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; 170*059032f7SStefano Zampini if (reuse) { 171a3df083aSStefano Zampini ctx->benign_zerodiag_subs = reuse->benign_zerodiag_subs; 172*059032f7SStefano Zampini } else { /* TODO: could be optimized for successive solves */ 173*059032f7SStefano Zampini ISLocalToGlobalMapping N_to_D; 174*059032f7SStefano Zampini PetscInt i; 175*059032f7SStefano Zampini 176*059032f7SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(pcis->is_I_local,&N_to_D);CHKERRQ(ierr); 177*059032f7SStefano Zampini ierr = PetscMalloc1(pcbddc->benign_n,&ctx->benign_zerodiag_subs);CHKERRQ(ierr); 178*059032f7SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 179*059032f7SStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(N_to_D,IS_GTOLM_DROP,pcbddc->benign_zerodiag_subs[i],&ctx->benign_zerodiag_subs[i]);CHKERRQ(ierr); 180*059032f7SStefano Zampini } 181*059032f7SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&N_to_D);CHKERRQ(ierr); 182*059032f7SStefano 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 { 195*059032f7SStefano Zampini PCBDDCReuseSolvers reuse = pcbddc->sub_schurs->reuse_solver; 196*059032f7SStefano 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; 204*059032f7SStefano Zampini if (!reuse) { 205*059032f7SStefano Zampini PetscInt i; 206*059032f7SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 207*059032f7SStefano Zampini ierr = ISDestroy(&ctx->benign_zerodiag_subs[i]);CHKERRQ(ierr); 208*059032f7SStefano Zampini } 209*059032f7SStefano Zampini ierr = PetscFree(ctx->benign_zerodiag_subs);CHKERRQ(ierr); 210*059032f7SStefano 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) { 798a3df083aSStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)pcbddc->local_mat),&pcbddc->benign_B0);CHKERRQ(ierr); 799a3df083aSStefano Zampini ierr = MatSetType(pcbddc->benign_B0,MATAIJ);CHKERRQ(ierr); 800a3df083aSStefano Zampini ierr = MatSetSizes(pcbddc->benign_B0,pcbddc->benign_n,n,PETSC_DECIDE,PETSC_DECIDE);CHKERRQ(ierr); 801331e053bSStefano Zampini if (!pcbddc->current_level) { 802331e053bSStefano Zampini /* this matrix is actually sparse: the nnz pattern is not known unless we do 2 sweeps of the next loop. 803331e053bSStefano Zampini Setting nnz=100 should be more than enough for the finest level */ 804a3df083aSStefano Zampini ierr = MatSeqAIJSetPreallocation(pcbddc->benign_B0,100,NULL);CHKERRQ(ierr); 805331e053bSStefano Zampini } else { 806331e053bSStefano Zampini PetscInt *nnz; 807331e053bSStefano Zampini ierr = PetscMalloc1(pcbddc->benign_n,&nnz);CHKERRQ(ierr); 808331e053bSStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 809331e053bSStefano Zampini ierr = ISGetLocalSize(pcbddc->benign_zerodiag_subs[i],&nnz[i]);CHKERRQ(ierr); 810331e053bSStefano Zampini nnz[i] = n - nnz[i]; 811331e053bSStefano Zampini } 812331e053bSStefano Zampini ierr = MatSeqAIJSetPreallocation(pcbddc->benign_B0,0,nnz);CHKERRQ(ierr); 813331e053bSStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 814331e053bSStefano Zampini } 815a3df083aSStefano Zampini } 816a3df083aSStefano Zampini 817a3df083aSStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 818a3df083aSStefano Zampini PetscScalar *array; 819a3df083aSStefano Zampini PetscInt *idxs,j,nz,cum; 820a3df083aSStefano Zampini 821a3df083aSStefano Zampini ierr = VecSet(matis->x,0.);CHKERRQ(ierr); 822a3df083aSStefano Zampini ierr = ISGetLocalSize(pcbddc->benign_zerodiag_subs[i],&nz);CHKERRQ(ierr); 823a3df083aSStefano Zampini ierr = ISGetIndices(pcbddc->benign_zerodiag_subs[i],(const PetscInt**)&idxs);CHKERRQ(ierr); 824a3df083aSStefano Zampini for (j=0;j<nz;j++) vals[j] = 1.; 825a3df083aSStefano Zampini ierr = VecSetValues(matis->x,nz,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 826a3df083aSStefano Zampini ierr = VecAssemblyBegin(matis->x);CHKERRQ(ierr); 827a3df083aSStefano Zampini ierr = VecAssemblyEnd(matis->x);CHKERRQ(ierr); 828a3df083aSStefano Zampini ierr = VecSet(matis->y,0.);CHKERRQ(ierr); 829a3df083aSStefano Zampini ierr = MatMult(matis->A,matis->x,matis->y);CHKERRQ(ierr); 830a3df083aSStefano Zampini ierr = VecGetArray(matis->y,&array);CHKERRQ(ierr); 831a3df083aSStefano Zampini cum = 0; 832a3df083aSStefano Zampini for (j=0;j<n;j++) { 833a3df083aSStefano Zampini if (PetscUnlikely(PetscAbsReal(array[j]) > PETSC_SMALL)) { 834a3df083aSStefano Zampini vals[cum] = array[j]; 835a3df083aSStefano Zampini idxs_ins[cum] = j; 836a3df083aSStefano Zampini cum++; 837a3df083aSStefano Zampini } 838a3df083aSStefano Zampini } 839a3df083aSStefano Zampini ierr = MatSetValues(pcbddc->benign_B0,1,&i,cum,idxs_ins,vals,INSERT_VALUES);CHKERRQ(ierr); 840a3df083aSStefano Zampini ierr = VecRestoreArray(matis->y,&array);CHKERRQ(ierr); 841a3df083aSStefano Zampini ierr = ISRestoreIndices(pcbddc->benign_zerodiag_subs[i],(const PetscInt**)&idxs);CHKERRQ(ierr); 842a3df083aSStefano Zampini } 843a3df083aSStefano Zampini ierr = MatAssemblyBegin(pcbddc->benign_B0,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 844a3df083aSStefano Zampini ierr = MatAssemblyEnd(pcbddc->benign_B0,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 845a3df083aSStefano Zampini ierr = PetscFree2(idxs_ins,vals);CHKERRQ(ierr); 846a3df083aSStefano Zampini } 847c263805aSStefano Zampini } else { /* push */ 848a3df083aSStefano Zampini if (pcbddc->benign_change_explicit) { 8494f1b2e48SStefano Zampini PetscInt i; 8504f1b2e48SStefano Zampini 8514f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 8524f1b2e48SStefano Zampini PetscScalar *B0_vals; 8534f1b2e48SStefano Zampini PetscInt *B0_cols,B0_ncol; 8544f1b2e48SStefano Zampini 8554f1b2e48SStefano Zampini ierr = MatGetRow(pcbddc->benign_B0,i,&B0_ncol,(const PetscInt**)&B0_cols,(const PetscScalar**)&B0_vals);CHKERRQ(ierr); 8564f1b2e48SStefano Zampini ierr = MatSetValues(pcbddc->local_mat,1,pcbddc->benign_p0_lidx+i,B0_ncol,B0_cols,B0_vals,INSERT_VALUES);CHKERRQ(ierr); 8577b034428SStefano Zampini ierr = MatSetValues(pcbddc->local_mat,B0_ncol,B0_cols,1,pcbddc->benign_p0_lidx+i,B0_vals,INSERT_VALUES);CHKERRQ(ierr); 8584f1b2e48SStefano Zampini ierr = MatSetValue(pcbddc->local_mat,pcbddc->benign_p0_lidx[i],pcbddc->benign_p0_lidx[i],0.0,INSERT_VALUES);CHKERRQ(ierr); 8594f1b2e48SStefano Zampini ierr = MatRestoreRow(pcbddc->benign_B0,i,&B0_ncol,(const PetscInt**)&B0_cols,(const PetscScalar**)&B0_vals);CHKERRQ(ierr); 8604f1b2e48SStefano Zampini } 861c263805aSStefano Zampini ierr = MatAssemblyBegin(pcbddc->local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 862c263805aSStefano Zampini ierr = MatAssemblyEnd(pcbddc->local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 863a3df083aSStefano Zampini } else { 864a3df083aSStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Cannot push B0!\n"); 865a3df083aSStefano Zampini } 866c263805aSStefano Zampini } 867c263805aSStefano Zampini PetscFunctionReturn(0); 868c263805aSStefano Zampini } 869c263805aSStefano Zampini 870c263805aSStefano Zampini #undef __FUNCT__ 871b1b3d7a2SStefano Zampini #define __FUNCT__ "PCBDDCAdaptiveSelection" 87208122e43SStefano Zampini PetscErrorCode PCBDDCAdaptiveSelection(PC pc) 873b1b3d7a2SStefano Zampini { 874b1b3d7a2SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 87508122e43SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 87608122e43SStefano Zampini PetscBLASInt B_dummyint,B_neigs,B_ierr,B_lwork; 87708122e43SStefano Zampini PetscBLASInt *B_iwork,*B_ifail; 87808122e43SStefano Zampini PetscScalar *work,lwork; 87908122e43SStefano Zampini PetscScalar *St,*S,*eigv; 88008122e43SStefano Zampini PetscScalar *Sarray,*Starray; 88108122e43SStefano Zampini PetscReal *eigs,thresh; 8821b968477SStefano Zampini PetscInt i,nmax,nmin,nv,cum,mss,cum2,cumarray,maxneigs; 883f6f667cfSStefano Zampini PetscBool allocated_S_St; 88408122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 88508122e43SStefano Zampini PetscReal *rwork; 88608122e43SStefano Zampini #endif 887b1b3d7a2SStefano Zampini PetscErrorCode ierr; 888b1b3d7a2SStefano Zampini 889b1b3d7a2SStefano Zampini PetscFunctionBegin; 890df4d28bfSStefano Zampini if (!sub_schurs->schur_explicit) { 891df4d28bfSStefano Zampini SETERRQ(PetscObjectComm((PetscObject)pc),PETSC_ERR_SUP,"Adaptive selection of constraints requires MUMPS and/or MKL_CPARDISO"); 89208122e43SStefano Zampini } 89308122e43SStefano Zampini 89406a4e24aSStefano Zampini if (sub_schurs->n_subs && (!sub_schurs->is_hermitian || !sub_schurs->is_posdef)) { 89506a4e24aSStefano 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); 89606a4e24aSStefano Zampini } 89706a4e24aSStefano Zampini 898fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 899fd14bc51SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 900fd14bc51SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 901fd14bc51SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Check adaptive selection of constraints\n");CHKERRQ(ierr); 9021575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 903fd14bc51SStefano Zampini } 904fd14bc51SStefano Zampini 905e496cd5dSStefano Zampini if (pcbddc->dbg_flag) { 906e496cd5dSStefano 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); 907e496cd5dSStefano Zampini } 908e496cd5dSStefano Zampini 90908122e43SStefano Zampini /* max size of subsets */ 91008122e43SStefano Zampini mss = 0; 91108122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 91208122e43SStefano Zampini PetscInt subset_size; 913862806e4SStefano Zampini 91408122e43SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 91508122e43SStefano Zampini mss = PetscMax(mss,subset_size); 91608122e43SStefano Zampini } 91708122e43SStefano Zampini 91808122e43SStefano Zampini /* min/max and threshold */ 91908122e43SStefano Zampini nmax = pcbddc->adaptive_nmax > 0 ? pcbddc->adaptive_nmax : mss; 920f6f667cfSStefano Zampini nmin = pcbddc->adaptive_nmin > 0 ? pcbddc->adaptive_nmin : 0; 92108122e43SStefano Zampini nmax = PetscMax(nmin,nmax); 922f6f667cfSStefano Zampini allocated_S_St = PETSC_FALSE; 923f6f667cfSStefano Zampini if (nmin) { 924f6f667cfSStefano Zampini allocated_S_St = PETSC_TRUE; 925f6f667cfSStefano Zampini } 92608122e43SStefano Zampini 92708122e43SStefano Zampini /* allocate lapack workspace */ 92808122e43SStefano Zampini cum = cum2 = 0; 92908122e43SStefano Zampini maxneigs = 0; 93008122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 93108122e43SStefano Zampini PetscInt n,subset_size; 932f6f667cfSStefano Zampini 93308122e43SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 93408122e43SStefano Zampini n = PetscMin(subset_size,nmax); 9359162d606SStefano Zampini cum += subset_size; 9369162d606SStefano Zampini cum2 += subset_size*n; 93708122e43SStefano Zampini maxneigs = PetscMax(maxneigs,n); 93808122e43SStefano Zampini } 93908122e43SStefano Zampini if (mss) { 9409ab7bb16SStefano Zampini if (sub_schurs->is_hermitian && sub_schurs->is_posdef) { 94108122e43SStefano Zampini PetscBLASInt B_itype = 1; 94208122e43SStefano Zampini PetscBLASInt B_N = mss; 9434c6709b3SStefano Zampini PetscReal zero = 0.0; 9444c6709b3SStefano Zampini PetscReal eps = 0.0; /* dlamch? */ 94508122e43SStefano Zampini 94608122e43SStefano Zampini B_lwork = -1; 94708122e43SStefano Zampini S = NULL; 94808122e43SStefano Zampini St = NULL; 949a58a30b4SStefano Zampini eigs = NULL; 950a58a30b4SStefano Zampini eigv = NULL; 951a58a30b4SStefano Zampini B_iwork = NULL; 952a58a30b4SStefano Zampini B_ifail = NULL; 953d1710679SStefano Zampini #if defined(PETSC_USE_COMPLEX) 954d1710679SStefano Zampini rwork = NULL; 955d1710679SStefano Zampini #endif 9568bec7fa6SStefano Zampini thresh = 1.0; 95708122e43SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 95808122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 95908122e43SStefano 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)); 96008122e43SStefano Zampini #else 96108122e43SStefano 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)); 96208122e43SStefano Zampini #endif 96308122e43SStefano Zampini if (B_ierr != 0) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to SYGVX Lapack routine %d",(int)B_ierr); 96408122e43SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 96508122e43SStefano Zampini } else { 96608122e43SStefano Zampini /* TODO */ 96708122e43SStefano Zampini } 96808122e43SStefano Zampini } else { 96908122e43SStefano Zampini lwork = 0; 97008122e43SStefano Zampini } 97108122e43SStefano Zampini 97208122e43SStefano Zampini nv = 0; 973d62866d3SStefano 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) */ 974d62866d3SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_vertices,&nv);CHKERRQ(ierr); 97508122e43SStefano Zampini } 9764c6709b3SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lwork),&B_lwork);CHKERRQ(ierr); 977f6f667cfSStefano Zampini if (allocated_S_St) { 978f6f667cfSStefano Zampini ierr = PetscMalloc2(mss*mss,&S,mss*mss,&St);CHKERRQ(ierr); 979f6f667cfSStefano Zampini } 980f6f667cfSStefano Zampini ierr = PetscMalloc5(mss*mss,&eigv,mss,&eigs,B_lwork,&work,5*mss,&B_iwork,mss,&B_ifail);CHKERRQ(ierr); 98108122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 98208122e43SStefano Zampini ierr = PetscMalloc1(7*mss,&rwork);CHKERRQ(ierr); 98308122e43SStefano Zampini #endif 9849162d606SStefano Zampini ierr = PetscMalloc5(nv+sub_schurs->n_subs,&pcbddc->adaptive_constraints_n, 9859162d606SStefano Zampini nv+sub_schurs->n_subs+1,&pcbddc->adaptive_constraints_idxs_ptr, 9869162d606SStefano Zampini nv+sub_schurs->n_subs+1,&pcbddc->adaptive_constraints_data_ptr, 98708122e43SStefano Zampini nv+cum,&pcbddc->adaptive_constraints_idxs, 9889162d606SStefano Zampini nv+cum2,&pcbddc->adaptive_constraints_data);CHKERRQ(ierr); 98908122e43SStefano Zampini ierr = PetscMemzero(pcbddc->adaptive_constraints_n,(nv+sub_schurs->n_subs)*sizeof(PetscInt));CHKERRQ(ierr); 99008122e43SStefano Zampini 99108122e43SStefano Zampini maxneigs = 0; 99208122e43SStefano Zampini cum = cum2 = cumarray = 0; 9939162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr[0] = 0; 9949162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr[0] = 0; 995d62866d3SStefano Zampini if (sub_schurs->is_vertices && pcbddc->use_vertices) { 99608122e43SStefano Zampini const PetscInt *idxs; 99708122e43SStefano Zampini 998d62866d3SStefano Zampini ierr = ISGetIndices(sub_schurs->is_vertices,&idxs);CHKERRQ(ierr); 99908122e43SStefano Zampini for (cum=0;cum<nv;cum++) { 100008122e43SStefano Zampini pcbddc->adaptive_constraints_n[cum] = 1; 100108122e43SStefano Zampini pcbddc->adaptive_constraints_idxs[cum] = idxs[cum]; 100208122e43SStefano Zampini pcbddc->adaptive_constraints_data[cum] = 1.0; 10039162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr[cum+1] = pcbddc->adaptive_constraints_idxs_ptr[cum]+1; 10049162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr[cum+1] = pcbddc->adaptive_constraints_data_ptr[cum]+1; 100508122e43SStefano Zampini } 100608122e43SStefano Zampini cum2 = cum; 1007d62866d3SStefano Zampini ierr = ISRestoreIndices(sub_schurs->is_vertices,&idxs);CHKERRQ(ierr); 100808122e43SStefano Zampini } 100908122e43SStefano Zampini 101008122e43SStefano Zampini if (mss) { /* multilevel */ 101108122e43SStefano Zampini ierr = MatSeqAIJGetArray(sub_schurs->sum_S_Ej_inv_all,&Sarray);CHKERRQ(ierr); 101208122e43SStefano Zampini ierr = MatSeqAIJGetArray(sub_schurs->sum_S_Ej_tilda_all,&Starray);CHKERRQ(ierr); 101308122e43SStefano Zampini } 101408122e43SStefano Zampini 1015ffd830a3SStefano Zampini thresh = pcbddc->adaptive_threshold; 101608122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 101708122e43SStefano Zampini const PetscInt *idxs; 1018f6f667cfSStefano Zampini PetscReal infty = PETSC_MAX_REAL; 1019862806e4SStefano Zampini PetscInt j,subset_size,eigs_start = 0; 102008122e43SStefano Zampini PetscBLASInt B_N; 1021aff50787SStefano Zampini PetscBool same_data = PETSC_FALSE; 102208122e43SStefano Zampini 1023862806e4SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 1024ffd830a3SStefano Zampini ierr = ISGetIndices(sub_schurs->is_subs[i],&idxs);CHKERRQ(ierr); 1025f6f667cfSStefano Zampini ierr = PetscBLASIntCast(subset_size,&B_N);CHKERRQ(ierr); 1026f6f667cfSStefano Zampini if (allocated_S_St) { /* S and S_t should be copied since we could need them later */ 10279ab7bb16SStefano Zampini if (sub_schurs->is_hermitian) { 1028aff50787SStefano Zampini PetscInt j,k; 1029aff50787SStefano Zampini if (sub_schurs->n_subs == 1) { /* zeroing memory to use PetscMemcmp later */ 1030aff50787SStefano Zampini ierr = PetscMemzero(S,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 1031aff50787SStefano Zampini ierr = PetscMemzero(St,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 103208122e43SStefano Zampini } 103308122e43SStefano Zampini for (j=0;j<subset_size;j++) { 1034aff50787SStefano Zampini for (k=j;k<subset_size;k++) { 1035aff50787SStefano Zampini S [j*subset_size+k] = Sarray [cumarray+j*subset_size+k]; 1036aff50787SStefano Zampini St[j*subset_size+k] = Starray[cumarray+j*subset_size+k]; 1037aff50787SStefano Zampini } 103808122e43SStefano Zampini } 103908122e43SStefano Zampini } else { 104008122e43SStefano Zampini ierr = PetscMemcpy(S,Sarray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 104108122e43SStefano Zampini ierr = PetscMemcpy(St,Starray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 104208122e43SStefano Zampini } 10438bec7fa6SStefano Zampini } else { 1044f6f667cfSStefano Zampini S = Sarray + cumarray; 1045f6f667cfSStefano Zampini St = Starray + cumarray; 10468bec7fa6SStefano Zampini } 104708122e43SStefano Zampini 1048aff50787SStefano Zampini /* see if we can save some work */ 1049aff50787SStefano Zampini if (sub_schurs->n_subs == 1) { 1050aff50787SStefano Zampini ierr = PetscMemcmp(S,St,subset_size*subset_size*sizeof(PetscScalar),&same_data);CHKERRQ(ierr); 1051aff50787SStefano Zampini } 1052aff50787SStefano Zampini 1053aff50787SStefano Zampini if (same_data) { /* there's no need of constraints here, deluxe scaling is enough */ 1054aff50787SStefano Zampini B_neigs = 0; 1055aff50787SStefano Zampini } else { 10569ab7bb16SStefano Zampini if (sub_schurs->is_hermitian && sub_schurs->is_posdef) { 105708122e43SStefano Zampini PetscBLASInt B_itype = 1; 1058f6f667cfSStefano Zampini PetscBLASInt B_IL, B_IU; 10594c6709b3SStefano Zampini PetscReal eps = -1.0; /* dlamch? */ 10609552c7c7SStefano Zampini PetscInt nmin_s; 106108122e43SStefano Zampini 1062fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 10638bec7fa6SStefano 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]]); 1064fd14bc51SStefano Zampini } 1065d16cbb6bSStefano Zampini 106608122e43SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 1067d16cbb6bSStefano Zampini if (thresh > 1.+PETSC_SMALL) { 1068d16cbb6bSStefano Zampini 1069d16cbb6bSStefano Zampini /* ask for eigenvalues larger than thresh */ 107008122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 1071f6f667cfSStefano 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)); 107208122e43SStefano Zampini #else 1073f6f667cfSStefano 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)); 107408122e43SStefano Zampini #endif 1075d16cbb6bSStefano Zampini } else { 1076d16cbb6bSStefano Zampini B_IU = PetscMax(1,PetscMin(B_N,nmax)); 1077d16cbb6bSStefano Zampini B_IL = 1; 1078d16cbb6bSStefano Zampini #if defined(PETSC_USE_COMPLEX) 1079d16cbb6bSStefano 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)); 1080d16cbb6bSStefano Zampini #else 1081d16cbb6bSStefano 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)); 1082d16cbb6bSStefano Zampini #endif 1083d16cbb6bSStefano Zampini } 108408122e43SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 108508122e43SStefano Zampini if (B_ierr) { 108608122e43SStefano Zampini if (B_ierr < 0 ) { 108708122e43SStefano Zampini SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYGVX Lapack routine: illegal value for argument %d",-(int)B_ierr); 108808122e43SStefano Zampini } else if (B_ierr <= B_N) { 108908122e43SStefano Zampini SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYGVX Lapack routine: %d eigenvalues failed to converge",(int)B_ierr); 109008122e43SStefano Zampini } else { 10919552c7c7SStefano 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); 109208122e43SStefano Zampini } 109308122e43SStefano Zampini } 109408122e43SStefano Zampini 109508122e43SStefano Zampini if (B_neigs > nmax) { 1096fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 1097fd14bc51SStefano Zampini PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," found %d eigs, more than maximum required %d.\n",B_neigs,nmax); 1098fd14bc51SStefano Zampini } 1099f6f667cfSStefano Zampini eigs_start = B_neigs -nmax; 110008122e43SStefano Zampini B_neigs = nmax; 110108122e43SStefano Zampini } 110208122e43SStefano Zampini 11039552c7c7SStefano Zampini nmin_s = PetscMin(nmin,B_N); 11049552c7c7SStefano Zampini if (B_neigs < nmin_s) { 110508122e43SStefano Zampini PetscBLASInt B_neigs2; 110608122e43SStefano Zampini 1107f6f667cfSStefano Zampini B_IU = B_N - B_neigs; 1108f6f667cfSStefano Zampini B_IL = B_N - nmin_s + 1; 1109fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 1110fd14bc51SStefano 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); 1111fd14bc51SStefano Zampini } 11129ab7bb16SStefano Zampini if (sub_schurs->is_hermitian) { 111308122e43SStefano Zampini PetscInt j; 111408122e43SStefano Zampini for (j=0;j<subset_size;j++) { 111508122e43SStefano Zampini ierr = PetscMemcpy(S+j*(subset_size+1),Sarray+cumarray+j*(subset_size+1),(subset_size-j)*sizeof(PetscScalar));CHKERRQ(ierr); 111608122e43SStefano Zampini } 111708122e43SStefano Zampini for (j=0;j<subset_size;j++) { 111808122e43SStefano Zampini ierr = PetscMemcpy(St+j*(subset_size+1),Starray+cumarray+j*(subset_size+1),(subset_size-j)*sizeof(PetscScalar));CHKERRQ(ierr); 111908122e43SStefano Zampini } 112008122e43SStefano Zampini } else { 112108122e43SStefano Zampini ierr = PetscMemcpy(S,Sarray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 112208122e43SStefano Zampini ierr = PetscMemcpy(St,Starray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 112308122e43SStefano Zampini } 112408122e43SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 112508122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 1126f6f667cfSStefano 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)); 112708122e43SStefano Zampini #else 1128f6f667cfSStefano 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)); 112908122e43SStefano Zampini #endif 113008122e43SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 113108122e43SStefano Zampini B_neigs += B_neigs2; 113208122e43SStefano Zampini } 113308122e43SStefano Zampini if (B_ierr) { 113408122e43SStefano Zampini if (B_ierr < 0 ) { 113508122e43SStefano Zampini SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYGVX Lapack routine: illegal value for argument %d",-(int)B_ierr); 113608122e43SStefano Zampini } else if (B_ierr <= B_N) { 113708122e43SStefano Zampini SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYGVX Lapack routine: %d eigenvalues failed to converge",(int)B_ierr); 113808122e43SStefano Zampini } else { 11399552c7c7SStefano 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); 114008122e43SStefano Zampini } 114108122e43SStefano Zampini } 1142fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 1143ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," -> Got %d eigs\n",B_neigs);CHKERRQ(ierr); 114408122e43SStefano Zampini for (j=0;j<B_neigs;j++) { 114508122e43SStefano Zampini if (eigs[j] == 0.0) { 1146ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," Inf\n");CHKERRQ(ierr); 114708122e43SStefano Zampini } else { 1148ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," %1.6e\n",eigs[j+eigs_start]);CHKERRQ(ierr); 1149fd14bc51SStefano Zampini } 115008122e43SStefano Zampini } 115108122e43SStefano Zampini } 115208122e43SStefano Zampini } else { 115308122e43SStefano Zampini /* TODO */ 115408122e43SStefano Zampini } 1155aff50787SStefano Zampini } 11568bec7fa6SStefano Zampini maxneigs = PetscMax(B_neigs,maxneigs); 11578bec7fa6SStefano Zampini pcbddc->adaptive_constraints_n[i+nv] = B_neigs; 11589162d606SStefano Zampini if (B_neigs) { 11599162d606SStefano 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); 1160fd14bc51SStefano Zampini 1161fd14bc51SStefano Zampini if (pcbddc->dbg_flag > 1) { 11629552c7c7SStefano Zampini PetscInt ii; 11639552c7c7SStefano Zampini for (ii=0;ii<B_neigs;ii++) { 1164ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," -> Eigenvector %d/%d (%d)\n",ii,B_neigs,B_N);CHKERRQ(ierr); 11659552c7c7SStefano Zampini for (j=0;j<B_N;j++) { 1166ac47001eSStefano Zampini #if defined(PETSC_USE_COMPLEX) 1167ac47001eSStefano Zampini PetscReal r = PetscRealPart(pcbddc->adaptive_constraints_data[ii*subset_size+j+pcbddc->adaptive_constraints_data_ptr[cum]]); 1168ac47001eSStefano Zampini PetscReal c = PetscImaginaryPart(pcbddc->adaptive_constraints_data[ii*subset_size+j+pcbddc->adaptive_constraints_data_ptr[cum]]); 1169ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," %1.4e + %1.4e i\n",r,c);CHKERRQ(ierr); 1170ac47001eSStefano Zampini #else 1171ac47001eSStefano 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); 1172ac47001eSStefano Zampini #endif 11739552c7c7SStefano Zampini } 11749552c7c7SStefano Zampini } 1175fd14bc51SStefano Zampini } 117608122e43SStefano Zampini #if 0 11779162d606SStefano Zampini for (j=0;j<B_neigs;j++) { 117808122e43SStefano Zampini PetscBLASInt Blas_N,Blas_one = 1.0; 117908122e43SStefano Zampini PetscScalar norm; 118008122e43SStefano Zampini ierr = PetscBLASIntCast(subset_size,&Blas_N);CHKERRQ(ierr); 11819162d606SStefano Zampini PetscStackCallBLAS("BLASdot",norm = BLASdot_(&Blas_N,pcbddc->adaptive_constraints_data+pcbddc->adaptive_constraints_data_ptr[cum]+j*subset_size, 11829162d606SStefano Zampini &Blas_one,pcbddc->adaptive_constraints_data+pcbddc->adaptive_constraints_data_ptr[cum]+j*subset_size,&Blas_one)); 118308122e43SStefano Zampini if (pcbddc->adaptive_constraints_data[cum2] > 0.0) { 118408122e43SStefano Zampini norm = 1.0/PetscSqrtReal(PetscRealPart(norm)); 118508122e43SStefano Zampini } else { 118608122e43SStefano Zampini norm = -1.0/PetscSqrtReal(PetscRealPart(norm)); 118708122e43SStefano Zampini } 11889162d606SStefano Zampini PetscStackCallBLAS("BLASscal",BLASscal_(&Blas_N,&norm,pcbddc->adaptive_constraints_data+pcbddc->adaptive_constraints_data_ptr[cum]+j*subset_size,&Blas_one)); 1189b1b3d7a2SStefano Zampini } 1190b1b3d7a2SStefano Zampini #endif 11919162d606SStefano Zampini ierr = PetscMemcpy(pcbddc->adaptive_constraints_idxs+pcbddc->adaptive_constraints_idxs_ptr[cum],idxs,subset_size*sizeof(PetscInt));CHKERRQ(ierr); 11929162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr[cum+1] = pcbddc->adaptive_constraints_idxs_ptr[cum] + subset_size; 11939162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr[cum+1] = pcbddc->adaptive_constraints_data_ptr[cum] + subset_size*B_neigs; 11949162d606SStefano Zampini cum++; 119508122e43SStefano Zampini } 119608122e43SStefano Zampini ierr = ISRestoreIndices(sub_schurs->is_subs[i],&idxs);CHKERRQ(ierr); 119708122e43SStefano Zampini /* shift for next computation */ 119808122e43SStefano Zampini cumarray += subset_size*subset_size; 119908122e43SStefano Zampini } 1200fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 1201fd14bc51SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1202fd14bc51SStefano Zampini } 120308122e43SStefano Zampini 120408122e43SStefano Zampini if (mss) { 120508122e43SStefano Zampini ierr = MatSeqAIJRestoreArray(sub_schurs->sum_S_Ej_inv_all,&Sarray);CHKERRQ(ierr); 120608122e43SStefano Zampini ierr = MatSeqAIJRestoreArray(sub_schurs->sum_S_Ej_tilda_all,&Starray);CHKERRQ(ierr); 1207f6f667cfSStefano Zampini /* destroy matrices (junk) */ 1208f6f667cfSStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_inv_all);CHKERRQ(ierr); 1209f6f667cfSStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_tilda_all);CHKERRQ(ierr); 121008122e43SStefano Zampini } 1211f6f667cfSStefano Zampini if (allocated_S_St) { 1212f6f667cfSStefano Zampini ierr = PetscFree2(S,St);CHKERRQ(ierr); 1213f6f667cfSStefano Zampini } 1214f6f667cfSStefano Zampini ierr = PetscFree5(eigv,eigs,work,B_iwork,B_ifail);CHKERRQ(ierr); 121508122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 121608122e43SStefano Zampini ierr = PetscFree(rwork);CHKERRQ(ierr); 121708122e43SStefano Zampini #endif 121808122e43SStefano Zampini if (pcbddc->dbg_flag) { 12191b968477SStefano Zampini PetscInt maxneigs_r; 122008122e43SStefano Zampini ierr = MPI_Allreduce(&maxneigs,&maxneigs_r,1,MPIU_INT,MPI_MAX,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 12219b28b941SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Maximum number of constraints per cc %d\n",maxneigs_r);CHKERRQ(ierr); 122208122e43SStefano Zampini } 122308122e43SStefano Zampini PetscFunctionReturn(0); 122408122e43SStefano Zampini } 1225b1b3d7a2SStefano Zampini 1226674ae819SStefano Zampini #undef __FUNCT__ 1227c8587f34SStefano Zampini #define __FUNCT__ "PCBDDCSetUpSolvers" 1228c8587f34SStefano Zampini PetscErrorCode PCBDDCSetUpSolvers(PC pc) 1229c8587f34SStefano Zampini { 1230c8587f34SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 12318629588bSStefano Zampini PetscScalar *coarse_submat_vals; 1232c8587f34SStefano Zampini PetscErrorCode ierr; 1233c8587f34SStefano Zampini 1234c8587f34SStefano Zampini PetscFunctionBegin; 1235f4ddd8eeSStefano Zampini /* Setup local scatters R_to_B and (optionally) R_to_D */ 12365e8657edSStefano Zampini /* PCBDDCSetUpLocalWorkVectors should be called first! */ 1237c8587f34SStefano Zampini ierr = PCBDDCSetUpLocalScatters(pc);CHKERRQ(ierr); 1238c8587f34SStefano Zampini 1239684f6988SStefano Zampini /* Setup local neumann solver ksp_R */ 12400fccc4e9SStefano Zampini /* PCBDDCSetUpLocalScatters should be called first! */ 1241684f6988SStefano Zampini ierr = PCBDDCSetUpLocalSolvers(pc,PETSC_FALSE,PETSC_TRUE);CHKERRQ(ierr); 1242c8587f34SStefano Zampini 1243c8587f34SStefano Zampini /* Change global null space passed in by the user if change of basis has been requested */ 1244b9b85e73SStefano Zampini if (pcbddc->NullSpace && pcbddc->ChangeOfBasisMatrix) { 1245c8587f34SStefano Zampini ierr = PCBDDCNullSpaceAdaptGlobal(pc);CHKERRQ(ierr); 1246c8587f34SStefano Zampini } 1247c8587f34SStefano Zampini 12488629588bSStefano Zampini /* 12498629588bSStefano Zampini Setup local correction and local part of coarse basis. 