHKERRQ(ierr); 519128143c3dSStefano Zampini ptr_idxs = recv_buffer_idxs_is; 519228143c3dSStefano Zampini psum = 0; 519328143c3dSStefano Zampini for (i=0;i<n_recvs;i++) { 519428143c3dSStefano Zampini for (j=0;j<nis;j++) { 519528143c3dSStefano Zampini PetscInt plen = *(ptr_idxs); /* first element is the local size of IS's indices */ 519628143c3dSStefano Zampini count_is[j] += plen; /* increment counting of buffer for j-th IS */ 519728143c3dSStefano Zampini psum += plen; 519828143c3dSStefano Zampini ptr_idxs += plen+1; /* shift pointer to received data */ 519928143c3dSStefano Zampini } 520028143c3dSStefano Zampini } 5201854ce69bSBarry Smith ierr = PetscMalloc1(nis,&temp_idxs);CHKERRQ(ierr); 5202854ce69bSBarry Smith ierr = PetscMalloc1(psum,&temp_idxs[0]);CHKERRQ(ierr); 520328143c3dSStefano Zampini for (i=1;i<nis;i++) { 520428143c3dSStefano Zampini temp_idxs[i] = temp_idxs[i-1]+count_is[i-1]; 520528143c3dSStefano Zampini } 520628143c3dSStefano Zampini ierr = PetscMemzero(count_is,nis*sizeof(PetscInt));CHKERRQ(ierr); 520728143c3dSStefano Zampini ptr_idxs = recv_buffer_idxs_is; 520828143c3dSStefano Zampini for (i=0;i<n_recvs;i++) { 520928143c3dSStefano Zampini for (j=0;j<nis;j++) { 521028143c3dSStefano Zampini PetscInt plen = *(ptr_idxs); /* first element is the local size of IS's indices */ 521128143c3dSStefano Zampini ierr = PetscMemcpy(&temp_idxs[j][count_is[j]],ptr_idxs+1,plen*sizeof(PetscInt));CHKERRQ(ierr); 521228143c3dSStefano Zampini count_is[j] += plen; /* increment starting point of buffer for j-th IS */ 521328143c3dSStefano Zampini ptr_idxs += plen+1; /* shift pointer to received data */ 521428143c3dSStefano Zampini } 521528143c3dSStefano Zampini } 521628143c3dSStefano Zampini for (i=0;i<nis;i++) { 521728143c3dSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 521828143c3dSStefano Zampini ierr = PetscSortRemoveDupsInt(&count_is[i],temp_idxs[i]);CHKERRQ(ierr);CHKERRQ(ierr); 521928143c3dSStefano Zampini ierr = ISCreateGeneral(comm_n,count_is[i],temp_idxs[i],PETSC_COPY_VALUES,&isarray[i]);CHKERRQ(ierr); 522028143c3dSStefano Zampini } 522128143c3dSStefano Zampini ierr = PetscFree(count_is);CHKERRQ(ierr); 522228143c3dSStefano Zampini ierr = PetscFree(temp_idxs[0]);CHKERRQ(ierr); 522328143c3dSStefano Zampini ierr = PetscFree(temp_idxs);CHKERRQ(ierr); 522428143c3dSStefano Zampini } 5225e7931f94SStefano Zampini /* free workspace */ 522628143c3dSStefano Zampini ierr = PetscFree(recv_buffer_idxs_is);CHKERRQ(ierr); 5227e7931f94SStefano Zampini ierr = MPI_Waitall(n_sends,send_req_idxs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 5228e7931f94SStefano Zampini ierr = PetscFree(send_buffer_idxs);CHKERRQ(ierr); 5229e7931f94SStefano Zampini ierr = MPI_Waitall(n_sends,send_req_vals,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 5230e7931f94SStefano Zampini if (isdense) { 5231e7931f94SStefano Zampini ierr = MatISGetLocalMat(mat,&local_mat);CHKERRQ(ierr); 5232e7931f94SStefano Zampini ierr = MatDenseRestoreArray(local_mat,&send_buffer_vals);CHKERRQ(ierr); 5233e7931f94SStefano Zampini } else { 5234e7931f94SStefano Zampini /* ierr = PetscFree(send_buffer_vals);CHKERRQ(ierr); */ 5235e7931f94SStefano Zampini } 523628143c3dSStefano Zampini if (nis) { 523728143c3dSStefano Zampini ierr = MPI_Waitall(n_sends,send_req_idxs_is,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 523828143c3dSStefano Zampini ierr = PetscFree(send_buffer_idxs_is);CHKERRQ(ierr); 523928143c3dSStefano Zampini } 5240e7931f94SStefano Zampini ierr = PetscFree(recv_req_idxs);CHKERRQ(ierr); 5241e7931f94SStefano Zampini ierr = PetscFree(recv_req_vals);CHKERRQ(ierr); 524228143c3dSStefano Zampini ierr = PetscFree(recv_req_idxs_is);CHKERRQ(ierr); 5243e7931f94SStefano Zampini ierr = PetscFree(send_req_idxs);CHKERRQ(ierr); 5244e7931f94SStefano Zampini ierr = PetscFree(send_req_vals);CHKERRQ(ierr); 524528143c3dSStefano Zampini ierr = PetscFree(send_req_idxs_is);CHKERRQ(ierr); 5246e7931f94SStefano Zampini ierr = PetscFree(ilengths_vals);CHKERRQ(ierr); 5247e7931f94SStefano Zampini ierr = PetscFree(ilengths_idxs);CHKERRQ(ierr); 5248e7931f94SStefano Zampini ierr = PetscFree(olengths_vals);CHKERRQ(ierr); 5249e7931f94SStefano Zampini ierr = PetscFree(olengths_idxs);CHKERRQ(ierr); 5250e7931f94SStefano Zampini ierr = PetscFree(onodes);CHKERRQ(ierr); 525128143c3dSStefano Zampini if (nis) { 525228143c3dSStefano Zampini ierr = PetscFree(ilengths_idxs_is);CHKERRQ(ierr); 525328143c3dSStefano Zampini ierr = PetscFree(olengths_idxs_is);CHKERRQ(ierr); 525428143c3dSStefano Zampini ierr = PetscFree(onodes_is);CHKERRQ(ierr); 525528143c3dSStefano Zampini } 525628143c3dSStefano Zampini ierr = PetscSubcommDestroy(&subcomm);CHKERRQ(ierr); 525728143c3dSStefano Zampini if (destroy_mat) { /* destroy mat is true only if restrict comm is true and process will not partecipate */ 525828143c3dSStefano Zampini ierr = MatDestroy(mat_n);CHKERRQ(ierr); 525928143c3dSStefano Zampini for (i=0;i<nis;i++) { 526028143c3dSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 526128143c3dSStefano Zampini } 526253a05cb3SStefano Zampini *mat_n = NULL; 526328143c3dSStefano Zampini } 5264e7931f94SStefano Zampini PetscFunctionReturn(0); 5265e7931f94SStefano Zampini } 5266a57a6d2fSStefano Zampini 526712edc857SStefano Zampini /* temporary hack into ksp private data structure */ 5268af0996ceSBarry Smith #include <petsc/private/kspimpl.h> 526912edc857SStefano Zampini 5270c8587f34SStefano Zampini #undef __FUNCT__ 5271c8587f34SStefano Zampini #define __FUNCT__ "PCBDDCSetUpCoarseSolver" 5272c8587f34SStefano Zampini PetscErrorCode PCBDDCSetUpCoarseSolver(PC pc,PetscScalar* coarse_submat_vals) 5273c8587f34SStefano Zampini { 5274c8587f34SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 5275c8587f34SStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 527620a2ab83SStefano Zampini Mat coarse_mat,coarse_mat_is,coarse_submat_dense; 52779881197aSStefano Zampini MatNullSpace CoarseNullSpace=NULL; 527820a2ab83SStefano Zampini ISLocalToGlobalMapping coarse_islg; 52796e683305SStefano Zampini IS coarse_is,*isarray; 52806e683305SStefano Zampini PetscInt i,im_active=-1,active_procs=-1; 528130368db7SStefano Zampini PetscInt nis,nisdofs,nisneu,nisvert; 5282f9eb5b7dSStefano Zampini PC pc_temp; 5283c8587f34SStefano Zampini PCType coarse_pc_type; 5284c8587f34SStefano Zampini KSPType coarse_ksp_type; 5285f9eb5b7dSStefano Zampini PetscBool multilevel_requested,multilevel_allowed; 52864f3a063dSStefano Zampini PetscBool isredundant,isbddc,isnn,coarse_reuse; 52876e683305SStefano Zampini Mat t_coarse_mat_is; 52886e683305SStefano Zampini PetscInt void_procs,ncoarse_ml,ncoarse_ds,ncoarse; 52896e683305SStefano Zampini PetscMPIInt all_procs; 529074e2c79eSStefano Zampini PetscBool csin_ml,csin_ds,csin,csin_type_simple,redist; 529168457ee5SStefano Zampini PetscBool compute_vecs = PETSC_FALSE; 529222bc73bbSStefano Zampini PetscScalar *array; 52939881197aSStefano Zampini PetscErrorCode ierr; 5294fdc09c96SStefano Zampini 5295c8587f34SStefano Zampini PetscFunctionBegin; 5296c8587f34SStefano Zampini /* Assign global numbering to coarse dofs */ 529768457ee5SStefano 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 */ 5298fa7f1dd8SStefano Zampini PetscInt ocoarse_size; 52995a75c04eSSatish Balay compute_vecs = PETSC_TRUE; 5300fa7f1dd8SStefano Zampini ocoarse_size = pcbddc->coarse_size; 5301f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->global_primal_indices);CHKERRQ(ierr); 5302f4ddd8eeSStefano Zampini ierr = PCBDDCComputePrimalNumbering(pc,&pcbddc->coarse_size,&pcbddc->global_primal_indices);CHKERRQ(ierr); 5303f4ddd8eeSStefano Zampini /* see if we can avoid some work */ 5304fa7f1dd8SStefano Zampini if (pcbddc->coarse_ksp) { /* coarse ksp has already been created */ 530551bea450SStefano Zampini /* if the coarse size is different or we are using adaptive selection, better to not reuse the coarse matrix */ 530651bea450SStefano Zampini if (ocoarse_size != pcbddc->coarse_size || pcbddc->adaptive_selection) { 5307dc4bcba2SStefano Zampini PC pc; 5308dc4bcba2SStefano Zampini PetscBool isbddc; 5309dc4bcba2SStefano Zampini 5310dc4bcba2SStefano Zampini /* temporary workaround since PCBDDC does not have a reset method so far */ 5311dc4bcba2SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&pc);CHKERRQ(ierr); 5312dc4bcba2SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc,PCBDDC,&isbddc);CHKERRQ(ierr); 5313dc4bcba2SStefano Zampini if (isbddc) { 531463c961adSStefano Zampini ierr = KSPDestroy(&pcbddc->coarse_ksp);CHKERRQ(ierr); 531563c961adSStefano Zampini } else { 5316727cdba6SStefano Zampini ierr = KSPReset(pcbddc->coarse_ksp);CHKERRQ(ierr); 531763c961adSStefano Zampini } 5318fa7f1dd8SStefano Zampini coarse_reuse = PETSC_FALSE; 5319fa7f1dd8SStefano Zampini } else { /* we can safely reuse already computed coarse matrix */ 5320fa7f1dd8SStefano Zampini coarse_reuse = PETSC_TRUE; 5321f4ddd8eeSStefano Zampini } 5322fa7f1dd8SStefano Zampini } else { /* there's no coarse ksp, so we need to create the coarse matrix too */ 5323fa7f1dd8SStefano Zampini coarse_reuse = PETSC_FALSE; 5324f4ddd8eeSStefano Zampini } 532570cf5478SStefano Zampini /* reset any subassembling information */ 532670cf5478SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling);CHKERRQ(ierr); 53276e683305SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 53286e683305SStefano Zampini } else { /* primal space is unchanged, so we can reuse coarse matrix */ 5329fa7f1dd8SStefano Zampini coarse_reuse = PETSC_TRUE; 5330f4ddd8eeSStefano Zampini } 5331c8587f34SStefano Zampini 53326e683305SStefano Zampini /* count "active" (i.e. with positive local size) and "void" processes */ 53332b510759SStefano Zampini im_active = !!