12508629588bSStefano Zampini Gives back the dense local part of the coarse matrix in column major ordering 12518629588bSStefano Zampini */ 125247f4ddc3SStefano Zampini ierr = PCBDDCSetUpCorrection(pc,&coarse_submat_vals);CHKERRQ(ierr); 12538629588bSStefano Zampini 12548629588bSStefano Zampini /* Compute total number of coarse nodes and setup coarse solver */ 12558629588bSStefano Zampini ierr = PCBDDCSetUpCoarseSolver(pc,coarse_submat_vals);CHKERRQ(ierr); 12568629588bSStefano Zampini 12578629588bSStefano Zampini /* free */ 12588629588bSStefano Zampini ierr = PetscFree(coarse_submat_vals);CHKERRQ(ierr); 1259c8587f34SStefano Zampini PetscFunctionReturn(0); 1260c8587f34SStefano Zampini } 1261c8587f34SStefano Zampini 1262c8587f34SStefano Zampini #undef __FUNCT__ 1263674ae819SStefano Zampini #define __FUNCT__ "PCBDDCResetCustomization" 1264674ae819SStefano Zampini PetscErrorCode PCBDDCResetCustomization(PC pc) 1265674ae819SStefano Zampini { 1266674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 1267674ae819SStefano Zampini PetscErrorCode ierr; 1268674ae819SStefano Zampini 1269674ae819SStefano Zampini PetscFunctionBegin; 1270674ae819SStefano Zampini ierr = PCBDDCGraphResetCSR(pcbddc->mat_graph);CHKERRQ(ierr); 1271674ae819SStefano Zampini ierr = ISDestroy(&pcbddc->user_primal_vertices);CHKERRQ(ierr); 127230368db7SStefano Zampini ierr = ISDestroy(&pcbddc->user_primal_vertices_local);CHKERRQ(ierr); 1273674ae819SStefano Zampini ierr = MatNullSpaceDestroy(&pcbddc->NullSpace);CHKERRQ(ierr); 1274674ae819SStefano Zampini ierr = ISDestroy(&pcbddc->NeumannBoundaries);CHKERRQ(ierr); 1275785d1243SStefano Zampini ierr = ISDestroy(&pcbddc->NeumannBoundariesLocal);CHKERRQ(ierr); 1276674ae819SStefano Zampini ierr = ISDestroy(&pcbddc->DirichletBoundaries);CHKERRQ(ierr); 1277f4ddd8eeSStefano Zampini ierr = MatNullSpaceDestroy(&pcbddc->onearnullspace);CHKERRQ(ierr); 1278f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->onearnullvecs_state);CHKERRQ(ierr); 1279785d1243SStefano Zampini ierr = ISDestroy(&pcbddc->DirichletBoundariesLocal);CHKERRQ(ierr); 128063602bcaSStefano Zampini ierr = PCBDDCSetDofsSplitting(pc,0,NULL);CHKERRQ(ierr); 128163602bcaSStefano Zampini ierr = PCBDDCSetDofsSplittingLocal(pc,0,NULL);CHKERRQ(ierr); 1282674ae819SStefano Zampini PetscFunctionReturn(0); 1283674ae819SStefano Zampini } 1284674ae819SStefano Zampini 1285674ae819SStefano Zampini #undef __FUNCT__ 1286674ae819SStefano Zampini #define __FUNCT__ "PCBDDCResetTopography" 1287674ae819SStefano Zampini PetscErrorCode PCBDDCResetTopography(PC pc) 1288674ae819SStefano Zampini { 1289674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 12904f1b2e48SStefano Zampini PetscInt i; 1291674ae819SStefano Zampini PetscErrorCode ierr; 1292674ae819SStefano Zampini 1293674ae819SStefano Zampini PetscFunctionBegin; 1294b9b85e73SStefano Zampini ierr = MatDestroy(&pcbddc->user_ChangeOfBasisMatrix);CHKERRQ(ierr); 1295674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 1296674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 1297674ae819SStefano Zampini ierr = PCBDDCGraphReset(pcbddc->mat_graph);CHKERRQ(ierr); 12984f1b2e48SStefano Zampini for (i=0;i<pcbddc->n_local_subs;i++) { 12994f1b2e48SStefano Zampini ierr = ISDestroy(&pcbddc->local_subs[i]);CHKERRQ(ierr); 13004f1b2e48SStefano Zampini } 13014f1b2e48SStefano Zampini ierr = PetscFree(pcbddc->local_subs);CHKERRQ(ierr); 1302b96c3477SStefano Zampini ierr = PCBDDCSubSchursReset(pcbddc->sub_schurs);CHKERRQ(ierr); 1303674ae819SStefano Zampini PetscFunctionReturn(0); 1304674ae819SStefano Zampini } 1305674ae819SStefano Zampini 1306674ae819SStefano Zampini #undef __FUNCT__ 1307674ae819SStefano Zampini #define __FUNCT__ "PCBDDCResetSolvers" 1308674ae819SStefano Zampini PetscErrorCode PCBDDCResetSolvers(PC pc) 1309674ae819SStefano Zampini { 1310674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 1311674ae819SStefano Zampini PetscErrorCode ierr; 1312674ae819SStefano Zampini 1313674ae819SStefano Zampini PetscFunctionBegin; 1314674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->coarse_vec);CHKERRQ(ierr); 131558da7f69SStefano Zampini if (pcbddc->coarse_phi_B) { 1316ca92afb2SStefano Zampini PetscScalar *array; 131706656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&array);CHKERRQ(ierr); 131806656605SStefano Zampini ierr = PetscFree(array);CHKERRQ(ierr); 131958da7f69SStefano Zampini } 1320674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_B);CHKERRQ(ierr); 1321674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_D);CHKERRQ(ierr); 132215aaf578SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_B);CHKERRQ(ierr); 132315aaf578SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_D);CHKERRQ(ierr); 1324674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec1_P);CHKERRQ(ierr); 1325674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec1_C);CHKERRQ(ierr); 1326674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat2);CHKERRQ(ierr); 132706656605SStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat1);CHKERRQ(ierr); 1328674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec1_R);CHKERRQ(ierr); 1329674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec2_R);CHKERRQ(ierr); 13308ce42a96SStefano Zampini ierr = ISDestroy(&pcbddc->is_R_local);CHKERRQ(ierr); 1331674ae819SStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_B);CHKERRQ(ierr); 1332674ae819SStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_D);CHKERRQ(ierr); 1333674ae819SStefano Zampini ierr = VecScatterDestroy(&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 1334f4ddd8eeSStefano Zampini ierr = KSPDestroy(&pcbddc->ksp_D);CHKERRQ(ierr); 1335f4ddd8eeSStefano Zampini ierr = KSPDestroy(&pcbddc->ksp_R);CHKERRQ(ierr); 1336f4ddd8eeSStefano Zampini ierr = KSPDestroy(&pcbddc->coarse_ksp);CHKERRQ(ierr); 1337f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 1338727cdba6SStefano Zampini ierr = PetscFree(pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 13390e6343abSStefano Zampini ierr = PetscFree2(pcbddc->local_primal_ref_node,pcbddc->local_primal_ref_mult);CHKERRQ(ierr); 1340f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->global_primal_indices);CHKERRQ(ierr); 134170cf5478SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling);CHKERRQ(ierr); 13426e683305SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 134381d14e9dSStefano Zampini ierr = MatDestroy(&pcbddc->benign_change);CHKERRQ(ierr); 13440369aaf7SStefano Zampini ierr = VecDestroy(&pcbddc->benign_vec);CHKERRQ(ierr); 13458b9f24d4SStefano Zampini ierr = MatDestroy(&pcbddc->benign_original_mat);CHKERRQ(ierr); 13464f1b2e48SStefano Zampini ierr = MatDestroy(&pcbddc->benign_B0);CHKERRQ(ierr); 13478b9f24d4SStefano Zampini ierr = PetscSFDestroy(&pcbddc->benign_sf);CHKERRQ(ierr); 1348ca92afb2SStefano Zampini if (pcbddc->benign_zerodiag_subs) { 1349ca92afb2SStefano Zampini PetscInt i; 1350ca92afb2SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 1351ca92afb2SStefano Zampini ierr = ISDestroy(&pcbddc->benign_zerodiag_subs[i]);CHKERRQ(ierr); 1352ca92afb2SStefano Zampini } 1353ca92afb2SStefano Zampini ierr = PetscFree(pcbddc->benign_zerodiag_subs);CHKERRQ(ierr); 1354ca92afb2SStefano Zampini } 13554f1b2e48SStefano Zampini ierr = PetscFree3(pcbddc->benign_p0_lidx,pcbddc->benign_p0_gidx,pcbddc->benign_p0);CHKERRQ(ierr); 1356674ae819SStefano Zampini PetscFunctionReturn(0); 1357674ae819SStefano Zampini } 1358674ae819SStefano Zampini 1359674ae819SStefano Zampini #undef __FUNCT__ 1360f4ddd8eeSStefano Zampini #define __FUNCT__ "PCBDDCSetUpLocalWorkVectors" 1361f4ddd8eeSStefano Zampini PetscErrorCode PCBDDCSetUpLocalWorkVectors(PC pc) 13626bfb1811SStefano Zampini { 13636bfb1811SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 13646bfb1811SStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 13656bfb1811SStefano Zampini VecType impVecType; 13664f1b2e48SStefano Zampini PetscInt n_constraints,n_R,old_size; 13676bfb1811SStefano Zampini PetscErrorCode ierr; 13686bfb1811SStefano Zampini 13696bfb1811SStefano Zampini PetscFunctionBegin; 1370f4ddd8eeSStefano Zampini if (!pcbddc->ConstraintMatrix) { 1371019a44ceSStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"BDDC Constraint matrix has not been created"); 1372f4ddd8eeSStefano Zampini } 1373e7b262bdSStefano Zampini /* get sizes */ 13744f1b2e48SStefano Zampini n_constraints = pcbddc->local_primal_size - pcbddc->benign_n - pcbddc->n_vertices; 1375b371cd4fSStefano Zampini n_R = pcis->n - pcbddc->n_vertices; 13766bfb1811SStefano Zampini ierr = VecGetType(pcis->vec1_N,&impVecType);CHKERRQ(ierr); 1377e7b262bdSStefano Zampini /* local work vectors (try to avoid unneeded work)*/ 1378e7b262bdSStefano Zampini /* R nodes */ 1379e7b262bdSStefano Zampini old_size = -1; 1380e7b262bdSStefano Zampini if (pcbddc->vec1_R) { 1381e7b262bdSStefano Zampini ierr = VecGetSize(pcbddc->vec1_R,&old_size);CHKERRQ(ierr); 1382e7b262bdSStefano Zampini } 1383e7b262bdSStefano Zampini if (n_R != old_size) { 1384e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec1_R);CHKERRQ(ierr); 1385e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec2_R);CHKERRQ(ierr); 13866bfb1811SStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_R);CHKERRQ(ierr); 13876bfb1811SStefano Zampini ierr = VecSetSizes(pcbddc->vec1_R,PETSC_DECIDE,n_R);CHKERRQ(ierr); 13886bfb1811SStefano Zampini ierr = VecSetType(pcbddc->vec1_R,impVecType);CHKERRQ(ierr); 13896bfb1811SStefano Zampini ierr = VecDuplicate(pcbddc->vec1_R,&pcbddc->vec2_R);CHKERRQ(ierr); 1390e7b262bdSStefano Zampini } 1391e7b262bdSStefano Zampini /* local primal dofs */ 1392e7b262bdSStefano Zampini old_size = -1; 1393e7b262bdSStefano Zampini if (pcbddc->vec1_P) { 1394e7b262bdSStefano Zampini ierr = VecGetSize(pcbddc->vec1_P,&old_size);CHKERRQ(ierr); 1395e7b262bdSStefano Zampini } 1396e9189074SStefano Zampini if (pcbddc->local_primal_size != old_size) { 1397e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec1_P);CHKERRQ(ierr); 139883b7ccabSStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_P);CHKERRQ(ierr); 1399e9189074SStefano Zampini ierr = VecSetSizes(pcbddc->vec1_P,PETSC_DECIDE,pcbddc->local_primal_size);CHKERRQ(ierr); 14006bfb1811SStefano Zampini ierr = VecSetType(pcbddc->vec1_P,impVecType);CHKERRQ(ierr); 1401e7b262bdSStefano Zampini } 1402e7b262bdSStefano Zampini /* local explicit constraints */ 1403e7b262bdSStefano Zampini old_size = -1; 1404e7b262bdSStefano Zampini if (pcbddc->vec1_C) { 1405e7b262bdSStefano Zampini ierr = VecGetSize(pcbddc->vec1_C,&old_size);CHKERRQ(ierr); 1406e7b262bdSStefano Zampini } 1407e7b262bdSStefano Zampini if (n_constraints && n_constraints != old_size) { 1408e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec1_C);CHKERRQ(ierr); 140983b7ccabSStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_C);CHKERRQ(ierr); 141083b7ccabSStefano Zampini ierr = VecSetSizes(pcbddc->vec1_C,PETSC_DECIDE,n_constraints);CHKERRQ(ierr); 141183b7ccabSStefano Zampini ierr = VecSetType(pcbddc->vec1_C,impVecType);CHKERRQ(ierr); 141283b7ccabSStefano Zampini } 14136bfb1811SStefano Zampini PetscFunctionReturn(0); 14146bfb1811SStefano Zampini } 14156bfb1811SStefano Zampini 14166bfb1811SStefano Zampini #undef __FUNCT__ 141747f4ddc3SStefano Zampini #define __FUNCT__ "PCBDDCSetUpCorrection" 141847f4ddc3SStefano Zampini PetscErrorCode PCBDDCSetUpCorrection(PC pc, PetscScalar **coarse_submat_vals_n) 141988ebb749SStefano Zampini { 142025084f0cSStefano Zampini PetscErrorCode ierr; 142125084f0cSStefano Zampini /* pointers to pcis and pcbddc */ 142288ebb749SStefano Zampini PC_IS* pcis = (PC_IS*)pc->data; 142388ebb749SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 1424d62866d3SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 142525084f0cSStefano Zampini /* submatrices of local problem */ 142680677318SStefano Zampini Mat A_RV,A_VR,A_VV,local_auxmat2_R; 1427019a44ceSStefano Zampini /* submatrices of benign trick */ 1428d16cbb6bSStefano Zampini Mat B0_V = NULL; 142906656605SStefano Zampini /* submatrices of local coarse problem */ 143006656605SStefano Zampini Mat S_VV,S_CV,S_VC,S_CC; 143125084f0cSStefano Zampini /* working matrices */ 143206656605SStefano Zampini Mat C_CR; 143325084f0cSStefano Zampini /* additional working stuff */ 143406656605SStefano Zampini PC pc_R; 14354f1b2e48SStefano Zampini Mat F; 1436a3df083aSStefano Zampini PetscBool isLU,isCHOL,isILU,need_benign_correction; 143706656605SStefano Zampini 143825084f0cSStefano Zampini PetscScalar *coarse_submat_vals; /* TODO: use a PETSc matrix */ 143906656605SStefano Zampini PetscScalar *work; 144006656605SStefano Zampini PetscInt *idx_V_B; 1441ffd830a3SStefano Zampini PetscInt lda_rhs,n,n_vertices,n_constraints,*p0_lidx_I; 144206656605SStefano Zampini PetscInt i,n_R,n_D,n_B; 1443ffd830a3SStefano Zampini 144425084f0cSStefano Zampini /* some shortcuts to scalars */ 144506656605SStefano Zampini PetscScalar one=1.0,m_one=-1.0; 144688ebb749SStefano Zampini 144788ebb749SStefano Zampini PetscFunctionBegin; 1448ffd830a3SStefano Zampini if (!pcbddc->symmetric_primal && pcbddc->benign_n) { 1449ffd830a3SStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Non-symmetric primal basis computation with benign trick not yet implemented"); 1450ffd830a3SStefano Zampini } 1451ffd830a3SStefano Zampini 1452ffd830a3SStefano Zampini /* Set Non-overlapping dimensions */ 1453b371cd4fSStefano Zampini n_vertices = pcbddc->n_vertices; 14544f1b2e48SStefano Zampini n_constraints = pcbddc->local_primal_size - pcbddc->benign_n - n_vertices; 1455b371cd4fSStefano Zampini n_B = pcis->n_B; 1456b371cd4fSStefano Zampini n_D = pcis->n - n_B; 145788ebb749SStefano Zampini n_R = pcis->n - n_vertices; 145888ebb749SStefano Zampini 145988ebb749SStefano Zampini /* vertices in boundary numbering */ 1460785e854fSJed Brown ierr = PetscMalloc1(n_vertices,&idx_V_B);CHKERRQ(ierr); 14610e6343abSStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,n_vertices,pcbddc->local_primal_ref_node,&i,idx_V_B);CHKERRQ(ierr); 146288ebb749SStefano Zampini if (i != n_vertices) { 146322d5777bSStefano Zampini SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Error in boundary numbering for BDDC vertices! %d != %d\n",n_vertices,i); 146488ebb749SStefano Zampini } 146588ebb749SStefano Zampini 146606656605SStefano Zampini /* Subdomain contribution (Non-overlapping) to coarse matrix */ 1467019a44ceSStefano Zampini ierr = PetscCalloc1(pcbddc->local_primal_size*pcbddc->local_primal_size,&coarse_submat_vals);CHKERRQ(ierr); 146806656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_vertices,n_vertices,coarse_submat_vals,&S_VV);CHKERRQ(ierr); 146906656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_VV,pcbddc->local_primal_size);CHKERRQ(ierr); 147006656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_constraints,n_vertices,coarse_submat_vals+n_vertices,&S_CV);CHKERRQ(ierr); 147106656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_CV,pcbddc->local_primal_size);CHKERRQ(ierr); 147206656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_vertices,n_constraints,coarse_submat_vals+pcbddc->local_primal_size*n_vertices,&S_VC);CHKERRQ(ierr); 147306656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_VC,pcbddc->local_primal_size);CHKERRQ(ierr); 147406656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_constraints,n_constraints,coarse_submat_vals+(pcbddc->local_primal_size+1)*n_vertices,&S_CC);CHKERRQ(ierr); 147506656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_CC,pcbddc->local_primal_size);CHKERRQ(ierr); 147606656605SStefano Zampini 147706656605SStefano Zampini /* determine if can use MatSolve routines instead of calling KSPSolve on ksp_R */ 147806656605SStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_R);CHKERRQ(ierr); 147906656605SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_R,PCLU,&isLU);CHKERRQ(ierr); 148006656605SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_R,PCILU,&isILU);CHKERRQ(ierr); 148106656605SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_R,PCCHOLESKY,&isCHOL);CHKERRQ(ierr); 1482ffd830a3SStefano Zampini lda_rhs = n_R; 1483a3df083aSStefano Zampini need_benign_correction = PETSC_FALSE; 148406656605SStefano Zampini if (isLU || isILU || isCHOL) { 148506656605SStefano Zampini ierr = PCFactorGetMatrix(pc_R,&F);CHKERRQ(ierr); 1486df4d28bfSStefano Zampini } else if (sub_schurs->reuse_solver) { 1487df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 1488d62866d3SStefano Zampini MatFactorType type; 1489d62866d3SStefano Zampini 1490df4d28bfSStefano Zampini F = reuse_solver->F; 14916816873aSStefano Zampini ierr = MatGetFactorType(F,&type);CHKERRQ(ierr); 1492d62866d3SStefano Zampini if (type == MAT_FACTOR_CHOLESKY) isCHOL = PETSC_TRUE; 1493ffd830a3SStefano Zampini ierr = MatGetSize(F,&lda_rhs,NULL);CHKERRQ(ierr); 1494df4d28bfSStefano Zampini need_benign_correction = !!reuse_solver->benign_n; 149506656605SStefano Zampini } else { 149606656605SStefano Zampini F = NULL; 149706656605SStefano Zampini } 149806656605SStefano Zampini 1499ffd830a3SStefano Zampini /* allocate workspace */ 1500ffd830a3SStefano Zampini n = 0; 1501ffd830a3SStefano Zampini if (n_constraints) { 1502ffd830a3SStefano Zampini n += lda_rhs*n_constraints; 1503ffd830a3SStefano Zampini } 1504ffd830a3SStefano Zampini if (n_vertices) { 1505ffd830a3SStefano Zampini n = PetscMax(2*lda_rhs*n_vertices,n); 1506ffd830a3SStefano Zampini n = PetscMax((lda_rhs+n_B)*n_vertices,n); 1507ffd830a3SStefano Zampini } 1508ffd830a3SStefano Zampini ierr = PetscMalloc1(n,&work);CHKERRQ(ierr); 1509ffd830a3SStefano Zampini 151088ebb749SStefano Zampini /* Precompute stuffs needed for preprocessing and application of BDDC*/ 151188ebb749SStefano Zampini if (n_constraints) { 151206656605SStefano Zampini Mat M1,M2,M3; 151380677318SStefano Zampini Mat auxmat; 151406656605SStefano Zampini IS is_aux; 151580677318SStefano Zampini PetscScalar *array,*array2; 151606656605SStefano Zampini 1517f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat1);CHKERRQ(ierr); 151880677318SStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat2);CHKERRQ(ierr); 151988ebb749SStefano Zampini 152025084f0cSStefano Zampini /* Extract constraints on R nodes: C_{CR} */ 152125084f0cSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,n_constraints,n_vertices,1,&is_aux);CHKERRQ(ierr); 15228ce42a96SStefano Zampini ierr = MatGetSubMatrix(pcbddc->ConstraintMatrix,is_aux,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&C_CR);CHKERRQ(ierr); 152380677318SStefano Zampini ierr = MatGetSubMatrix(pcbddc->ConstraintMatrix,is_aux,pcis->is_B_local,MAT_INITIAL_MATRIX,&auxmat);CHKERRQ(ierr); 152488ebb749SStefano Zampini 152580677318SStefano Zampini /* Assemble local_auxmat2_R = (- A_{RR}^{-1} C^T_{CR}) needed by BDDC setup */ 152680677318SStefano Zampini /* Assemble pcbddc->local_auxmat2 = R_to_B (- A_{RR}^{-1} C^T_{CR}) needed by BDDC application */ 1527ffd830a3SStefano Zampini ierr = PetscMemzero(work,lda_rhs*n_constraints*sizeof(PetscScalar));CHKERRQ(ierr); 152888ebb749SStefano Zampini for (i=0;i<n_constraints;i++) { 152906656605SStefano Zampini const PetscScalar *row_cmat_values; 153006656605SStefano Zampini const PetscInt *row_cmat_indices; 153106656605SStefano Zampini PetscInt size_of_constraint,j; 153288ebb749SStefano Zampini 153306656605SStefano Zampini ierr = MatGetRow(C_CR,i,&size_of_constraint,&row_cmat_indices,&row_cmat_values);CHKERRQ(ierr); 153406656605SStefano Zampini for (j=0;j<size_of_constraint;j++) { 1535ffd830a3SStefano Zampini work[row_cmat_indices[j]+i*lda_rhs] = -row_cmat_values[j]; 153606656605SStefano Zampini } 153706656605SStefano Zampini ierr = MatRestoreRow(C_CR,i,&size_of_constraint,&row_cmat_indices,&row_cmat_values);CHKERRQ(ierr); 153806656605SStefano Zampini } 1539ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_constraints,NULL,&local_auxmat2_R);CHKERRQ(ierr); 154006656605SStefano Zampini if (F) { 154106656605SStefano Zampini Mat B; 154206656605SStefano Zampini 1543ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_constraints,work,&B);CHKERRQ(ierr); 1544a3df083aSStefano Zampini if (need_benign_correction) { 1545a3df083aSStefano Zampini PetscScalar *marr; 1546df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 1547a3df083aSStefano Zampini 1548a3df083aSStefano Zampini ierr = MatDenseGetArray(B,&marr);CHKERRQ(ierr); 1549a3df083aSStefano Zampini for (i=0;i<n_constraints;i++) { 1550a3df083aSStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,marr+i*lda_rhs);CHKERRQ(ierr); 1551df4d28bfSStefano Zampini ierr = PCBDDCReuseSolversChangeInterior(reuse_solver,pcbddc->vec1_R,NULL,PETSC_FALSE);CHKERRQ(ierr); 1552a3df083aSStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 1553a3df083aSStefano Zampini } 1554a3df083aSStefano Zampini ierr = MatDenseRestoreArray(B,&marr);CHKERRQ(ierr); 1555a3df083aSStefano Zampini } 155680677318SStefano Zampini ierr = MatMatSolve(F,B,local_auxmat2_R);CHKERRQ(ierr); 1557a3df083aSStefano Zampini if (need_benign_correction) { 1558a3df083aSStefano Zampini PetscScalar *marr; 1559df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 1560a3df083aSStefano Zampini 1561a3df083aSStefano Zampini ierr = MatDenseGetArray(local_auxmat2_R,&marr);CHKERRQ(ierr); 1562a3df083aSStefano Zampini for (i=0;i<n_constraints;i++) { 1563a3df083aSStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,marr+i*lda_rhs);CHKERRQ(ierr); 1564df4d28bfSStefano Zampini ierr = PCBDDCReuseSolversChangeInterior(reuse_solver,pcbddc->vec1_R,NULL,PETSC_TRUE);CHKERRQ(ierr); 1565a3df083aSStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 1566a3df083aSStefano Zampini } 1567a3df083aSStefano Zampini ierr = MatDenseRestoreArray(local_auxmat2_R,&marr);CHKERRQ(ierr); 1568a3df083aSStefano Zampini } 156906656605SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 157006656605SStefano Zampini } else { 157180677318SStefano Zampini PetscScalar *marr; 157280677318SStefano Zampini 157380677318SStefano Zampini ierr = MatDenseGetArray(local_auxmat2_R,&marr);CHKERRQ(ierr); 157406656605SStefano Zampini for (i=0;i<n_constraints;i++) { 1575ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*lda_rhs);CHKERRQ(ierr); 1576ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,marr+i*lda_rhs);CHKERRQ(ierr); 157706656605SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 157806656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 157906656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 158006656605SStefano Zampini } 158180677318SStefano Zampini ierr = MatDenseRestoreArray(local_auxmat2_R,&marr);CHKERRQ(ierr); 158206656605SStefano Zampini } 158380677318SStefano Zampini if (!pcbddc->switch_static) { 158480677318SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,n_constraints,NULL,&pcbddc->local_auxmat2);CHKERRQ(ierr); 158580677318SStefano Zampini ierr = MatDenseGetArray(pcbddc->local_auxmat2,&array);CHKERRQ(ierr); 158680677318SStefano Zampini ierr = MatDenseGetArray(local_auxmat2_R,&array2);CHKERRQ(ierr); 158780677318SStefano Zampini for (i=0;i<n_constraints;i++) { 1588ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,array2+i*lda_rhs);CHKERRQ(ierr); 158980677318SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,array+i*n_B);CHKERRQ(ierr); 159080677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 159180677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 159280677318SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 159380677318SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 159480677318SStefano Zampini } 159580677318SStefano Zampini ierr = MatDenseRestoreArray(local_auxmat2_R,&array2);CHKERRQ(ierr); 159680677318SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->local_auxmat2,&array);CHKERRQ(ierr); 159780677318SStefano Zampini ierr = MatMatMult(auxmat,pcbddc->local_auxmat2,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&M3);CHKERRQ(ierr); 159880677318SStefano Zampini } else { 1599ffd830a3SStefano Zampini if (lda_rhs != n_R) { 1600ffd830a3SStefano Zampini IS dummy; 1601ffd830a3SStefano Zampini 1602ffd830a3SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,n_R,0,1,&dummy);CHKERRQ(ierr); 1603ffd830a3SStefano Zampini ierr = MatGetSubMatrix(local_auxmat2_R,dummy,dummy,MAT_INITIAL_MATRIX,&pcbddc->local_auxmat2);CHKERRQ(ierr); 1604ffd830a3SStefano Zampini ierr = ISDestroy(&dummy);CHKERRQ(ierr); 1605ffd830a3SStefano Zampini } else { 160680677318SStefano Zampini ierr = PetscObjectReference((PetscObject)local_auxmat2_R);CHKERRQ(ierr); 160780677318SStefano Zampini pcbddc->local_auxmat2 = local_auxmat2_R; 1608ffd830a3SStefano Zampini } 160925084f0cSStefano Zampini ierr = MatMatMult(C_CR,pcbddc->local_auxmat2,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&M3);CHKERRQ(ierr); 161080677318SStefano Zampini } 161180677318SStefano Zampini ierr = ISDestroy(&is_aux);CHKERRQ(ierr); 161280677318SStefano Zampini /* Assemble explicitly S_CC = ( C_{CR} A_{RR}^{-1} C^T_{CR} )^{-1} */ 161380677318SStefano Zampini ierr = MatScale(M3,m_one);CHKERRQ(ierr); 161406656605SStefano Zampini ierr = MatDuplicate(M3,MAT_DO_NOT_COPY_VALUES,&M1);CHKERRQ(ierr); 161506656605SStefano Zampini ierr = MatDuplicate(M3,MAT_DO_NOT_COPY_VALUES,&M2);CHKERRQ(ierr); 161680677318SStefano Zampini if (isCHOL) { 161780677318SStefano Zampini ierr = MatCholeskyFactor(M3,NULL,NULL);CHKERRQ(ierr); 161880677318SStefano Zampini } else { 161925084f0cSStefano Zampini ierr = MatLUFactor(M3,NULL,NULL,NULL);CHKERRQ(ierr); 162080677318SStefano Zampini } 162180677318SStefano Zampini ierr = VecSet(pcbddc->vec1_C,one);CHKERRQ(ierr); 162206656605SStefano Zampini ierr = MatDiagonalSet(M2,pcbddc->vec1_C,INSERT_VALUES);CHKERRQ(ierr); 162325084f0cSStefano Zampini ierr = MatMatSolve(M3,M2,M1);CHKERRQ(ierr); 162425084f0cSStefano Zampini ierr = MatDestroy(&M2);CHKERRQ(ierr); 162525084f0cSStefano Zampini ierr = MatDestroy(&M3);CHKERRQ(ierr); 162680677318SStefano Zampini /* Assemble local_auxmat1 = S_CC*C_{CB} needed by BDDC application in KSP and in preproc */ 162780677318SStefano Zampini ierr = MatMatMult(M1,auxmat,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&pcbddc->local_auxmat1);CHKERRQ(ierr); 162880677318SStefano Zampini ierr = MatDestroy(&auxmat);CHKERRQ(ierr); 162906656605SStefano Zampini ierr = MatCopy(M1,S_CC,SAME_NONZERO_PATTERN);CHKERRQ(ierr); /* S_CC can have a different LDA, MatMatSolve doesn't support it */ 163006656605SStefano Zampini ierr = MatDestroy(&M1);CHKERRQ(ierr); 1631f4ddd8eeSStefano Zampini } 163288ebb749SStefano Zampini /* Get submatrices from subdomain matrix */ 16334f1b2e48SStefano Zampini if (pcbddc->benign_n) { 1634d16cbb6bSStefano Zampini IS dummy; 1635d16cbb6bSStefano Zampini Mat B0_R; 1636d16cbb6bSStefano Zampini PetscReal norm; 1637d16cbb6bSStefano Zampini 16384f1b2e48SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcbddc->benign_n,0,1,&dummy);CHKERRQ(ierr); 16394f1b2e48SStefano Zampini ierr = MatGetSubMatrix(pcbddc->benign_B0,dummy,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&B0_R);CHKERRQ(ierr); 1640d16cbb6bSStefano Zampini ierr = MatNorm(B0_R,NORM_INFINITY,&norm);CHKERRQ(ierr); 1641d16cbb6bSStefano Zampini if (norm > PETSC_SMALL) { 1642d16cbb6bSStefano Zampini SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_SUP,"Benign trick can not be applied! ||B0_R|| = %f (should be numerically 0.)",norm); 1643d16cbb6bSStefano Zampini } 1644d16cbb6bSStefano Zampini ierr = MatDestroy(&B0_R);CHKERRQ(ierr); 1645d16cbb6bSStefano Zampini ierr = ISDestroy(&dummy);CHKERRQ(ierr); 1646d16cbb6bSStefano Zampini } 1647d16cbb6bSStefano Zampini 164888ebb749SStefano Zampini if (n_vertices) { 164906656605SStefano Zampini IS is_aux; 16503a50541eSStefano Zampini 1651df4d28bfSStefano Zampini if (sub_schurs->reuse_solver) { /* is_R_local is not sorted, ISComplement doesn't like it */ 16526816873aSStefano Zampini IS tis; 16536816873aSStefano Zampini 16546816873aSStefano Zampini ierr = ISDuplicate(pcbddc->is_R_local,&tis);CHKERRQ(ierr); 16556816873aSStefano Zampini ierr = ISSort(tis);CHKERRQ(ierr); 16566816873aSStefano Zampini ierr = ISComplement(tis,0,pcis->n,&is_aux);CHKERRQ(ierr); 16576816873aSStefano Zampini ierr = ISDestroy(&tis);CHKERRQ(ierr); 16586816873aSStefano Zampini } else { 16593a50541eSStefano Zampini ierr = ISComplement(pcbddc->is_R_local,0,pcis->n,&is_aux);CHKERRQ(ierr); 16606816873aSStefano Zampini } 16619577ea80SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,pcbddc->is_R_local,is_aux,MAT_INITIAL_MATRIX,&A_RV);CHKERRQ(ierr); 16629577ea80SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,is_aux,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&A_VR);CHKERRQ(ierr); 166304708bb6SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,is_aux,is_aux,MAT_INITIAL_MATRIX,&A_VV);CHKERRQ(ierr); 16644f1b2e48SStefano Zampini if (pcbddc->benign_n) { 1665019a44ceSStefano Zampini IS dummy; 1666019a44ceSStefano Zampini 16674f1b2e48SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcbddc->benign_n,0,1,&dummy);CHKERRQ(ierr); 16684f1b2e48SStefano Zampini ierr = MatGetSubMatrix(pcbddc->benign_B0,dummy,is_aux,MAT_INITIAL_MATRIX,&B0_V);CHKERRQ(ierr); 1669019a44ceSStefano Zampini ierr = ISDestroy(&dummy);CHKERRQ(ierr); 1670019a44ceSStefano Zampini } 167125084f0cSStefano Zampini ierr = ISDestroy(&is_aux);CHKERRQ(ierr); 167288ebb749SStefano Zampini } 167388ebb749SStefano Zampini 167488ebb749SStefano Zampini /* Matrix of coarse basis functions (local) */ 1675f4ddd8eeSStefano Zampini if (pcbddc->coarse_phi_B) { 167606656605SStefano Zampini PetscInt on_B,on_primal,on_D=n_D; 167706656605SStefano Zampini if (pcbddc->coarse_phi_D) { 167806656605SStefano Zampini ierr = MatGetSize(pcbddc->coarse_phi_D,&on_D,NULL);CHKERRQ(ierr); 167906656605SStefano Zampini } 1680f4ddd8eeSStefano Zampini ierr = MatGetSize(pcbddc->coarse_phi_B,&on_B,&on_primal);CHKERRQ(ierr); 168106656605SStefano Zampini if (on_B != n_B || on_primal != pcbddc->local_primal_size || on_D != n_D) { 168206656605SStefano Zampini PetscScalar *marray; 168306656605SStefano Zampini 168406656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&marray);CHKERRQ(ierr); 168506656605SStefano Zampini ierr = PetscFree(marray);CHKERRQ(ierr); 1686f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_B);CHKERRQ(ierr); 1687f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_B);CHKERRQ(ierr); 1688f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_D);CHKERRQ(ierr); 1689f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_D);CHKERRQ(ierr); 1690f4ddd8eeSStefano Zampini } 1691f4ddd8eeSStefano Zampini } 169206656605SStefano Zampini 1693f4ddd8eeSStefano Zampini if (!pcbddc->coarse_phi_B) { 169406656605SStefano Zampini PetscScalar *marray; 169588ebb749SStefano Zampini 169606656605SStefano Zampini n = n_B*pcbddc->local_primal_size; 16978eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 169806656605SStefano Zampini n += n_D*pcbddc->local_primal_size; 169988ebb749SStefano Zampini } 17003301b35fSStefano Zampini if (!pcbddc->symmetric_primal) { 170106656605SStefano Zampini n *= 2; 170288ebb749SStefano Zampini } 170306656605SStefano Zampini ierr = PetscCalloc1(n,&marray);CHKERRQ(ierr); 170406656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,pcbddc->local_primal_size,marray,&pcbddc->coarse_phi_B);CHKERRQ(ierr); 170506656605SStefano Zampini n = n_B*pcbddc->local_primal_size; 17068eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 170706656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_D,pcbddc->local_primal_size,marray+n,&pcbddc->coarse_phi_D);CHKERRQ(ierr); 170806656605SStefano Zampini n += n_D*pcbddc->local_primal_size; 170988ebb749SStefano Zampini } 17103301b35fSStefano Zampini if (!pcbddc->symmetric_primal) { 171106656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,pcbddc->local_primal_size,marray+n,&pcbddc->coarse_psi_B);CHKERRQ(ierr); 17128eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 171306656605SStefano Zampini n = n_B*pcbddc->local_primal_size; 171406656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_D,pcbddc->local_primal_size,marray+n,&pcbddc->coarse_psi_D);CHKERRQ(ierr); 171588ebb749SStefano Zampini } 171688ebb749SStefano Zampini } else { 1717c0553b1fSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->coarse_phi_B);CHKERRQ(ierr); 1718c0553b1fSStefano Zampini pcbddc->coarse_psi_B = pcbddc->coarse_phi_B; 17191b968477SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 1720c0553b1fSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->coarse_phi_D);CHKERRQ(ierr); 1721c0553b1fSStefano Zampini pcbddc->coarse_psi_D = pcbddc->coarse_phi_D; 1722c0553b1fSStefano Zampini } 172388ebb749SStefano Zampini } 172406656605SStefano Zampini } 1725019a44ceSStefano Zampini 172606656605SStefano Zampini /* We are now ready to evaluate coarse basis functions and subdomain contribution to coarse problem */ 17274f1b2e48SStefano Zampini p0_lidx_I = NULL; 17284f1b2e48SStefano Zampini if (pcbddc->benign_n && (pcbddc->switch_static || pcbddc->dbg_flag)) { 1729d12edf2fSStefano Zampini const PetscInt *idxs; 1730d12edf2fSStefano Zampini 1731d12edf2fSStefano Zampini ierr = ISGetIndices(pcis->is_I_local,&idxs);CHKERRQ(ierr); 17324f1b2e48SStefano Zampini ierr = PetscMalloc1(pcbddc->benign_n,&p0_lidx_I);CHKERRQ(ierr); 17334f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 17344f1b2e48SStefano Zampini ierr = PetscFindInt(pcbddc->benign_p0_lidx[i],pcis->n-pcis->n_B,idxs,&p0_lidx_I[i]);CHKERRQ(ierr); 17354f1b2e48SStefano Zampini } 1736d12edf2fSStefano Zampini ierr = ISRestoreIndices(pcis->is_I_local,&idxs);CHKERRQ(ierr); 1737d12edf2fSStefano Zampini } 1738d16cbb6bSStefano Zampini 173906656605SStefano Zampini /* vertices */ 174006656605SStefano Zampini if (n_vertices) { 174116f15bc4SStefano Zampini 1742ffd830a3SStefano Zampini ierr = MatConvert(A_VV,MATDENSE,MAT_REUSE_MATRIX,&A_VV);CHKERRQ(ierr); 174304708bb6SStefano Zampini 174416f15bc4SStefano Zampini if (n_R) { 174514393ed6SStefano Zampini Mat A_RRmA_RV,A_RV_bcorr=NULL,S_VVt; /* S_VVt with LDA=N */ 174606656605SStefano Zampini PetscBLASInt B_N,B_one = 1; 174716f15bc4SStefano Zampini PetscScalar *x,*y; 174804708bb6SStefano Zampini PetscBool isseqaij; 174906656605SStefano Zampini 175021eccb56SStefano Zampini ierr = MatScale(A_RV,m_one);CHKERRQ(ierr); 175114393ed6SStefano Zampini if (need_benign_correction) { 175214393ed6SStefano Zampini ISLocalToGlobalMapping RtoN; 175314393ed6SStefano Zampini IS is_p0; 175414393ed6SStefano Zampini PetscInt *idxs_p0,n; 175514393ed6SStefano Zampini 175614393ed6SStefano Zampini ierr = PetscMalloc1(pcbddc->benign_n,&idxs_p0);CHKERRQ(ierr); 175714393ed6SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(pcbddc->is_R_local,&RtoN);CHKERRQ(ierr); 175814393ed6SStefano Zampini ierr = ISGlobalToLocalMappingApply(RtoN,IS_GTOLM_DROP,pcbddc->benign_n,pcbddc->benign_p0_lidx,&n,idxs_p0);CHKERRQ(ierr); 175914393ed6SStefano Zampini if (n != pcbddc->benign_n) { 176014393ed6SStefano Zampini SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Error in R numbering for benign p0! %d != %d\n",n,pcbddc->benign_n); 176114393ed6SStefano Zampini } 176214393ed6SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&RtoN);CHKERRQ(ierr); 176314393ed6SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,n,idxs_p0,PETSC_OWN_POINTER,&is_p0);CHKERRQ(ierr); 176414393ed6SStefano Zampini ierr = MatGetSubMatrix(A_RV,is_p0,NULL,MAT_INITIAL_MATRIX,&A_RV_bcorr);CHKERRQ(ierr); 176514393ed6SStefano Zampini ierr = ISDestroy(&is_p0);CHKERRQ(ierr); 176614393ed6SStefano Zampini } 176714393ed6SStefano Zampini 1768ffd830a3SStefano Zampini if (lda_rhs == n_R) { 1769ffd830a3SStefano Zampini ierr = MatConvert(A_RV,MATDENSE,MAT_REUSE_MATRIX,&A_RV);CHKERRQ(ierr); 1770ffd830a3SStefano Zampini } else { 1771ca92afb2SStefano Zampini PetscScalar *av,*array; 1772ca92afb2SStefano Zampini const PetscInt *xadj,*adjncy; 1773ca92afb2SStefano Zampini PetscInt n; 1774ca92afb2SStefano Zampini PetscBool flg_row; 1775ffd830a3SStefano Zampini 1776ca92afb2SStefano Zampini array = work+lda_rhs*n_vertices; 1777ca92afb2SStefano Zampini ierr = PetscMemzero(array,lda_rhs*n_vertices*sizeof(PetscScalar));CHKERRQ(ierr); 1778ca92afb2SStefano Zampini ierr = MatConvert(A_RV,MATSEQAIJ,MAT_REUSE_MATRIX,&A_RV);CHKERRQ(ierr); 1779ca92afb2SStefano Zampini ierr = MatGetRowIJ(A_RV,0,PETSC_FALSE,PETSC_FALSE,&n,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 1780ca92afb2SStefano Zampini ierr = MatSeqAIJGetArray(A_RV,&av);CHKERRQ(ierr); 1781ca92afb2SStefano Zampini for (i=0;i<n;i++) { 1782ca92afb2SStefano Zampini PetscInt j; 1783ca92afb2SStefano Zampini for (j=xadj[i];j<xadj[i+1];j++) array[lda_rhs*adjncy[j]+i] = av[j]; 1784ffd830a3SStefano Zampini } 1785ca92afb2SStefano Zampini ierr = MatRestoreRowIJ(A_RV,0,PETSC_FALSE,PETSC_FALSE,&n,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 1786ca92afb2SStefano Zampini ierr = MatDestroy(&A_RV);CHKERRQ(ierr); 1787ca92afb2SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_vertices,array,&A_RV);CHKERRQ(ierr); 1788ffd830a3SStefano Zampini } 1789ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_vertices,work,&A_RRmA_RV);CHKERRQ(ierr); 179014393ed6SStefano Zampini if (F) { 1791a3df083aSStefano Zampini if (need_benign_correction) { 1792df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 1793a3df083aSStefano Zampini PetscScalar *marr; 1794a3df083aSStefano Zampini 1795a3df083aSStefano Zampini ierr = MatDenseGetArray(A_RV,&marr);CHKERRQ(ierr); 179614393ed6SStefano Zampini /* need \Phi^T A_RV = (I+L)A_RV, L given by 179714393ed6SStefano Zampini 179814393ed6SStefano Zampini | 0 0 0 | (V) 179914393ed6SStefano Zampini L = | 0 0 -1 | (P-p0) 180014393ed6SStefano Zampini | 0 0 -1 | (p0) 180114393ed6SStefano Zampini 180214393ed6SStefano Zampini */ 1803df4d28bfSStefano Zampini for (i=0;i<reuse_solver->benign_n;i++) { 180414393ed6SStefano Zampini const PetscScalar *vals; 180514393ed6SStefano Zampini const PetscInt *idxs,*idxs_zero; 180614393ed6SStefano Zampini PetscInt n,j,nz; 180714393ed6SStefano Zampini 1808df4d28bfSStefano Zampini ierr = ISGetLocalSize(reuse_solver->benign_zerodiag_subs[i],&nz);CHKERRQ(ierr); 1809df4d28bfSStefano Zampini ierr = ISGetIndices(reuse_solver->benign_zerodiag_subs[i],&idxs_zero);CHKERRQ(ierr); 181014393ed6SStefano Zampini ierr = MatGetRow(A_RV_bcorr,i,&n,&idxs,&vals);CHKERRQ(ierr); 181114393ed6SStefano Zampini for (j=0;j<n;j++) { 181214393ed6SStefano Zampini PetscScalar val = vals[j]; 181314393ed6SStefano Zampini PetscInt k,col = idxs[j]; 181414393ed6SStefano Zampini for (k=0;k<nz;k++) marr[idxs_zero[k]+lda_rhs*col] -= val; 181514393ed6SStefano Zampini } 181614393ed6SStefano Zampini ierr = MatRestoreRow(A_RV_bcorr,i,&n,&idxs,&vals);CHKERRQ(ierr); 1817df4d28bfSStefano Zampini ierr = ISRestoreIndices(reuse_solver->benign_zerodiag_subs[i],&idxs_zero);CHKERRQ(ierr); 181814393ed6SStefano Zampini } 181914393ed6SStefano Zampini /* need to correct the rhs */ 1820a3df083aSStefano Zampini for (i=0;i<n_vertices;i++) { 1821a3df083aSStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,marr+i*lda_rhs);CHKERRQ(ierr); 1822df4d28bfSStefano Zampini ierr = PCBDDCReuseSolversChangeInterior(reuse_solver,pcbddc->vec1_R,NULL,PETSC_FALSE);CHKERRQ(ierr); 1823a3df083aSStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 1824a3df083aSStefano Zampini } 1825a3df083aSStefano Zampini ierr = MatDenseRestoreArray(A_RV,&marr);CHKERRQ(ierr); 1826a3df083aSStefano Zampini } 182706656605SStefano Zampini ierr = MatMatSolve(F,A_RV,A_RRmA_RV);CHKERRQ(ierr); 182814393ed6SStefano Zampini /* need to correct the solution */ 1829a3df083aSStefano Zampini if (need_benign_correction) { 1830df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 1831a3df083aSStefano Zampini PetscScalar *marr; 1832a3df083aSStefano Zampini 1833a3df083aSStefano Zampini ierr = MatDenseGetArray(A_RRmA_RV,&marr);CHKERRQ(ierr); 1834a3df083aSStefano Zampini for (i=0;i<n_vertices;i++) { 1835a3df083aSStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,marr+i*lda_rhs);CHKERRQ(ierr); 1836df4d28bfSStefano Zampini ierr = PCBDDCReuseSolversChangeInterior(reuse_solver,pcbddc->vec1_R,NULL,PETSC_TRUE);CHKERRQ(ierr); 1837a3df083aSStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 1838a3df083aSStefano Zampini } 1839a3df083aSStefano Zampini ierr = MatDenseRestoreArray(A_RRmA_RV,&marr);CHKERRQ(ierr); 1840a3df083aSStefano Zampini } 184106656605SStefano Zampini } else { 184206656605SStefano Zampini ierr = MatDenseGetArray(A_RV,&y);CHKERRQ(ierr); 184306656605SStefano Zampini for (i=0;i<n_vertices;i++) { 1844ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,y+i*lda_rhs);CHKERRQ(ierr); 1845ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,work+i*lda_rhs);CHKERRQ(ierr); 184606656605SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 184706656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 184806656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 184906656605SStefano Zampini } 185006656605SStefano Zampini ierr = MatDenseRestoreArray(A_RV,&y);CHKERRQ(ierr); 185106656605SStefano Zampini } 185280677318SStefano Zampini ierr = MatDestroy(&A_RV);CHKERRQ(ierr); 1853ffd830a3SStefano Zampini /* S_VV and S_CV */ 185406656605SStefano Zampini if (n_constraints) { 185506656605SStefano Zampini Mat B; 185680677318SStefano Zampini 1857ffd830a3SStefano Zampini ierr = PetscMemzero(work+lda_rhs*n_vertices,n_B*n_vertices*sizeof(PetscScalar));CHKERRQ(ierr); 185880677318SStefano Zampini for (i=0;i<n_vertices;i++) { 1859ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*lda_rhs);CHKERRQ(ierr); 1860ffd830a3SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,work+lda_rhs*n_vertices+i*n_B);CHKERRQ(ierr); 186180677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 186280677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 186380677318SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 186480677318SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 186580677318SStefano Zampini } 1866ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,n_vertices,work+lda_rhs*n_vertices,&B);CHKERRQ(ierr); 186780677318SStefano Zampini ierr = MatMatMult(pcbddc->local_auxmat1,B,MAT_REUSE_MATRIX,PETSC_DEFAULT,&S_CV);CHKERRQ(ierr); 186880677318SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 1869ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_vertices,work+lda_rhs*n_vertices,&B);CHKERRQ(ierr); 187080677318SStefano Zampini ierr = MatMatMult(local_auxmat2_R,S_CV,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B);CHKERRQ(ierr); 187106656605SStefano Zampini ierr = MatScale(S_CV,m_one);CHKERRQ(ierr); 1872ffd830a3SStefano Zampini ierr = PetscBLASIntCast(lda_rhs*n_vertices,&B_N);CHKERRQ(ierr); 1873ffd830a3SStefano Zampini PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&B_N,&one,work+lda_rhs*n_vertices,&B_one,work,&B_one)); 187406656605SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 187506656605SStefano Zampini } 187604708bb6SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)A_VR,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 187704708bb6SStefano Zampini if (!isseqaij) { /* MatMatMult with SEQ(S)BAIJ below will raise an error */ 187804708bb6SStefano Zampini ierr = MatConvert(A_VR,MATSEQAIJ,MAT_REUSE_MATRIX,&A_VR);CHKERRQ(ierr); 187904708bb6SStefano Zampini } 1880ffd830a3SStefano Zampini if (lda_rhs != n_R) { 1881ffd830a3SStefano Zampini ierr = MatDestroy(&A_RRmA_RV);CHKERRQ(ierr); 1882ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_vertices,work,&A_RRmA_RV);CHKERRQ(ierr); 1883ffd830a3SStefano Zampini ierr = MatSeqDenseSetLDA(A_RRmA_RV,lda_rhs);CHKERRQ(ierr); 1884ffd830a3SStefano Zampini } 188506656605SStefano Zampini ierr = MatMatMult(A_VR,A_RRmA_RV,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&S_VVt);CHKERRQ(ierr); 188614393ed6SStefano Zampini /* need A_VR * \Phi * A_RRmA_RV = A_VR * (I+L)^T * A_RRmA_RV, L given as before */ 188714393ed6SStefano Zampini if (need_benign_correction) { 1888df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 188914393ed6SStefano Zampini PetscScalar *marr,*sums; 189014393ed6SStefano Zampini 189114393ed6SStefano Zampini ierr = PetscMalloc1(n_vertices,&sums);CHKERRQ(ierr); 189214393ed6SStefano Zampini ierr = MatDenseGetArray(S_VVt,&marr); 1893df4d28bfSStefano Zampini for (i=0;i<reuse_solver->benign_n;i++) { 189414393ed6SStefano Zampini const PetscScalar *vals; 189514393ed6SStefano Zampini const PetscInt *idxs,*idxs_zero; 189614393ed6SStefano Zampini PetscInt n,j,nz; 189714393ed6SStefano Zampini 1898df4d28bfSStefano Zampini ierr = ISGetLocalSize(reuse_solver->benign_zerodiag_subs[i],&nz);CHKERRQ(ierr); 1899df4d28bfSStefano Zampini ierr = ISGetIndices(reuse_solver->benign_zerodiag_subs[i],&idxs_zero);CHKERRQ(ierr); 190014393ed6SStefano Zampini for (j=0;j<n_vertices;j++) { 190114393ed6SStefano Zampini PetscInt k; 190214393ed6SStefano Zampini sums[j] = 0.; 190314393ed6SStefano Zampini for (k=0;k<nz;k++) sums[j] += work[idxs_zero[k]+j*lda_rhs]; 190414393ed6SStefano Zampini } 190514393ed6SStefano Zampini ierr = MatGetRow(A_RV_bcorr,i,&n,&idxs,&vals);CHKERRQ(ierr); 190614393ed6SStefano Zampini for (j=0;j<n;j++) { 190714393ed6SStefano Zampini PetscScalar val = vals[j]; 190814393ed6SStefano Zampini PetscInt k; 190914393ed6SStefano Zampini for (k=0;k<n_vertices;k++) { 191014393ed6SStefano Zampini marr[idxs[j]+k*n_vertices] += val*sums[k]; 191114393ed6SStefano Zampini } 191214393ed6SStefano Zampini } 191314393ed6SStefano Zampini ierr = MatRestoreRow(A_RV_bcorr,i,&n,&idxs,&vals);CHKERRQ(ierr); 1914df4d28bfSStefano Zampini ierr = ISRestoreIndices(reuse_solver->benign_zerodiag_subs[i],&idxs_zero);CHKERRQ(ierr); 191514393ed6SStefano Zampini } 191614393ed6SStefano Zampini ierr = PetscFree(sums);CHKERRQ(ierr); 191714393ed6SStefano Zampini ierr = MatDenseRestoreArray(S_VVt,&marr); 191814393ed6SStefano Zampini ierr = MatDestroy(&A_RV_bcorr);CHKERRQ(ierr); 191914393ed6SStefano Zampini } 192080677318SStefano Zampini ierr = MatDestroy(&A_RRmA_RV);CHKERRQ(ierr); 192106656605SStefano Zampini ierr = PetscBLASIntCast(n_vertices*n_vertices,&B_N);CHKERRQ(ierr); 192206656605SStefano Zampini ierr = MatDenseGetArray(A_VV,&x);CHKERRQ(ierr); 192306656605SStefano Zampini ierr = MatDenseGetArray(S_VVt,&y);CHKERRQ(ierr); 192406656605SStefano Zampini PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&B_N,&one,x,&B_one,y,&B_one)); 192506656605SStefano Zampini ierr = MatDenseRestoreArray(A_VV,&x);CHKERRQ(ierr); 192606656605SStefano Zampini ierr = MatDenseRestoreArray(S_VVt,&y);CHKERRQ(ierr); 192706656605SStefano Zampini ierr = MatCopy(S_VVt,S_VV,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 1928d16cbb6bSStefano Zampini ierr = MatDestroy(&S_VVt);CHKERRQ(ierr); 1929019a44ceSStefano Zampini } else { 1930d16cbb6bSStefano Zampini ierr = MatCopy(A_VV,S_VV,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 1931d16cbb6bSStefano Zampini } 19324f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 1933019a44ceSStefano Zampini const PetscScalar *vals; 1934019a44ceSStefano Zampini const PetscInt *idxs; 19354f1b2e48SStefano Zampini PetscInt n,j,primal_idx; 1936019a44ceSStefano Zampini 19374f1b2e48SStefano Zampini ierr = MatGetRow(B0_V,i,&n,&idxs,&vals);CHKERRQ(ierr); 19384f1b2e48SStefano Zampini primal_idx = pcbddc->local_primal_size - pcbddc->benign_n + i; 1939d16cbb6bSStefano Zampini for (j=0;j<n;j++) { 19404f1b2e48SStefano Zampini coarse_submat_vals[primal_idx*pcbddc->local_primal_size+idxs[j]] = vals[j]; 19414f1b2e48SStefano Zampini coarse_submat_vals[idxs[j]*pcbddc->local_primal_size+primal_idx] = vals[j]; 1942019a44ceSStefano Zampini } 19434f1b2e48SStefano Zampini ierr = MatRestoreRow(B0_V,i,&n,&idxs,&vals);CHKERRQ(ierr); 194416f15bc4SStefano Zampini } 194521eccb56SStefano Zampini ierr = MatDestroy(&A_VV);CHKERRQ(ierr); 1946d16cbb6bSStefano Zampini 194706656605SStefano Zampini /* coarse basis functions */ 194806656605SStefano Zampini for (i=0;i<n_vertices;i++) { 194916f15bc4SStefano Zampini PetscScalar *y; 195016f15bc4SStefano Zampini 1951ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+lda_rhs*i);CHKERRQ(ierr); 195206656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 195306656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*i);CHKERRQ(ierr); 195406656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 195506656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 195606656605SStefano Zampini y[n_B*i+idx_V_B[i]] = 1.0; 195706656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 195806656605SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 195906656605SStefano Zampini 196006656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 19614f1b2e48SStefano Zampini PetscInt j; 19624f1b2e48SStefano Zampini 196306656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 196406656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*i);CHKERRQ(ierr); 196506656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 196606656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 196706656605SStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 19684f1b2e48SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) y[n_D*i+p0_lidx_I[j]] = 0.0; 196906656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 197006656605SStefano Zampini } 197106656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 197206656605SStefano Zampini } 197304708bb6SStefano Zampini /* if n_R == 0 the object is not destroyed */ 197404708bb6SStefano Zampini ierr = MatDestroy(&A_RV);CHKERRQ(ierr); 197506656605SStefano Zampini } 197606656605SStefano Zampini 197706656605SStefano Zampini if (n_constraints) { 197806656605SStefano Zampini Mat B; 197906656605SStefano Zampini 1980ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_constraints,work,&B);CHKERRQ(ierr); 198106656605SStefano Zampini ierr = MatScale(S_CC,m_one);CHKERRQ(ierr); 198280677318SStefano Zampini ierr = MatMatMult(local_auxmat2_R,S_CC,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B);CHKERRQ(ierr); 198306656605SStefano Zampini ierr = MatScale(S_CC,m_one);CHKERRQ(ierr); 198406656605SStefano Zampini if (n_vertices) { 198580677318SStefano Zampini if (isCHOL) { /* if we can solve the interior problem with cholesky, we should also be fine with transposing here */ 198680677318SStefano Zampini ierr = MatTranspose(S_CV,MAT_REUSE_MATRIX,&S_VC);CHKERRQ(ierr); 198780677318SStefano Zampini } else { 198880677318SStefano Zampini Mat S_VCt; 198980677318SStefano Zampini 1990ffd830a3SStefano Zampini if (lda_rhs != n_R) { 1991ffd830a3SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 1992ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_vertices,work,&B);CHKERRQ(ierr); 1993ffd830a3SStefano