(pcis->n); 53342b510759SStefano Zampini ierr = MPI_Allreduce(&im_active,&active_procs,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 53356e683305SStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)pc),&all_procs);CHKERRQ(ierr); 53366e683305SStefano Zampini void_procs = all_procs-active_procs; 53376e683305SStefano Zampini csin_type_simple = PETSC_TRUE; 533874e2c79eSStefano Zampini redist = PETSC_FALSE; 533922bc73bbSStefano Zampini if (pcbddc->current_level && void_procs) { 53406e683305SStefano Zampini csin_ml = PETSC_TRUE; 53416e683305SStefano Zampini ncoarse_ml = void_procs; 5342779c1cceSStefano Zampini /* it has no sense to redistribute on a set of processors larger than the number of active processes */ 5343779c1cceSStefano Zampini if (pcbddc->redistribute_coarse > 0 && pcbddc->redistribute_coarse < active_procs) { 53446e683305SStefano Zampini csin_ds = PETSC_TRUE; 534518a45a71SStefano Zampini ncoarse_ds = pcbddc->redistribute_coarse; 534618a45a71SStefano Zampini redist = PETSC_TRUE; 534718a45a71SStefano Zampini } else { 53486e683305SStefano Zampini csin_ds = PETSC_TRUE; 5349779c1cceSStefano Zampini ncoarse_ds = active_procs; 5350779c1cceSStefano Zampini redist = PETSC_TRUE; 535118a45a71SStefano Zampini } 53526e683305SStefano Zampini } else { 53536e683305SStefano Zampini csin_ml = PETSC_FALSE; 53546e683305SStefano Zampini ncoarse_ml = all_procs; 53556e683305SStefano Zampini if (void_procs) { 53566e683305SStefano Zampini csin_ds = PETSC_TRUE; 53576e683305SStefano Zampini ncoarse_ds = void_procs; 53586e683305SStefano Zampini csin_type_simple = PETSC_FALSE; 53596e683305SStefano Zampini } else { 5360779c1cceSStefano Zampini if (pcbddc->redistribute_coarse > 0 && pcbddc->redistribute_coarse < all_procs) { 536174e2c79eSStefano Zampini csin_ds = PETSC_TRUE; 536274e2c79eSStefano Zampini ncoarse_ds = pcbddc->redistribute_coarse; 536374e2c79eSStefano Zampini redist = PETSC_TRUE; 536474e2c79eSStefano Zampini } else { 53656e683305SStefano Zampini csin_ds = PETSC_FALSE; 53666e683305SStefano Zampini ncoarse_ds = all_procs; 53676e683305SStefano Zampini } 53686e683305SStefano Zampini } 536974e2c79eSStefano Zampini } 53706e683305SStefano Zampini 53716e683305SStefano Zampini /* 53726e683305SStefano Zampini test if we can go multilevel: three conditions must be satisfied: 53736e683305SStefano Zampini - we have not exceeded the number of levels requested 53746e683305SStefano Zampini - we can actually subassemble the active processes 53756e683305SStefano Zampini - we can find a suitable number of MPI processes where we can place the subassembled problem 53766e683305SStefano Zampini */ 53776e683305SStefano Zampini multilevel_allowed = PETSC_FALSE; 53786e683305SStefano Zampini multilevel_requested = PETSC_FALSE; 53796e683305SStefano Zampini if (pcbddc->current_level < pcbddc->max_levels) { 53806e683305SStefano Zampini multilevel_requested = PETSC_TRUE; 53816e683305SStefano Zampini if (active_procs/pcbddc->coarsening_ratio < 2 || ncoarse_ml/pcbddc->coarsening_ratio < 2) { 5382f9eb5b7dSStefano Zampini multilevel_allowed = PETSC_FALSE; 53832b510759SStefano Zampini } else { 5384f9eb5b7dSStefano Zampini multilevel_allowed = PETSC_TRUE; 5385c8587f34SStefano Zampini } 5386c8587f34SStefano Zampini } 53876e683305SStefano Zampini /* determine number of process partecipating to coarse solver */ 53886e683305SStefano Zampini if (multilevel_allowed) { 53896e683305SStefano Zampini ncoarse = ncoarse_ml; 53906e683305SStefano Zampini csin = csin_ml; 539158da7f69SStefano Zampini redist = PETSC_FALSE; 53926e683305SStefano Zampini } else { 53936e683305SStefano Zampini ncoarse = ncoarse_ds; 53946e683305SStefano Zampini csin = csin_ds; 53956e683305SStefano Zampini } 5396e7931f94SStefano Zampini 5397abbbba34SStefano Zampini /* creates temporary l2gmap and IS for coarse indexes */ 5398abbbba34SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),pcbddc->local_primal_size,pcbddc->global_primal_indices,PETSC_COPY_VALUES,&coarse_is);CHKERRQ(ierr); 5399abbbba34SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(coarse_is,&coarse_islg);CHKERRQ(ierr); 5400abbbba34SStefano Zampini 5401abbbba34SStefano Zampini /* creates temporary MATIS object for coarse matrix */ 540222bc73bbSStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,pcbddc->local_primal_size,pcbddc->local_primal_size,NULL,&coarse_submat_dense);CHKERRQ(ierr); 540322bc73bbSStefano Zampini ierr = MatDenseGetArray(coarse_submat_dense,&array);CHKERRQ(ierr); 540422bc73bbSStefano Zampini ierr = PetscMemcpy(array,coarse_submat_vals,sizeof(*coarse_submat_vals)*pcbddc->local_primal_size*pcbddc->local_primal_size);CHKERRQ(ierr); 540522bc73bbSStefano Zampini ierr = MatDenseRestoreArray(coarse_submat_dense,&array);CHKERRQ(ierr); 5406e176bc59SStefano 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); 54076e683305SStefano Zampini ierr = MatISSetLocalMat(t_coarse_mat_is,coarse_submat_dense);CHKERRQ(ierr); 54086e683305SStefano Zampini ierr = MatAssemblyBegin(t_coarse_mat_is,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 54096e683305SStefano Zampini ierr = MatAssemblyEnd(t_coarse_mat_is,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5410abbbba34SStefano Zampini ierr = MatDestroy(&coarse_submat_dense);CHKERRQ(ierr); 5411abbbba34SStefano Zampini 54126e683305SStefano Zampini /* compute dofs splitting and neumann boundaries for coarse dofs */ 541330368db7SStefano Zampini if (multilevel_allowed && (pcbddc->n_ISForDofsLocal || pcbddc->NeumannBoundariesLocal || (pcbddc->benign_saddle_point && pcbddc->user_primal_vertices_local))) { /* protects from unneded computations */ 54146e683305SStefano Zampini PetscInt *tidxs,*tidxs2,nout,tsize,i; 54156e683305SStefano Zampini const PetscInt *idxs; 54166e683305SStefano Zampini ISLocalToGlobalMapping tmap; 54176e683305SStefano Zampini 54186e683305SStefano Zampini /* create map between primal indices (in local representative ordering) and local primal numbering */ 54190be93d54SStefano Zampini ierr = ISLocalToGlobalMappingCreate(PETSC_COMM_SELF,1,pcbddc->local_primal_size,pcbddc->primal_indices_local_idxs,PETSC_COPY_VALUES,&tmap);CHKERRQ(ierr); 54206e683305SStefano Zampini /* allocate space for temporary storage */ 5421854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->local_primal_size,&tidxs);CHKERRQ(ierr); 5422854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->local_primal_size,&tidxs2);CHKERRQ(ierr); 54236e683305SStefano Zampini /* allocate for IS array */ 54246e683305SStefano Zampini nisdofs = pcbddc->n_ISForDofsLocal; 54256e683305SStefano Zampini nisneu = !!pcbddc->NeumannBoundariesLocal; 542630368db7SStefano Zampini nisvert = 0; 542730368db7SStefano Zampini if (pcbddc->benign_saddle_point && pcbddc->user_primal_vertices_local) { 542830368db7SStefano Zampini nisvert = 1; 542930368db7SStefano Zampini } 543030368db7SStefano Zampini nis = nisdofs + nisneu + nisvert; 5431854ce69bSBarry Smith ierr = PetscMalloc1(nis,&isarray);CHKERRQ(ierr); 54326e683305SStefano Zampini /* dofs splitting */ 54336e683305SStefano Zampini for (i=0;i<nisdofs;i++) { 54346e683305SStefano Zampini /* ierr = ISView(pcbddc->ISForDofsLocal[i],0);CHKERRQ(ierr); */ 54356e683305SStefano Zampini ierr = ISGetLocalSize(pcbddc->ISForDofsLocal[i],&tsize);CHKERRQ(ierr); 54366e683305SStefano Zampini ierr = ISGetIndices(pcbddc->ISForDofsLocal[i],&idxs);CHKERRQ(ierr); 54376e683305SStefano Zampini ierr = ISGlobalToLocalMappingApply(tmap,IS_GTOLM_DROP,tsize,idxs,&nout,tidxs);CHKERRQ(ierr); 54386e683305SStefano Zampini ierr = ISRestoreIndices(pcbddc->ISForDofsLocal[i],&idxs);CHKERRQ(ierr); 54396e683305SStefano Zampini ierr = ISLocalToGlobalMappingApply(coarse_islg,nout,tidxs,tidxs2);CHKERRQ(ierr); 544030368db7SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),nout,tidxs2,PETSC_COPY_VALUES,&isarray[i]);CHKERRQ(ierr); 54416e683305SStefano Zampini /* ierr = ISView(isarray[i],0);CHKERRQ(ierr); */ 54426e683305SStefano Zampini } 54436e683305SStefano Zampini /* neumann boundaries */ 54446e683305SStefano Zampini if (pcbddc->NeumannBoundariesLocal) { 54456e683305SStefano Zampini /* ierr = ISView(pcbddc->NeumannBoundariesLocal,0);CHKERRQ(ierr); */ 54466e683305SStefano Zampini ierr = ISGetLocalSize(pcbddc->NeumannBoundariesLocal,&tsize);CHKERRQ(ierr); 54476e683305SStefano Zampini ierr = ISGetIndices(pcbddc->NeumannBoundariesLocal,&idxs);CHKERRQ(ierr); 54486e683305SStefano Zampini ierr = ISGlobalToLocalMappingApply(tmap,IS_GTOLM_DROP,tsize,idxs,&nout,tidxs);CHKERRQ(ierr); 54496e683305SStefano Zampini ierr = ISRestoreIndices(pcbddc->NeumannBoundariesLocal,&idxs);CHKERRQ(ierr); 54506e683305SStefano Zampini ierr = ISLocalToGlobalMappingApply(coarse_islg,nout,tidxs,tidxs2);CHKERRQ(ierr); 545130368db7SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),nout,tidxs2,PETSC_COPY_VALUES,&isarray[nisdofs]);CHKERRQ(ierr); 54526e683305SStefano Zampini /* ierr = ISView(isarray[nisdofs],0);CHKERRQ(ierr); */ 54536e683305SStefano Zampini } 545430368db7SStefano Zampini /* primal vertices (benign) */ 545530368db7SStefano Zampini if (pcbddc->benign_saddle_point && pcbddc->user_primal_vertices_local) { 545630368db7SStefano Zampini ierr = ISGetLocalSize(pcbddc->user_primal_vertices_local,&tsize);CHKERRQ(ierr); 545730368db7SStefano Zampini ierr = ISGetIndices(pcbddc->user_primal_vertices_local,&idxs);CHKERRQ(ierr); 545830368db7SStefano Zampini ierr = ISGlobalToLocalMappingApply(tmap,IS_GTOLM_DROP,tsize,idxs,&nout,tidxs);CHKERRQ(ierr); 545930368db7SStefano Zampini ierr = ISRestoreIndices(pcbddc->user_primal_vertices_local,&idxs);CHKERRQ(ierr); 546030368db7SStefano Zampini ierr = ISLocalToGlobalMappingApply(coarse_islg,nout,tidxs,tidxs2);CHKERRQ(ierr); 546130368db7SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),nout,tidxs2,PETSC_COPY_VALUES,&isarray[nis-1]);CHKERRQ(ierr); 546230368db7SStefano Zampini } 54636e683305SStefano Zampini /* free memory */ 54646e683305SStefano Zampini ierr = PetscFree(tidxs);CHKERRQ(ierr); 54656e683305SStefano Zampini ierr = PetscFree(tidxs2);CHKERRQ(ierr); 54666e683305SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&tmap);CHKERRQ(ierr); 54676e683305SStefano Zampini } else { 54686e683305SStefano Zampini nis = 0; 54696e683305SStefano Zampini nisdofs = 0; 54706e683305SStefano Zampini nisneu = 0; 547130368db7SStefano Zampini nisvert = 0; 54726e683305SStefano Zampini isarray = NULL; 54736e683305SStefano Zampini } 54746e683305SStefano Zampini /* destroy no longer needed map */ 54756e683305SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&coarse_islg);CHKERRQ(ierr); 54766e683305SStefano Zampini 54776e683305SStefano Zampini /* restrict on coarse candidates (if needed) */ 54786e683305SStefano Zampini coarse_mat_is = NULL; 54796e683305SStefano Zampini if (csin) { 54806e683305SStefano Zampini if (!pcbddc->coarse_subassembling_init ) { /* creates subassembling init pattern if not present */ 548174e2c79eSStefano Zampini if (redist) { 548274e2c79eSStefano Zampini PetscMPIInt rank; 5483779c1cceSStefano Zampini PetscInt spc,n_spc_p1,dest[1],destsize; 548474e2c79eSStefano Zampini 548574e2c79eSStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)pc),&rank);CHKERRQ(ierr); 548658da7f69SStefano Zampini spc = active_procs/ncoarse; 548758da7f69SStefano Zampini n_spc_p1 = active_procs%ncoarse; 5488779c1cceSStefano Zampini if (im_active) { 5489779c1cceSStefano Zampini destsize = 1; 549074e2c79eSStefano Zampini if (rank > n_spc_p1*(spc+1)-1) { 549174e2c79eSStefano Zampini dest[0] = n_spc_p1+(rank-(n_spc_p1*(spc+1)))/spc; 549274e2c79eSStefano Zampini } else { 549374e2c79eSStefano Zampini dest[0] = rank/(spc+1); 549474e2c79eSStefano Zampini } 549574e2c79eSStefano Zampini } else { 5496779c1cceSStefano Zampini destsize = 0; 54976e683305SStefano Zampini } 5498779c1cceSStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),destsize,dest,PETSC_COPY_VALUES,&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 5499779c1cceSStefano Zampini } else if (csin_type_simple) { 55006e683305SStefano Zampini PetscMPIInt rank; 55016e683305SStefano Zampini PetscInt issize,isidx; 5502779c1cceSStefano Zampini 55036e683305SStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)pc),&rank);CHKERRQ(ierr); 