Zampini ierr = MatSeqDenseSetLDA(B,lda_rhs);CHKERRQ(ierr); 1994ffd830a3SStefano Zampini } 199580677318SStefano Zampini ierr = MatMatMult(A_VR,B,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&S_VCt);CHKERRQ(ierr); 199680677318SStefano Zampini ierr = MatCopy(S_VCt,S_VC,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 199780677318SStefano Zampini ierr = MatDestroy(&S_VCt);CHKERRQ(ierr); 199880677318SStefano Zampini } 199906656605SStefano Zampini } 200006656605SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 200106656605SStefano Zampini /* coarse basis functions */ 200206656605SStefano Zampini for (i=0;i<n_constraints;i++) { 200306656605SStefano Zampini PetscScalar *y; 200406656605SStefano Zampini 2005ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+lda_rhs*i);CHKERRQ(ierr); 200606656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 200706656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*(i+n_vertices));CHKERRQ(ierr); 200806656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 200906656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 201006656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 201106656605SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 201206656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 20134f1b2e48SStefano Zampini PetscInt j; 20144f1b2e48SStefano Zampini 201506656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 201606656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*(i+n_vertices));CHKERRQ(ierr); 201706656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 201806656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 201906656605SStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 20204f1b2e48SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) y[n_D*i+p0_lidx_I[j]] = 0.0; 202106656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 202206656605SStefano Zampini } 202306656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 202406656605SStefano Zampini } 202506656605SStefano Zampini } 202680677318SStefano Zampini if (n_constraints) { 202780677318SStefano Zampini ierr = MatDestroy(&local_auxmat2_R);CHKERRQ(ierr); 202880677318SStefano Zampini } 20294f1b2e48SStefano Zampini ierr = PetscFree(p0_lidx_I);CHKERRQ(ierr); 2030019a44ceSStefano Zampini ierr = MatDestroy(&B0_V);CHKERRQ(ierr); 2031019a44ceSStefano Zampini 203206656605SStefano Zampini /* compute other basis functions for non-symmetric problems */ 20333301b35fSStefano Zampini if (!pcbddc->symmetric_primal) { 2034ffd830a3SStefano Zampini Mat B_V=NULL,B_C=NULL; 2035ffd830a3SStefano Zampini PetscScalar *marray; 203606656605SStefano Zampini 203706656605SStefano Zampini if (n_constraints) { 2038ffd830a3SStefano Zampini Mat S_CCT,C_CRT; 203906656605SStefano Zampini 2040ffd830a3SStefano Zampini ierr = MatTranspose(C_CR,MAT_INITIAL_MATRIX,&C_CRT);CHKERRQ(ierr); 204106656605SStefano Zampini ierr = MatTranspose(S_CC,MAT_INITIAL_MATRIX,&S_CCT);CHKERRQ(ierr); 2042ffd830a3SStefano Zampini ierr = MatMatMult(C_CRT,S_CCT,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&B_C);CHKERRQ(ierr); 204316f15bc4SStefano Zampini ierr = MatDestroy(&S_CCT);CHKERRQ(ierr); 204406656605SStefano Zampini if (n_vertices) { 2045ffd830a3SStefano Zampini Mat S_VCT; 204606656605SStefano Zampini 204706656605SStefano Zampini ierr = MatTranspose(S_VC,MAT_INITIAL_MATRIX,&S_VCT);CHKERRQ(ierr); 2048ffd830a3SStefano Zampini ierr = MatMatMult(C_CRT,S_VCT,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&B_V);CHKERRQ(ierr); 204916f15bc4SStefano Zampini ierr = MatDestroy(&S_VCT);CHKERRQ(ierr); 205006656605SStefano Zampini } 2051ffd830a3SStefano Zampini ierr = MatDestroy(&C_CRT);CHKERRQ(ierr); 205206656605SStefano Zampini } 205316f15bc4SStefano Zampini if (n_vertices && n_R) { 2054ffd830a3SStefano Zampini PetscScalar *av,*marray; 2055ffd830a3SStefano Zampini const PetscInt *xadj,*adjncy; 2056ffd830a3SStefano Zampini PetscInt n; 2057ffd830a3SStefano Zampini PetscBool flg_row; 205806656605SStefano Zampini 2059ffd830a3SStefano Zampini /* B_V = B_V - A_VR^T */ 2060ffd830a3SStefano Zampini ierr = MatConvert(A_VR,MATSEQAIJ,MAT_REUSE_MATRIX,&A_VR);CHKERRQ(ierr); 2061ffd830a3SStefano Zampini ierr = MatGetRowIJ(A_VR,0,PETSC_FALSE,PETSC_FALSE,&n,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 2062ffd830a3SStefano Zampini ierr = MatSeqAIJGetArray(A_VR,&av);CHKERRQ(ierr); 2063ffd830a3SStefano Zampini ierr = MatDenseGetArray(B_V,&marray);CHKERRQ(ierr); 2064ffd830a3SStefano Zampini for (i=0;i<n;i++) { 2065ffd830a3SStefano Zampini PetscInt j; 2066ffd830a3SStefano Zampini for (j=xadj[i];j<xadj[i+1];j++) marray[i*n_R + adjncy[j]] -= av[j]; 2067ffd830a3SStefano Zampini } 2068ffd830a3SStefano Zampini ierr = MatDenseRestoreArray(B_V,&marray);CHKERRQ(ierr); 2069ffd830a3SStefano Zampini ierr = MatRestoreRowIJ(A_VR,0,PETSC_FALSE,PETSC_FALSE,&n,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 2070ffd830a3SStefano Zampini ierr = MatDestroy(&A_VR);CHKERRQ(ierr); 207106656605SStefano Zampini } 207206656605SStefano Zampini 2073ffd830a3SStefano Zampini /* currently there's no support for MatTransposeMatSolve(F,B,X) */ 2074ffd830a3SStefano Zampini ierr = MatDenseGetArray(B_V,&marray);CHKERRQ(ierr); 2075ffd830a3SStefano Zampini for (i=0;i<n_vertices;i++) { 2076ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,marray+i*n_R);CHKERRQ(ierr); 2077ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,work+i*n_R);CHKERRQ(ierr); 207806656605SStefano Zampini ierr = KSPSolveTranspose(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 207906656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 208006656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 208106656605SStefano Zampini } 2082ffd830a3SStefano Zampini ierr = MatDenseRestoreArray(B_V,&marray);CHKERRQ(ierr); 2083ffd830a3SStefano Zampini ierr = MatDenseGetArray(B_C,&marray);CHKERRQ(ierr); 2084ffd830a3SStefano Zampini for (i=n_vertices;i<n_constraints+n_vertices;i++) { 2085ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,marray+(i-n_vertices)*n_R);CHKERRQ(ierr); 2086ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,work+i*n_R);CHKERRQ(ierr); 2087ffd830a3SStefano Zampini ierr = KSPSolveTranspose(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 2088ffd830a3SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 2089ffd830a3SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 209006656605SStefano Zampini } 2091ffd830a3SStefano Zampini ierr = MatDenseRestoreArray(B_C,&marray);CHKERRQ(ierr); 209206656605SStefano Zampini /* coarse basis functions */ 209306656605SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 209406656605SStefano Zampini PetscScalar *y; 209506656605SStefano Zampini 2096ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*n_R);CHKERRQ(ierr); 209706656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_psi_B,&y);CHKERRQ(ierr); 209806656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*i);CHKERRQ(ierr); 209906656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 210006656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 210106656605SStefano Zampini if (i<n_vertices) { 210206656605SStefano Zampini y[n_B*i+idx_V_B[i]] = 1.0; 210306656605SStefano Zampini } 210406656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_psi_B,&y);CHKERRQ(ierr); 210506656605SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 210606656605SStefano Zampini 210706656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 210806656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_psi_D,&y);CHKERRQ(ierr); 210906656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*i);CHKERRQ(ierr); 211006656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 211106656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 211206656605SStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 211306656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_psi_D,&y);CHKERRQ(ierr); 211406656605SStefano Zampini } 211506656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 211606656605SStefano Zampini } 2117ffd830a3SStefano Zampini ierr = MatDestroy(&B_V);CHKERRQ(ierr); 2118ffd830a3SStefano Zampini ierr = MatDestroy(&B_C);CHKERRQ(ierr); 211906656605SStefano Zampini } 2120d62866d3SStefano Zampini /* free memory */ 212188ebb749SStefano Zampini ierr = PetscFree(idx_V_B);CHKERRQ(ierr); 212206656605SStefano Zampini ierr = MatDestroy(&S_VV);CHKERRQ(ierr); 212306656605SStefano Zampini ierr = MatDestroy(&S_CV);CHKERRQ(ierr); 212406656605SStefano Zampini ierr = MatDestroy(&S_VC);CHKERRQ(ierr); 212506656605SStefano Zampini ierr = MatDestroy(&S_CC);CHKERRQ(ierr); 2126d62866d3SStefano Zampini ierr = PetscFree(work);CHKERRQ(ierr); 2127d62866d3SStefano Zampini if (n_vertices) { 2128d62866d3SStefano Zampini ierr = MatDestroy(&A_VR);CHKERRQ(ierr); 2129d62866d3SStefano Zampini } 2130d62866d3SStefano Zampini if (n_constraints) { 2131d62866d3SStefano Zampini ierr = MatDestroy(&C_CR);CHKERRQ(ierr); 2132d62866d3SStefano Zampini } 213388ebb749SStefano Zampini /* Checking coarse_sub_mat and coarse basis functios */ 213488ebb749SStefano Zampini /* Symmetric case : It should be \Phi^{(j)^T} A^{(j)} \Phi^{(j)}=coarse_sub_mat */ 213588ebb749SStefano Zampini /* Non-symmetric case : It should be \Psi^{(j)^T} A^{(j)} \Phi^{(j)}=coarse_sub_mat */ 2136d12edf2fSStefano Zampini if (pcbddc->dbg_flag) { 213788ebb749SStefano Zampini Mat coarse_sub_mat; 213825084f0cSStefano Zampini Mat AUXMAT,TM1,TM2,TM3,TM4; 213988ebb749SStefano Zampini Mat coarse_phi_D,coarse_phi_B; 214088ebb749SStefano Zampini Mat coarse_psi_D,coarse_psi_B; 214188ebb749SStefano Zampini Mat A_II,A_BB,A_IB,A_BI; 21428bec7fa6SStefano Zampini Mat C_B,CPHI; 21438bec7fa6SStefano Zampini IS is_dummy; 21448bec7fa6SStefano Zampini Vec mones; 214588ebb749SStefano Zampini MatType checkmattype=MATSEQAIJ; 214688ebb749SStefano Zampini PetscReal real_value; 214788ebb749SStefano Zampini 2148a3df083aSStefano Zampini if (pcbddc->benign_n && !pcbddc->benign_change_explicit) { 2149a3df083aSStefano Zampini Mat A; 2150a3df083aSStefano Zampini ierr = PCBDDCBenignProject(pc,NULL,NULL,&A);CHKERRQ(ierr); 2151a3df083aSStefano Zampini ierr = MatGetSubMatrix(A,pcis->is_I_local,pcis->is_I_local,MAT_INITIAL_MATRIX,&A_II);CHKERRQ(ierr); 2152a3df083aSStefano Zampini ierr = MatGetSubMatrix(A,pcis->is_I_local,pcis->is_B_local,MAT_INITIAL_MATRIX,&A_IB);CHKERRQ(ierr); 2153a3df083aSStefano Zampini ierr = MatGetSubMatrix(A,pcis->is_B_local,pcis->is_I_local,MAT_INITIAL_MATRIX,&A_BI);CHKERRQ(ierr); 2154a3df083aSStefano Zampini ierr = MatGetSubMatrix(A,pcis->is_B_local,pcis->is_B_local,MAT_INITIAL_MATRIX,&A_BB);CHKERRQ(ierr); 2155a3df083aSStefano Zampini ierr = MatDestroy(&A);CHKERRQ(ierr); 2156a3df083aSStefano Zampini } else { 215788ebb749SStefano Zampini ierr = MatConvert(pcis->A_II,checkmattype,MAT_INITIAL_MATRIX,&A_II);CHKERRQ(ierr); 215888ebb749SStefano Zampini ierr = MatConvert(pcis->A_IB,checkmattype,MAT_INITIAL_MATRIX,&A_IB);CHKERRQ(ierr); 215988ebb749SStefano Zampini ierr = MatConvert(pcis->A_BI,checkmattype,MAT_INITIAL_MATRIX,&A_BI);CHKERRQ(ierr); 216088ebb749SStefano Zampini ierr = MatConvert(pcis->A_BB,checkmattype,MAT_INITIAL_MATRIX,&A_BB);CHKERRQ(ierr); 2161a3df083aSStefano Zampini } 216288ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_phi_D,checkmattype,MAT_INITIAL_MATRIX,&coarse_phi_D);CHKERRQ(ierr); 216388ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_phi_B,checkmattype,MAT_INITIAL_MATRIX,&coarse_phi_B);CHKERRQ(ierr); 2164ffd830a3SStefano Zampini if (!pcbddc->symmetric_primal) { 216588ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_psi_D,checkmattype,MAT_INITIAL_MATRIX,&coarse_psi_D);CHKERRQ(ierr); 216688ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_psi_B,checkmattype,MAT_INITIAL_MATRIX,&coarse_psi_B);CHKERRQ(ierr); 216788ebb749SStefano Zampini } 216888ebb749SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,pcbddc->local_primal_size,pcbddc->local_primal_size,coarse_submat_vals,&coarse_sub_mat);CHKERRQ(ierr); 216988ebb749SStefano Zampini 217025084f0cSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 21713301b35fSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Check coarse sub mat computation (symmetric %d)\n",pcbddc->symmetric_primal);CHKERRQ(ierr); 217225084f0cSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 2173ffd830a3SStefano Zampini if (!pcbddc->symmetric_primal) { 217488ebb749SStefano Zampini ierr = MatMatMult(A_II,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 217588ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM1);CHKERRQ(ierr); 217688ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 217788ebb749SStefano Zampini ierr = MatMatMult(A_BB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 217888ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM2);CHKERRQ(ierr); 217988ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 218088ebb749SStefano Zampini ierr = MatMatMult(A_IB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 218188ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM3);CHKERRQ(ierr); 218288ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 218388ebb749SStefano Zampini ierr = MatMatMult(A_BI,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 218488ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM4);CHKERRQ(ierr); 218588ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 218688ebb749SStefano Zampini } else { 218788ebb749SStefano Zampini ierr = MatPtAP(A_II,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&TM1);CHKERRQ(ierr); 218888ebb749SStefano Zampini ierr = MatPtAP(A_BB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&TM2);CHKERRQ(ierr); 218988ebb749SStefano Zampini ierr = MatMatMult(A_IB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 219088ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_phi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM3);CHKERRQ(ierr); 219188ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 219288ebb749SStefano Zampini ierr = MatMatMult(A_BI,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 219388ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_phi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM4);CHKERRQ(ierr); 219488ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 219588ebb749SStefano Zampini } 219688ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM2,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 219788ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM3,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 219888ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM4,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 219988ebb749SStefano Zampini ierr = MatConvert(TM1,MATSEQDENSE,MAT_REUSE_MATRIX,&TM1);CHKERRQ(ierr); 22004f1b2e48SStefano Zampini if (pcbddc->benign_n) { 2201d12edf2fSStefano Zampini Mat B0_I,B0_B,B0_BPHI,B0_IPHI; 2202d12edf2fSStefano Zampini PetscScalar *data,*data2; 22034f1b2e48SStefano Zampini PetscInt j; 2204d12edf2fSStefano Zampini 22054f1b2e48SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcbddc->benign_n,0,1,&is_dummy);CHKERRQ(ierr); 22064f1b2e48SStefano Zampini ierr = MatGetSubMatrix(pcbddc->benign_B0,is_dummy,pcis->is_B_local,MAT_INITIAL_MATRIX,&B0_B); 22074f1b2e48SStefano Zampini ierr = MatGetSubMatrix(pcbddc->benign_B0,is_dummy,pcis->is_I_local,MAT_INITIAL_MATRIX,&B0_I); 2208d12edf2fSStefano Zampini ierr = MatMatMult(B0_B,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&B0_BPHI);CHKERRQ(ierr); 2209d12edf2fSStefano Zampini ierr = MatConvert(B0_BPHI,MATSEQDENSE,MAT_REUSE_MATRIX,&B0_BPHI);CHKERRQ(ierr); 2210d12edf2fSStefano Zampini ierr = MatDenseGetArray(TM1,&data);CHKERRQ(ierr); 2211d12edf2fSStefano Zampini ierr = MatDenseGetArray(B0_BPHI,&data2);CHKERRQ(ierr); 22124f1b2e48SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) { 22134f1b2e48SStefano Zampini PetscInt primal_idx = pcbddc->local_primal_size - pcbddc->benign_n + j; 2214d12edf2fSStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 22154f1b2e48SStefano Zampini data[primal_idx*pcbddc->local_primal_size+i] += data2[i*pcbddc->benign_n+j]; 22164f1b2e48SStefano Zampini data[i*pcbddc->local_primal_size+primal_idx] += data2[i*pcbddc->benign_n+j]; 22174f1b2e48SStefano Zampini } 2218d12edf2fSStefano Zampini } 2219d12edf2fSStefano Zampini ierr = MatDenseRestoreArray(TM1,&data);CHKERRQ(ierr); 2220d12edf2fSStefano Zampini ierr = MatDenseRestoreArray(B0_BPHI,&data2);CHKERRQ(ierr); 2221d12edf2fSStefano Zampini ierr = MatDestroy(&B0_B);CHKERRQ(ierr); 2222d12edf2fSStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 2223d12edf2fSStefano Zampini ierr = MatDestroy(&B0_BPHI);CHKERRQ(ierr); 2224d12edf2fSStefano Zampini ierr = MatMatMult(B0_I,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&B0_IPHI);CHKERRQ(ierr); 2225d12edf2fSStefano Zampini ierr = MatDestroy(&B0_I);CHKERRQ(ierr); 2226d12edf2fSStefano Zampini ierr = MatNorm(B0_IPHI,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 2227d12edf2fSStefano Zampini ierr = MatDestroy(&B0_IPHI);CHKERRQ(ierr); 2228d12edf2fSStefano Zampini } 2229d12edf2fSStefano Zampini #if 0 2230d12edf2fSStefano Zampini { 2231d12edf2fSStefano Zampini PetscViewer viewer; 2232d12edf2fSStefano Zampini char filename[256]; 2233ffd830a3SStefano Zampini sprintf(filename,"details_local_coarse_mat%d_level%d.m",PetscGlobalRank,pcbddc->current_level); 2234d12edf2fSStefano Zampini ierr = PetscViewerASCIIOpen(PETSC_COMM_SELF,filename,&viewer);CHKERRQ(ierr); 2235d12edf2fSStefano Zampini ierr = PetscViewerSetFormat(viewer,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 2236ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)coarse_sub_mat,"computed");CHKERRQ(ierr); 2237ffd830a3SStefano Zampini ierr = MatView(coarse_sub_mat,viewer);CHKERRQ(ierr); 2238ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)TM1,"projected");CHKERRQ(ierr); 2239d12edf2fSStefano Zampini ierr = MatView(TM1,viewer);CHKERRQ(ierr); 2240ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)pcbddc->coarse_phi_B,"phi_B");CHKERRQ(ierr); 2241ffd830a3SStefano Zampini ierr = MatView(pcbddc->coarse_phi_B,viewer);CHKERRQ(ierr); 2242ffd830a3SStefano Zampini if (pcbddc->coarse_phi_D) { 2243ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)pcbddc->coarse_phi_D,"phi_D");CHKERRQ(ierr); 2244ffd830a3SStefano Zampini ierr = MatView(pcbddc->coarse_phi_D,viewer);CHKERRQ(ierr); 2245ffd830a3SStefano Zampini } 2246ffd830a3SStefano Zampini if (pcbddc->coarse_psi_B) { 2247ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)pcbddc->coarse_psi_B,"psi_B");CHKERRQ(ierr); 2248ffd830a3SStefano Zampini ierr = MatView(pcbddc->coarse_psi_B,viewer);CHKERRQ(ierr); 2249ffd830a3SStefano Zampini } 2250ffd830a3SStefano Zampini if (pcbddc->coarse_phi_B) { 2251ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)pcbddc->coarse_psi_D,"psi_D");CHKERRQ(ierr); 2252ffd830a3SStefano Zampini ierr = MatView(pcbddc->coarse_psi_D,viewer);CHKERRQ(ierr); 2253ffd830a3SStefano Zampini } 2254d12edf2fSStefano Zampini ierr = PetscViewerDestroy(&viewer);CHKERRQ(ierr); 2255d12edf2fSStefano Zampini } 2256d12edf2fSStefano Zampini #endif 225781d9aea3SBarry Smith ierr = MatAXPY(TM1,m_one,coarse_sub_mat,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 22588bec7fa6SStefano Zampini ierr = MatNorm(TM1,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 22591575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 226006656605SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d matrix error % 1.14e\n",PetscGlobalRank,real_value);CHKERRQ(ierr); 22618bec7fa6SStefano Zampini 22628bec7fa6SStefano Zampini /* check constraints */ 22634f1b2e48SStefano Zampini if (!pcbddc->benign_n) { /* TODO: add benign case */ 22648bec7fa6SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcbddc->local_primal_size,0,1,&is_dummy);CHKERRQ(ierr); 22658bec7fa6SStefano Zampini ierr = MatGetSubMatrix(pcbddc->ConstraintMatrix,is_dummy,pcis->is_B_local,MAT_INITIAL_MATRIX,&C_B); 22668bec7fa6SStefano Zampini ierr = MatMatMult(C_B,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&CPHI);CHKERRQ(ierr); 22678bec7fa6SStefano Zampini ierr = MatCreateVecs(CPHI,&mones,NULL);CHKERRQ(ierr); 22688bec7fa6SStefano Zampini ierr = VecSet(mones,-1.0);CHKERRQ(ierr); 22698bec7fa6SStefano Zampini ierr = MatDiagonalSet(CPHI,mones,ADD_VALUES);CHKERRQ(ierr); 22708bec7fa6SStefano Zampini ierr = MatNorm(CPHI,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 2271bdae7319SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d phi constraints error % 1.14e\n",PetscGlobalRank,real_value);CHKERRQ(ierr); 2272ffd830a3SStefano Zampini if (!pcbddc->symmetric_primal) { 2273bdae7319SStefano Zampini ierr = MatMatMult(C_B,coarse_psi_B,MAT_REUSE_MATRIX,1.0,&CPHI);CHKERRQ(ierr); 2274bdae7319SStefano Zampini ierr = VecSet(mones,-1.0);CHKERRQ(ierr); 2275bdae7319SStefano Zampini ierr = MatDiagonalSet(CPHI,mones,ADD_VALUES);CHKERRQ(ierr); 2276bdae7319SStefano Zampini ierr = MatNorm(CPHI,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 2277bdae7319SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d psi constraints error % 1.14e\n",PetscGlobalRank,real_value);CHKERRQ(ierr); 227888ebb749SStefano Zampini } 22798bec7fa6SStefano Zampini ierr = MatDestroy(&C_B);CHKERRQ(ierr); 22808bec7fa6SStefano Zampini ierr = MatDestroy(&CPHI);CHKERRQ(ierr); 22818bec7fa6SStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 22828bec7fa6SStefano Zampini ierr = VecDestroy(&mones);CHKERRQ(ierr); 2283d12edf2fSStefano Zampini } 228425084f0cSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 228588ebb749SStefano Zampini ierr = MatDestroy(&A_II);CHKERRQ(ierr); 228688ebb749SStefano Zampini ierr = MatDestroy(&A_BB);CHKERRQ(ierr); 228788ebb749SStefano Zampini ierr = MatDestroy(&A_IB);CHKERRQ(ierr); 228888ebb749SStefano Zampini ierr = MatDestroy(&A_BI);CHKERRQ(ierr); 228988ebb749SStefano Zampini ierr = MatDestroy(&TM1);CHKERRQ(ierr); 229088ebb749SStefano Zampini ierr = MatDestroy(&TM2);CHKERRQ(ierr); 229188ebb749SStefano Zampini ierr = MatDestroy(&TM3);CHKERRQ(ierr); 229288ebb749SStefano Zampini ierr = MatDestroy(&TM4);CHKERRQ(ierr); 229388ebb749SStefano Zampini ierr = MatDestroy(&coarse_phi_D);CHKERRQ(ierr); 229488ebb749SStefano Zampini ierr = MatDestroy(&coarse_phi_B);CHKERRQ(ierr); 2295ffd830a3SStefano Zampini if (!pcbddc->symmetric_primal) { 229688ebb749SStefano Zampini ierr = MatDestroy(&coarse_psi_D);CHKERRQ(ierr); 229788ebb749SStefano Zampini ierr = MatDestroy(&coarse_psi_B);CHKERRQ(ierr); 229888ebb749SStefano Zampini } 229988ebb749SStefano Zampini ierr = MatDestroy(&coarse_sub_mat);CHKERRQ(ierr); 230088ebb749SStefano Zampini } 23018629588bSStefano Zampini /* get back data */ 23028629588bSStefano Zampini *coarse_submat_vals_n = coarse_submat_vals; 230388ebb749SStefano Zampini PetscFunctionReturn(0); 230488ebb749SStefano Zampini } 230588ebb749SStefano Zampini 230688ebb749SStefano Zampini #undef __FUNCT__ 2307d65f70fdSStefano Zampini #define __FUNCT__ "MatGetSubMatrixUnsorted" 2308d648f858SStefano Zampini PetscErrorCode MatGetSubMatrixUnsorted(Mat A, IS isrow, IS iscol, Mat* B) 2309aa0d41d4SStefano Zampini { 2310d65f70fdSStefano Zampini Mat *work_mat; 2311d65f70fdSStefano Zampini IS isrow_s,iscol_s; 2312d65f70fdSStefano Zampini PetscBool rsorted,csorted; 2313d65f70fdSStefano Zampini PetscInt rsize,*idxs_perm_r,csize,*idxs_perm_c; 2314aa0d41d4SStefano Zampini PetscErrorCode ierr; 2315aa0d41d4SStefano Zampini 2316aa0d41d4SStefano Zampini PetscFunctionBegin; 2317d65f70fdSStefano Zampini ierr = ISSorted(isrow,&rsorted);CHKERRQ(ierr); 2318d65f70fdSStefano Zampini ierr = ISSorted(iscol,&csorted);CHKERRQ(ierr); 2319d65f70fdSStefano Zampini ierr = ISGetLocalSize(isrow,&rsize);CHKERRQ(ierr); 2320d65f70fdSStefano Zampini ierr = ISGetLocalSize(iscol,&csize);CHKERRQ(ierr); 2321aa0d41d4SStefano Zampini 2322d65f70fdSStefano Zampini if (!rsorted) { 2323906d46d4SStefano Zampini const PetscInt *idxs; 2324906d46d4SStefano Zampini PetscInt *idxs_sorted,i; 2325aa0d41d4SStefano Zampini 2326d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_perm_r);CHKERRQ(ierr); 2327d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_sorted);CHKERRQ(ierr); 2328d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 2329d65f70fdSStefano Zampini idxs_perm_r[i] = i; 2330aa0d41d4SStefano Zampini } 2331d65f70fdSStefano Zampini ierr = ISGetIndices(isrow,&idxs);CHKERRQ(ierr); 2332d65f70fdSStefano Zampini ierr = PetscSortIntWithPermutation(rsize,idxs,idxs_perm_r);CHKERRQ(ierr); 2333d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 2334d65f70fdSStefano Zampini idxs_sorted[i] = idxs[idxs_perm_r[i]]; 2335aa0d41d4SStefano Zampini } 2336d65f70fdSStefano Zampini ierr = ISRestoreIndices(isrow,&idxs);CHKERRQ(ierr); 2337d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,rsize,idxs_sorted,PETSC_OWN_POINTER,&isrow_s);CHKERRQ(ierr); 2338d65f70fdSStefano Zampini } else { 2339d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)isrow);CHKERRQ(ierr); 2340d65f70fdSStefano Zampini isrow_s = isrow; 2341aa0d41d4SStefano Zampini } 2342906d46d4SStefano Zampini 2343d65f70fdSStefano Zampini if (!csorted) { 2344d65f70fdSStefano Zampini if (isrow == iscol) { 2345d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)isrow_s);CHKERRQ(ierr); 2346d65f70fdSStefano Zampini iscol_s = isrow_s; 2347d65f70fdSStefano Zampini } else { 2348d65f70fdSStefano Zampini const PetscInt *idxs; 2349d65f70fdSStefano Zampini PetscInt *idxs_sorted,i; 2350906d46d4SStefano Zampini 2351d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_perm_c);CHKERRQ(ierr); 2352d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_sorted);CHKERRQ(ierr); 2353d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 2354d65f70fdSStefano Zampini idxs_perm_c[i] = i; 2355d65f70fdSStefano Zampini } 2356d65f70fdSStefano Zampini ierr = ISGetIndices(iscol,&idxs);CHKERRQ(ierr); 2357d65f70fdSStefano Zampini ierr = PetscSortIntWithPermutation(csize,idxs,idxs_perm_c);CHKERRQ(ierr); 2358d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 2359d65f70fdSStefano Zampini idxs_sorted[i] = idxs[idxs_perm_c[i]]; 2360d65f70fdSStefano Zampini } 2361d65f70fdSStefano Zampini ierr = ISRestoreIndices(iscol,&idxs);CHKERRQ(ierr); 2362d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,csize,idxs_sorted,PETSC_OWN_POINTER,&iscol_s);CHKERRQ(ierr); 2363d65f70fdSStefano Zampini } 2364d65f70fdSStefano Zampini } else { 2365d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)iscol);CHKERRQ(ierr); 2366d65f70fdSStefano Zampini iscol_s = iscol; 2367d65f70fdSStefano Zampini } 2368d65f70fdSStefano Zampini 2369d648f858SStefano Zampini ierr = MatGetSubMatrices(A,1,&isrow_s,&iscol_s,MAT_INITIAL_MATRIX,&work_mat);CHKERRQ(ierr); 2370d65f70fdSStefano Zampini 2371d65f70fdSStefano Zampini if (!rsorted || !csorted) { 2372906d46d4SStefano Zampini Mat new_mat; 2373d65f70fdSStefano Zampini IS is_perm_r,is_perm_c; 2374906d46d4SStefano Zampini 2375d65f70fdSStefano Zampini if (!rsorted) { 2376d65f70fdSStefano Zampini PetscInt *idxs_r,i; 2377d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_r);CHKERRQ(ierr); 2378d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 2379d65f70fdSStefano Zampini idxs_r[idxs_perm_r[i]] = i; 2380906d46d4SStefano Zampini } 2381d65f70fdSStefano Zampini ierr = PetscFree(idxs_perm_r);CHKERRQ(ierr); 2382d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,rsize,idxs_r,PETSC_OWN_POINTER,&is_perm_r);CHKERRQ(ierr); 2383d65f70fdSStefano Zampini } else { 2384d65f70fdSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,rsize,0,1,&is_perm_r);CHKERRQ(ierr); 2385906d46d4SStefano Zampini } 2386d65f70fdSStefano Zampini ierr = ISSetPermutation(is_perm_r);CHKERRQ(ierr); 2387d65f70fdSStefano Zampini 2388d65f70fdSStefano Zampini if (!csorted) { 2389d65f70fdSStefano Zampini if (isrow_s == iscol_s) { 2390d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)is_perm_r);CHKERRQ(ierr); 2391d65f70fdSStefano Zampini is_perm_c = is_perm_r; 2392d65f70fdSStefano Zampini } else { 2393d65f70fdSStefano Zampini PetscInt *idxs_c,i; 2394d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_c);CHKERRQ(ierr); 2395d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 2396d65f70fdSStefano Zampini idxs_c[idxs_perm_c[i]] = i; 2397d65f70fdSStefano Zampini } 2398d65f70fdSStefano Zampini ierr = PetscFree(idxs_perm_c);CHKERRQ(ierr); 2399d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,csize,idxs_c,PETSC_OWN_POINTER,&is_perm_c);CHKERRQ(ierr); 2400d65f70fdSStefano Zampini } 2401d65f70fdSStefano Zampini } else { 2402d65f70fdSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,csize,0,1,&is_perm_c);CHKERRQ(ierr); 2403d65f70fdSStefano Zampini } 2404d65f70fdSStefano Zampini ierr = ISSetPermutation(is_perm_c);CHKERRQ(ierr); 2405d65f70fdSStefano Zampini 2406d65f70fdSStefano Zampini ierr = MatPermute(work_mat[0],is_perm_r,is_perm_c,&new_mat);CHKERRQ(ierr); 2407d65f70fdSStefano Zampini ierr = MatDestroy(&work_mat[0]);CHKERRQ(ierr); 2408d65f70fdSStefano Zampini work_mat[0] = new_mat; 2409d65f70fdSStefano Zampini ierr = ISDestroy(&is_perm_r);CHKERRQ(ierr); 2410d65f70fdSStefano Zampini ierr = ISDestroy(&is_perm_c);CHKERRQ(ierr); 2411d65f70fdSStefano Zampini } 2412d65f70fdSStefano Zampini 2413d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)work_mat[0]);CHKERRQ(ierr); 2414d65f70fdSStefano Zampini *B = work_mat[0]; 2415d65f70fdSStefano Zampini ierr = MatDestroyMatrices(1,&work_mat);CHKERRQ(ierr); 2416d65f70fdSStefano Zampini ierr = ISDestroy(&isrow_s);CHKERRQ(ierr); 2417d65f70fdSStefano Zampini ierr = ISDestroy(&iscol_s);CHKERRQ(ierr); 2418d65f70fdSStefano Zampini PetscFunctionReturn(0); 2419d65f70fdSStefano Zampini } 2420d65f70fdSStefano Zampini 2421d65f70fdSStefano Zampini #undef __FUNCT__ 24225e8657edSStefano Zampini #define __FUNCT__ "PCBDDCComputeLocalMatrix" 24235e8657edSStefano Zampini PetscErrorCode PCBDDCComputeLocalMatrix(PC pc, Mat ChangeOfBasisMatrix) 2424aa0d41d4SStefano Zampini { 2425aa0d41d4SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 24265e8657edSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 2427d65f70fdSStefano Zampini Mat new_mat; 24285e8657edSStefano Zampini IS is_local,is_global; 2429d65f70fdSStefano Zampini PetscInt local_size; 2430d65f70fdSStefano Zampini PetscBool isseqaij; 2431aa0d41d4SStefano Zampini PetscErrorCode ierr; 2432aa0d41d4SStefano Zampini 2433aa0d41d4SStefano Zampini PetscFunctionBegin; 2434aa0d41d4SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 24355e8657edSStefano Zampini ierr = MatGetSize(matis->A,&local_size,NULL);CHKERRQ(ierr); 24365e8657edSStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)matis->A),local_size,0,1,&is_local);CHKERRQ(ierr); 2437b087196eSStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(pc->pmat->rmap->mapping,is_local,&is_global);CHKERRQ(ierr); 2438aa0d41d4SStefano Zampini ierr = ISDestroy(&is_local);CHKERRQ(ierr); 2439d648f858SStefano Zampini ierr = MatGetSubMatrixUnsorted(ChangeOfBasisMatrix,is_global,is_global,&new_mat);CHKERRQ(ierr); 2440aa0d41d4SStefano Zampini ierr = ISDestroy(&is_global);CHKERRQ(ierr); 2441906d46d4SStefano Zampini 2442906d46d4SStefano Zampini /* check */ 2443906d46d4SStefano Zampini if (pcbddc->dbg_flag) { 2444906d46d4SStefano Zampini Vec x,x_change; 2445906d46d4SStefano Zampini PetscReal error; 2446906d46d4SStefano Zampini 24475e8657edSStefano Zampini ierr = MatCreateVecs(ChangeOfBasisMatrix,&x,&x_change);CHKERRQ(ierr); 2448906d46d4SStefano Zampini ierr = VecSetRandom(x,NULL);CHKERRQ(ierr); 24495e8657edSStefano Zampini ierr = MatMult(ChangeOfBasisMatrix,x,x_change);CHKERRQ(ierr); 2450e176bc59SStefano Zampini ierr = VecScatterBegin(matis->cctx,x,matis->x,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2451e176bc59SStefano Zampini ierr = VecScatterEnd(matis->cctx,x,matis->x,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2452d65f70fdSStefano Zampini ierr = MatMult(new_mat,matis->x,matis->y);CHKERRQ(ierr); 2453e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,matis->y,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2454e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,matis->y,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2455906d46d4SStefano Zampini ierr = VecAXPY(x,-1.0,x_change);CHKERRQ(ierr); 2456906d46d4SStefano Zampini ierr = VecNorm(x,NORM_INFINITY,&error);CHKERRQ(ierr); 2457906d46d4SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 2458906d46d4SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Error global vs local change on N: %1.6e\n",error);CHKERRQ(ierr); 2459906d46d4SStefano Zampini ierr = VecDestroy(&x);CHKERRQ(ierr); 2460906d46d4SStefano Zampini ierr = VecDestroy(&x_change);CHKERRQ(ierr); 2461906d46d4SStefano Zampini } 2462906d46d4SStefano Zampini 246322d5777bSStefano Zampini /* TODO: HOW TO WORK WITH BAIJ and SBAIJ and SEQDENSE? */ 24649b28b3ffSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)matis->A,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 246522d5777bSStefano Zampini if (isseqaij) { 24661cf9b237SStefano Zampini Mat M; 24671cf9b237SStefano Zampini 24681cf9b237SStefano Zampini ierr = MatPtAP(matis->A,new_mat,MAT_INITIAL_MATRIX,2.0,&M);CHKERRQ(ierr); 24691cf9b237SStefano Zampini ierr = MatSeqAIJCompress(M,&pcbddc->local_mat);CHKERRQ(ierr); 24701cf9b237SStefano Zampini ierr = MatDestroy(&M);CHKERRQ(ierr); 2471aa0d41d4SStefano Zampini } else { 24721cf9b237SStefano Zampini Mat work_mat,M; 24731cf9b237SStefano Zampini 2474aa0d41d4SStefano Zampini ierr = MatConvert(matis->A,MATSEQAIJ,MAT_INITIAL_MATRIX,&work_mat);CHKERRQ(ierr); 24751cf9b237SStefano Zampini ierr = MatPtAP(work_mat,new_mat,MAT_INITIAL_MATRIX,2.0,&M);CHKERRQ(ierr); 24761cf9b237SStefano Zampini ierr = MatSeqAIJCompress(M,&pcbddc->local_mat);CHKERRQ(ierr); 24771cf9b237SStefano Zampini ierr = MatDestroy(&M);CHKERRQ(ierr); 2478aa0d41d4SStefano Zampini ierr = MatDestroy(&work_mat);CHKERRQ(ierr); 2479aa0d41d4SStefano Zampini } 24803301b35fSStefano Zampini if (matis->A->symmetric_set) { 24813301b35fSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_SYMMETRIC,matis->A->symmetric);CHKERRQ(ierr); 2482e496cd5dSStefano Zampini #if !defined(PETSC_USE_COMPLEX) 24833301b35fSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_HERMITIAN,matis->A->symmetric);CHKERRQ(ierr); 2484e496cd5dSStefano Zampini #endif 24853301b35fSStefano Zampini } 248645a1bb75SStefano Zampini /* 248745a1bb75SStefano Zampini ierr = PetscViewerSetFormat(PETSC_VIEWER_STDOUT_SELF,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 2488d65f70fdSStefano Zampini ierr = MatView(new_mat,(PetscViewer)0);CHKERRQ(ierr); 248945a1bb75SStefano Zampini */ 2490d65f70fdSStefano Zampini ierr = MatDestroy(&new_mat);CHKERRQ(ierr); 2491aa0d41d4SStefano Zampini PetscFunctionReturn(0); 2492aa0d41d4SStefano Zampini } 2493aa0d41d4SStefano Zampini 2494aa0d41d4SStefano Zampini #undef __FUNCT__ 2495a64d13efSStefano Zampini #define __FUNCT__ "PCBDDCSetUpLocalScatters" 24968ce42a96SStefano Zampini PetscErrorCode PCBDDCSetUpLocalScatters(PC pc) 2497a64d13efSStefano Zampini { 2498a64d13efSStefano Zampini PC_IS* pcis = (PC_IS*)(pc->data); 2499a64d13efSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 2500d62866d3SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 250153892102SStefano Zampini PetscInt *idx_R_local=NULL; 25023a50541eSStefano Zampini PetscInt n_vertices,i,j,n_R,n_D,n_B; 25033a50541eSStefano Zampini PetscInt vbs,bs; 25046816873aSStefano Zampini PetscBT bitmask=NULL; 2505a64d13efSStefano Zampini PetscErrorCode ierr; 2506a64d13efSStefano Zampini 2507a64d13efSStefano Zampini PetscFunctionBegin; 2508b23d619eSStefano Zampini /* 2509b23d619eSStefano Zampini No need to setup local scatters if 2510b23d619eSStefano Zampini - primal space is unchanged 2511b23d619eSStefano Zampini AND 2512b23d619eSStefano Zampini - we actually have locally some primal dofs (could not be true in multilevel or for isolated subdomains) 2513b23d619eSStefano Zampini AND 2514b23d619eSStefano Zampini - we are not in debugging mode (this is needed since there are Synchronized prints at the end of the subroutine 2515b23d619eSStefano Zampini */ 2516b23d619eSStefano Zampini if (!pcbddc->new_primal_space_local && pcbddc->local_primal_size && !pcbddc->dbg_flag) { 2517f4ddd8eeSStefano Zampini PetscFunctionReturn(0); 2518f4ddd8eeSStefano Zampini } 2519f4ddd8eeSStefano Zampini /* destroy old objects */ 2520f4ddd8eeSStefano Zampini ierr = ISDestroy(&pcbddc->is_R_local);CHKERRQ(ierr); 2521f4ddd8eeSStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_B);CHKERRQ(ierr); 2522f4ddd8eeSStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_D);CHKERRQ(ierr); 2523a64d13efSStefano Zampini /* Set Non-overlapping dimensions */ 2524b371cd4fSStefano Zampini n_B = pcis->n_B; 2525b371cd4fSStefano Zampini n_D = pcis->n - n_B; 2526b371cd4fSStefano Zampini n_vertices = pcbddc->n_vertices; 25273a50541eSStefano Zampini 2528a64d13efSStefano Zampini /* Dohrmann's notation: dofs splitted in R (Remaining: all dofs but the vertices) and V (Vertices) */ 25296816873aSStefano Zampini 253053892102SStefano Zampini /* create auxiliary bitmask and allocate workspace */ 2531df4d28bfSStefano Zampini if (!sub_schurs->reuse_solver) { 2532854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n-n_vertices,&idx_R_local);CHKERRQ(ierr); 2533a64d13efSStefano Zampini ierr = PetscBTCreate(pcis->n,&bitmask);CHKERRQ(ierr); 2534a64d13efSStefano Zampini for (i=0;i<n_vertices;i++) { 25350e6343abSStefano Zampini ierr = PetscBTSet(bitmask,pcbddc->local_primal_ref_node[i]);CHKERRQ(ierr); 2536a64d13efSStefano Zampini } 2537a64d13efSStefano Zampini 2538a64d13efSStefano Zampini for (i=0, n_R=0; i<pcis->n; i++) { 25394641a718SStefano Zampini if (!PetscBTLookup(bitmask,i)) { 25406816873aSStefano Zampini idx_R_local[n_R++] = i; 2541a64d13efSStefano Zampini } 2542a64d13efSStefano Zampini } 2543df4d28bfSStefano Zampini } else { /* A different ordering (already computed) is present if we are reusing the Schur solver */ 2544df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 25456816873aSStefano Zampini 2546df4d28bfSStefano Zampini ierr = ISGetIndices(reuse_solver->is_R,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 2547df4d28bfSStefano Zampini ierr = ISGetLocalSize(reuse_solver->is_R,&n_R);CHKERRQ(ierr); 25486816873aSStefano Zampini } 25493a50541eSStefano Zampini 25503a50541eSStefano Zampini /* Block code */ 25513a50541eSStefano Zampini vbs = 1; 25523a50541eSStefano Zampini ierr = MatGetBlockSize(pcbddc->local_mat,&bs);CHKERRQ(ierr); 25533a50541eSStefano Zampini if (bs>1 && !(n_vertices%bs)) { 25543a50541eSStefano Zampini PetscBool is_blocked = PETSC_TRUE; 25553a50541eSStefano Zampini PetscInt *vary; 2556df4d28bfSStefano Zampini if (!sub_schurs->reuse_solver) { 2557785e854fSJed Brown ierr = PetscMalloc1(pcis->n/bs,&vary);CHKERRQ(ierr); 25583a50541eSStefano Zampini ierr = PetscMemzero(vary,pcis->n/bs*sizeof(PetscInt));CHKERRQ(ierr); 2559d3df7717SStefano Zampini /* Verify that the vertex indices correspond to each element in a block (code taken from sbaij2.c) */ 2560d3df7717SStefano 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 */ 25610e6343abSStefano Zampini for (i=0; i<n_vertices; i++) vary[pcbddc->local_primal_ref_node[i]/bs]++; 2562d3df7717SStefano Zampini for (i=0; i<pcis->n/bs; i++) { 25633a50541eSStefano Zampini if (vary[i]!=0 && vary[i]!=bs) { 25643a50541eSStefano Zampini is_blocked = PETSC_FALSE; 25653a50541eSStefano Zampini break; 25663a50541eSStefano Zampini } 25673a50541eSStefano Zampini } 2568d3df7717SStefano Zampini ierr = PetscFree(vary);CHKERRQ(ierr); 2569d3df7717SStefano Zampini } else { 2570d3df7717SStefano Zampini /* Verify directly the R set */ 2571d3df7717SStefano Zampini for (i=0; i<n_R/bs; i++) { 2572d3df7717SStefano Zampini PetscInt j,node=idx_R_local[bs*i]; 2573d3df7717SStefano Zampini for (j=1; j<bs; j++) { 2574d3df7717SStefano Zampini if (node != idx_R_local[bs*i+j]-j) { 2575d3df7717SStefano Zampini is_blocked = PETSC_FALSE; 2576d3df7717SStefano Zampini break; 2577d3df7717SStefano Zampini } 2578d3df7717SStefano Zampini } 2579d3df7717SStefano Zampini } 2580d3df7717SStefano Zampini } 25813a50541eSStefano Zampini if (is_blocked) { /* build compressed IS for R nodes (complement of vertices) */ 25823a50541eSStefano Zampini vbs = bs; 25833a50541eSStefano Zampini for (i=0;i<n_R/vbs;i++) { 25843a50541eSStefano Zampini idx_R_local[i] = idx_R_local[vbs*i]/vbs; 25853a50541eSStefano Zampini } 25863a50541eSStefano Zampini } 25873a50541eSStefano Zampini } 25883a50541eSStefano Zampini ierr = ISCreateBlock(PETSC_COMM_SELF,vbs,n_R/vbs,idx_R_local,PETSC_COPY_VALUES,&pcbddc->is_R_local);CHKERRQ(ierr); 2589df4d28bfSStefano Zampini if (sub_schurs->reuse_solver) { 2590df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 259153892102SStefano Zampini 2592df4d28bfSStefano Zampini ierr = ISRestoreIndices(reuse_solver->is_R,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 2593df4d28bfSStefano Zampini ierr = ISDestroy(&reuse_solver->is_R);CHKERRQ(ierr); 259453892102SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->is_R_local);CHKERRQ(ierr); 2595df4d28bfSStefano Zampini reuse_solver->is_R = pcbddc->is_R_local; 259653892102SStefano Zampini } else { 25973a50541eSStefano Zampini ierr = PetscFree(idx_R_local);CHKERRQ(ierr); 259853892102SStefano Zampini } 2599a64d13efSStefano Zampini 2600a64d13efSStefano Zampini /* print some info if requested */ 2601a64d13efSStefano Zampini if (pcbddc->dbg_flag) { 2602a64d13efSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 2603a64d13efSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 26041575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 2605a64d13efSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d local dimensions\n",PetscGlobalRank);CHKERRQ(ierr); 2606a64d13efSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"local_size = %d, dirichlet_size = %d, boundary_size = %d\n",pcis->n,n_D,n_B);CHKERRQ(ierr); 26074f1b2e48SStefano 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); 2608a64d13efSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 2609a64d13efSStefano Zampini } 2610a64d13efSStefano Zampini 2611a64d13efSStefano Zampini /* VecScatters pcbddc->R_to_B and (optionally) pcbddc->R_to_D */ 2612df4d28bfSStefano Zampini if (!sub_schurs->reuse_solver) { 26136816873aSStefano Zampini IS is_aux1,is_aux2; 26146816873aSStefano Zampini PetscInt *aux_array1,*aux_array2,*is_indices,*idx_R_local; 26156816873aSStefano Zampini 26163a50541eSStefano Zampini ierr = ISGetIndices(pcbddc->is_R_local,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 2617854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n_B-n_vertices,&aux_array1);CHKERRQ(ierr); 2618854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n_B-n_vertices,&aux_array2);CHKERRQ(ierr); 2619a64d13efSStefano Zampini ierr = ISGetIndices(pcis->is_I_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 26204641a718SStefano Zampini for (i=0; i<n_D; i++) { 26214641a718SStefano Zampini ierr = PetscBTSet(bitmask,is_indices[i]);CHKERRQ(ierr); 26224641a718SStefano Zampini } 2623a64d13efSStefano Zampini ierr = ISRestoreIndices(pcis->is_I_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 2624a64d13efSStefano Zampini for (i=0, j=0; i<n_R; i++) { 26254641a718SStefano Zampini if (!PetscBTLookup(bitmask,idx_R_local[i])) { 26264641a718SStefano Zampini aux_array1[j++] = i; 2627a64d13efSStefano Zampini } 2628a64d13efSStefano Zampini } 2629a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array1,PETSC_OWN_POINTER,&is_aux1);CHKERRQ(ierr); 2630a64d13efSStefano Zampini ierr = ISGetIndices(pcis->is_B_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 2631a64d13efSStefano Zampini for (i=0, j=0; i<n_B; i++) { 26324641a718SStefano Zampini if (!PetscBTLookup(bitmask,is_indices[i])) { 26334641a718SStefano Zampini aux_array2[j++] = i; 2634a64d13efSStefano Zampini } 2635a64d13efSStefano Zampini } 2636a64d13efSStefano Zampini ierr = ISRestoreIndices(pcis->is_B_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 2637a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array2,PETSC_OWN_POINTER,&is_aux2);CHKERRQ(ierr); 2638a64d13efSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,is_aux1,pcis->vec1_B,is_aux2,&pcbddc->R_to_B);CHKERRQ(ierr); 2639a64d13efSStefano Zampini ierr = ISDestroy(&is_aux1);CHKERRQ(ierr); 2640a64d13efSStefano Zampini ierr = ISDestroy(&is_aux2);CHKERRQ(ierr); 2641a64d13efSStefano Zampini 26428eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 2643785e854fSJed Brown ierr = PetscMalloc1(n_D,&aux_array1);CHKERRQ(ierr); 2644a64d13efSStefano Zampini for (i=0, j=0; i<n_R; i++) { 26454641a718SStefano Zampini if (PetscBTLookup(bitmask,idx_R_local[i])) { 26464641a718SStefano Zampini aux_array1[j++] = i; 2647a64d13efSStefano Zampini } 2648a64d13efSStefano Zampini } 2649a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array1,PETSC_OWN_POINTER,&is_aux1);CHKERRQ(ierr); 2650a64d13efSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,is_aux1,pcis->vec1_D,(IS)0,&pcbddc->R_to_D);CHKERRQ(ierr); 2651a64d13efSStefano Zampini ierr = ISDestroy(&is_aux1);CHKERRQ(ierr); 2652a64d13efSStefano Zampini } 26534641a718SStefano Zampini ierr = PetscBTDestroy(&bitmask);CHKERRQ(ierr); 26543a50541eSStefano Zampini ierr = ISRestoreIndices(pcbddc->is_R_local,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 2655d62866d3SStefano Zampini } else { 2656df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 26576816873aSStefano Zampini IS tis; 26586816873aSStefano Zampini PetscInt schur_size; 26596816873aSStefano Zampini 2660df4d28bfSStefano Zampini ierr = ISGetLocalSize(reuse_solver->is_B,&schur_size);CHKERRQ(ierr); 26616816873aSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,schur_size,n_D,1,&tis);CHKERRQ(ierr); 2662df4d28bfSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,tis,pcis->vec1_B,reuse_solver->is_B,&pcbddc->R_to_B);CHKERRQ(ierr); 26636816873aSStefano Zampini ierr = ISDestroy(&tis);CHKERRQ(ierr); 26646816873aSStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 26656816873aSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,n_D,0,1,&tis);CHKERRQ(ierr); 26666816873aSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,tis,pcis->vec1_D,(IS)0,&pcbddc->R_to_D);CHKERRQ(ierr); 26676816873aSStefano Zampini ierr = ISDestroy(&tis);CHKERRQ(ierr); 2668d62866d3SStefano Zampini } 2669d62866d3SStefano Zampini } 2670a64d13efSStefano Zampini PetscFunctionReturn(0); 2671a64d13efSStefano Zampini } 2672a64d13efSStefano Zampini 2673304d26faSStefano Zampini 2674304d26faSStefano Zampini #undef __FUNCT__ 2675304d26faSStefano Zampini #define __FUNCT__ "PCBDDCSetUpLocalSolvers" 2676684f6988SStefano Zampini PetscErrorCode PCBDDCSetUpLocalSolvers(PC pc, PetscBool dirichlet, PetscBool neumann) 2677304d26faSStefano Zampini { 2678304d26faSStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 2679304d26faSStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 2680304d26faSStefano Zampini PC pc_temp; 2681304d26faSStefano Zampini Mat A_RR; 2682f4ddd8eeSStefano Zampini MatReuse reuse; 2683304d26faSStefano Zampini PetscScalar m_one = -1.0; 2684304d26faSStefano Zampini PetscReal value; 268504708bb6SStefano Zampini PetscInt n_D,n_R; 26869577ea80SStefano Zampini PetscBool use_exact,use_exact_reduced,issbaij; 2687304d26faSStefano Zampini PetscErrorCode ierr; 2688e604994aSStefano Zampini /* prefixes stuff */ 2689312be037SStefano Zampini char dir_prefix[256],neu_prefix[256],str_level[16]; 2690e604994aSStefano Zampini size_t len; 2691304d26faSStefano Zampini 2692304d26faSStefano Zampini PetscFunctionBegin; 2693304d26faSStefano Zampini 2694e604994aSStefano Zampini /* compute prefixes */ 2695e604994aSStefano Zampini ierr = PetscStrcpy(dir_prefix,"");CHKERRQ(ierr); 2696e604994aSStefano Zampini ierr = PetscStrcpy(neu_prefix,"");CHKERRQ(ierr); 2697e604994aSStefano Zampini if (!pcbddc->current_level) { 2698e604994aSStefano Zampini ierr = PetscStrcpy(dir_prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 2699e604994aSStefano Zampini ierr = PetscStrcpy(neu_prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 2700e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,"pc_bddc_dirichlet_");CHKERRQ(ierr); 2701e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,"pc_bddc_neumann_");CHKERRQ(ierr); 2702e604994aSStefano Zampini } else { 2703e604994aSStefano Zampini ierr = PetscStrcpy(str_level,"");CHKERRQ(ierr); 2704312be037SStefano Zampini sprintf(str_level,"l%d_",(int)(pcbddc->current_level)); 2705e604994aSStefano Zampini ierr = PetscStrlen(((PetscObject)pc)->prefix,&len);CHKERRQ(ierr); 2706e604994aSStefano Zampini len -= 15; /* remove "pc_bddc_coarse_" */ 2707312be037SStefano Zampini if (pcbddc->current_level>1) len -= 3; /* remove "lX_" with X level number */ 2708312be037SStefano Zampini if (pcbddc->current_level>10) len -= 1; /* remove another char from level number */ 270934d6797cSStefano Zampini ierr = PetscStrncpy(dir_prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 271034d6797cSStefano Zampini ierr = PetscStrncpy(neu_prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 2711e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,"pc_bddc_dirichlet_");CHKERRQ(ierr); 2712e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,"pc_bddc_neumann_");CHKERRQ(ierr); 2713e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,str_level);CHKERRQ(ierr); 2714e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,str_level);CHKERRQ(ierr); 2715e604994aSStefano Zampini } 2716e604994aSStefano Zampini 2717304d26faSStefano Zampini /* DIRICHLET PROBLEM */ 2718684f6988SStefano Zampini if (dirichlet) { 2719d5574798SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 2720a3df083aSStefano Zampini if (pcbddc->benign_n && !pcbddc->benign_change_explicit && pcbddc->dbg_flag) { 2721a3df083aSStefano Zampini Mat A_IIn; 2722a3df083aSStefano Zampini 2723a3df083aSStefano Zampini ierr = PCBDDCBenignProject(pc,pcis->is_I_local,pcis->is_I_local,&A_IIn);CHKERRQ(ierr); 2724a3df083aSStefano Zampini ierr = MatDestroy(&pcis->A_II);CHKERRQ(ierr); 2725a3df083aSStefano Zampini pcis->A_II = A_IIn; 2726a3df083aSStefano Zampini } 27273301b35fSStefano Zampini if (pcbddc->local_mat->symmetric_set) { 27283301b35fSStefano Zampini ierr = MatSetOption(pcis->A_II,MAT_SYMMETRIC,pcbddc->local_mat->symmetric_set);CHKERRQ(ierr); 2729964fefecSStefano Zampini } 2730ac78edfcSStefano Zampini /* Matrix for Dirichlet problem is pcis->A_II */ 2731964fefecSStefano Zampini n_D = pcis->n - pcis->n_B; 2732304d26faSStefano Zampini if (!pcbddc->ksp_D) { /* create object if not yet build */ 2733304d26faSStefano Zampini ierr = KSPCreate(PETSC_COMM_SELF,&pcbddc->ksp_D);CHKERRQ(ierr); 2734304d26faSStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->ksp_D,(PetscObject)pc,1);CHKERRQ(ierr); 2735304d26faSStefano Zampini /* default */ 2736304d26faSStefano Zampini ierr = KSPSetType(pcbddc->ksp_D,KSPPREONLY);CHKERRQ(ierr); 2737e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->ksp_D,dir_prefix);CHKERRQ(ierr); 27389577ea80SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcis->A_II,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 2739304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_D,&pc_temp);CHKERRQ(ierr); 27409577ea80SStefano Zampini if (issbaij) { 27419577ea80SStefano Zampini ierr = PCSetType(pc_temp,PCCHOLESKY);CHKERRQ(ierr); 27429577ea80SStefano Zampini } else { 2743304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCLU);CHKERRQ(ierr); 27449577ea80SStefano Zampini } 2745304d26faSStefano Zampini /* Allow user's customization */ 2746304d26faSStefano Zampini ierr = KSPSetFromOptions(pcbddc->ksp_D);CHKERRQ(ierr); 2747304d26faSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 2748304d26faSStefano Zampini } 2749d1e9a80fSBarry Smith ierr = KSPSetOperators(pcbddc->ksp_D,pcis->A_II,pcis->A_II);CHKERRQ(ierr); 2750df4d28bfSStefano Zampini if (sub_schurs->reuse_solver) { 2751df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 2752d62866d3SStefano Zampini 2753df4d28bfSStefano Zampini ierr = KSPSetPC(pcbddc->ksp_D,reuse_solver->interior_solver);CHKERRQ(ierr); 2754d5574798SStefano Zampini } 2755304d26faSStefano Zampini /* umfpack interface has a bug when matrix dimension is zero. TODO solve from umfpack interface */ 2756304d26faSStefano Zampini if (!n_D) { 2757304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_D,&pc_temp);CHKERRQ(ierr); 2758304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCNONE);CHKERRQ(ierr); 2759304d26faSStefano Zampini } 2760304d26faSStefano Zampini /* Set Up KSP for Dirichlet problem of BDDC */ 2761304d26faSStefano Zampini ierr = KSPSetUp(pcbddc->ksp_D);CHKERRQ(ierr); 2762304d26faSStefano Zampini /* set ksp_D into pcis data */ 2763304d26faSStefano Zampini ierr = KSPDestroy(&pcis->ksp_D);CHKERRQ(ierr); 2764304d26faSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->ksp_D);CHKERRQ(ierr); 2765304d26faSStefano Zampini pcis->ksp_D = pcbddc->ksp_D; 2766684f6988SStefano Zampini } 2767304d26faSStefano Zampini 2768304d26faSStefano Zampini /* NEUMANN PROBLEM */ 2769684f6988SStefano Zampini A_RR = 0; 2770684f6988SStefano Zampini if (neumann) { 2771d62866d3SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 277204708bb6SStefano Zampini PetscInt ibs,mbs; 277304708bb6SStefano Zampini PetscBool issbaij; 277404708bb6SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 2775f4ddd8eeSStefano Zampini /* Matrix for Neumann problem is A_RR -> we need to create/reuse it at this point */ 27768ce42a96SStefano Zampini ierr = ISGetSize(pcbddc->is_R_local,&n_R);CHKERRQ(ierr); 2777f4ddd8eeSStefano Zampini if (pcbddc->ksp_R) { /* already created ksp */ 2778f4ddd8eeSStefano Zampini PetscInt nn_R; 277981d9aea3SBarry Smith ierr = KSPGetOperators(pcbddc->ksp_R,NULL,&A_RR);CHKERRQ(ierr); 2780f4ddd8eeSStefano Zampini ierr = PetscObjectReference((PetscObject)A_RR);CHKERRQ(ierr); 2781f4ddd8eeSStefano Zampini ierr = MatGetSize(A_RR,&nn_R,NULL);CHKERRQ(ierr); 2782f4ddd8eeSStefano Zampini if (nn_R != n_R) { /* old ksp is not reusable, so reset it */ 2783f4ddd8eeSStefano Zampini ierr = KSPReset(pcbddc->ksp_R);CHKERRQ(ierr); 2784f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 2785f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 2786f4ddd8eeSStefano Zampini } else { /* same sizes, but nonzero pattern depend on primal vertices so it can be changed */ 2787727cdba6SStefano Zampini if (pcbddc->new_primal_space_local) { /* we are not sure the matrix will have the same nonzero pattern */ 2788f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 2789f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 2790f4ddd8eeSStefano Zampini } else { /* safe to reuse the matrix */ 2791f4ddd8eeSStefano Zampini reuse = MAT_REUSE_MATRIX; 2792f4ddd8eeSStefano Zampini } 2793f4ddd8eeSStefano Zampini } 2794f4ddd8eeSStefano Zampini /* last check */ 2795d1e9a80fSBarry Smith if (pc->flag == DIFFERENT_NONZERO_PATTERN) { 2796f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 2797f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 2798f4ddd8eeSStefano Zampini } 2799f4ddd8eeSStefano Zampini } else { /* first time, so we need to create the matrix */ 2800f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 2801f4ddd8eeSStefano Zampini } 2802f4ddd8eeSStefano Zampini /* extract A_RR */ 2803af732b37SStefano Zampini ierr = MatGetBlockSize(pcbddc->local_mat,&mbs);CHKERRQ(ierr); 2804af732b37SStefano Zampini ierr = ISGetBlockSize(pcbddc->is_R_local,&ibs);CHKERRQ(ierr); 280504708bb6SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcbddc->local_mat,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 280604708bb6SStefano Zampini if (ibs != mbs) { /* need to convert to SEQAIJ to extract any submatrix with is_R_local */ 280704708bb6SStefano Zampini if (matis->A == pcbddc->local_mat) { 280804708bb6SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 280904708bb6SStefano Zampini ierr = MatConvert(matis->A,MATSEQAIJ,MAT_INITIAL_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 2810af732b37SStefano Zampini } else { 281104708bb6SStefano Zampini ierr = MatConvert(pcbddc->local_mat,MATSEQAIJ,MAT_REUSE_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 28126816873aSStefano Zampini } 281304708bb6SStefano Zampini } else if (issbaij) { /* need to convert to BAIJ to get offdiagonal blocks */ 281404708bb6SStefano Zampini if (matis->A == pcbddc->local_mat) { 281504708bb6SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 281604708bb6SStefano Zampini ierr = MatConvert(matis->A,MATSEQBAIJ,MAT_INITIAL_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 281704708bb6SStefano Zampini } else { 281804708bb6SStefano Zampini ierr = MatConvert(pcbddc->local_mat,MATSEQBAIJ,MAT_REUSE_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 281904708bb6SStefano Zampini } 282004708bb6SStefano Zampini } 2821f4ddd8eeSStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,pcbddc->is_R_local,pcbddc->is_R_local,reuse,&A_RR);CHKERRQ(ierr); 28223301b35fSStefano Zampini if (pcbddc->local_mat->symmetric_set) { 28233301b35fSStefano Zampini ierr = MatSetOption(A_RR,MAT_SYMMETRIC,pcbddc->local_mat->symmetric_set);CHKERRQ(ierr); 28246816873aSStefano Zampini } 2825a3df083aSStefano Zampini if (pcbddc->benign_n && !pcbddc->benign_change_explicit && pcbddc->dbg_flag) { 2826a3df083aSStefano Zampini Mat A_RRn; 2827a3df083aSStefano Zampini ierr = PCBDDCBenignProject(pc,pcbddc->is_R_local,pcbddc->is_R_local,&A_RRn);CHKERRQ(ierr); 2828a3df083aSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 2829a3df083aSStefano Zampini A_RR = A_RRn; 2830a3df083aSStefano Zampini } 2831f4ddd8eeSStefano Zampini if (!pcbddc->ksp_R) { /* create object if not present */ 2832304d26faSStefano Zampini ierr = KSPCreate(PETSC_COMM_SELF,&pcbddc->ksp_R);CHKERRQ(ierr); 2833304d26faSStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->ksp_R,(PetscObject)pc,1);CHKERRQ(ierr); 2834304d26faSStefano Zampini /* default */ 2835304d26faSStefano Zampini ierr = KSPSetType(pcbddc->ksp_R,KSPPREONLY);CHKERRQ(ierr); 2836e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->ksp_R,neu_prefix);CHKERRQ(ierr); 2837304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_temp);CHKERRQ(ierr); 28389577ea80SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)A_RR,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 28399577ea80SStefano Zampini if (issbaij) { 28409577ea80SStefano Zampini ierr = PCSetType(pc_temp,PCCHOLESKY);CHKERRQ(ierr); 28419577ea80SStefano Zampini } else { 2842304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCLU);CHKERRQ(ierr); 28439577ea80SStefano Zampini } 2844304d26faSStefano Zampini /* Allow user's customization */ 2845304d26faSStefano Zampini ierr = KSPSetFromOptions(pcbddc->ksp_R);CHKERRQ(ierr); 2846304d26faSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 2847304d26faSStefano Zampini } 2848304d26faSStefano Zampini /* umfpack interface has a bug when matrix dimension is zero. TODO solve from umfpack interface */ 2849304d26faSStefano Zampini if (!n_R) { 2850304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_temp);CHKERRQ(ierr); 2851304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCNONE);CHKERRQ(ierr); 2852304d26faSStefano Zampini } 2853df4d28bfSStefano Zampini /* Reuse solver if it is present */ 2854df4d28bfSStefano Zampini if (sub_schurs->reuse_solver) { 2855df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 2856d62866d3SStefano Zampini 2857df4d28bfSStefano Zampini ierr = KSPSetPC(pcbddc->ksp_R,reuse_solver->correction_solver);CHKERRQ(ierr); 2858d62866d3SStefano Zampini } 2859ffd830a3SStefano Zampini ierr = KSPSetOperators(pcbddc->ksp_R,A_RR,A_RR);CHKERRQ(ierr); 2860304d26faSStefano Zampini /* Set Up KSP for Neumann problem of BDDC */ 2861304d26faSStefano Zampini ierr = KSPSetUp(pcbddc->ksp_R);CHKERRQ(ierr); 2862684f6988SStefano Zampini } 28636816873aSStefano Zampini /* free Neumann problem's matrix */ 28646816873aSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 2865304d26faSStefano Zampini 2866304d26faSStefano Zampini /* check Dirichlet and Neumann solvers and adapt them if a nullspace correction is needed */ 28670fccc4e9SStefano Zampini if (pcbddc->NullSpace || pcbddc->dbg_flag) { 2868684f6988SStefano Zampini if (pcbddc->dbg_flag) { 2869684f6988SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 28701575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 2871684f6988SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 2872684f6988SStefano Zampini } 2873684f6988SStefano Zampini if (dirichlet) { /* Dirichlet */ 28740fccc4e9SStefano Zampini ierr = VecSetRandom(pcis->vec1_D,NULL);CHKERRQ(ierr); 28750fccc4e9SStefano Zampini ierr = MatMult(pcis->A_II,pcis->vec1_D,pcis->vec2_D);CHKERRQ(ierr); 28760fccc4e9SStefano Zampini ierr = KSPSolve(pcbddc->ksp_D,pcis->vec2_D,pcis->vec2_D);CHKERRQ(ierr); 28770fccc4e9SStefano Zampini ierr = VecAXPY(pcis->vec1_D,m_one,pcis->vec2_D);CHKERRQ(ierr); 28780fccc4e9SStefano Zampini ierr = VecNorm(pcis->vec1_D,NORM_INFINITY,&value);CHKERRQ(ierr); 2879304d26faSStefano Zampini /* need to be adapted? */ 2880b8ffe317SStefano Zampini use_exact = (PetscAbsReal(value) > 1.e-4 ? PETSC_FALSE : PETSC_TRUE); 2881b8ffe317SStefano Zampini ierr = MPI_Allreduce(&use_exact,&use_exact_reduced,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 2882b8ffe317SStefano Zampini ierr = PCBDDCSetUseExactDirichlet(pc,use_exact_reduced);CHKERRQ(ierr); 2883304d26faSStefano Zampini /* print info */ 2884304d26faSStefano Zampini if (pcbddc->dbg_flag) { 2885e604994aSStefano 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); 2886304d26faSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 2887304d26faSStefano Zampini } 2888b8ffe317SStefano Zampini if (pcbddc->NullSpace && !use_exact_reduced && !pcbddc->switch_static) { 2889298c0119SStefano Zampini ierr = PCBDDCNullSpaceAssembleCorrection(pc,PETSC_TRUE,pcis->is_I_local);CHKERRQ(ierr); 2890304d26faSStefano Zampini } 2891684f6988SStefano Zampini } 2892684f6988SStefano Zampini if (neumann) { /* Neumann */ 28936816873aSStefano Zampini ierr = KSPGetOperators(pcbddc->ksp_R,&A_RR,NULL);CHKERRQ(ierr); 28940fccc4e9SStefano Zampini ierr = VecSetRandom(pcbddc->vec1_R,NULL);CHKERRQ(ierr); 28950fccc4e9SStefano Zampini ierr = MatMult(A_RR,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 28960fccc4e9SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec2_R,pcbddc->vec2_R);CHKERRQ(ierr); 28970fccc4e9SStefano Zampini ierr = VecAXPY(pcbddc->vec1_R,m_one,pcbddc->vec2_R);CHKERRQ(ierr); 28980fccc4e9SStefano Zampini ierr = VecNorm(pcbddc->vec1_R,NORM_INFINITY,&value);CHKERRQ(ierr); 2899304d26faSStefano Zampini /* need to be adapted? */ 2900b8ffe317SStefano Zampini use_exact = (PetscAbsReal(value) > 1.e-4 ? PETSC_FALSE : PETSC_TRUE); 2901b8ffe317SStefano Zampini ierr = MPI_Allreduce(&use_exact,&use_exact_reduced,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 2902304d26faSStefano Zampini /* print info */ 2903304d26faSStefano Zampini if (pcbddc->dbg_flag) { 2904e604994aSStefano 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); 2905304d26faSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 2906304d26faSStefano Zampini } 2907b8ffe317SStefano Zampini if (pcbddc->NullSpace && !use_exact_reduced) { /* is it the right logic? */ 2908298c0119SStefano Zampini ierr = PCBDDCNullSpaceAssembleCorrection(pc,PETSC_FALSE,pcbddc->is_R_local);CHKERRQ(ierr); 2909304d26faSStefano Zampini } 29100fccc4e9SStefano Zampini } 2911684f6988SStefano Zampini } 2912304d26faSStefano Zampini PetscFunctionReturn(0); 2913304d26faSStefano Zampini } 2914304d26faSStefano Zampini 2915304d26faSStefano Zampini #undef __FUNCT__ 2916ba15a52eSStefano Zampini #define __FUNCT__ "PCBDDCSolveSubstructureCorrection" 291780677318SStefano Zampini static PetscErrorCode PCBDDCSolveSubstructureCorrection(PC pc, Vec inout_B, Vec inout_D, PetscBool applytranspose) 2918674ae819SStefano Zampini { 2919674ae819SStefano Zampini PetscErrorCode ierr; 2920674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 2921be83ff47SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 2922674ae819SStefano Zampini 2923674ae819SStefano Zampini PetscFunctionBegin; 2924df4d28bfSStefano Zampini if (!sub_schurs->reuse_solver) { 292580677318SStefano Zampini ierr = VecSet(pcbddc->vec1_R,0.);CHKERRQ(ierr); 292620c7b377SStefano Zampini } 292780677318SStefano Zampini if (!pcbddc->switch_static) { 292880677318SStefano Zampini if (applytranspose && pcbddc->local_auxmat1) { 292980677318SStefano Zampini ierr = MatMultTranspose(pcbddc->local_auxmat2,inout_B,pcbddc->vec1_C);CHKERRQ(ierr); 293080677318SStefano Zampini ierr = MatMultTransposeAdd(pcbddc->local_auxmat1,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr); 293120c7b377SStefano Zampini } 2932df4d28bfSStefano Zampini if (!sub_schurs->reuse_solver) { 293380677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 293480677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 293520c7b377SStefano Zampini } else { 2936df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 2937be83ff47SStefano Zampini 2938df4d28bfSStefano Zampini ierr = VecScatterBegin(reuse_solver->correction_scatter_B,inout_B,reuse_solver->rhs_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2939df4d28bfSStefano Zampini ierr = VecScatterEnd(reuse_solver->correction_scatter_B,inout_B,reuse_solver->rhs_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 294020c7b377SStefano Zampini } 2941be83ff47SStefano Zampini } else { 294280677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 294380677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 294480677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,inout_D,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 294580677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,inout_D,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 294680677318SStefano Zampini if (applytranspose && pcbddc->local_auxmat1) { 294780677318SStefano Zampini ierr = MatMultTranspose(pcbddc->local_auxmat2,pcbddc->vec1_R,pcbddc->vec1_C);CHKERRQ(ierr); 294880677318SStefano Zampini ierr = MatMultTransposeAdd(pcbddc->local_auxmat1,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr); 294980677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 295080677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2951674ae819SStefano Zampini } 2952674ae819SStefano Zampini } 2953df4d28bfSStefano Zampini if (!sub_schurs->reuse_solver) { 295480677318SStefano Zampini if (applytranspose) { 295580677318SStefano Zampini ierr = KSPSolveTranspose(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 295680677318SStefano Zampini } else { 295780677318SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 295880677318SStefano Zampini } 2959be83ff47SStefano Zampini } else { 2960df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 2961be83ff47SStefano Zampini 2962be83ff47SStefano Zampini if (applytranspose) { 2963df4d28bfSStefano Zampini ierr = MatFactorSolveSchurComplementTranspose(reuse_solver->F,reuse_solver->rhs_B,reuse_solver->sol_B);CHKERRQ(ierr); 2964be83ff47SStefano Zampini } else { 2965df4d28bfSStefano Zampini ierr = MatFactorSolveSchurComplement(reuse_solver->F,reuse_solver->rhs_B,reuse_solver->sol_B);CHKERRQ(ierr); 2966be83ff47SStefano Zampini } 2967be83ff47SStefano Zampini } 296880677318SStefano Zampini ierr = VecSet(inout_B,0.);CHKERRQ(ierr); 296980677318SStefano Zampini if (!pcbddc->switch_static) { 2970df4d28bfSStefano Zampini if (!sub_schurs->reuse_solver) { 297180677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 297280677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2973be83ff47SStefano Zampini } else { 2974df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 2975be83ff47SStefano Zampini 2976df4d28bfSStefano Zampini ierr = VecScatterBegin(reuse_solver->correction_scatter_B,reuse_solver->sol_B,inout_B,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2977df4d28bfSStefano Zampini ierr = VecScatterEnd(reuse_solver->correction_scatter_B,reuse_solver->sol_B,inout_B,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2978be83ff47SStefano Zampini } 297980677318SStefano Zampini if (!applytranspose && pcbddc->local_auxmat1) { 298080677318SStefano Zampini ierr = MatMult(pcbddc->local_auxmat1,inout_B,pcbddc->vec1_C);CHKERRQ(ierr); 298180677318SStefano Zampini ierr = MatMultAdd(pcbddc->local_auxmat2,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr); 298280677318SStefano Zampini } 298380677318SStefano Zampini } else { 298480677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 298580677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 298680677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 298780677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 298880677318SStefano Zampini if (!applytranspose && pcbddc->local_auxmat1) { 298980677318SStefano Zampini ierr = MatMult(pcbddc->local_auxmat1,inout_B,pcbddc->vec1_C);CHKERRQ(ierr); 299080677318SStefano Zampini ierr = MatMultAdd(pcbddc->local_auxmat2,pcbddc->vec1_C,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 299180677318SStefano Zampini } 299280677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 299380677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 299480677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 299580677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2996674ae819SStefano Zampini } 2997674ae819SStefano Zampini PetscFunctionReturn(0); 2998674ae819SStefano Zampini } 2999674ae819SStefano Zampini 3000dc359a40SStefano Zampini /* parameter apply transpose determines if the interface preconditioner should be applied transposed or not */ 3001674ae819SStefano Zampini #undef __FUNCT__ 3002674ae819SStefano Zampini #define __FUNCT__ "PCBDDCApplyInterfacePreconditioner" 3003dc359a40SStefano Zampini PetscErrorCode PCBDDCApplyInterfacePreconditioner(PC pc, PetscBool applytranspose) 3004674ae819SStefano Zampini { 3005674ae819SStefano Zampini PetscErrorCode ierr; 3006674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 3007674ae819SStefano Zampini PC_IS* pcis = (PC_IS*) (pc->data); 3008674ae819SStefano Zampini const PetscScalar zero = 0.0; 3009674ae819SStefano Zampini 3010674ae819SStefano Zampini PetscFunctionBegin; 3011dc359a40SStefano Zampini /* Application of PSI^T or PHI^T (depending on applytranspose, see comment above) */ 30124fee134fSStefano Zampini if (!pcbddc->benign_apply_coarse_only) { 3013dc359a40SStefano Zampini if (applytranspose) { 3014674ae819SStefano Zampini ierr = MatMultTranspose(pcbddc->coarse_phi_B,pcis->vec1_B,pcbddc->vec1_P);CHKERRQ(ierr); 30158eeda7d8SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultTransposeAdd(pcbddc->coarse_phi_D,pcis->vec1_D,pcbddc->vec1_P,pcbddc->vec1_P);CHKERRQ(ierr); } 3016dc359a40SStefano Zampini } else { 3017674ae819SStefano Zampini ierr = MatMultTranspose(pcbddc->coarse_psi_B,pcis->vec1_B,pcbddc->vec1_P);CHKERRQ(ierr); 3018674ae819SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultTransposeAdd(pcbddc->coarse_psi_D,pcis->vec1_D,pcbddc->vec1_P,pcbddc->vec1_P);CHKERRQ(ierr); } 301915aaf578SStefano Zampini } 30204fee134fSStefano Zampini } else { 30214fee134fSStefano Zampini ierr = VecSet(pcbddc->vec1_P,zero);CHKERRQ(ierr); 30224fee134fSStefano Zampini } 3023efc2fbd9SStefano Zampini 3024efc2fbd9SStefano Zampini /* add p0 to the last value of vec1_P holding the coarse dof relative to p0 */ 30254f1b2e48SStefano Zampini if (pcbddc->benign_n) { 3026efc2fbd9SStefano Zampini PetscScalar *array; 30274f1b2e48SStefano Zampini PetscInt j; 3028efc2fbd9SStefano Zampini 3029efc2fbd9SStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 30304f1b2e48SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) array[pcbddc->local_primal_size-pcbddc->benign_n+j] += pcbddc->benign_p0[j]; 3031efc2fbd9SStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 3032efc2fbd9SStefano Zampini } 3033efc2fbd9SStefano Zampini 303412edc857SStefano Zampini /* start communications from local primal nodes to rhs of coarse solver */ 303512edc857SStefano Zampini ierr = VecSet(pcbddc->coarse_vec,zero);CHKERRQ(ierr); 303612edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataBegin(pc,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 303712edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataEnd(pc,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 303812edc857SStefano Zampini 30399f00e9b4SStefano Zampini /* Coarse solution -> rhs and sol updated inside PCBDDCScattarCoarseDataBegin/End */ 304012edc857SStefano Zampini /* TODO remove null space when doing multilevel */ 304112edc857SStefano Zampini if (pcbddc->coarse_ksp) { 304251694757SStefano Zampini Mat coarse_mat; 3043964fefecSStefano Zampini Vec rhs,sol; 304451694757SStefano Zampini MatNullSpace nullsp; 3045964fefecSStefano Zampini 3046964fefecSStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&rhs);CHKERRQ(ierr); 3047964fefecSStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&sol);CHKERRQ(ierr); 304851694757SStefano Zampini ierr = KSPGetOperators(pcbddc->coarse_ksp,&coarse_mat,NULL);CHKERRQ(ierr); 304951694757SStefano Zampini ierr = MatGetNullSpace(coarse_mat,&nullsp);CHKERRQ(ierr); 305051694757SStefano Zampini if (nullsp) { 305151694757SStefano Zampini ierr = MatNullSpaceRemove(nullsp,rhs);CHKERRQ(ierr); 305251694757SStefano Zampini } 305312edc857SStefano Zampini if (applytranspose) { 3054964fefecSStefano Zampini ierr = KSPSolveTranspose(pcbddc->coarse_ksp,rhs,sol);CHKERRQ(ierr); 305512edc857SStefano Zampini } else { 3056964fefecSStefano Zampini ierr = KSPSolve(pcbddc->coarse_ksp,rhs,sol);CHKERRQ(ierr); 305712edc857SStefano Zampini } 305851694757SStefano Zampini if (nullsp) { 305951694757SStefano Zampini ierr = MatNullSpaceRemove(nullsp,sol);CHKERRQ(ierr); 306051694757SStefano Zampini } 306112edc857SStefano Zampini } 3062674ae819SStefano Zampini 3063674ae819SStefano Zampini /* Local solution on R nodes */ 30644fee134fSStefano Zampini if (pcis->n && !pcbddc->benign_apply_coarse_only) { 306580677318SStefano Zampini ierr = PCBDDCSolveSubstructureCorrection(pc,pcis->vec1_B,pcis->vec1_D,applytranspose);CHKERRQ(ierr); 30669f00e9b4SStefano Zampini } 30679f00e9b4SStefano Zampini /* communications from coarse sol to local primal nodes */ 30689f00e9b4SStefano Zampini ierr = PCBDDCScatterCoarseDataBegin(pc,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 306912edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataEnd(pc,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 3070674ae819SStefano Zampini 30714fee134fSStefano Zampini /* Sum contributions from the two levels */ 30724fee134fSStefano Zampini if (!pcbddc->benign_apply_coarse_only) { 3073dc359a40SStefano Zampini if (applytranspose) { 3074dc359a40SStefano Zampini ierr = MatMultAdd(pcbddc->coarse_psi_B,pcbddc->vec1_P,pcis->vec1_B,pcis->vec1_B);CHKERRQ(ierr); 3075dc359a40SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultAdd(pcbddc->coarse_psi_D,pcbddc->vec1_P,pcis->vec1_D,pcis->vec1_D);CHKERRQ(ierr); } 3076dc359a40SStefano Zampini } else { 3077674ae819SStefano Zampini ierr = MatMultAdd(pcbddc->coarse_phi_B,pcbddc->vec1_P,pcis->vec1_B,pcis->vec1_B);CHKERRQ(ierr); 30788eeda7d8SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultAdd(pcbddc->coarse_phi_D,pcbddc->vec1_P,pcis->vec1_D,pcis->vec1_D);CHKERRQ(ierr); } 3079dc359a40SStefano Zampini } 3080efc2fbd9SStefano Zampini /* store p0 */ 30814f1b2e48SStefano Zampini if (pcbddc->benign_n) { 3082efc2fbd9SStefano Zampini PetscScalar *array; 30834f1b2e48SStefano Zampini PetscInt j; 3084efc2fbd9SStefano Zampini 3085efc2fbd9SStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 30864f1b2e48SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) pcbddc->benign_p0[j] = array[pcbddc->local_primal_size-pcbddc->benign_n+j]; 3087efc2fbd9SStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 3088efc2fbd9SStefano Zampini } 30894fee134fSStefano Zampini } else { /* expand the coarse solution */ 30904fee134fSStefano Zampini if (applytranspose) { 30914fee134fSStefano Zampini ierr = MatMult(pcbddc->coarse_psi_B,pcbddc->vec1_P,pcis->vec1_B);CHKERRQ(ierr); 30924fee134fSStefano Zampini } else { 30934fee134fSStefano Zampini ierr = MatMult(pcbddc->coarse_phi_B,pcbddc->vec1_P,pcis->vec1_B);CHKERRQ(ierr); 30944fee134fSStefano Zampini } 30954fee134fSStefano Zampini } 3096674ae819SStefano Zampini PetscFunctionReturn(0); 3097674ae819SStefano Zampini } 3098674ae819SStefano Zampini 3099674ae819SStefano Zampini #undef __FUNCT__ 3100674ae819SStefano Zampini #define __FUNCT__ "PCBDDCScatterCoarseDataBegin" 310112edc857SStefano Zampini PetscErrorCode PCBDDCScatterCoarseDataBegin(PC pc,InsertMode imode, ScatterMode smode) 3102674ae819SStefano Zampini { 3103674ae819SStefano Zampini PetscErrorCode ierr; 3104674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 310558da7f69SStefano Zampini PetscScalar *array; 310612edc857SStefano Zampini Vec from,to; 3107674ae819SStefano Zampini 3108674ae819SStefano Zampini PetscFunctionBegin; 310912edc857SStefano Zampini if (smode == SCATTER_REVERSE) { /* from global to local -> get data from coarse solution */ 311012edc857SStefano Zampini from = pcbddc->coarse_vec; 311112edc857SStefano Zampini to = pcbddc->vec1_P; 311212edc857SStefano Zampini if (pcbddc->coarse_ksp) { /* get array from coarse processes */ 311312edc857SStefano Zampini Vec tvec; 311458da7f69SStefano Zampini 311558da7f69SStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 311658da7f69SStefano Zampini ierr = VecResetArray(tvec);CHKERRQ(ierr); 311712edc857SStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 311858da7f69SStefano Zampini ierr = VecGetArray(tvec,&array);CHKERRQ(ierr); 311958da7f69SStefano Zampini ierr = VecPlaceArray(from,array);CHKERRQ(ierr); 312058da7f69SStefano Zampini ierr = VecRestoreArray(tvec,&array);CHKERRQ(ierr); 312112edc857SStefano Zampini } 312212edc857SStefano Zampini } else { /* from local to global -> put data in coarse right hand side */ 312312edc857SStefano Zampini from = pcbddc->vec1_P; 312412edc857SStefano Zampini to = pcbddc->coarse_vec; 312512edc857SStefano Zampini } 312612edc857SStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,from,to,imode,smode);CHKERRQ(ierr); 3127674ae819SStefano Zampini PetscFunctionReturn(0); 3128674ae819SStefano Zampini } 3129674ae819SStefano Zampini 3130674ae819SStefano Zampini #undef __FUNCT__ 3131674ae819SStefano Zampini #define __FUNCT__ "PCBDDCScatterCoarseDataEnd" 313212edc857SStefano Zampini PetscErrorCode PCBDDCScatterCoarseDataEnd(PC pc, InsertMode imode, ScatterMode smode) 3133674ae819SStefano Zampini { 3134674ae819SStefano Zampini PetscErrorCode ierr; 3135674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 313658da7f69SStefano Zampini PetscScalar *array; 313712edc857SStefano Zampini Vec from,to; 3138674ae819SStefano Zampini 3139674ae819SStefano Zampini PetscFunctionBegin; 314012edc857SStefano Zampini if (smode == SCATTER_REVERSE) { /* from global to local -> get data from coarse solution */ 314112edc857SStefano Zampini from = pcbddc->coarse_vec; 314212edc857SStefano Zampini to = pcbddc->vec1_P; 314312edc857SStefano Zampini } else { /* from local to global -> put data in coarse right hand side */ 314412edc857SStefano Zampini from = pcbddc->vec1_P; 314512edc857SStefano Zampini to = pcbddc->coarse_vec; 314612edc857SStefano Zampini } 314712edc857SStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,from,to,imode,smode);CHKERRQ(ierr); 314812edc857SStefano Zampini if (smode == SCATTER_FORWARD) { 314912edc857SStefano Zampini if (pcbddc->coarse_ksp) { /* get array from coarse processes */ 315012edc857SStefano Zampini Vec tvec; 315158da7f69SStefano Zampini 315212edc857SStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 315358da7f69SStefano Zampini ierr = VecGetArray(to,&array);CHKERRQ(ierr); 315458da7f69SStefano Zampini ierr = VecPlaceArray(tvec,array);CHKERRQ(ierr); 315558da7f69SStefano Zampini ierr = VecRestoreArray(to,&array);CHKERRQ(ierr); 315658da7f69SStefano Zampini } 315758da7f69SStefano Zampini } else { 315858da7f69SStefano Zampini if (pcbddc->coarse_ksp) { /* restore array of pcbddc->coarse_vec */ 315958da7f69SStefano Zampini ierr = VecResetArray(from);CHKERRQ(ierr); 316012edc857SStefano Zampini } 316112edc857SStefano Zampini } 3162674ae819SStefano Zampini PetscFunctionReturn(0); 3163674ae819SStefano Zampini } 3164674ae819SStefano Zampini 3165984c4197SStefano Zampini /* uncomment for testing purposes */ 3166984c4197SStefano Zampini /* #define PETSC_MISSING_LAPACK_GESVD 1 */ 3167674ae819SStefano Zampini #undef __FUNCT__ 3168674ae819SStefano Zampini #define __FUNCT__ "PCBDDCConstraintsSetUp" 3169674ae819SStefano Zampini PetscErrorCode PCBDDCConstraintsSetUp(PC pc) 3170674ae819SStefano Zampini { 3171674ae819SStefano Zampini PetscErrorCode ierr; 3172674ae819SStefano Zampini PC_IS* pcis = (PC_IS*)(pc->data); 3173674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 3174674ae819SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 3175984c4197SStefano