55046e683305SStefano Zampini if (im_active) { 55056e683305SStefano Zampini issize = 1; 55066e683305SStefano Zampini isidx = (PetscInt)rank; 55076e683305SStefano Zampini } else { 55086e683305SStefano Zampini issize = 0; 55096e683305SStefano Zampini isidx = -1; 55106e683305SStefano Zampini } 55116e683305SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),issize,&isidx,PETSC_COPY_VALUES,&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 5512779c1cceSStefano Zampini } else { /* get a suitable subassembling pattern from MATIS code */ 5513b0c7d250SStefano Zampini ierr = MatISGetSubassemblingPattern(t_coarse_mat_is,ncoarse,pcbddc->coarse_adj_red,&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 55146e683305SStefano Zampini } 5515779c1cceSStefano Zampini 5516779c1cceSStefano Zampini /* we need to shift on coarse candidates either if we are not redistributing or we are redistributing and we have enough void processes */ 5517779c1cceSStefano Zampini if (!redist || ncoarse <= void_procs) { 5518779c1cceSStefano Zampini PetscInt ncoarse_cand,tissize,*nisindices; 5519779c1cceSStefano Zampini PetscInt *coarse_candidates; 5520779c1cceSStefano Zampini const PetscInt* tisindices; 5521779c1cceSStefano Zampini 5522779c1cceSStefano Zampini /* get coarse candidates' ranks in pc communicator */ 5523779c1cceSStefano Zampini ierr = PetscMalloc1(all_procs,&coarse_candidates);CHKERRQ(ierr); 5524779c1cceSStefano Zampini ierr = MPI_Allgather(&im_active,1,MPIU_INT,coarse_candidates,1,MPIU_INT,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 5525779c1cceSStefano Zampini for (i=0,ncoarse_cand=0;i<all_procs;i++) { 5526779c1cceSStefano Zampini if (!coarse_candidates[i]) { 5527779c1cceSStefano Zampini coarse_candidates[ncoarse_cand++]=i; 5528779c1cceSStefano Zampini } 5529779c1cceSStefano Zampini } 5530779c1cceSStefano Zampini if (ncoarse_cand < ncoarse) SETERRQ2(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"This should not happen! %d < %d",ncoarse_cand,ncoarse); 5531779c1cceSStefano Zampini 5532779c1cceSStefano Zampini 55336e683305SStefano Zampini if (pcbddc->dbg_flag) { 55346e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 55356e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Subassembling pattern init (before shift)\n");CHKERRQ(ierr); 55366e683305SStefano Zampini ierr = ISView(pcbddc->coarse_subassembling_init,pcbddc->dbg_viewer);CHKERRQ(ierr); 55376e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Coarse candidates\n");CHKERRQ(ierr); 5538779c1cceSStefano Zampini for (i=0;i<ncoarse_cand;i++) { 55396e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"%d ",coarse_candidates[i]);CHKERRQ(ierr); 55406e683305SStefano Zampini } 55416e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"\n");CHKERRQ(ierr); 55426e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 55436e683305SStefano Zampini } 55446e683305SStefano Zampini /* shift the pattern on coarse candidates */ 55456e683305SStefano Zampini ierr = ISGetLocalSize(pcbddc->coarse_subassembling_init,&tissize);CHKERRQ(ierr); 55466e683305SStefano Zampini ierr = ISGetIndices(pcbddc->coarse_subassembling_init,&tisindices);CHKERRQ(ierr); 5547854ce69bSBarry Smith ierr = PetscMalloc1(tissize,&nisindices);CHKERRQ(ierr); 55486e683305SStefano Zampini for (i=0;i<tissize;i++) nisindices[i] = coarse_candidates[tisindices[i]]; 55496e683305SStefano Zampini ierr = ISRestoreIndices(pcbddc->coarse_subassembling_init,&tisindices);CHKERRQ(ierr); 55506e683305SStefano Zampini ierr = ISGeneralSetIndices(pcbddc->coarse_subassembling_init,tissize,nisindices,PETSC_OWN_POINTER);CHKERRQ(ierr); 55516e683305SStefano Zampini ierr = PetscFree(coarse_candidates);CHKERRQ(ierr); 55526e683305SStefano Zampini } 55536e683305SStefano Zampini if (pcbddc->dbg_flag) { 55546e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 55556e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Subassembling pattern init\n");CHKERRQ(ierr); 55566e683305SStefano Zampini ierr = ISView(pcbddc->coarse_subassembling_init,pcbddc->dbg_viewer);CHKERRQ(ierr); 55576e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 55586e683305SStefano Zampini } 5559779c1cceSStefano Zampini } 55606e683305SStefano Zampini /* get temporary coarse mat in IS format restricted on coarse procs (plus additional index sets of isarray) */ 556153a05cb3SStefano Zampini if (multilevel_allowed) { /* we need to keep tracking of void processes for future placements */ 556253a05cb3SStefano 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); 556353a05cb3SStefano Zampini } else { /* this is the last level, so use just receiving processes in subcomm */ 556453a05cb3SStefano 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); 556553a05cb3SStefano Zampini } 55666e683305SStefano Zampini } else { 55676e683305SStefano Zampini if (pcbddc->dbg_flag) { 55686e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 55696e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Subassembling pattern init not needed\n");CHKERRQ(ierr); 55706e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 55716e683305SStefano Zampini } 55726e683305SStefano Zampini ierr = PetscObjectReference((PetscObject)t_coarse_mat_is);CHKERRQ(ierr); 55736e683305SStefano Zampini coarse_mat_is = t_coarse_mat_is; 55746e683305SStefano Zampini } 55756e683305SStefano Zampini 55766e683305SStefano Zampini /* create local to global scatters for coarse problem */ 557768457ee5SStefano Zampini if (compute_vecs) { 55786e683305SStefano Zampini PetscInt lrows; 55796e683305SStefano Zampini ierr = VecDestroy(&pcbddc->coarse_vec);CHKERRQ(ierr); 55806e683305SStefano Zampini if (coarse_mat_is) { 55816e683305SStefano Zampini ierr = MatGetLocalSize(coarse_mat_is,&lrows,NULL);CHKERRQ(ierr); 55826e683305SStefano Zampini } else { 55836e683305SStefano Zampini lrows = 0; 55846e683305SStefano Zampini } 55856e683305SStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pc),&pcbddc->coarse_vec);CHKERRQ(ierr); 55866e683305SStefano Zampini ierr = VecSetSizes(pcbddc->coarse_vec,lrows,PETSC_DECIDE);CHKERRQ(ierr); 55876e683305SStefano Zampini ierr = VecSetType(pcbddc->coarse_vec,VECSTANDARD);CHKERRQ(ierr); 55886e683305SStefano Zampini ierr = VecScatterDestroy(&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 55896e683305SStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_P,NULL,pcbddc->coarse_vec,coarse_is,&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 55906e683305SStefano Zampini } 55916e683305SStefano Zampini ierr = ISDestroy(&coarse_is);CHKERRQ(ierr); 55926e683305SStefano Zampini ierr = MatDestroy(&t_coarse_mat_is);CHKERRQ(ierr); 5593c8587f34SStefano Zampini 5594f9eb5b7dSStefano Zampini /* set defaults for coarse KSP and PC */ 5595f9eb5b7dSStefano Zampini if (multilevel_allowed) { 5596f9eb5b7dSStefano Zampini coarse_ksp_type = KSPRICHARDSON; 5597f9eb5b7dSStefano Zampini coarse_pc_type = PCBDDC; 5598f9eb5b7dSStefano Zampini } else { 5599f9eb5b7dSStefano Zampini coarse_ksp_type = KSPPREONLY; 5600f9eb5b7dSStefano Zampini coarse_pc_type = PCREDUNDANT; 5601c8587f34SStefano Zampini } 5602c8587f34SStefano Zampini 56036e683305SStefano Zampini /* print some info if requested */ 56046e683305SStefano Zampini if (pcbddc->dbg_flag) { 56056e683305SStefano Zampini if (!multilevel_allowed) { 56066e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 56076e683305SStefano Zampini if (multilevel_requested) { 56086e683305SStefano 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); 56096e683305SStefano Zampini } else if (pcbddc->max_levels) { 56106e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Maximum number of requested levels reached (%d)\n",pcbddc->max_levels);CHKERRQ(ierr); 56116e683305SStefano Zampini } 56126e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 56136e683305SStefano Zampini } 56146e683305SStefano Zampini } 56156e683305SStefano Zampini 5616f9eb5b7dSStefano Zampini /* create the coarse KSP object only once with defaults */ 56176e683305SStefano Zampini if (coarse_mat_is) { 56186e683305SStefano Zampini MatReuse coarse_mat_reuse; 56196a1308c2SStefano Zampini PetscViewer dbg_viewer = NULL; 56206e683305SStefano Zampini if (pcbddc->dbg_flag) { 56216e683305SStefano Zampini dbg_viewer = PETSC_VIEWER_STDOUT_(PetscObjectComm((PetscObject)coarse_mat_is)); 56226e683305SStefano Zampini ierr = PetscViewerASCIIAddTab(dbg_viewer,2*pcbddc->current_level);CHKERRQ(ierr); 56236e683305SStefano Zampini } 5624f9eb5b7dSStefano Zampini if (!pcbddc->coarse_ksp) { 5625312be037SStefano Zampini char prefix[256],str_level[16]; 5626e604994aSStefano Zampini size_t len; 56276e683305SStefano Zampini ierr = KSPCreate(PetscObjectComm((PetscObject)coarse_mat_is),&pcbddc->coarse_ksp);CHKERRQ(ierr); 5628422a814eSBarry Smith ierr = KSPSetErrorIfNotConverged(pcbddc->coarse_ksp,pc->erroriffailure);CHKERRQ(ierr); 5629c8587f34SStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->coarse_ksp,(PetscObject)pc,1);CHKERRQ(ierr); 5630f9eb5b7dSStefano Zampini ierr = KSPSetTolerances(pcbddc->coarse_ksp,PETSC_DEFAULT,PETSC_DEFAULT,PETSC_DEFAULT,1);CHKERRQ(ierr); 56315f76c7aeSStefano Zampini ierr = KSPSetOperators(pcbddc->coarse_ksp,coarse_mat_is,coarse_mat_is);CHKERRQ(ierr); 5632c8587f34SStefano Zampini ierr = KSPSetType(pcbddc->coarse_ksp,coarse_ksp_type);CHKERRQ(ierr); 56336e683305SStefano Zampini ierr = KSPSetNormType(pcbddc->coarse_ksp,KSP_NORM_NONE);CHKERRQ(ierr); 5634c8587f34SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&pc_temp);CHKERRQ(ierr); 5635c8587f34SStefano Zampini ierr = PCSetType(pc_temp,coarse_pc_type);CHKERRQ(ierr); 5636e604994aSStefano Zampini /* prefix */ 5637e604994aSStefano Zampini ierr = PetscStrcpy(prefix,"");CHKERRQ(ierr); 5638e604994aSStefano Zampini ierr = PetscStrcpy(str_level,"");CHKERRQ(ierr); 5639e604994aSStefano Zampini if (!pcbddc->current_level) { 5640e604994aSStefano Zampini ierr = PetscStrcpy(prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 5641e604994aSStefano Zampini ierr = PetscStrcat(prefix,"pc_bddc_coarse_");CHKERRQ(ierr); 5642c8587f34SStefano Zampini } else { 5643e604994aSStefano Zampini ierr = PetscStrlen(((PetscObject)pc)->prefix,&len);CHKERRQ(ierr); 5644312be037SStefano Zampini if (pcbddc->current_level>1) len -= 3; /* remove "lX_" with X level number */ 5645312be037SStefano Zampini if (pcbddc->current_level>10) len -= 1; /* remove another char from level number */ 564634d6797cSStefano Zampini ierr = PetscStrncpy(prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 5647312be037SStefano Zampini sprintf(str_level,"l%d_",(int)(pcbddc->current_level)); 5648e604994aSStefano Zampini ierr = PetscStrcat(prefix,str_level);CHKERRQ(ierr); 5649e604994aSStefano Zampini } 5650e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->coarse_ksp,prefix);CHKERRQ(ierr); 56513e3c6dadSStefano Zampini /* propagate BDDC info to the next level (these are dummy calls if pc_temp is not of type PCBDDC) */ 56523e3c6dadSStefano Zampini