Zampini /* one and zero */ 3176984c4197SStefano Zampini PetscScalar one=1.0,zero=0.0; 3177984c4197SStefano Zampini /* space to store constraints and their local indices */ 31789162d606SStefano Zampini PetscScalar *constraints_data; 31799162d606SStefano Zampini PetscInt *constraints_idxs,*constraints_idxs_B; 31809162d606SStefano Zampini PetscInt *constraints_idxs_ptr,*constraints_data_ptr; 31819162d606SStefano Zampini PetscInt *constraints_n; 3182984c4197SStefano Zampini /* iterators */ 3183b3d85658SStefano Zampini PetscInt i,j,k,total_counts,total_counts_cc,cum; 3184984c4197SStefano Zampini /* BLAS integers */ 3185e310c8b4SStefano Zampini PetscBLASInt lwork,lierr; 3186e310c8b4SStefano Zampini PetscBLASInt Blas_N,Blas_M,Blas_K,Blas_one=1; 3187c4303822SStefano Zampini PetscBLASInt Blas_LDA,Blas_LDB,Blas_LDC; 3188727cdba6SStefano Zampini /* reuse */ 31890e6343abSStefano Zampini PetscInt olocal_primal_size,olocal_primal_size_cc; 31900e6343abSStefano Zampini PetscInt *olocal_primal_ref_node,*olocal_primal_ref_mult; 3191984c4197SStefano Zampini /* change of basis */ 3192b3d85658SStefano Zampini PetscBool qr_needed; 31939162d606SStefano Zampini PetscBT change_basis,qr_needed_idx; 3194984c4197SStefano Zampini /* auxiliary stuff */ 319564efe560SStefano Zampini PetscInt *nnz,*is_indices; 31968a0068c3SStefano Zampini PetscInt ncc; 3197984c4197SStefano Zampini /* some quantities */ 319845a1bb75SStefano Zampini PetscInt n_vertices,total_primal_vertices,valid_constraints; 3199a58a30b4SStefano Zampini PetscInt size_of_constraint,max_size_of_constraint=0,max_constraints,temp_constraints; 3200984c4197SStefano Zampini 3201674ae819SStefano Zampini PetscFunctionBegin; 32028e61c736SStefano Zampini /* Destroy Mat objects computed previously */ 32038e61c736SStefano Zampini ierr = MatDestroy(&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 32048e61c736SStefano Zampini ierr = MatDestroy(&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 3205088faed8SStefano Zampini /* save info on constraints from previous setup (if any) */ 3206088faed8SStefano Zampini olocal_primal_size = pcbddc->local_primal_size; 32070e6343abSStefano Zampini olocal_primal_size_cc = pcbddc->local_primal_size_cc; 32080e6343abSStefano Zampini ierr = PetscMalloc2(olocal_primal_size_cc,&olocal_primal_ref_node,olocal_primal_size_cc,&olocal_primal_ref_mult);CHKERRQ(ierr); 32090e6343abSStefano Zampini ierr = PetscMemcpy(olocal_primal_ref_node,pcbddc->local_primal_ref_node,olocal_primal_size_cc*sizeof(PetscInt));CHKERRQ(ierr); 32100e6343abSStefano Zampini ierr = PetscMemcpy(olocal_primal_ref_mult,pcbddc->local_primal_ref_mult,olocal_primal_size_cc*sizeof(PetscInt));CHKERRQ(ierr); 32110e6343abSStefano Zampini ierr = PetscFree2(pcbddc->local_primal_ref_node,pcbddc->local_primal_ref_mult);CHKERRQ(ierr); 3212088faed8SStefano Zampini ierr = PetscFree(pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 3213cf5a6209SStefano Zampini 3214cf5a6209SStefano Zampini /* print some info */ 3215cf5a6209SStefano Zampini if (pcbddc->dbg_flag) { 3216cf5a6209SStefano Zampini IS vertices; 3217cf5a6209SStefano Zampini PetscInt nv,nedges,nfaces; 3218cf5a6209SStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,&nfaces,NULL,&nedges,NULL,&vertices);CHKERRQ(ierr); 3219cf5a6209SStefano Zampini ierr = ISGetSize(vertices,&nv);CHKERRQ(ierr); 3220cf5a6209SStefano Zampini ierr = ISDestroy(&vertices);CHKERRQ(ierr); 32211575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 3222cf5a6209SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"--------------------------------------------------------------\n");CHKERRQ(ierr); 3223cf5a6209SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate vertices (%d)\n",PetscGlobalRank,nv,pcbddc->use_vertices);CHKERRQ(ierr); 3224fd14bc51SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate edges (%d)\n",PetscGlobalRank,nedges,pcbddc->use_edges);CHKERRQ(ierr); 3225fd14bc51SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate faces (%d)\n",PetscGlobalRank,nfaces,pcbddc->use_faces);CHKERRQ(ierr); 3226cf5a6209SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 32271575c14dSBarry Smith ierr = PetscViewerASCIIPopSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 3228cf5a6209SStefano Zampini } 3229cf5a6209SStefano Zampini 3230cf5a6209SStefano Zampini if (!pcbddc->adaptive_selection) { 32319162d606SStefano Zampini IS ISForVertices,*ISForFaces,*ISForEdges; 3232cf5a6209SStefano Zampini MatNullSpace nearnullsp; 3233cf5a6209SStefano Zampini const Vec *nearnullvecs; 3234cf5a6209SStefano Zampini Vec *localnearnullsp; 3235cf5a6209SStefano Zampini PetscScalar *array; 3236cf5a6209SStefano Zampini PetscInt n_ISForFaces,n_ISForEdges,nnsp_size; 3237cf5a6209SStefano Zampini PetscBool nnsp_has_cnst; 3238674ae819SStefano Zampini /* LAPACK working arrays for SVD or POD */ 3239b3d85658SStefano Zampini PetscBool skip_lapack,boolforchange; 3240674ae819SStefano Zampini PetscScalar *work; 3241674ae819SStefano Zampini PetscReal *singular_vals; 3242674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX) 3243674ae819SStefano Zampini PetscReal *rwork; 3244674ae819SStefano Zampini #endif 3245674ae819SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 3246674ae819SStefano Zampini PetscScalar *temp_basis,*correlation_mat; 3247674ae819SStefano Zampini #else 3248964fefecSStefano Zampini PetscBLASInt dummy_int=1; 3249964fefecSStefano Zampini PetscScalar dummy_scalar=1.; 3250674ae819SStefano Zampini #endif 3251674ae819SStefano Zampini 3252674ae819SStefano Zampini /* Get index sets for faces, edges and vertices from graph */ 3253d06fc5fdSStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,&n_ISForFaces,&ISForFaces,&n_ISForEdges,&ISForEdges,&ISForVertices);CHKERRQ(ierr); 3254d06fc5fdSStefano Zampini /* free unneeded index sets */ 3255d06fc5fdSStefano Zampini if (!pcbddc->use_vertices) { 3256d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForVertices);CHKERRQ(ierr); 3257674ae819SStefano Zampini } 3258d06fc5fdSStefano Zampini if (!pcbddc->use_edges) { 3259d06fc5fdSStefano Zampini for (i=0;i<n_ISForEdges;i++) { 3260d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForEdges[i]);CHKERRQ(ierr); 3261d06fc5fdSStefano Zampini } 3262d06fc5fdSStefano Zampini ierr = PetscFree(ISForEdges);CHKERRQ(ierr); 3263d06fc5fdSStefano Zampini n_ISForEdges = 0; 3264d06fc5fdSStefano Zampini } 3265d06fc5fdSStefano Zampini if (!pcbddc->use_faces) { 3266d06fc5fdSStefano Zampini for (i=0;i<n_ISForFaces;i++) { 3267d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForFaces[i]);CHKERRQ(ierr); 3268d06fc5fdSStefano Zampini } 3269d06fc5fdSStefano Zampini ierr = PetscFree(ISForFaces);CHKERRQ(ierr); 3270d06fc5fdSStefano Zampini n_ISForFaces = 0; 3271d06fc5fdSStefano Zampini } 327270022509SStefano Zampini 327370022509SStefano Zampini #if defined(PETSC_USE_DEBUG) 327470022509SStefano Zampini /* HACK: when solving singular problems not using vertices, a change of basis is mandatory. 327570022509SStefano Zampini Also use_change_of_basis should be consistent among processors */ 327670022509SStefano Zampini if (pcbddc->NullSpace) { 327770022509SStefano Zampini PetscBool tbool[2],gbool[2]; 327870022509SStefano Zampini 327970022509SStefano Zampini if (!ISForVertices && !pcbddc->user_ChangeOfBasisMatrix) { 3280b8ffe317SStefano Zampini pcbddc->use_change_of_basis = PETSC_TRUE; 3281d06fc5fdSStefano Zampini if (!ISForEdges) { 3282d06fc5fdSStefano Zampini pcbddc->use_change_on_faces = PETSC_TRUE; 3283d06fc5fdSStefano Zampini } 3284b8ffe317SStefano Zampini } 3285d06fc5fdSStefano Zampini tbool[0] = pcbddc->use_change_of_basis; 3286d06fc5fdSStefano Zampini tbool[1] = pcbddc->use_change_on_faces; 3287d06fc5fdSStefano Zampini ierr = MPI_Allreduce(tbool,gbool,2,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 3288d06fc5fdSStefano Zampini pcbddc->use_change_of_basis = gbool[0]; 3289d06fc5fdSStefano Zampini pcbddc->use_change_on_faces = gbool[1]; 329098a51de6SStefano Zampini } 329170022509SStefano Zampini #endif 329208122e43SStefano Zampini 3293674ae819SStefano Zampini /* check if near null space is attached to global mat */ 3294674ae819SStefano Zampini ierr = MatGetNearNullSpace(pc->pmat,&nearnullsp);CHKERRQ(ierr); 3295674ae819SStefano Zampini if (nearnullsp) { 3296674ae819SStefano Zampini ierr = MatNullSpaceGetVecs(nearnullsp,&nnsp_has_cnst,&nnsp_size,&nearnullvecs);CHKERRQ(ierr); 3297f4ddd8eeSStefano Zampini /* remove any stored info */ 3298f4ddd8eeSStefano Zampini ierr = MatNullSpaceDestroy(&pcbddc->onearnullspace);CHKERRQ(ierr); 3299f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->onearnullvecs_state);CHKERRQ(ierr); 3300f4ddd8eeSStefano Zampini /* store information for BDDC solver reuse */ 3301f4ddd8eeSStefano Zampini ierr = PetscObjectReference((PetscObject)nearnullsp);CHKERRQ(ierr); 3302f4ddd8eeSStefano Zampini pcbddc->onearnullspace = nearnullsp; 3303473ba861SJed Brown ierr = PetscMalloc1(nnsp_size,&pcbddc->onearnullvecs_state);CHKERRQ(ierr); 3304f4ddd8eeSStefano Zampini for (i=0;i<nnsp_size;i++) { 3305f4ddd8eeSStefano Zampini ierr = PetscObjectStateGet((PetscObject)nearnullvecs[i],&pcbddc->onearnullvecs_state[i]);CHKERRQ(ierr); 3306f4ddd8eeSStefano Zampini } 3307984c4197SStefano Zampini } else { /* if near null space is not provided BDDC uses constants by default */ 3308984c4197SStefano Zampini nnsp_size = 0; 3309674ae819SStefano Zampini nnsp_has_cnst = PETSC_TRUE; 3310674ae819SStefano Zampini } 3311984c4197SStefano Zampini /* get max number of constraints on a single cc */ 3312984c4197SStefano Zampini max_constraints = nnsp_size; 3313984c4197SStefano Zampini if (nnsp_has_cnst) max_constraints++; 3314984c4197SStefano Zampini 3315674ae819SStefano Zampini /* 3316674ae819SStefano Zampini Evaluate maximum storage size needed by the procedure 33179162d606SStefano Zampini - Indices for connected component i stored at "constraints_idxs + constraints_idxs_ptr[i]" 33189162d606SStefano Zampini - Values for constraints on connected component i stored at "constraints_data + constraints_data_ptr[i]" 33199162d606SStefano Zampini There can be multiple constraints per connected component 3320674ae819SStefano Zampini */ 3321674ae819SStefano Zampini n_vertices = 0; 3322674ae819SStefano Zampini if (ISForVertices) { 3323674ae819SStefano Zampini ierr = ISGetSize(ISForVertices,&n_vertices);CHKERRQ(ierr); 3324674ae819SStefano Zampini } 33259162d606SStefano Zampini ncc = n_vertices+n_ISForFaces+n_ISForEdges; 33269162d606SStefano Zampini ierr = PetscMalloc3(ncc+1,&constraints_idxs_ptr,ncc+1,&constraints_data_ptr,ncc,&constraints_n);CHKERRQ(ierr); 33279162d606SStefano Zampini 33289162d606SStefano Zampini total_counts = n_ISForFaces+n_ISForEdges; 33299162d606SStefano Zampini total_counts *= max_constraints; 3330674ae819SStefano Zampini total_counts += n_vertices; 33314641a718SStefano Zampini ierr = PetscBTCreate(total_counts,&change_basis);CHKERRQ(ierr); 33329162d606SStefano Zampini 3333674ae819SStefano Zampini total_counts = 0; 3334674ae819SStefano Zampini max_size_of_constraint = 0; 3335674ae819SStefano Zampini for (i=0;i<n_ISForEdges+n_ISForFaces;i++) { 33369162d606SStefano Zampini IS used_is; 3337674ae819SStefano Zampini if (i<n_ISForEdges) { 33389162d606SStefano Zampini used_is = ISForEdges[i]; 3339674ae819SStefano Zampini } else { 33409162d606SStefano Zampini used_is = ISForFaces[i-n_ISForEdges]; 3341674ae819SStefano Zampini } 33429162d606SStefano Zampini ierr = ISGetSize(used_is,&j);CHKERRQ(ierr); 3343674ae819SStefano Zampini total_counts += j; 3344674ae819SStefano Zampini max_size_of_constraint = PetscMax(j,max_size_of_constraint); 3345674ae819SStefano Zampini } 33469162d606SStefano 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); 33479162d606SStefano Zampini 3348984c4197SStefano Zampini /* get local part of global near null space vectors */ 3349785e854fSJed Brown ierr = PetscMalloc1(nnsp_size,&localnearnullsp);CHKERRQ(ierr); 3350984c4197SStefano Zampini for (k=0;k<nnsp_size;k++) { 3351984c4197SStefano Zampini ierr = VecDuplicate(pcis->vec1_N,&localnearnullsp[k]);CHKERRQ(ierr); 3352e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,nearnullvecs[k],localnearnullsp[k],INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 3353e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,nearnullvecs[k],localnearnullsp[k],INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 3354984c4197SStefano Zampini } 3355674ae819SStefano Zampini 3356242a89d7SStefano Zampini /* whether or not to skip lapack calls */ 3357242a89d7SStefano Zampini skip_lapack = PETSC_TRUE; 3358a773dcb8SStefano Zampini if (n_ISForFaces+n_ISForEdges && max_constraints > 1 && !pcbddc->use_nnsp_true) skip_lapack = PETSC_FALSE; 3359242a89d7SStefano Zampini 3360984c4197SStefano Zampini /* First we issue queries to allocate optimal workspace for LAPACKgesvd (or LAPACKsyev if SVD is missing) */ 3361a773dcb8SStefano Zampini if (!skip_lapack) { 3362674ae819SStefano Zampini PetscScalar temp_work; 3363911cabfeSStefano Zampini 3364674ae819SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 3365984c4197SStefano Zampini /* Proper Orthogonal Decomposition (POD) using the snapshot method */ 3366785e854fSJed Brown ierr = PetscMalloc1(max_constraints*max_constraints,&correlation_mat);CHKERRQ(ierr); 3367785e854fSJed Brown ierr = PetscMalloc1(max_constraints,&singular_vals);CHKERRQ(ierr); 3368785e854fSJed Brown ierr = PetscMalloc1(max_size_of_constraint*max_constraints,&temp_basis);CHKERRQ(ierr); 3369674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX) 3370785e854fSJed Brown ierr = PetscMalloc1(3*max_constraints,&rwork);CHKERRQ(ierr); 3371674ae819SStefano Zampini #endif 3372674ae819SStefano Zampini /* now we evaluate the optimal workspace using query with lwork=-1 */ 3373c8244a33SStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_N);CHKERRQ(ierr); 3374c8244a33SStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_LDA);CHKERRQ(ierr); 3375674ae819SStefano Zampini lwork = -1; 3376674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 3377674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 3378c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,&temp_work,&lwork,&lierr)); 3379674ae819SStefano Zampini #else 3380c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,&temp_work,&lwork,rwork,&lierr)); 3381674ae819SStefano Zampini #endif 3382674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 3383984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to SYEV Lapack routine %d",(int)lierr); 3384674ae819SStefano Zampini #else /* on missing GESVD */ 3385674ae819SStefano Zampini /* SVD */ 3386674ae819SStefano Zampini PetscInt max_n,min_n; 3387674ae819SStefano Zampini max_n = max_size_of_constraint; 3388984c4197SStefano Zampini min_n = max_constraints; 3389984c4197SStefano Zampini if (max_size_of_constraint < max_constraints) { 3390674ae819SStefano Zampini min_n = max_size_of_constraint; 3391984c4197SStefano Zampini max_n = max_constraints; 3392674ae819SStefano Zampini } 3393785e854fSJed Brown ierr = PetscMalloc1(min_n,&singular_vals);CHKERRQ(ierr); 3394674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX) 3395785e854fSJed Brown ierr = PetscMalloc1(5*min_n,&rwork);CHKERRQ(ierr); 3396674ae819SStefano Zampini #endif 3397674ae819SStefano Zampini /* now we evaluate the optimal workspace using query with lwork=-1 */ 3398674ae819SStefano Zampini lwork = -1; 3399e310c8b4SStefano Zampini ierr = PetscBLASIntCast(max_n,&Blas_M);CHKERRQ(ierr); 3400e310c8b4SStefano Zampini ierr = PetscBLASIntCast(min_n,&Blas_N);CHKERRQ(ierr); 3401b7d8b9f8SStefano Zampini ierr = PetscBLASIntCast(max_n,&Blas_LDA);CHKERRQ(ierr); 3402674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 3403674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 34049162d606SStefano 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)); 3405674ae819SStefano Zampini #else 34069162d606SStefano 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)); 3407674ae819SStefano Zampini #endif 3408674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 3409984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to GESVD Lapack routine %d",(int)lierr); 3410984c4197SStefano Zampini #endif /* on missing GESVD */ 3411674ae819SStefano Zampini /* Allocate optimal workspace */ 3412674ae819SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(temp_work),&lwork);CHKERRQ(ierr); 3413854ce69bSBarry Smith ierr = PetscMalloc1(lwork,&work);CHKERRQ(ierr); 3414674ae819SStefano Zampini } 3415674ae819SStefano Zampini /* Now we can loop on constraining sets */ 3416674ae819SStefano Zampini total_counts = 0; 34179162d606SStefano Zampini constraints_idxs_ptr[0] = 0; 34189162d606SStefano Zampini constraints_data_ptr[0] = 0; 3419674ae819SStefano Zampini /* vertices */ 34209162d606SStefano Zampini if (n_vertices) { 3421674ae819SStefano Zampini ierr = ISGetIndices(ISForVertices,(const PetscInt**)&is_indices);CHKERRQ(ierr); 34229162d606SStefano Zampini ierr = PetscMemcpy(constraints_idxs,is_indices,n_vertices*sizeof(PetscInt));CHKERRQ(ierr); 3423674ae819SStefano Zampini for (i=0;i<n_vertices;i++) { 34249162d606SStefano Zampini constraints_n[total_counts] = 1; 34259162d606SStefano Zampini constraints_data[total_counts] = 1.0; 34269162d606SStefano Zampini constraints_idxs_ptr[total_counts+1] = constraints_idxs_ptr[total_counts]+1; 34279162d606SStefano Zampini constraints_data_ptr[total_counts+1] = constraints_data_ptr[total_counts]+1; 3428674ae819SStefano Zampini total_counts++; 3429674ae819SStefano Zampini } 3430674ae819SStefano Zampini ierr = ISRestoreIndices(ISForVertices,(const PetscInt**)&is_indices);CHKERRQ(ierr); 3431674ae819SStefano Zampini n_vertices = total_counts; 3432674ae819SStefano Zampini } 3433984c4197SStefano Zampini 3434674ae819SStefano Zampini /* edges and faces */ 34359162d606SStefano Zampini total_counts_cc = total_counts; 3436911cabfeSStefano Zampini for (ncc=0;ncc<n_ISForEdges+n_ISForFaces;ncc++) { 34379162d606SStefano Zampini IS used_is; 34389162d606SStefano Zampini PetscBool idxs_copied = PETSC_FALSE; 34399162d606SStefano Zampini 3440911cabfeSStefano Zampini if (ncc<n_ISForEdges) { 34419162d606SStefano Zampini used_is = ISForEdges[ncc]; 3442984c4197SStefano Zampini boolforchange = pcbddc->use_change_of_basis; /* change or not the basis on the edge */ 3443674ae819SStefano Zampini } else { 34449162d606SStefano Zampini used_is = ISForFaces[ncc-n_ISForEdges]; 3445984c4197SStefano Zampini boolforchange = (PetscBool)(pcbddc->use_change_of_basis && pcbddc->use_change_on_faces); /* change or not the basis on the face */ 3446674ae819SStefano Zampini } 3447674ae819SStefano Zampini temp_constraints = 0; /* zero the number of constraints I have on this conn comp */ 34489162d606SStefano Zampini 34499162d606SStefano Zampini ierr = ISGetSize(used_is,&size_of_constraint);CHKERRQ(ierr); 34509162d606SStefano Zampini ierr = ISGetIndices(used_is,(const PetscInt**)&is_indices);CHKERRQ(ierr); 3451984c4197SStefano Zampini /* change of basis should not be performed on local periodic nodes */ 3452984c4197SStefano Zampini if (pcbddc->mat_graph->mirrors && pcbddc->mat_graph->mirrors[is_indices[0]]) boolforchange = PETSC_FALSE; 3453674ae819SStefano Zampini if (nnsp_has_cnst) { 34545b08dc53SStefano Zampini PetscScalar quad_value; 34559162d606SStefano Zampini 34569162d606SStefano Zampini ierr = PetscMemcpy(constraints_idxs + constraints_idxs_ptr[total_counts_cc],is_indices,size_of_constraint*sizeof(PetscInt));CHKERRQ(ierr); 34579162d606SStefano Zampini idxs_copied = PETSC_TRUE; 34589162d606SStefano Zampini 3459a773dcb8SStefano Zampini if (!pcbddc->use_nnsp_true) { 3460674ae819SStefano Zampini quad_value = (PetscScalar)(1.0/PetscSqrtReal((PetscReal)size_of_constraint)); 3461a773dcb8SStefano Zampini } else { 3462a773dcb8SStefano Zampini quad_value = 1.0; 3463a773dcb8SStefano Zampini } 3464674ae819SStefano Zampini for (j=0;j<size_of_constraint;j++) { 34659162d606SStefano Zampini constraints_data[constraints_data_ptr[total_counts_cc]+j] = quad_value; 3466674ae819SStefano Zampini } 34679162d606SStefano Zampini temp_constraints++; 3468674ae819SStefano Zampini total_counts++; 3469674ae819SStefano Zampini } 3470674ae819SStefano Zampini for (k=0;k<nnsp_size;k++) { 3471984c4197SStefano Zampini PetscReal real_value; 34729162d606SStefano Zampini PetscScalar *ptr_to_data; 34739162d606SStefano Zampini 3474984c4197SStefano Zampini ierr = VecGetArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 34759162d606SStefano Zampini ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]+temp_constraints*size_of_constraint]; 3476674ae819SStefano Zampini for (j=0;j<size_of_constraint;j++) { 34779162d606SStefano Zampini ptr_to_data[j] = array[is_indices[j]]; 3478674ae819SStefano Zampini } 3479984c4197SStefano Zampini ierr = VecRestoreArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 3480984c4197SStefano Zampini /* check if array is null on the connected component */ 3481e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 34829162d606SStefano Zampini PetscStackCallBLAS("BLASasum",real_value = BLASasum_(&Blas_N,ptr_to_data,&Blas_one)); 34835b08dc53SStefano Zampini if (real_value > 0.0) { /* keep indices and values */ 3484674ae819SStefano Zampini temp_constraints++; 3485674ae819SStefano Zampini total_counts++; 34869162d606SStefano Zampini if (!idxs_copied) { 34879162d606SStefano Zampini ierr = PetscMemcpy(constraints_idxs + constraints_idxs_ptr[total_counts_cc],is_indices,size_of_constraint*sizeof(PetscInt));CHKERRQ(ierr); 34889162d606SStefano Zampini idxs_copied = PETSC_TRUE; 3489674ae819SStefano Zampini } 3490674ae819SStefano Zampini } 34919162d606SStefano Zampini } 34929162d606SStefano Zampini ierr = ISRestoreIndices(used_is,(const PetscInt**)&is_indices);CHKERRQ(ierr); 349345a1bb75SStefano Zampini valid_constraints = temp_constraints; 3494eb97c9d2SStefano Zampini if (!pcbddc->use_nnsp_true && temp_constraints) { 3495a773dcb8SStefano Zampini if (temp_constraints == 1) { /* just normalize the constraint */ 34969162d606SStefano Zampini PetscScalar norm,*ptr_to_data; 34979162d606SStefano Zampini 34989162d606SStefano Zampini ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]]; 3499a773dcb8SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 35009162d606SStefano Zampini PetscStackCallBLAS("BLASdot",norm = BLASdot_(&Blas_N,ptr_to_data,&Blas_one,ptr_to_data,&Blas_one)); 3501a773dcb8SStefano Zampini norm = 1.0/PetscSqrtReal(PetscRealPart(norm)); 35029162d606SStefano Zampini PetscStackCallBLAS("BLASscal",BLASscal_(&Blas_N,&norm,ptr_to_data,&Blas_one)); 3503a773dcb8SStefano Zampini } else { /* perform SVD */ 3504984c4197SStefano Zampini PetscReal tol = 1.0e-8; /* tolerance for retaining eigenmodes */ 35059162d606SStefano Zampini PetscScalar *ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]]; 3506674ae819SStefano Zampini 3507674ae819SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 3508984c4197SStefano Zampini /* SVD: Y = U*S*V^H -> U (eigenvectors of Y*Y^H) = Y*V*(S)^\dag 3509984c4197SStefano Zampini POD: Y^H*Y = V*D*V^H, D = S^H*S -> U = Y*V*D^(-1/2) 3510984c4197SStefano Zampini -> When PETSC_USE_COMPLEX and PETSC_MISSING_LAPACK_GESVD are defined 3511984c4197SStefano Zampini the constraints basis will differ (by a complex factor with absolute value equal to 1) 3512984c4197SStefano Zampini from that computed using LAPACKgesvd 3513984c4197SStefano Zampini -> This is due to a different computation of eigenvectors in LAPACKheev 3514984c4197SStefano Zampini -> The quality of the POD-computed basis will be the same */ 3515984c4197SStefano Zampini ierr = PetscMemzero(correlation_mat,temp_constraints*temp_constraints*sizeof(PetscScalar));CHKERRQ(ierr); 3516674ae819SStefano Zampini /* Store upper triangular part of correlation matrix */ 3517e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 3518984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 3519674ae819SStefano Zampini for (j=0;j<temp_constraints;j++) { 3520674ae819SStefano Zampini for (k=0;k<j+1;k++) { 35219162d606SStefano 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)); 3522674ae819SStefano Zampini } 3523674ae819SStefano Zampini } 3524e310c8b4SStefano Zampini /* compute eigenvalues and eigenvectors of correlation matrix */ 3525e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 3526e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_LDA);CHKERRQ(ierr); 3527674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 3528c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,work,&lwork,&lierr)); 3529674ae819SStefano Zampini #else 3530c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,work,&lwork,rwork,&lierr)); 3531674ae819SStefano Zampini #endif 3532674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 3533984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYEV Lapack routine %d",(int)lierr); 3534984c4197SStefano Zampini /* retain eigenvalues greater than tol: note that LAPACKsyev gives eigs in ascending order */ 3535674ae819SStefano Zampini j = 0; 3536984c4197SStefano Zampini while (j < temp_constraints && singular_vals[j] < tol) j++; 3537674ae819SStefano Zampini total_counts = total_counts-j; 353845a1bb75SStefano Zampini valid_constraints = temp_constraints-j; 3539e310c8b4SStefano Zampini /* scale and copy POD basis into used quadrature memory */ 3540c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 3541c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 3542c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_K);CHKERRQ(ierr); 3543c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 3544c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_LDB);CHKERRQ(ierr); 3545c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDC);CHKERRQ(ierr); 3546674ae819SStefano Zampini if (j<temp_constraints) { 3547984c4197SStefano Zampini PetscInt ii; 3548984c4197SStefano Zampini for (k=j;k<temp_constraints;k++) singular_vals[k] = 1.0/PetscSqrtReal(singular_vals[k]); 3549674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 35509162d606SStefano 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)); 3551674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 3552984c4197SStefano Zampini for (k=0;k<temp_constraints-j;k++) { 3553674ae819SStefano Zampini for (ii=0;ii<size_of_constraint;ii++) { 35549162d606SStefano 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]; 3555674ae819SStefano Zampini } 3556674ae819SStefano Zampini } 3557674ae819SStefano Zampini } 3558674ae819SStefano Zampini #else /* on missing GESVD */ 3559e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 3560e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 3561b7d8b9f8SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 3562674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 3563674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 35649162d606SStefano 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)); 3565674ae819SStefano Zampini #else 35669162d606SStefano 