ierr = PCBDDCSetLevel(pc_temp,pcbddc->current_level+1);CHKERRQ(ierr); 56533e3c6dadSStefano Zampini ierr = PCBDDCSetCoarseningRatio(pc_temp,pcbddc->coarsening_ratio);CHKERRQ(ierr); 56543e3c6dadSStefano Zampini ierr = PCBDDCSetLevels(pc_temp,pcbddc->max_levels);CHKERRQ(ierr); 5655f9eb5b7dSStefano Zampini /* allow user customization */ 5656f9eb5b7dSStefano Zampini ierr = KSPSetFromOptions(pcbddc->coarse_ksp);CHKERRQ(ierr); 56573e3c6dadSStefano Zampini } 56583e3c6dadSStefano Zampini /* propagate BDDC info to the next level (these are dummy calls if pc_temp is not of type PCBDDC) */ 565951bea450SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&pc_temp);CHKERRQ(ierr); 56603e3c6dadSStefano Zampini if (nisdofs) { 56613e3c6dadSStefano Zampini ierr = PCBDDCSetDofsSplitting(pc_temp,nisdofs,isarray);CHKERRQ(ierr); 56623e3c6dadSStefano Zampini for (i=0;i<nisdofs;i++) { 56633e3c6dadSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 56643e3c6dadSStefano Zampini } 56653e3c6dadSStefano Zampini } 56663e3c6dadSStefano Zampini if (nisneu) { 56673e3c6dadSStefano Zampini ierr = PCBDDCSetNeumannBoundaries(pc_temp,isarray[nisdofs]);CHKERRQ(ierr); 56683e3c6dadSStefano Zampini ierr = ISDestroy(&isarray[nisdofs]);CHKERRQ(ierr); 5669312be037SStefano Zampini } 567030368db7SStefano Zampini if (nisvert) { 567130368db7SStefano Zampini ierr = PCBDDCSetPrimalVerticesIS(pc_temp,isarray[nis-1]);CHKERRQ(ierr); 567230368db7SStefano Zampini ierr = ISDestroy(&isarray[nis-1]);CHKERRQ(ierr); 567330368db7SStefano Zampini } 5674f9eb5b7dSStefano Zampini 5675f9eb5b7dSStefano Zampini /* get some info after set from options */ 5676f9eb5b7dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCNN,&isnn);CHKERRQ(ierr); 5677f9eb5b7dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCBDDC,&isbddc);CHKERRQ(ierr); 56784f3a063dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCREDUNDANT,&isredundant);CHKERRQ(ierr); 56796e683305SStefano Zampini if (isbddc && !multilevel_allowed) { /* multilevel can only be requested via pc_bddc_set_levels */ 5680f9eb5b7dSStefano Zampini ierr = PCSetType(pc_temp,coarse_pc_type);CHKERRQ(ierr); 5681f9eb5b7dSStefano Zampini isbddc = PETSC_FALSE; 5682f9eb5b7dSStefano Zampini } 568339f0f02cSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 56844f3a063dSStefano Zampini if (isredundant) { 56854f3a063dSStefano Zampini KSP inner_ksp; 56864f3a063dSStefano Zampini PC inner_pc; 56874f3a063dSStefano Zampini ierr = PCRedundantGetKSP(pc_temp,&inner_ksp);CHKERRQ(ierr); 56884f3a063dSStefano Zampini ierr = KSPGetPC(inner_ksp,&inner_pc);CHKERRQ(ierr); 56894f3a063dSStefano Zampini ierr = PCFactorSetReuseFill(inner_pc,PETSC_TRUE);CHKERRQ(ierr); 56904f3a063dSStefano Zampini } 5691f9eb5b7dSStefano Zampini 5692f9eb5b7dSStefano Zampini /* assemble coarse matrix */ 5693fa7f1dd8SStefano Zampini if (coarse_reuse) { 569481d9aea3SBarry Smith ierr = KSPGetOperators(pcbddc->coarse_ksp,&coarse_mat,NULL);CHKERRQ(ierr); 5695fa7f1dd8SStefano Zampini ierr = PetscObjectReference((PetscObject)coarse_mat);CHKERRQ(ierr); 56966e683305SStefano Zampini coarse_mat_reuse = MAT_REUSE_MATRIX; 5697fa7f1dd8SStefano Zampini } else { 56986e683305SStefano Zampini coarse_mat_reuse = MAT_INITIAL_MATRIX; 5699fa7f1dd8SStefano Zampini } 5700c8587f34SStefano Zampini if (isbddc || isnn) { 5701d4d8cf7bSStefano Zampini if (isbddc) { /* currently there's no API for this */ 5702d4d8cf7bSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc_temp->data; 5703d4d8cf7bSStefano Zampini pcbddc->detect_disconnected = PETSC_TRUE; 5704d4d8cf7bSStefano Zampini } 570522bc73bbSStefano Zampini if (pcbddc->coarsening_ratio > 1) { 570670cf5478SStefano Zampini if (!pcbddc->coarse_subassembling) { /* subassembling info is not present */ 5707b0c7d250SStefano Zampini ierr = MatISGetSubassemblingPattern(coarse_mat_is,active_procs/pcbddc->coarsening_ratio,pcbddc->coarse_adj_red,&pcbddc->coarse_subassembling);CHKERRQ(ierr); 570822b6e8a2SStefano Zampini if (pcbddc->dbg_flag) { 57096e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 57106e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Subassembling pattern\n");CHKERRQ(ierr); 57116e683305SStefano Zampini ierr = ISView(pcbddc->coarse_subassembling,dbg_viewer);CHKERRQ(ierr); 57126e683305SStefano Zampini ierr = PetscViewerFlush(dbg_viewer);CHKERRQ(ierr); 571322b6e8a2SStefano Zampini } 571470cf5478SStefano Zampini } 571553a05cb3SStefano Zampini ierr = MatISSubassemble(coarse_mat_is,pcbddc->coarse_subassembling,0,PETSC_FALSE,PETSC_FALSE,coarse_mat_reuse,&coarse_mat,0,NULL);CHKERRQ(ierr); 571670cf5478SStefano Zampini } else { 571722bc73bbSStefano Zampini ierr = PetscObjectReference((PetscObject)coarse_mat_is);CHKERRQ(ierr); 571822bc73bbSStefano Zampini coarse_mat = coarse_mat_is; 571922bc73bbSStefano Zampini } 572022bc73bbSStefano Zampini } else { 57212e1e5fa4SStefano Zampini ierr = MatISGetMPIXAIJ(coarse_mat_is,coarse_mat_reuse,&coarse_mat);CHKERRQ(ierr); 5722c8587f34SStefano Zampini } 5723c8587f34SStefano Zampini ierr = MatDestroy(&coarse_mat_is);CHKERRQ(ierr); 5724c8587f34SStefano Zampini 57253301b35fSStefano Zampini /* propagate symmetry info of coarse matrix */ 57265a16e3a0SStefano Zampini ierr = MatSetOption(coarse_mat,MAT_STRUCTURALLY_SYMMETRIC,PETSC_TRUE);CHKERRQ(ierr); 57273301b35fSStefano Zampini if (pc->pmat->symmetric_set) { 57283301b35fSStefano Zampini ierr = MatSetOption(coarse_mat,MAT_SYMMETRIC,pc->pmat->symmetric);CHKERRQ(ierr); 57293301b35fSStefano Zampini } 57303301b35fSStefano Zampini if (pc->pmat->hermitian_set) { 57313301b35fSStefano Zampini ierr = MatSetOption(coarse_mat,MAT_HERMITIAN,pc->pmat->hermitian);CHKERRQ(ierr); 57323301b35fSStefano Zampini } 57333301b35fSStefano Zampini if (pc->pmat->spd_set) { 57343301b35fSStefano Zampini ierr = MatSetOption(coarse_mat,MAT_SPD,pc->pmat->spd);CHKERRQ(ierr); 57353301b35fSStefano Zampini } 57366e683305SStefano Zampini /* set operators */ 57375f76c7aeSStefano Zampini ierr = KSPSetOperators(pcbddc->coarse_ksp,coarse_mat,coarse_mat);CHKERRQ(ierr); 57386e683305SStefano Zampini if (pcbddc->dbg_flag) { 57396e683305SStefano Zampini ierr = PetscViewerASCIISubtractTab(dbg_viewer,2*pcbddc->current_level);CHKERRQ(ierr); 57406e683305SStefano Zampini } 57416e683305SStefano Zampini } else { /* processes non partecipating to coarse solver (if any) */ 57426e683305SStefano Zampini coarse_mat = 0; 57436e683305SStefano Zampini } 57446e683305SStefano Zampini ierr = PetscFree(isarray);CHKERRQ(ierr); 5745b1ecc7b1SStefano Zampini #if 0 5746b9b85e73SStefano Zampini { 5747b9b85e73SStefano Zampini PetscViewer viewer; 5748b9b85e73SStefano Zampini char filename[256]; 5749b1ecc7b1SStefano Zampini sprintf(filename,"coarse_mat_level%d.m",pcbddc->current_level); 5750b1ecc7b1SStefano Zampini ierr = PetscViewerASCIIOpen(PetscObjectComm((PetscObject)coarse_mat),filename,&viewer);CHKERRQ(ierr); 5751b9b85e73SStefano Zampini ierr = PetscViewerSetFormat(viewer,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 5752b9b85e73SStefano Zampini ierr = MatView(coarse_mat,viewer);CHKERRQ(ierr); 5753b9b85e73SStefano Zampini ierr = PetscViewerDestroy(&viewer);CHKERRQ(ierr); 5754b9b85e73SStefano Zampini } 5755b9b85e73SStefano Zampini #endif 5756c8587f34SStefano Zampini 5757c8587f34SStefano Zampini /* Compute coarse null space (special handling by BDDC only) */ 5758298c0119SStefano Zampini #if 0 5759c8587f34SStefano Zampini if (pcbddc->NullSpace) { 5760c8587f34SStefano Zampini ierr = PCBDDCNullSpaceAssembleCoarse(pc,coarse_mat,&CoarseNullSpace);CHKERRQ(ierr); 576198a51de6SStefano Zampini } 5762298c0119SStefano Zampini #endif 5763b0f5fe93SStefano Zampini /* hack */ 576498a51de6SStefano Zampini if (pcbddc->coarse_ksp) { 576598a51de6SStefano Zampini Vec crhs,csol; 576604708bb6SStefano Zampini 5767f347579bSStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&csol);CHKERRQ(ierr); 5768f347579bSStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&crhs);CHKERRQ(ierr); 5769f347579bSStefano Zampini if (!csol) { 57702a7a6963SBarry Smith ierr = MatCreateVecs(coarse_mat,&((pcbddc->coarse_ksp)->vec_sol),NULL);CHKERRQ(ierr); 5771f9eb5b7dSStefano Zampini } 5772f347579bSStefano Zampini if (!crhs) { 57732a7a6963SBarry Smith ierr = MatCreateVecs(coarse_mat,NULL,&((pcbddc->coarse_ksp)->vec_rhs));CHKERRQ(ierr); 5774f347579bSStefano Zampini } 5775b0f5fe93SStefano Zampini } 5776b0f5fe93SStefano Zampini 5777b0f5fe93SStefano Zampini /* compute null space for coarse solver if the benign trick has been requested */ 5778b0f5fe93SStefano Zampini if (pcbddc->benign_null) { 5779b0f5fe93SStefano Zampini 5780b0f5fe93SStefano Zampini ierr = VecSet(pcbddc->vec1_P,0.);CHKERRQ(ierr); 57814f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 57824f1b2e48SStefano Zampini ierr = VecSetValue(pcbddc->vec1_P,pcbddc->local_primal_size-pcbddc->benign_n+i,1.0,INSERT_VALUES);CHKERRQ(ierr); 57834f1b2e48SStefano Zampini } 5784b0f5fe93SStefano Zampini ierr = VecAssemblyBegin(pcbddc->vec1_P);CHKERRQ(ierr); 5785b0f5fe93SStefano Zampini ierr = VecAssemblyEnd(pcbddc->vec1_P);CHKERRQ(ierr); 5786b0f5fe93SStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 5787b0f5fe93SStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 5788b0f5fe93SStefano Zampini if (coarse_mat) { 5789b0f5fe93SStefano Zampini Vec nullv; 5790b0f5fe93SStefano Zampini PetscScalar *array,*array2; 5791b0f5fe93SStefano Zampini PetscInt nl; 5792b0f5fe93SStefano Zampini 5793b0f5fe93SStefano Zampini ierr = MatCreateVecs(coarse_mat,&nullv,NULL);CHKERRQ(ierr); 5794b0f5fe93SStefano Zampini ierr = VecGetLocalSize(nullv,&nl);CHKERRQ(ierr); 5795b0f5fe93SStefano Zampini ierr = VecGetArrayRead(pcbddc->coarse_vec,(const PetscScalar**)&array);CHKERRQ(ierr); 5796b0f5fe93SStefano Zampini ierr = VecGetArray(nullv,&array2);CHKERRQ(ierr); 5797b0f5fe93SStefano Zampini ierr = PetscMemcpy(array2,array,nl*sizeof(*array));CHKERRQ(ierr); 5798b0f5fe93SStefano Zampini ierr = VecRestoreArray(nullv,&array2);CHKERRQ(ierr); 5799b0f5fe93SStefano Zampini ierr = VecRestoreArrayRead(pcbddc->coarse_vec,(const PetscScalar**)&array);CHKERRQ(ierr); 5800b0f5fe93SStefano Zampini ierr = VecNormalize(nullv,NULL);CHKERRQ(ierr); 5801b0f5fe93SStefano Zampini ierr = MatNullSpaceCreate(PetscObjectComm((PetscObject)coarse_mat),PETSC_FALSE,1,&nullv,&CoarseNullSpace);CHKERRQ(ierr); 5802b0f5fe93SStefano Zampini ierr = VecDestroy(&nullv);CHKERRQ(ierr); 5803b0f5fe93SStefano Zampini } 5804b0f5fe93SStefano Zampini } 5805b0f5fe93SStefano Zampini 5806b0f5fe93SStefano Zampini if (pcbddc->coarse_ksp) { 5807b0f5fe93SStefano Zampini PetscBool ispreonly; 5808b0f5fe93SStefano Zampini 5809b0f5fe93SStefano Zampini if (CoarseNullSpace) { 5810b0f5fe93SStefano Zampini PetscBool isnull; 5811b0f5fe93SStefano Zampini ierr = MatNullSpaceTest(CoarseNullSpace,coarse_mat,&isnull);CHKERRQ(ierr); 5812d4d8cf7bSStefano Zampini if (0) { 581330368db7SStefano Zampini if (isbddc && !pcbddc->benign_saddle_point) { 5814b0f5fe93SStefano Zampini ierr = PCBDDCSetNullSpace(pc_temp,CoarseNullSpace);CHKERRQ(ierr); 5815b0f5fe93SStefano Zampini } else { 5816b0f5fe93SStefano Zampini ierr = MatSetNullSpace(coarse_mat,CoarseNullSpace);CHKERRQ(ierr); 5817b0f5fe93SStefano Zampini } 5818b0f5fe93SStefano Zampini } else { 5819b0f5fe93SStefano Zampini ierr = MatNullSpaceDestroy(&CoarseNullSpace);CHKERRQ(ierr); 5820b0f5fe93SStefano Zampini } 5821b0f5fe93SStefano Zampini } 5822b0f5fe93SStefano Zampini /* setup coarse ksp */ 5823b0f5fe93SStefano Zampini ierr = KSPSetUp(pcbddc->coarse_ksp);CHKERRQ(ierr); 5824cbcc2c2aSStefano Zampini /* Check coarse problem if in debug mode or if solving with an iterative method */ 5825cbcc2c2aSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcbddc->coarse_ksp,KSPPREONLY,&ispreonly);CHKERRQ(ierr); 58266e683305SStefano Zampini if (pcbddc->dbg_flag || (!ispreonly && pcbddc->use_coarse_estimates) ) { 5827c8587f34SStefano Zampini KSP check_ksp; 58282b510759SStefano Zampini KSPType check_ksp_type; 5829c8587f34SStefano Zampini PC check_pc; 58306e683305SStefano Zampini Vec check_vec,coarse_vec; 58316a1308c2SStefano Zampini PetscReal abs_infty_error,infty_error,lambda_min=1.0,lambda_max=1.0; 58322b510759SStefano Zampini PetscInt its; 58336e683305SStefano Zampini PetscBool compute_eigs; 58346e683305SStefano Zampini PetscReal *eigs_r,*eigs_c; 58356e683305SStefano Zampini PetscInt neigs; 58368e185a42SStefano Zampini const char *prefix; 5837c8587f34SStefano Zampini 58382b510759SStefano Zampini /* Create ksp object suitable for estimation of extreme eigenvalues */ 58396e683305SStefano Zampini ierr = KSPCreate(PetscObjectComm((PetscObject)pcbddc->coarse_ksp),&check_ksp);CHKERRQ(ierr); 5840422a814eSBarry Smith ierr = KSPSetErrorIfNotConverged(pcbddc->coarse_ksp,pc->erroriffailure);CHKERRQ(ierr); 584123ee1639SBarry Smith ierr = KSPSetOperators(check_ksp,coarse_mat,coarse_mat);CHKERRQ(ierr); 5842f4ddd8eeSStefano Zampini ierr = KSPSetTolerances(check_ksp,1.e-12,1.e-12,PETSC_DEFAULT,pcbddc->coarse_size);CHKERRQ(ierr); 58432b510759SStefano Zampini if (ispreonly) { 58442b510759SStefano Zampini check_ksp_type = KSPPREONLY; 58456e683305SStefano Zampini compute_eigs = PETSC_FALSE; 58462b510759SStefano Zampini } else { 5847cbcc2c2aSStefano Zampini check_ksp_type = KSPGMRES; 58486e683305SStefano Zampini compute_eigs = PETSC_TRUE; 5849c8587f34SStefano Zampini } 5850c8587f34SStefano Zampini ierr = KSPSetType(check_ksp,check_ksp_type);CHKERRQ(ierr); 58516e683305SStefano Zampini ierr = KSPSetComputeSingularValues(check_ksp,compute_eigs);CHKERRQ(ierr); 58526e683305SStefano Zampini ierr = KSPSetComputeEigenvalues(check_ksp,compute_eigs);CHKERRQ(ierr); 58536e683305SStefano Zampini ierr = KSPGMRESSetRestart(check_ksp,pcbddc->coarse_size+1);CHKERRQ(ierr); 5854a7dc3881SStefano Zampini ierr = KSPGetOptionsPrefix(pcbddc->coarse_ksp,&prefix);CHKERRQ(ierr); 5855a7dc3881SStefano Zampini ierr = KSPSetOptionsPrefix(check_ksp,prefix);CHKERRQ(ierr); 5856a7dc3881SStefano Zampini ierr = KSPAppendOptionsPrefix(check_ksp,"check_");CHKERRQ(ierr); 5857a7dc3881SStefano Zampini ierr = KSPSetFromOptions(check_ksp);CHKERRQ(ierr); 5858c8587f34SStefano Zampini ierr = KSPSetUp(check_ksp);CHKERRQ(ierr); 5859c8587f34SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&check_pc);CHKERRQ(ierr); 5860c8587f34SStefano Zampini ierr = KSPSetPC(check_ksp,check_pc);CHKERRQ(ierr); 5861c8587f34SStefano Zampini /* create random vec */ 58626e683305SStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&coarse_vec);CHKERRQ(ierr); 58636e683305SStefano Zampini ierr = VecDuplicate(coarse_vec,&check_vec);CHKERRQ(ierr); 5864c8587f34SStefano Zampini ierr = VecSetRandom(check_vec,NULL);CHKERRQ(ierr); 5865c8587f34SStefano Zampini if (CoarseNullSpace) { 5866c8587f34SStefano Zampini ierr = MatNullSpaceRemove(CoarseNullSpace,check_vec);CHKERRQ(ierr); 5867c8587f34SStefano Zampini } 58686e683305SStefano Zampini ierr = MatMult(coarse_mat,check_vec,coarse_vec);CHKERRQ(ierr); 5869c8587f34SStefano Zampini /* solve coarse problem */ 58706e683305SStefano Zampini ierr = KSPSolve(check_ksp,coarse_vec,coarse_vec);CHKERRQ(ierr); 5871c8587f34SStefano Zampini if (CoarseNullSpace) { 58726e683305SStefano Zampini ierr = MatNullSpaceRemove(CoarseNullSpace,coarse_vec);CHKERRQ(ierr); 5873c8587f34SStefano Zampini } 5874cbcc2c2aSStefano Zampini /* set eigenvalue estimation if preonly has not been requested */ 58756e683305SStefano Zampini if (compute_eigs) { 5876854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->coarse_size+1,&eigs_r);CHKERRQ(ierr); 5877854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->coarse_size+1,&eigs_c);CHKERRQ(ierr); 58786e683305SStefano Zampini ierr = KSPComputeEigenvalues(check_ksp,pcbddc->coarse_size+1,eigs_r,eigs_c,&neigs);CHKERRQ(ierr); 58796e683305SStefano Zampini lambda_max = eigs_r[neigs-1]; 58806e683305SStefano Zampini lambda_min = eigs_r[0]; 58816e683305SStefano Zampini if (pcbddc->use_coarse_estimates) { 58826e683305SStefano Zampini if (lambda_max>lambda_min) { 5883cbcc2c2aSStefano Zampini ierr = KSPChebyshevSetEigenvalues(pcbddc->coarse_ksp,lambda_max,lambda_min);CHKERRQ(ierr); 5884cbcc2c2aSStefano Zampini ierr = KSPRichardsonSetScale(pcbddc->coarse_ksp,2.0/(lambda_max+lambda_min));CHKERRQ(ierr); 5885cbcc2c2aSStefano Zampini } 5886c8587f34SStefano Zampini } 5887c8587f34SStefano Zampini } 5888cbcc2c2aSStefano Zampini 5889c8587f34SStefano Zampini /* check coarse problem residual error */ 58906e683305SStefano Zampini if (pcbddc->dbg_flag) { 58916e683305SStefano Zampini PetscViewer dbg_viewer = PETSC_VIEWER_STDOUT_(PetscObjectComm((PetscObject)pcbddc->coarse_ksp)); 58926e683305SStefano Zampini ierr = PetscViewerASCIIAddTab(dbg_viewer,2*(pcbddc->current_level+1));CHKERRQ(ierr); 58936e683305SStefano Zampini ierr = VecAXPY(check_vec,-1.0,coarse_vec);CHKERRQ(ierr); 5894c8587f34SStefano Zampini ierr = VecNorm(check_vec,NORM_INFINITY,&infty_error);CHKERRQ(ierr); 58956e683305SStefano Zampini ierr = MatMult(coarse_mat,check_vec,coarse_vec);CHKERRQ(ierr); 58966e683305SStefano Zampini ierr = VecNorm(coarse_vec,NORM_INFINITY,&abs_infty_error);CHKERRQ(ierr); 5897c8587f34SStefano Zampini ierr = VecDestroy(&check_vec);CHKERRQ(ierr); 5898779c1cceSStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem details (use estimates %d)\n",pcbddc->use_coarse_estimates);CHKERRQ(ierr); 58996e683305SStefano Zampini ierr = PetscObjectPrintClassNamePrefixType((PetscObject)(pcbddc->coarse_ksp),dbg_viewer);CHKERRQ(ierr); 59006e683305SStefano Zampini ierr = PetscObjectPrintClassNamePrefixType((PetscObject)(check_pc),dbg_viewer);CHKERRQ(ierr); 59016e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem exact infty_error : %1.6e\n",infty_error);CHKERRQ(ierr); 59026e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem residual infty_error: %1.6e\n",abs_infty_error);CHKERRQ(ierr); 5903b0f5fe93SStefano Zampini if (CoarseNullSpace) { 5904b0f5fe93SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem is singular\n");CHKERRQ(ierr); 5905b0f5fe93SStefano Zampini } 59066e683305SStefano Zampini if (compute_eigs) { 59076e683305SStefano Zampini PetscReal lambda_max_s,lambda_min_s; 5908deec49d1SStefano Zampini ierr = KSPGetType(check_ksp,&check_ksp_type);CHKERRQ(ierr); 5909c8587f34SStefano Zampini ierr = KSPGetIterationNumber(check_ksp,&its);CHKERRQ(ierr); 59106e683305SStefano Zampini ierr = KSPComputeExtremeSingularValues(check_ksp,&lambda_max_s,&lambda_min_s);CHKERRQ(ierr); 59116e683305SStefano 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); 59126e683305SStefano Zampini for (i=0;i<neigs;i++) { 59136e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"%1.6e %1.6ei\n",eigs_r[i],eigs_c[i]);CHKERRQ(ierr); 5914c8587f34SStefano Zampini } 59156e683305SStefano Zampini } 59166e683305SStefano Zampini ierr = PetscViewerFlush(dbg_viewer);CHKERRQ(ierr); 59176e683305SStefano Zampini ierr = PetscViewerASCIISubtractTab(dbg_viewer,2*(pcbddc->current_level+1));CHKERRQ(ierr); 59186e683305SStefano Zampini } 5919c8587f34SStefano Zampini ierr = KSPDestroy(&check_ksp);CHKERRQ(ierr); 59206e683305SStefano Zampini if (compute_eigs) { 59216e683305SStefano Zampini ierr = PetscFree(eigs_r);CHKERRQ(ierr); 59226e683305SStefano Zampini ierr = PetscFree(eigs_c);CHKERRQ(ierr); 5923c8587f34SStefano Zampini } 59246e683305SStefano Zampini } 59256e683305SStefano Zampini } 5926cbcc2c2aSStefano Zampini /* print additional info */ 5927cbcc2c2aSStefano Zampini if (pcbddc->dbg_flag) { 59286e683305SStefano Zampini /* waits until all processes reaches this point */ 59296e683305SStefano Zampini ierr = PetscBarrier((PetscObject)pc);CHKERRQ(ierr); 5930cbcc2c2aSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Coarse solver setup completed at level %d\n",pcbddc->current_level);CHKERRQ(ierr); 5931cbcc2c2aSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 5932cbcc2c2aSStefano Zampini } 5933cbcc2c2aSStefano Zampini 59342b510759SStefano Zampini /* free memory */ 5935c8587f34SStefano Zampini ierr = MatNullSpaceDestroy(&CoarseNullSpace);CHKERRQ(ierr); 5936fdc635d7SStefano Zampini ierr = MatDestroy(&coarse_mat);CHKERRQ(ierr); 5937c8587f34SStefano Zampini PetscFunctionReturn(0); 5938c8587f34SStefano Zampini } 5939674ae819SStefano Zampini 5940f34684f1SStefano Zampini #undef __FUNCT__ 5941f34684f1SStefano Zampini #define __FUNCT__ "PCBDDCComputePrimalNumbering" 5942f34684f1SStefano Zampini PetscErrorCode PCBDDCComputePrimalNumbering(PC pc,PetscInt* coarse_size_n,PetscInt** local_primal_indices_n) 5943f34684f1SStefano Zampini { 5944f34684f1SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 5945f34684f1SStefano Zampini PC_IS* pcis = (PC_IS*)pc->data; 5946f34684f1SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 5947dc456d91SStefano Zampini IS subset,subset_mult,subset_n; 5948dc456d91SStefano Zampini PetscInt local_size,coarse_size=0; 594973be2a3aSStefano Zampini PetscInt *local_primal_indices=NULL; 5950dc456d91SStefano Zampini const PetscInt *t_local_primal_indices; 5951f34684f1SStefano Zampini PetscErrorCode ierr; 5952f34684f1SStefano Zampini 5953f34684f1SStefano Zampini PetscFunctionBegin; 5954f34684f1SStefano Zampini /* Compute global number of coarse dofs */ 59550e6343abSStefano Zampini if (pcbddc->local_primal_size && !pcbddc->local_primal_ref_node) { 59560e6343abSStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"BDDC ConstraintsSetUp should be called first"); 5957727cdba6SStefano Zampini } 5958dc456d91SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)(pc->pmat)),pcbddc->local_primal_size_cc,pcbddc->local_primal_ref_node,PETSC_COPY_VALUES,&subset_n);CHKERRQ(ierr); 59593bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(pcis->mapping,subset_n,&subset);CHKERRQ(ierr); 5960dc456d91SStefano Zampini ierr = ISDestroy(&subset_n);CHKERRQ(ierr); 5961dc456d91SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)(pc->pmat)),pcbddc->local_primal_size_cc,pcbddc->local_primal_ref_mult,PETSC_COPY_VALUES,&subset_mult);CHKERRQ(ierr); 5962dc456d91SStefano Zampini ierr = PCBDDCSubsetNumbering(subset,subset_mult,&coarse_size,&subset_n);CHKERRQ(ierr); 5963dc456d91SStefano Zampini ierr = ISDestroy(&subset);CHKERRQ(ierr); 5964dc456d91SStefano Zampini ierr = ISDestroy(&subset_mult);CHKERRQ(ierr); 5965dc456d91SStefano Zampini ierr = ISGetLocalSize(subset_n,&local_size);CHKERRQ(ierr); 59660e6343abSStefano Zampini if (local_size != pcbddc->local_primal_size) { 59670e6343abSStefano Zampini SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Invalid number of local primal indices computed %d != %d",local_size,pcbddc->local_primal_size); 59680e6343abSStefano Zampini } 5969dc456d91SStefano Zampini ierr = PetscMalloc1(local_size,&local_primal_indices);CHKERRQ(ierr); 5970dc456d91SStefano Zampini ierr = ISGetIndices(subset_n,&t_local_primal_indices);CHKERRQ(ierr); 5971dc456d91SStefano Zampini ierr = PetscMemcpy(local_primal_indices,t_local_primal_indices,local_size*sizeof(PetscInt));CHKERRQ(ierr); 5972dc456d91SStefano Zampini ierr = ISRestoreIndices(subset_n,&t_local_primal_indices);CHKERRQ(ierr); 5973dc456d91SStefano Zampini ierr = ISDestroy(&subset_n);CHKERRQ(ierr); 5974f34684f1SStefano Zampini 5975f34684f1SStefano Zampini /* check numbering */ 5976f34684f1SStefano Zampini if (pcbddc->dbg_flag) { 5977019a44ceSStefano Zampini PetscScalar coarsesum,*array,*array2; 5978dc456d91SStefano Zampini PetscInt i; 5979b9b85e73SStefano Zampini PetscBool set_error = PETSC_FALSE,set_error_reduced = PETSC_FALSE; 5980f34684f1SStefano Zampini 5981f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 5982f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 5983f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Check coarse indices\n");CHKERRQ(ierr); 59841575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 5985019a44ceSStefano Zampini /* counter */ 5986019a44ceSStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 5987019a44ceSStefano Zampini ierr = VecSet(pcis->vec1_N,1.0);CHKERRQ(ierr); 5988019a44ceSStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 5989019a44ceSStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 5990019a44ceSStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_global,pcis->vec2_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 5991019a44ceSStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_global,pcis->vec2_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 5992f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_N,0.0);CHKERRQ(ierr); 5993f34684f1SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 5994727cdba6SStefano Zampini ierr = VecSetValue(pcis->vec1_N,pcbddc->primal_indices_local_idxs[i],1.0,INSERT_VALUES);CHKERRQ(ierr); 5995f34684f1SStefano Zampini } 5996f34684f1SStefano Zampini ierr = VecAssemblyBegin(pcis->vec1_N);CHKERRQ(ierr); 5997f34684f1SStefano Zampini ierr = VecAssemblyEnd(pcis->vec1_N);CHKERRQ(ierr); 5998f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 5999e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6000e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6001e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_global,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6002e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_global,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6003f34684f1SStefano Zampini ierr = VecGetArray(pcis->vec1_N,&array);CHKERRQ(ierr); 6004019a44ceSStefano Zampini ierr = VecGetArray(pcis->vec2_N,&array2);CHKERRQ(ierr); 6005f34684f1SStefano Zampini for (i=0;i<pcis->n;i++) { 6006019a44ceSStefano Zampini if (array[i] != 0.0 && array[i] != array2[i]) { 600775c01103SStefano Zampini PetscInt owned = (PetscInt)PetscRealPart(array[i]); 600875c01103SStefano Zampini PetscInt neigh = (PetscInt)PetscRealPart(array2[i]); 6009b9b85e73SStefano Zampini set_error = PETSC_TRUE; 6010019a44ceSStefano 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); 6011f34684f1SStefano Zampini } 6012f34684f1SStefano Zampini } 6013019a44ceSStefano Zampini ierr = VecRestoreArray(pcis->vec2_N,&array2);CHKERRQ(ierr); 6014b9b85e73SStefano Zampini ierr = MPI_Allreduce(&set_error,&set_error_reduced,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 6015f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 6016f34684f1SStefano Zampini for (i=0;i<pcis->n;i++) { 6017f34684f1SStefano Zampini if (PetscRealPart(array[i]) > 0.0) array[i] = 1.0/PetscRealPart(array[i]); 6018f34684f1SStefano Zampini } 6019f34684f1SStefano Zampini ierr = VecRestoreArray(pcis->vec1_N,&array);CHKERRQ(ierr); 6020f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 6021e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6022e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6023f34684f1SStefano Zampini ierr = VecSum(pcis->vec1_global,&coarsesum);CHKERRQ(ierr); 6024f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Size of coarse problem is %d (%lf)\n",coarse_size,PetscRealPart(coarsesum));CHKERRQ(ierr); 6025b9b85e73SStefano Zampini if (pcbddc->dbg_flag > 1 || set_error_reduced) { 6026ca8b9ea9SStefano Zampini PetscInt *gidxs; 6027ca8b9ea9SStefano Zampini 6028ca8b9ea9SStefano Zampini ierr = PetscMalloc1(pcbddc->local_primal_size,&gidxs);CHKERRQ(ierr); 60293bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingApply(pcis->mapping,pcbddc->local_primal_size,pcbddc->primal_indices_local_idxs,gidxs);CHKERRQ(ierr); 6030f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Distribution of local primal indices\n");CHKERRQ(ierr); 6031f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 6032f34684f1SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d\n",PetscGlobalRank);CHKERRQ(ierr); 6033f34684f1SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 60344bc2dc4bSStefano 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); 6035f34684f1SStefano Zampini } 6036f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 6037ca8b9ea9SStefano Zampini ierr = PetscFree(gidxs);CHKERRQ(ierr); 6038f34684f1SStefano Zampini } 6039f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 60401575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 6041302440fdSBarry Smith if (set_error_reduced) SETERRQ(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"BDDC Numbering of coarse dofs failed"); 6042f34684f1SStefano Zampini } 60438bec7fa6SStefano Zampini /* ierr = PetscPrintf(PetscObjectComm((PetscObject)pc),"Size of coarse problem is %d\n",coarse_size);CHKERRQ(ierr); */ 6044f34684f1SStefano Zampini /* get back data */ 6045f34684f1SStefano Zampini *coarse_size_n = coarse_size; 6046f34684f1SStefano Zampini *local_primal_indices_n = local_primal_indices; 6047674ae819SStefano Zampini PetscFunctionReturn(0); 6048674ae819SStefano Zampini } 6049674ae819SStefano Zampini 6050e456f2a8SStefano Zampini #undef __FUNCT__ 6051e456f2a8SStefano Zampini #define __FUNCT__ "PCBDDCGlobalToLocal" 6052a7dc3881SStefano Zampini PetscErrorCode PCBDDCGlobalToLocal(VecScatter g2l_ctx,Vec gwork, Vec lwork, IS globalis, IS* localis) 6053e456f2a8SStefano Zampini { 6054e456f2a8SStefano Zampini IS localis_t; 6055a7dc3881SStefano Zampini PetscInt i,lsize,*idxs,n; 6056e456f2a8SStefano Zampini PetscScalar *vals; 6057e456f2a8SStefano Zampini PetscErrorCode ierr; 6058e456f2a8SStefano Zampini 6059e456f2a8SStefano Zampini PetscFunctionBegin; 6060a7dc3881SStefano Zampini /* get indices in local ordering exploiting local to global map */ 6061e456f2a8SStefano Zampini ierr = ISGetLocalSize(globalis,&lsize);CHKERRQ(ierr); 6062854ce69bSBarry Smith ierr = PetscMalloc1(lsize,&vals);CHKERRQ(ierr); 6063e456f2a8SStefano Zampini for (i=0;i<lsize;i++) vals[i] = 1.0; 6064e456f2a8SStefano Zampini ierr = ISGetIndices(globalis,(const PetscInt**)&idxs);CHKERRQ(ierr); 6065a7dc3881SStefano Zampini ierr = VecSet(gwork,0.0);CHKERRQ(ierr); 6066a7dc3881SStefano Zampini ierr = VecSet(lwork,0.0);CHKERRQ(ierr); 60671035eff8SStefano Zampini if (idxs) { /* multilevel guard */ 6068a7dc3881SStefano Zampini ierr = VecSetValues(gwork,lsize,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 60691035eff8SStefano Zampini } 6070a7dc3881SStefano Zampini ierr = VecAssemblyBegin(gwork);CHKERRQ(ierr); 6071e456f2a8SStefano Zampini ierr = ISRestoreIndices(globalis,(const PetscInt**)&idxs);CHKERRQ(ierr); 6072e456f2a8SStefano Zampini ierr = PetscFree(vals);CHKERRQ(ierr); 6073a7dc3881SStefano Zampini ierr = VecAssemblyEnd(gwork);CHKERRQ(ierr); 6074a7dc3881SStefano Zampini /* now compute set in local ordering */ 6075a7dc3881SStefano Zampini ierr = VecScatterBegin(g2l_ctx,gwork,lwork,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6076a7dc3881SStefano Zampini ierr = VecScatterEnd(g2l_ctx,gwork,lwork,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6077a7dc3881SStefano Zampini ierr = VecGetArrayRead(lwork,(const PetscScalar**)&vals);CHKERRQ(ierr); 6078a7dc3881SStefano Zampini ierr = VecGetSize(lwork,&n);CHKERRQ(ierr); 6079a7dc3881SStefano Zampini for (i=0,lsize=0;i<n;i++) { 6080ff92baa0SMatthew G. Knepley if (PetscRealPart(vals[i]) > 0.5) { 6081e456f2a8SStefano Zampini lsize++; 6082e456f2a8SStefano Zampini } 6083e456f2a8SStefano Zampini } 6084854ce69bSBarry Smith ierr = PetscMalloc1(lsize,&idxs);CHKERRQ(ierr); 6085a7dc3881SStefano Zampini for (i=0,lsize=0;i<n;i++) { 6086ff92baa0SMatthew G. Knepley if (PetscRealPart(vals[i]) > 0.5) { 6087e456f2a8SStefano Zampini idxs[lsize++] = i; 6088e456f2a8SStefano Zampini } 6089e456f2a8SStefano Zampini } 6090a7dc3881SStefano Zampini ierr = VecRestoreArrayRead(lwork,(const PetscScalar**)&vals);CHKERRQ(ierr); 6091a7dc3881SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)gwork),lsize,idxs,PETSC_OWN_POINTER,&localis_t);CHKERRQ(ierr); 6092e456f2a8SStefano Zampini *localis = localis_t; 6093e456f2a8SStefano Zampini PetscFunctionReturn(0); 6094e456f2a8SStefano Zampini } 6095906d46d4SStefano Zampini 6096906d46d4SStefano Zampini /* the next two functions will be called in KSPMatMult if a change of basis has been requested */ 6097906d46d4SStefano Zampini #undef __FUNCT__ 6098906d46d4SStefano Zampini #define __FUNCT__ "PCBDDCMatMult_Private" 6099906d46d4SStefano Zampini static PetscErrorCode PCBDDCMatMult_Private(Mat A, Vec x, Vec y) 6100906d46d4SStefano Zampini { 6101906d46d4SStefano Zampini PCBDDCChange_ctx change_ctx; 6102906d46d4SStefano Zampini PetscErrorCode ierr; 6103906d46d4SStefano Zampini 6104906d46d4SStefano Zampini PetscFunctionBegin; 6105906d46d4SStefano Zampini ierr = MatShellGetContext(A,&change_ctx);CHKERRQ(ierr); 6106906d46d4SStefano Zampini ierr = MatMult(change_ctx->global_change,x,change_ctx->work[0]);CHKERRQ(ierr); 6107906d46d4SStefano Zampini ierr = MatMult(change_ctx->original_mat,change_ctx->work[0],change_ctx->work[1]);CHKERRQ(ierr); 6108906d46d4SStefano Zampini ierr = MatMultTranspose(change_ctx->global_change,change_ctx->work[1],y);CHKERRQ(ierr); 6109906d46d4SStefano Zampini PetscFunctionReturn(0); 6110906d46d4SStefano Zampini } 6111906d46d4SStefano Zampini 6112906d46d4SStefano Zampini #undef __FUNCT__ 6113906d46d4SStefano Zampini #define __FUNCT__ "PCBDDCMatMultTranspose_Private" 6114906d46d4SStefano Zampini static PetscErrorCode PCBDDCMatMultTranspose_Private(Mat A, Vec x, Vec y) 6115906d46d4SStefano Zampini { 6116906d46d4SStefano Zampini PCBDDCChange_ctx change_ctx; 6117906d46d4SStefano Zampini PetscErrorCode ierr; 6118906d46d4SStefano Zampini 6119906d46d4SStefano Zampini PetscFunctionBegin; 6120906d46d4SStefano Zampini ierr = MatShellGetContext(A,&change_ctx);CHKERRQ(ierr); 6121906d46d4SStefano Zampini ierr = MatMult(change_ctx->global_change,x,change_ctx->work[0]);CHKERRQ(ierr); 6122906d46d4SStefano Zampini ierr = MatMultTranspose(change_ctx->original_mat,change_ctx->work[0],change_ctx->work[1]);CHKERRQ(ierr); 6123906d46d4SStefano Zampini ierr = MatMultTranspose(change_ctx->global_change,change_ctx->work[1],y);CHKERRQ(ierr); 6124906d46d4SStefano Zampini PetscFunctionReturn(0); 6125906d46d4SStefano Zampini } 6126b96c3477SStefano Zampini 6127b96c3477SStefano Zampini #undef __FUNCT__ 6128b96c3477SStefano Zampini #define __FUNCT__ "PCBDDCSetUpSubSchurs" 612908122e43SStefano Zampini PetscErrorCode PCBDDCSetUpSubSchurs(PC pc) 6130b96c3477SStefano Zampini { 6131a64f4aa4SStefano Zampini PC_IS *pcis=(PC_IS*)pc->data; 6132b96c3477SStefano Zampini PC_BDDC *pcbddc=(PC_BDDC*)pc->data; 6133b96c3477SStefano Zampini PCBDDCSubSchurs sub_schurs=pcbddc->sub_schurs; 6134a64f4aa4SStefano Zampini Mat S_j; 6135b96c3477SStefano Zampini PetscInt *used_xadj,*used_adjncy; 6136b96c3477SStefano Zampini PetscBool free_used_adj; 6137b96c3477SStefano Zampini PetscErrorCode ierr; 6138b96c3477SStefano Zampini 6139b96c3477SStefano Zampini PetscFunctionBegin; 6140b96c3477SStefano Zampini /* decide the adjacency to be used for determining internal problems for local schur on subsets */ 6141b96c3477SStefano Zampini free_used_adj = PETSC_FALSE; 614208122e43SStefano Zampini if (pcbddc->sub_schurs_layers == -1) { 6143b96c3477SStefano Zampini used_xadj = NULL; 6144b96c3477SStefano Zampini used_adjncy = NULL; 6145b96c3477SStefano Zampini } else { 614608122e43SStefano Zampini if (pcbddc->sub_schurs_use_useradj && pcbddc->mat_graph->xadj) { 614708122e43SStefano Zampini used_xadj = pcbddc->mat_graph->xadj; 614808122e43SStefano Zampini used_adjncy = pcbddc->mat_graph->adjncy; 614908122e43SStefano Zampini } else if (pcbddc->computed_rowadj) { 6150b96c3477SStefano Zampini used_xadj = pcbddc->mat_graph->xadj; 6151b96c3477SStefano Zampini used_adjncy = pcbddc->mat_graph->adjncy; 6152b96c3477SStefano Zampini } else { 61532fffb893SStefano Zampini PetscBool flg_row=PETSC_FALSE; 6154b96c3477SStefano Zampini const PetscInt *xadj,*adjncy; 6155b96c3477SStefano Zampini PetscInt nvtxs; 6156b96c3477SStefano Zampini 61572fffb893SStefano Zampini ierr = MatGetRowIJ(pcbddc->local_mat,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 61582fffb893SStefano Zampini if (flg_row) { 6159b96c3477SStefano Zampini ierr = PetscMalloc2(nvtxs+1,&used_xadj,xadj[nvtxs],&used_adjncy);CHKERRQ(ierr); 6160b96c3477SStefano Zampini ierr = PetscMemcpy(used_xadj,xadj,(nvtxs+1)*sizeof(*xadj));CHKERRQ(ierr); 6161b96c3477SStefano Zampini ierr = PetscMemcpy(used_adjncy,adjncy,(xadj[nvtxs])*sizeof(*adjncy));CHKERRQ(ierr); 6162b96c3477SStefano Zampini free_used_adj = PETSC_TRUE; 61632fffb893SStefano Zampini } else { 61642fffb893SStefano Zampini pcbddc->sub_schurs_layers = -1; 61652fffb893SStefano Zampini used_xadj = NULL; 61662fffb893SStefano Zampini used_adjncy = NULL; 61672fffb893SStefano Zampini } 61682fffb893SStefano Zampini ierr = MatRestoreRowIJ(pcbddc->local_mat,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 6169b96c3477SStefano Zampini } 6170b96c3477SStefano Zampini } 6171d5574798SStefano Zampini 6172d5574798SStefano Zampini /* setup sub_schurs data */ 6173a64f4aa4SStefano Zampini ierr = MatCreateSchurComplement(pcis->A_II,pcis->A_II,pcis->A_IB,pcis->A_BI,pcis->A_BB,&S_j);CHKERRQ(ierr); 6174df4d28bfSStefano Zampini if (!sub_schurs->schur_explicit) { 6175df4d28bfSStefano Zampini /* pcbddc->ksp_D up to date only if not using MatFactor with Schur complement support */ 6176a64f4aa4SStefano Zampini ierr = MatSchurComplementSetKSP(S_j,pcbddc->ksp_D);CHKERRQ(ierr); 6177ca92afb2SStefano Zampini ierr = PCBDDCSubSchursSetUp(sub_schurs,NULL,S_j,PETSC_FALSE,used_xadj,used_adjncy,pcbddc->sub_schurs_layers,pcbddc->faster_deluxe,pcbddc->adaptive_selection,PETSC_FALSE,PETSC_FALSE,0,NULL,NULL);CHKERRQ(ierr); 6178a64f4aa4SStefano Zampini } else { 61796816873aSStefano Zampini PetscBool reuse_solvers = (PetscBool)!pcbddc->use_change_of_basis; 618004708bb6SStefano Zampini PetscBool isseqaij; 6181a3df083aSStefano Zampini PetscInt benign_n; 6182a3df083aSStefano Zampini 61835feab87aSStefano Zampini if (!pcbddc->use_vertices && reuse_solvers) { 61845feab87aSStefano Zampini PetscInt n_vertices; 61855feab87aSStefano Zampini 61865feab87aSStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_vertices,&n_vertices);CHKERRQ(ierr); 61872034aafcSStefano Zampini reuse_solvers = (PetscBool)!n_vertices; 61885feab87aSStefano Zampini } 618904708bb6SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcbddc->local_mat,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 619004708bb6SStefano Zampini if (!isseqaij) { 619104708bb6SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 619204708bb6SStefano Zampini if (matis->A == pcbddc->local_mat) { 619304708bb6SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 619404708bb6SStefano Zampini ierr = MatConvert(matis->A,MATSEQAIJ,MAT_INITIAL_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 619504708bb6SStefano Zampini } else { 619604708bb6SStefano Zampini ierr = MatConvert(pcbddc->local_mat,MATSEQAIJ,MAT_REUSE_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 619704708bb6SStefano Zampini } 619804708bb6SStefano Zampini } 6199a3df083aSStefano Zampini if (!pcbddc->benign_change_explicit) { 6200a3df083aSStefano Zampini benign_n = pcbddc->benign_n; 6201ca92afb2SStefano Zampini } else { 6202a3df083aSStefano Zampini benign_n = 0; 6203ca92afb2SStefano Zampini } 6204a3df083aSStefano Zampini ierr = PCBDDCSubSchursSetUp(sub_schurs,pcbddc->local_mat,S_j,pcbddc->sub_schurs_exact_schur,used_xadj,used_adjncy,pcbddc->sub_schurs_layers,pcbddc->faster_deluxe,pcbddc->adaptive_selection,reuse_solvers,pcbddc->benign_saddle_point,benign_n,pcbddc->benign_p0_lidx,pcbddc->benign_zerodiag_subs);CHKERRQ(ierr); 6205*bf3a8328SStefano Zampini /* If the user defines additional constraints, we import them here. 6206*bf3a8328SStefano 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 */ 6207*bf3a8328SStefano Zampini if (pcbddc->adaptive_userdefined) { 620888c03ad3SStefano Zampini PC_IS *pcisf; 620988c03ad3SStefano Zampini PC_BDDC *pcbddcf; 621088c03ad3SStefano Zampini PC pcf; 621188c03ad3SStefano Zampini 621288c03ad3SStefano Zampini if (pcbddc->sub_schurs_rebuild) { 621388c03ad3SStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Cannot compute change of basis with a different graph"); 621488c03ad3SStefano Zampini } 621588c03ad3SStefano Zampini ierr = PCCreate(PetscObjectComm((PetscObject)pc),&pcf);CHKERRQ(ierr); 621688c03ad3SStefano Zampini ierr = PCSetOperators(pcf,pc->mat,pc->pmat);CHKERRQ(ierr); 621788c03ad3SStefano Zampini ierr = PCSetType(pcf,PCBDDC);CHKERRQ(ierr); 621888c03ad3SStefano Zampini /* hacks */ 621988c03ad3SStefano Zampini pcisf = (PC_IS*)pcf->data; 622088c03ad3SStefano Zampini pcbddcf = (PC_BDDC*)pcf->data; 622188c03ad3SStefano Zampini *pcisf = *pcis; 622288c03ad3SStefano Zampini ierr = PetscFree(pcbddcf->mat_graph);CHKERRQ(ierr); 622388c03ad3SStefano Zampini pcbddcf->mat_graph = pcbddc->mat_graph; 622488c03ad3SStefano Zampini ierr = PetscFree(pcbddcf->sub_schurs);CHKERRQ(ierr); 622588c03ad3SStefano Zampini pcbddcf->sub_schurs = sub_schurs; 622688c03ad3SStefano Zampini pcbddcf->use_faces = PETSC_TRUE; 622788c03ad3SStefano Zampini pcbddcf->use_change_of_basis = PETSC_TRUE; 622888c03ad3SStefano Zampini pcbddcf->use_change_on_faces = PETSC_TRUE; 622988c03ad3SStefano Zampini pcbddcf->use_deluxe_scaling = PETSC_TRUE; 6230*bf3a8328SStefano Zampini pcbddcf->deluxe_zerorows = PETSC_TRUE; 623188c03ad3SStefano Zampini pcbddcf->fake_change = PETSC_TRUE; 623288c03ad3SStefano Zampini ierr = PCBDDCConstraintsSetUp(pcf);CHKERRQ(ierr); 623388c03ad3SStefano Zampini /* free unneeded memory allocated in PCBDDCConstraintsSetUp */ 623488c03ad3SStefano Zampini pcbddcf->sub_schurs = NULL; 623588c03ad3SStefano Zampini ierr = MatDestroy(&pcbddcf->ConstraintMatrix);CHKERRQ(ierr); 623688c03ad3SStefano Zampini ierr = MatNullSpaceDestroy(&pcbddcf->onearnullspace);CHKERRQ(ierr); 623788c03ad3SStefano Zampini ierr = PetscFree2(pcbddcf->local_primal_ref_node,pcbddcf->local_primal_ref_mult);CHKERRQ(ierr); 623888c03ad3SStefano Zampini ierr = PetscFree(pcbddcf->primal_indices_local_idxs);CHKERRQ(ierr); 623988c03ad3SStefano Zampini ierr = PetscFree(pcbddcf->onearnullvecs_state);CHKERRQ(ierr); 624088c03ad3SStefano Zampini ierr = PetscFree(pcf->data);CHKERRQ(ierr); 624188c03ad3SStefano Zampini pcf->ops->destroy = NULL; 624288c03ad3SStefano Zampini ierr = PCDestroy(&pcf);CHKERRQ(ierr); 624388c03ad3SStefano Zampini } 6244ca92afb2SStefano Zampini } 6245d12d3064SStefano Zampini ierr = MatDestroy(&S_j);CHKERRQ(ierr); 6246b96c3477SStefano Zampini 6247b96c3477SStefano Zampini /* free adjacency */ 6248b96c3477SStefano Zampini if (free_used_adj) { 6249b96c3477SStefano Zampini ierr = PetscFree2(used_xadj,used_adjncy);CHKERRQ(ierr); 6250b96c3477SStefano Zampini } 6251b96c3477SStefano Zampini PetscFunctionReturn(0); 6252b96c3477SStefano Zampini } 6253b96c3477SStefano Zampini 6254b96c3477SStefano Zampini #undef __FUNCT__ 6255b96c3477SStefano Zampini #define __FUNCT__ "PCBDDCInitSubSchurs" 625608122e43SStefano Zampini PetscErrorCode PCBDDCInitSubSchurs(PC pc) 6257b96c3477SStefano Zampini { 6258b96c3477SStefano Zampini PC_IS *pcis=(PC_IS*)pc->data; 6259b96c3477SStefano Zampini PC_BDDC *pcbddc=(PC_BDDC*)pc->data; 6260b96c3477SStefano Zampini PCBDDCSubSchurs sub_schurs=pcbddc->sub_schurs; 6261b96c3477SStefano Zampini PCBDDCGraph graph; 6262b96c3477SStefano Zampini PetscErrorCode ierr; 6263b96c3477SStefano Zampini 6264b96c3477SStefano Zampini PetscFunctionBegin; 6265b96c3477SStefano Zampini /* attach interface graph for determining subsets */ 626608122e43SStefano Zampini if (pcbddc->sub_schurs_rebuild) { /* in case rebuild has been requested, it uses a graph generated only by the neighbouring information */ 62673301b35fSStefano Zampini IS verticesIS,verticescomm; 62683301b35fSStefano Zampini PetscInt vsize,*idxs; 6269b96c3477SStefano Zampini 6270b96c3477SStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,NULL,NULL,NULL,NULL,&verticesIS);CHKERRQ(ierr); 62713301b35fSStefano Zampini ierr = ISGetSize(verticesIS,&vsize);CHKERRQ(ierr); 62723301b35fSStefano Zampini ierr = ISGetIndices(verticesIS,(const PetscInt**)&idxs);CHKERRQ(ierr); 62733301b35fSStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),vsize,idxs,PETSC_COPY_VALUES,&verticescomm);CHKERRQ(ierr); 62743301b35fSStefano Zampini ierr = ISRestoreIndices(verticesIS,(const PetscInt**)&idxs);CHKERRQ(ierr); 62753301b35fSStefano Zampini ierr = ISDestroy(&verticesIS);CHKERRQ(ierr); 6276b96c3477SStefano Zampini ierr = PCBDDCGraphCreate(&graph);CHKERRQ(ierr); 62777fb0e2dbSStefano Zampini ierr = PCBDDCGraphInit(graph,pcbddc->mat_graph->l2gmap,pcbddc->mat_graph->nvtxs_global);CHKERRQ(ierr); 62783301b35fSStefano Zampini ierr = PCBDDCGraphSetUp(graph,0,NULL,pcbddc->DirichletBoundariesLocal,0,NULL,verticescomm);CHKERRQ(ierr); 62793301b35fSStefano Zampini ierr = ISDestroy(&verticescomm);CHKERRQ(ierr); 6280b96c3477SStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(graph);CHKERRQ(ierr); 6281b96c3477SStefano Zampini /* 6282b96c3477SStefano Zampini if (pcbddc->dbg_flag) { 6283b96c3477SStefano Zampini ierr = PCBDDCGraphASCIIView(graph,pcbddc->dbg_flag,pcbddc->dbg_viewer);CHKERRQ(ierr); 6284b96c3477SStefano Zampini } 6285b96c3477SStefano Zampini */ 6286b96c3477SStefano Zampini } else { 6287b96c3477SStefano Zampini graph = pcbddc->mat_graph; 6288b96c3477SStefano Zampini } 6289b96c3477SStefano Zampini 6290b96c3477SStefano Zampini /* sub_schurs init */ 6291a64f4aa4SStefano Zampini ierr = PCBDDCSubSchursInit(sub_schurs,pcis->is_I_local,pcis->is_B_local,graph,pcis->BtoNmap);CHKERRQ(ierr); 6292a64f4aa4SStefano Zampini 6293b96c3477SStefano Zampini /* free graph struct */ 629408122e43SStefano Zampini if (pcbddc->sub_schurs_rebuild) { 6295b96c3477SStefano Zampini ierr = PCBDDCGraphDestroy(&graph);CHKERRQ(ierr); 6296b96c3477SStefano Zampini } 6297b96c3477SStefano Zampini PetscFunctionReturn(0); 6298b96c3477SStefano Zampini } 6299fa34dd3eSStefano Zampini 6300fa34dd3eSStefano Zampini #undef __FUNCT__ 6301fa34dd3eSStefano Zampini #define __FUNCT__ "PCBDDCCheckOperator" 6302fa34dd3eSStefano Zampini PetscErrorCode PCBDDCCheckOperator(PC pc) 6303fa34dd3eSStefano Zampini { 6304fa34dd3eSStefano Zampini PC_IS *pcis=(PC_IS*)pc->data; 6305fa34dd3eSStefano Zampini PC_BDDC *pcbddc=(PC_BDDC*)pc->data; 6306fa34dd3eSStefano Zampini PetscErrorCode ierr; 6307fa34dd3eSStefano Zampini 6308fa34dd3eSStefano Zampini PetscFunctionBegin; 6309fa34dd3eSStefano Zampini if (pcbddc->n_vertices == pcbddc->local_primal_size) { 6310fa34dd3eSStefano Zampini IS zerodiag = NULL; 63114f1b2e48SStefano Zampini Mat S_j,B0_B=NULL; 6312fa34dd3eSStefano Zampini Vec dummy_vec=NULL,vec_check_B,vec_scale_P; 63134f1b2e48SStefano Zampini PetscScalar *p0_check,*array,*array2; 631475c01103SStefano Zampini PetscReal norm; 6315fa34dd3eSStefano Zampini PetscInt i; 6316fa34dd3eSStefano Zampini 6317fa34dd3eSStefano Zampini /* B0 and B0_B */ 6318fa34dd3eSStefano Zampini if (zerodiag) { 6319fa34dd3eSStefano Zampini IS dummy; 6320fa34dd3eSStefano Zampini 63214f1b2e48SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcbddc->benign_n,0,1,&dummy);CHKERRQ(ierr); 63224f1b2e48SStefano Zampini ierr = MatGetSubMatrix(pcbddc->benign_B0,dummy,pcis->is_B_local,MAT_INITIAL_MATRIX,&B0_B);CHKERRQ(ierr); 6323fa34dd3eSStefano Zampini ierr = MatCreateVecs(B0_B,NULL,&dummy_vec);CHKERRQ(ierr); 6324fa34dd3eSStefano Zampini ierr = ISDestroy(&dummy);CHKERRQ(ierr); 6325fa34dd3eSStefano Zampini } 6326fa34dd3eSStefano Zampini /* I need a primal vector to scale primal nodes since BDDC sums contibutions */ 6327fa34dd3eSStefano Zampini ierr = VecDuplicate(pcbddc->vec1_P,&vec_scale_P);CHKERRQ(ierr); 6328fa34dd3eSStefano Zampini ierr = VecSet(pcbddc->vec1_P,1.0);CHKERRQ(ierr); 6329fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6330fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6331fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,vec_scale_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6332fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,vec_scale_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6333fa34dd3eSStefano Zampini ierr = VecReciprocal(vec_scale_P);CHKERRQ(ierr); 6334fa34dd3eSStefano Zampini /* S_j */ 6335fa34dd3eSStefano Zampini ierr = MatCreateSchurComplement(pcis->A_II,pcis->A_II,pcis->A_IB,pcis->A_BI,pcis->A_BB,&S_j);CHKERRQ(ierr); 6336fa34dd3eSStefano Zampini ierr = MatSchurComplementSetKSP(S_j,pcbddc->ksp_D);CHKERRQ(ierr); 6337fa34dd3eSStefano Zampini 6338fa34dd3eSStefano Zampini /* mimic vector in \widetilde{W}_\Gamma */ 6339fa34dd3eSStefano Zampini ierr = VecSetRandom(pcis->vec1_N,NULL);CHKERRQ(ierr); 6340fa34dd3eSStefano Zampini /* continuous in primal space */ 6341fa34dd3eSStefano Zampini ierr = VecSetRandom(pcbddc->coarse_vec,NULL);CHKERRQ(ierr); 6342fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,pcbddc->vec1_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6343fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,pcbddc->vec1_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6344fa34dd3eSStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 63454f1b2e48SStefano Zampini ierr = PetscCalloc1(pcbddc->benign_n,&p0_check);CHKERRQ(ierr); 63464f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) p0_check[i] = array[pcbddc->local_primal_size-pcbddc->benign_n+i]; 6347fa34dd3eSStefano Zampini ierr = VecSetValues(pcis->vec1_N,pcbddc->local_primal_size,pcbddc->local_primal_ref_node,array,INSERT_VALUES);CHKERRQ(ierr); 6348fa34dd3eSStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 6349fa34dd3eSStefano Zampini ierr = VecAssemblyBegin(pcis->vec1_N);CHKERRQ(ierr); 6350fa34dd3eSStefano Zampini ierr = VecAssemblyEnd(pcis->vec1_N);CHKERRQ(ierr); 6351fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcis->N_to_B,pcis->vec1_N,pcis->vec2_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6352fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcis->N_to_B,pcis->vec1_N,pcis->vec2_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6353fa34dd3eSStefano Zampini ierr = VecDuplicate(pcis->vec2_B,&vec_check_B);CHKERRQ(ierr); 6354fa34dd3eSStefano Zampini ierr = VecCopy(pcis->vec2_B,vec_check_B);CHKERRQ(ierr); 6355fa34dd3eSStefano Zampini 6356fa34dd3eSStefano Zampini /* assemble rhs for coarse problem */ 6357fa34dd3eSStefano Zampini /* widetilde{S}_\Gamma w_\Gamma + \widetilde{B0}^T_B p0 */ 6358fa34dd3eSStefano Zampini /* local with Schur */ 6359fa34dd3eSStefano Zampini ierr = MatMult(S_j,pcis->vec2_B,pcis->vec1_B);CHKERRQ(ierr); 6360fa34dd3eSStefano Zampini if (zerodiag) { 6361fa34dd3eSStefano Zampini ierr = VecGetArray(dummy_vec,&array);CHKERRQ(ierr); 63624f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) array[i] = p0_check[i]; 6363fa34dd3eSStefano Zampini ierr = VecRestoreArray(dummy_vec,&array);CHKERRQ(ierr); 6364fa34dd3eSStefano Zampini ierr = MatMultTransposeAdd(B0_B,dummy_vec,pcis->vec1_B,pcis->vec1_B);CHKERRQ(ierr); 6365fa34dd3eSStefano Zampini } 6366fa34dd3eSStefano Zampini /* sum on primal nodes the local contributions */ 6367fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcis->N_to_B,pcis->vec1_B,pcis->vec1_N,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6368fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcis->N_to_B,pcis->vec1_B,pcis->vec1_N,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6369fa34dd3eSStefano Zampini ierr = VecGetArray(pcis->vec1_N,&array);CHKERRQ(ierr); 6370fa34dd3eSStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array2);CHKERRQ(ierr); 6371fa34dd3eSStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) array2[i] = array[pcbddc->local_primal_ref_node[i]]; 6372fa34dd3eSStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array2);CHKERRQ(ierr); 6373fa34dd3eSStefano Zampini ierr = VecRestoreArray(pcis->vec1_N,&array);CHKERRQ(ierr); 6374fa34dd3eSStefano Zampini ierr = VecSet(pcbddc->coarse_vec,0.);CHKERRQ(ierr); 6375fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6376fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6377fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,pcbddc->vec1_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6378fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,pcbddc->vec1_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6379fa34dd3eSStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 6380fa34dd3eSStefano Zampini /* scale primal nodes (BDDC sums contibutions) */ 6381fa34dd3eSStefano Zampini ierr = VecPointwiseMult(pcbddc->vec1_P,vec_scale_P,pcbddc->vec1_P);CHKERRQ(ierr); 6382fa34dd3eSStefano Zampini ierr = VecSetValues(pcis->vec1_N,pcbddc->local_primal_size,pcbddc->local_primal_ref_node,array,INSERT_VALUES);CHKERRQ(ierr); 6383fa34dd3eSStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 6384fa34dd3eSStefano Zampini ierr = VecAssemblyBegin(pcis->vec1_N);CHKERRQ(ierr); 6385fa34dd3eSStefano Zampini ierr = VecAssemblyEnd(pcis->vec1_N);CHKERRQ(ierr); 6386fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcis->N_to_B,pcis->vec1_N,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6387fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcis->N_to_B,pcis->vec1_N,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6388fa34dd3eSStefano Zampini /* global: \widetilde{B0}_B w_\Gamma */ 6389fa34dd3eSStefano Zampini if (zerodiag) { 6390fa34dd3eSStefano Zampini ierr = MatMult(B0_B,pcis->vec2_B,dummy_vec);CHKERRQ(ierr); 6391fa34dd3eSStefano Zampini ierr = VecGetArray(dummy_vec,&array);CHKERRQ(ierr); 63924f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) pcbddc->benign_p0[i] = array[i]; 6393fa34dd3eSStefano Zampini ierr = VecRestoreArray(dummy_vec,&array);CHKERRQ(ierr); 6394fa34dd3eSStefano Zampini } 6395fa34dd3eSStefano Zampini /* BDDC */ 6396fa34dd3eSStefano Zampini ierr = VecSet(pcis->vec1_D,0.);CHKERRQ(ierr); 6397fa34dd3eSStefano Zampini ierr = PCBDDCApplyInterfacePreconditioner(pc,PETSC_FALSE);CHKERRQ(ierr); 6398fa34dd3eSStefano Zampini 6399fa34dd3eSStefano Zampini ierr = VecCopy(pcis->vec1_B,pcis->vec2_B);CHKERRQ(ierr); 6400fa34dd3eSStefano Zampini ierr = VecAXPY(pcis->vec1_B,-1.0,vec_check_B);CHKERRQ(ierr); 6401fa34dd3eSStefano Zampini ierr = VecNorm(pcis->vec1_B,NORM_INFINITY,&norm);CHKERRQ(ierr); 6402fa34dd3eSStefano Zampini PetscPrintf(PETSC_COMM_SELF,"[%d] BDDC local error is %1.4e\n",PetscGlobalRank,norm); 64034f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 64044f1b2e48SStefano Zampini PetscPrintf(PETSC_COMM_SELF,"[%d] BDDC p0[%d] error is %1.4e\n",PetscGlobalRank,i,PetscAbsScalar(pcbddc->benign_p0[i]-p0_check[i])); 6405fa34dd3eSStefano Zampini } 64064f1b2e48SStefano Zampini ierr = PetscFree(p0_check);CHKERRQ(ierr); 6407fa34dd3eSStefano Zampini ierr = VecDestroy(&vec_scale_P);CHKERRQ(ierr); 6408fa34dd3eSStefano Zampini ierr = VecDestroy(&vec_check_B);CHKERRQ(ierr); 6409fa34dd3eSStefano Zampini ierr = VecDestroy(&dummy_vec);CHKERRQ(ierr); 6410fa34dd3eSStefano Zampini ierr = MatDestroy(&S_j);CHKERRQ(ierr); 6411fa34dd3eSStefano Zampini ierr = MatDestroy(&B0_B);CHKERRQ(ierr); 6412fa34dd3eSStefano Zampini } 6413fa34dd3eSStefano Zampini PetscFunctionReturn(0); 6414fa34dd3eSStefano Zampini } 6415