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)); 3567674ae819SStefano Zampini #endif 3568984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GESVD Lapack routine %d",(int)lierr); 3569674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 3570984c4197SStefano Zampini /* retain eigenvalues greater than tol: note that LAPACKgesvd gives eigs in descending order */ 3571e310c8b4SStefano Zampini k = temp_constraints; 3572e310c8b4SStefano Zampini if (k > size_of_constraint) k = size_of_constraint; 3573674ae819SStefano Zampini j = 0; 3574e310c8b4SStefano Zampini while (j < k && singular_vals[k-j-1] < tol) j++; 357545a1bb75SStefano Zampini valid_constraints = k-j; 3576911cabfeSStefano Zampini total_counts = total_counts-temp_constraints+valid_constraints; 3577984c4197SStefano Zampini #endif /* on missing GESVD */ 3578674ae819SStefano Zampini } 3579a773dcb8SStefano Zampini } 35809162d606SStefano Zampini /* update pointers information */ 35819162d606SStefano Zampini if (valid_constraints) { 35829162d606SStefano Zampini constraints_n[total_counts_cc] = valid_constraints; 35839162d606SStefano Zampini constraints_idxs_ptr[total_counts_cc+1] = constraints_idxs_ptr[total_counts_cc]+size_of_constraint; 35849162d606SStefano Zampini constraints_data_ptr[total_counts_cc+1] = constraints_data_ptr[total_counts_cc]+size_of_constraint*valid_constraints; 35859162d606SStefano Zampini /* set change_of_basis flag */ 358645a1bb75SStefano Zampini if (boolforchange) { 3587b3d85658SStefano Zampini PetscBTSet(change_basis,total_counts_cc); 35889162d606SStefano Zampini } 3589b3d85658SStefano Zampini total_counts_cc++; 359045a1bb75SStefano Zampini } 359145a1bb75SStefano Zampini } 3592984c4197SStefano Zampini /* free workspace */ 35938f1c130eSStefano Zampini if (!skip_lapack) { 3594984c4197SStefano Zampini ierr = PetscFree(work);CHKERRQ(ierr); 3595984c4197SStefano Zampini #if defined(PETSC_USE_COMPLEX) 3596984c4197SStefano Zampini ierr = PetscFree(rwork);CHKERRQ(ierr); 3597984c4197SStefano Zampini #endif 3598984c4197SStefano Zampini ierr = PetscFree(singular_vals);CHKERRQ(ierr); 3599984c4197SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 3600984c4197SStefano Zampini ierr = PetscFree(correlation_mat);CHKERRQ(ierr); 3601984c4197SStefano Zampini ierr = PetscFree(temp_basis);CHKERRQ(ierr); 3602984c4197SStefano Zampini #endif 3603984c4197SStefano Zampini } 3604984c4197SStefano Zampini for (k=0;k<nnsp_size;k++) { 3605984c4197SStefano Zampini ierr = VecDestroy(&localnearnullsp[k]);CHKERRQ(ierr); 3606984c4197SStefano Zampini } 3607984c4197SStefano Zampini ierr = PetscFree(localnearnullsp);CHKERRQ(ierr); 3608cf5a6209SStefano Zampini /* free index sets of faces, edges and vertices */ 3609cf5a6209SStefano Zampini for (i=0;i<n_ISForFaces;i++) { 3610cf5a6209SStefano Zampini ierr = ISDestroy(&ISForFaces[i]);CHKERRQ(ierr); 3611cf5a6209SStefano Zampini } 3612cf5a6209SStefano Zampini if (n_ISForFaces) { 3613cf5a6209SStefano Zampini ierr = PetscFree(ISForFaces);CHKERRQ(ierr); 3614cf5a6209SStefano Zampini } 3615cf5a6209SStefano Zampini for (i=0;i<n_ISForEdges;i++) { 3616cf5a6209SStefano Zampini ierr = ISDestroy(&ISForEdges[i]);CHKERRQ(ierr); 3617cf5a6209SStefano Zampini } 3618cf5a6209SStefano Zampini if (n_ISForEdges) { 3619cf5a6209SStefano Zampini ierr = PetscFree(ISForEdges);CHKERRQ(ierr); 3620cf5a6209SStefano Zampini } 3621cf5a6209SStefano Zampini ierr = ISDestroy(&ISForVertices);CHKERRQ(ierr); 362208122e43SStefano Zampini } else { 362308122e43SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 3624984c4197SStefano Zampini 362508122e43SStefano Zampini total_counts = 0; 362608122e43SStefano Zampini n_vertices = 0; 3627d62866d3SStefano Zampini if (sub_schurs->is_vertices && pcbddc->use_vertices) { 3628d62866d3SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_vertices,&n_vertices);CHKERRQ(ierr); 362908122e43SStefano Zampini } 363008122e43SStefano Zampini max_constraints = 0; 36319162d606SStefano Zampini total_counts_cc = 0; 363208122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs+n_vertices;i++) { 363308122e43SStefano Zampini total_counts += pcbddc->adaptive_constraints_n[i]; 36349162d606SStefano Zampini if (pcbddc->adaptive_constraints_n[i]) total_counts_cc++; 363508122e43SStefano Zampini max_constraints = PetscMax(max_constraints,pcbddc->adaptive_constraints_n[i]); 363608122e43SStefano Zampini } 36379162d606SStefano Zampini constraints_idxs_ptr = pcbddc->adaptive_constraints_idxs_ptr; 36389162d606SStefano Zampini constraints_data_ptr = pcbddc->adaptive_constraints_data_ptr; 36399162d606SStefano Zampini constraints_idxs = pcbddc->adaptive_constraints_idxs; 36409162d606SStefano Zampini constraints_data = pcbddc->adaptive_constraints_data; 364174d5cdf7SStefano Zampini /* constraints_n differs from pcbddc->adaptive_constraints_n */ 36429162d606SStefano Zampini ierr = PetscMalloc1(total_counts_cc,&constraints_n);CHKERRQ(ierr); 36439162d606SStefano Zampini total_counts_cc = 0; 36449162d606SStefano Zampini for (i=0;i<sub_schurs->n_subs+n_vertices;i++) { 36459162d606SStefano Zampini if (pcbddc->adaptive_constraints_n[i]) { 36469162d606SStefano Zampini constraints_n[total_counts_cc++] = pcbddc->adaptive_constraints_n[i]; 364708122e43SStefano Zampini } 364808122e43SStefano Zampini } 36499162d606SStefano Zampini #if 0 36509162d606SStefano Zampini printf("Found %d totals (%d)\n",total_counts_cc,total_counts); 36519162d606SStefano Zampini for (i=0;i<total_counts_cc;i++) { 36529162d606SStefano Zampini printf("const %d, start %d",i,constraints_idxs_ptr[i]); 36539162d606SStefano Zampini printf(" end %d:\n",constraints_idxs_ptr[i+1]); 36549162d606SStefano Zampini for (j=constraints_idxs_ptr[i];j<constraints_idxs_ptr[i+1];j++) { 36559162d606SStefano Zampini printf(" %d",constraints_idxs[j]); 36569162d606SStefano Zampini } 36579162d606SStefano Zampini printf("\n"); 36589162d606SStefano Zampini printf("number of cc: %d\n",constraints_n[i]); 36599162d606SStefano Zampini } 36601b968477SStefano Zampini for (i=0;i<n_vertices;i++) { 36618bec7fa6SStefano Zampini PetscPrintf(PETSC_COMM_SELF,"[%d] vertex %d, n %d\n",PetscGlobalRank,i,pcbddc->adaptive_constraints_n[i]); 36621b968477SStefano Zampini } 36631b968477SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 36648bec7fa6SStefano 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]); 36651b968477SStefano Zampini } 366608122e43SStefano Zampini #endif 366708122e43SStefano Zampini 36688bec7fa6SStefano Zampini max_size_of_constraint = 0; 36699162d606SStefano 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]); 36709162d606SStefano Zampini ierr = PetscMalloc1(constraints_idxs_ptr[total_counts_cc],&constraints_idxs_B);CHKERRQ(ierr); 367108122e43SStefano Zampini /* Change of basis */ 3672b3d85658SStefano Zampini ierr = PetscBTCreate(total_counts_cc,&change_basis);CHKERRQ(ierr); 367308122e43SStefano Zampini if (pcbddc->use_change_of_basis) { 367408122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 367508122e43SStefano Zampini if (PetscBTLookup(sub_schurs->is_edge,i) || pcbddc->use_change_on_faces) { 3676b3d85658SStefano Zampini ierr = PetscBTSet(change_basis,i+n_vertices);CHKERRQ(ierr); 367708122e43SStefano Zampini } 367808122e43SStefano Zampini } 367908122e43SStefano Zampini } 368008122e43SStefano Zampini } 3681984c4197SStefano Zampini pcbddc->local_primal_size = total_counts; 36824f1b2e48SStefano Zampini ierr = PetscMalloc1(pcbddc->local_primal_size+pcbddc->benign_n,&pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 368308122e43SStefano Zampini 36849162d606SStefano Zampini /* map constraints_idxs in boundary numbering */ 36859162d606SStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,constraints_idxs_ptr[total_counts_cc],constraints_idxs,&i,constraints_idxs_B);CHKERRQ(ierr); 36869162d606SStefano Zampini if (i != constraints_idxs_ptr[total_counts_cc]) { 36879162d606SStefano 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); 368808122e43SStefano Zampini } 3689674ae819SStefano Zampini 3690674ae819SStefano Zampini /* Create constraint matrix */ 3691674ae819SStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 369216f15bc4SStefano Zampini ierr = MatSetType(pcbddc->ConstraintMatrix,MATAIJ);CHKERRQ(ierr); 3693984c4197SStefano Zampini ierr = MatSetSizes(pcbddc->ConstraintMatrix,pcbddc->local_primal_size,pcis->n,pcbddc->local_primal_size,pcis->n);CHKERRQ(ierr); 3694984c4197SStefano Zampini 3695984c4197SStefano Zampini /* find primal_dofs: subdomain corners plus dofs selected as primal after change of basis */ 3696a717540cSStefano Zampini /* determine if a QR strategy is needed for change of basis */ 3697a717540cSStefano Zampini qr_needed = PETSC_FALSE; 369874d5cdf7SStefano Zampini ierr = PetscBTCreate(total_counts_cc,&qr_needed_idx);CHKERRQ(ierr); 3699984c4197SStefano Zampini total_primal_vertices=0; 3700b3d85658SStefano Zampini pcbddc->local_primal_size_cc = 0; 37019162d606SStefano Zampini for (i=0;i<total_counts_cc;i++) { 37029162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 3703984c4197SStefano Zampini if (size_of_constraint == 1) { 37049162d606SStefano Zampini pcbddc->primal_indices_local_idxs[total_primal_vertices++] = constraints_idxs[constraints_idxs_ptr[i]]; 3705b3d85658SStefano Zampini pcbddc->local_primal_size_cc += 1; 370664efe560SStefano Zampini } else if (PetscBTLookup(change_basis,i)) { 37079162d606SStefano Zampini for (k=0;k<constraints_n[i];k++) { 37089162d606SStefano Zampini pcbddc->primal_indices_local_idxs[total_primal_vertices++] = constraints_idxs[constraints_idxs_ptr[i]+k]; 3709a717540cSStefano Zampini } 3710b3d85658SStefano Zampini pcbddc->local_primal_size_cc += constraints_n[i]; 371174d5cdf7SStefano Zampini if (constraints_n[i] > 1 || pcbddc->use_qr_single || pcbddc->faster_deluxe) { 3712a717540cSStefano Zampini PetscBTSet(qr_needed_idx,i); 3713a717540cSStefano Zampini qr_needed = PETSC_TRUE; 3714a717540cSStefano Zampini } 3715fa434743SStefano Zampini } else { 3716b3d85658SStefano Zampini pcbddc->local_primal_size_cc += 1; 3717fa434743SStefano Zampini } 3718a717540cSStefano Zampini } 3719b371cd4fSStefano Zampini /* note that the local variable n_vertices used below stores the number of pointwise constraints */ 3720b371cd4fSStefano Zampini pcbddc->n_vertices = total_primal_vertices; 3721674ae819SStefano Zampini /* permute indices in order to have a sorted set of vertices */ 372270022509SStefano Zampini ierr = PetscSortInt(total_primal_vertices,pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 3723b3d85658SStefano Zampini 37244f1b2e48SStefano 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); 37250e6343abSStefano Zampini ierr = PetscMemcpy(pcbddc->local_primal_ref_node,pcbddc->primal_indices_local_idxs,total_primal_vertices*sizeof(PetscInt));CHKERRQ(ierr); 37260e6343abSStefano Zampini for (i=0;i<total_primal_vertices;i++) pcbddc->local_primal_ref_mult[i] = 1; 3727984c4197SStefano Zampini 3728984c4197SStefano Zampini /* nonzero structure of constraint matrix */ 372974d5cdf7SStefano Zampini /* and get reference dof for local constraints */ 3730785e854fSJed Brown ierr = PetscMalloc1(pcbddc->local_primal_size,&nnz);CHKERRQ(ierr); 3731984c4197SStefano Zampini for (i=0;i<total_primal_vertices;i++) nnz[i] = 1; 373274d5cdf7SStefano Zampini 3733984c4197SStefano Zampini j = total_primal_vertices; 373474d5cdf7SStefano Zampini total_counts = total_primal_vertices; 3735b3d85658SStefano Zampini cum = total_primal_vertices; 37369162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 37374641a718SStefano Zampini if (!PetscBTLookup(change_basis,i)) { 3738b3d85658SStefano Zampini pcbddc->local_primal_ref_node[cum] = constraints_idxs[constraints_idxs_ptr[i]]; 3739b3d85658SStefano Zampini pcbddc->local_primal_ref_mult[cum] = constraints_n[i]; 3740b3d85658SStefano Zampini cum++; 37419162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 374274d5cdf7SStefano Zampini for (k=0;k<constraints_n[i];k++) { 374374d5cdf7SStefano Zampini pcbddc->primal_indices_local_idxs[total_counts++] = constraints_idxs[constraints_idxs_ptr[i]+k]; 374474d5cdf7SStefano Zampini nnz[j+k] = size_of_constraint; 374574d5cdf7SStefano Zampini } 37469162d606SStefano Zampini j += constraints_n[i]; 3747674ae819SStefano Zampini } 3748674ae819SStefano Zampini } 3749674ae819SStefano Zampini ierr = MatSeqAIJSetPreallocation(pcbddc->ConstraintMatrix,0,nnz);CHKERRQ(ierr); 3750674ae819SStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 3751088faed8SStefano Zampini 3752674ae819SStefano Zampini /* set values in constraint matrix */ 3753984c4197SStefano Zampini for (i=0;i<total_primal_vertices;i++) { 37540e6343abSStefano Zampini ierr = MatSetValue(pcbddc->ConstraintMatrix,i,pcbddc->local_primal_ref_node[i],1.0,INSERT_VALUES);CHKERRQ(ierr); 3755674ae819SStefano Zampini } 3756984c4197SStefano Zampini total_counts = total_primal_vertices; 37579162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 37584641a718SStefano Zampini if (!PetscBTLookup(change_basis,i)) { 37599162d606SStefano Zampini PetscInt *cols; 37609162d606SStefano Zampini 37619162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 37629162d606SStefano Zampini cols = constraints_idxs+constraints_idxs_ptr[i]; 37639162d606SStefano Zampini for (k=0;k<constraints_n[i];k++) { 37649162d606SStefano Zampini PetscInt row = total_counts+k; 37659162d606SStefano Zampini PetscScalar *vals; 37669162d606SStefano Zampini 37679162d606SStefano Zampini vals = constraints_data+constraints_data_ptr[i]+k*size_of_constraint; 37689162d606SStefano Zampini ierr = MatSetValues(pcbddc->ConstraintMatrix,1,&row,size_of_constraint,cols,vals,INSERT_VALUES);CHKERRQ(ierr); 37699162d606SStefano Zampini } 37709162d606SStefano Zampini total_counts += constraints_n[i]; 3771674ae819SStefano Zampini } 3772674ae819SStefano Zampini } 3773674ae819SStefano Zampini /* assembling */ 3774674ae819SStefano Zampini ierr = MatAssemblyBegin(pcbddc->ConstraintMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3775674ae819SStefano Zampini ierr = MatAssemblyEnd(pcbddc->ConstraintMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3776088faed8SStefano Zampini 3777984c4197SStefano Zampini /* 377845a1bb75SStefano Zampini ierr = PetscViewerSetFormat(PETSC_VIEWER_STDOUT_SELF,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 3779984c4197SStefano Zampini ierr = MatView(pcbddc->ConstraintMatrix,(PetscViewer)0);CHKERRQ(ierr); 3780984c4197SStefano Zampini */ 3781674ae819SStefano Zampini /* Create matrix for change of basis. We don't need it in case pcbddc->use_change_of_basis is FALSE */ 3782674ae819SStefano Zampini if (pcbddc->use_change_of_basis) { 3783026de310SStefano Zampini /* dual and primal dofs on a single cc */ 3784984c4197SStefano Zampini PetscInt dual_dofs,primal_dofs; 3785984c4197SStefano Zampini /* working stuff for GEQRF */ 378681d9aea3SBarry Smith PetscScalar *qr_basis,*qr_tau = NULL,*qr_work,lqr_work_t; 3787984c4197SStefano Zampini PetscBLASInt lqr_work; 3788984c4197SStefano Zampini /* working stuff for UNGQR */ 3789984c4197SStefano Zampini PetscScalar *gqr_work,lgqr_work_t; 3790984c4197SStefano Zampini PetscBLASInt lgqr_work; 3791984c4197SStefano Zampini /* working stuff for TRTRS */ 3792984c4197SStefano Zampini PetscScalar *trs_rhs; 37933f08241aSStefano Zampini PetscBLASInt Blas_NRHS; 3794984c4197SStefano Zampini /* pointers for values insertion into change of basis matrix */ 3795984c4197SStefano Zampini PetscInt *start_rows,*start_cols; 3796984c4197SStefano Zampini PetscScalar *start_vals; 3797984c4197SStefano Zampini /* working stuff for values insertion */ 37984641a718SStefano Zampini PetscBT is_primal; 379964efe560SStefano Zampini PetscInt *aux_primal_numbering_B; 3800906d46d4SStefano Zampini /* matrix sizes */ 3801906d46d4SStefano Zampini PetscInt global_size,local_size; 3802906d46d4SStefano Zampini /* temporary change of basis */ 3803906d46d4SStefano Zampini Mat localChangeOfBasisMatrix; 3804cf5a6209SStefano Zampini /* extra space for debugging */ 3805cf5a6209SStefano Zampini PetscScalar *dbg_work; 3806984c4197SStefano Zampini 3807906d46d4SStefano Zampini /* local temporary change of basis acts on local interfaces -> dimension is n_B x n_B */ 3808906d46d4SStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&localChangeOfBasisMatrix);CHKERRQ(ierr); 380916f15bc4SStefano Zampini ierr = MatSetType(localChangeOfBasisMatrix,MATAIJ);CHKERRQ(ierr); 3810bbb9e6c6SStefano Zampini ierr = MatSetSizes(localChangeOfBasisMatrix,pcis->n,pcis->n,pcis->n,pcis->n);CHKERRQ(ierr); 3811906d46d4SStefano Zampini /* nonzeros for local mat */ 3812bbb9e6c6SStefano Zampini ierr = PetscMalloc1(pcis->n,&nnz);CHKERRQ(ierr); 3813bbb9e6c6SStefano Zampini for (i=0;i<pcis->n;i++) nnz[i]=1; 38149162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 3815a717540cSStefano Zampini if (PetscBTLookup(change_basis,i)) { 38169162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 3817a717540cSStefano Zampini if (PetscBTLookup(qr_needed_idx,i)) { 38189162d606SStefano Zampini for (j=0;j<size_of_constraint;j++) nnz[constraints_idxs[constraints_idxs_ptr[i]+j]] = size_of_constraint; 3819a717540cSStefano Zampini } else { 38209162d606SStefano Zampini nnz[constraints_idxs[constraints_idxs_ptr[i]]] = size_of_constraint; 38219162d606SStefano Zampini for (j=1;j<size_of_constraint;j++) nnz[constraints_idxs[constraints_idxs_ptr[i]+j]] = 2; 3822a717540cSStefano Zampini } 3823a717540cSStefano Zampini } 3824a717540cSStefano Zampini } 3825906d46d4SStefano Zampini ierr = MatSeqAIJSetPreallocation(localChangeOfBasisMatrix,0,nnz);CHKERRQ(ierr); 3826bbb9e6c6SStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 3827a717540cSStefano Zampini /* Set initial identity in the matrix */ 3828bbb9e6c6SStefano Zampini for (i=0;i<pcis->n;i++) { 3829906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,i,i,1.0,INSERT_VALUES);CHKERRQ(ierr); 3830a717540cSStefano Zampini } 3831a717540cSStefano Zampini 3832a717540cSStefano Zampini if (pcbddc->dbg_flag) { 3833a717540cSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"--------------------------------------------------------------\n");CHKERRQ(ierr); 3834a717540cSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Checking change of basis computation for subdomain %04d\n",PetscGlobalRank);CHKERRQ(ierr); 3835a717540cSStefano Zampini } 3836a717540cSStefano Zampini 3837a717540cSStefano Zampini 3838a717540cSStefano Zampini /* Now we loop on the constraints which need a change of basis */ 3839a717540cSStefano Zampini /* 3840a717540cSStefano Zampini Change of basis matrix is evaluated similarly to the FIRST APPROACH in 3841a717540cSStefano Zampini Klawonn and Widlund, Dual-primal FETI-DP methods for linear elasticity, (see Sect 6.2.1) 3842a717540cSStefano Zampini 3843a6b551f4SStefano Zampini Basic blocks of change of basis matrix T computed by 3844a717540cSStefano Zampini 3845a6b551f4SStefano 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) 3846a6b551f4SStefano Zampini 3847a6b551f4SStefano Zampini | 1 0 ... 0 s_1/S | 3848a6b551f4SStefano Zampini | 0 1 ... 0 s_2/S | 3849a717540cSStefano Zampini | ... | 3850a6b551f4SStefano Zampini | 0 ... 1 s_{n-1}/S | 3851a6b551f4SStefano Zampini | -s_1/s_n ... -s_{n-1}/s_n s_n/S | 3852a717540cSStefano Zampini 3853a6b551f4SStefano Zampini with S = \sum_{i=1}^n s_i^2 3854a6b551f4SStefano Zampini NOTE: in the above example, the primal dof is the last one of the edge in LOCAL ordering 3855a6b551f4SStefano Zampini in the current implementation, the primal dof is the first one of the edge in GLOBAL ordering 3856a6b551f4SStefano Zampini 3857a6b551f4SStefano Zampini - QR decomposition of constraints otherwise 3858a717540cSStefano Zampini */ 3859a717540cSStefano Zampini if (qr_needed) { 3860984c4197SStefano Zampini /* space to store Q */ 3861854ce69bSBarry Smith ierr = PetscMalloc1(max_size_of_constraint*max_size_of_constraint,&qr_basis);CHKERRQ(ierr); 3862984c4197SStefano Zampini /* first we issue queries for optimal work */ 38633f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_M);CHKERRQ(ierr); 38643f08241aSStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_N);CHKERRQ(ierr); 38653f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 3866984c4197SStefano Zampini lqr_work = -1; 38673f08241aSStefano Zampini PetscStackCallBLAS("LAPACKgeqrf",LAPACKgeqrf_(&Blas_M,&Blas_N,qr_basis,&Blas_LDA,qr_tau,&lqr_work_t,&lqr_work,&lierr)); 3868984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to GEQRF Lapack routine %d",(int)lierr); 3869984c4197SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lqr_work_t),&lqr_work);CHKERRQ(ierr); 3870785e854fSJed Brown ierr = PetscMalloc1((PetscInt)PetscRealPart(lqr_work_t),&qr_work);CHKERRQ(ierr); 3871984c4197SStefano Zampini lgqr_work = -1; 38723f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_M);CHKERRQ(ierr); 38733f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_N);CHKERRQ(ierr); 38743f08241aSStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_K);CHKERRQ(ierr); 38753f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 38763f08241aSStefano Zampini if (Blas_K>Blas_M) Blas_K=Blas_M; /* adjust just for computing optimal work */ 38773f08241aSStefano Zampini PetscStackCallBLAS("LAPACKungqr",LAPACKungqr_(&Blas_M,&Blas_N,&Blas_K,qr_basis,&Blas_LDA,qr_tau,&lgqr_work_t,&lgqr_work,&lierr)); 3878984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to UNGQR Lapack routine %d",(int)lierr); 3879984c4197SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lgqr_work_t),&lgqr_work);CHKERRQ(ierr); 3880785e854fSJed Brown ierr = PetscMalloc1((PetscInt)PetscRealPart(lgqr_work_t),&gqr_work);CHKERRQ(ierr); 3881984c4197SStefano Zampini /* array to store scaling factors for reflectors */ 3882785e854fSJed Brown ierr = PetscMalloc1(max_constraints,&qr_tau);CHKERRQ(ierr); 3883984c4197SStefano Zampini /* array to store rhs and solution of triangular solver */ 3884785e854fSJed Brown ierr = PetscMalloc1(max_constraints*max_constraints,&trs_rhs);CHKERRQ(ierr); 3885a717540cSStefano Zampini /* allocating workspace for check */ 3886a717540cSStefano Zampini if (pcbddc->dbg_flag) { 3887cf5a6209SStefano Zampini ierr = PetscMalloc1(max_size_of_constraint*(max_constraints+max_size_of_constraint),&dbg_work);CHKERRQ(ierr); 3888a717540cSStefano Zampini } 3889a717540cSStefano Zampini } 3890984c4197SStefano Zampini /* array to store whether a node is primal or not */ 38914641a718SStefano Zampini ierr = PetscBTCreate(pcis->n_B,&is_primal);CHKERRQ(ierr); 3892473ba861SJed Brown ierr = PetscMalloc1(total_primal_vertices,&aux_primal_numbering_B);CHKERRQ(ierr); 38930e6343abSStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,total_primal_vertices,pcbddc->local_primal_ref_node,&i,aux_primal_numbering_B);CHKERRQ(ierr); 389439e2fb2aSStefano Zampini if (i != total_primal_vertices) { 389539e2fb2aSStefano Zampini SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_SUP,"Error in boundary numbering for BDDC vertices! %d != %d\n",total_primal_vertices,i); 38964641a718SStefano Zampini } 389739e2fb2aSStefano Zampini for (i=0;i<total_primal_vertices;i++) { 389839e2fb2aSStefano Zampini ierr = PetscBTSet(is_primal,aux_primal_numbering_B[i]);CHKERRQ(ierr); 389939e2fb2aSStefano Zampini } 390039e2fb2aSStefano Zampini ierr = PetscFree(aux_primal_numbering_B);CHKERRQ(ierr); 3901984c4197SStefano Zampini 3902a717540cSStefano Zampini /* loop on constraints and see whether or not they need a change of basis and compute it */ 39039162d606SStefano Zampini for (total_counts=n_vertices;total_counts<total_counts_cc;total_counts++) { 39049162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[total_counts+1]-constraints_idxs_ptr[total_counts]; 39054641a718SStefano Zampini if (PetscBTLookup(change_basis,total_counts)) { 3906984c4197SStefano Zampini /* get constraint info */ 39079162d606SStefano Zampini primal_dofs = constraints_n[total_counts]; 3908984c4197SStefano Zampini dual_dofs = size_of_constraint-primal_dofs; 3909984c4197SStefano Zampini 3910984c4197SStefano Zampini if (pcbddc->dbg_flag) { 39119162d606SStefano 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); 3912674ae819SStefano Zampini } 3913984c4197SStefano Zampini 3914fa434743SStefano Zampini if (PetscBTLookup(qr_needed_idx,total_counts)) { /* QR */ 3915a717540cSStefano Zampini 3916a717540cSStefano Zampini /* copy quadrature constraints for change of basis check */ 3917a717540cSStefano Zampini if (pcbddc->dbg_flag) { 39189162d606SStefano Zampini ierr = PetscMemcpy(dbg_work,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 3919a717540cSStefano Zampini } 3920984c4197SStefano Zampini /* copy temporary constraints into larger work vector (in order to store all columns of Q) */ 39219162d606SStefano Zampini ierr = PetscMemcpy(qr_basis,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 3922984c4197SStefano Zampini 3923984c4197SStefano Zampini /* compute QR decomposition of constraints */ 39243f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 39253f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 39263f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 3927674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 39283f08241aSStefano Zampini PetscStackCallBLAS("LAPACKgeqrf",LAPACKgeqrf_(&Blas_M,&Blas_N,qr_basis,&Blas_LDA,qr_tau,qr_work,&lqr_work,&lierr)); 3929984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GEQRF Lapack routine %d",(int)lierr); 3930674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 3931984c4197SStefano Zampini 3932984c4197SStefano Zampini /* explictly compute R^-T */ 3933984c4197SStefano Zampini ierr = PetscMemzero(trs_rhs,primal_dofs*primal_dofs*sizeof(*trs_rhs));CHKERRQ(ierr); 3934984c4197SStefano Zampini for (j=0;j<primal_dofs;j++) trs_rhs[j*(primal_dofs+1)] = 1.0; 39353f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 39363f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_NRHS);CHKERRQ(ierr); 39373f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 39383f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDB);CHKERRQ(ierr); 3939984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 39403f08241aSStefano Zampini PetscStackCallBLAS("LAPACKtrtrs",LAPACKtrtrs_("U","T","N",&Blas_N,&Blas_NRHS,qr_basis,&Blas_LDA,trs_rhs,&Blas_LDB,&lierr)); 3941984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in TRTRS Lapack routine %d",(int)lierr); 3942984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 3943984c4197SStefano Zampini 3944a717540cSStefano Zampini /* explicitly compute all columns of Q (Q = [Q1 | Q2] ) overwriting QR factorization in qr_basis */ 39453f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 39463f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 39473f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_K);CHKERRQ(ierr); 39483f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 3949984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 39503f08241aSStefano Zampini PetscStackCallBLAS("LAPACKungqr",LAPACKungqr_(&Blas_M,&Blas_N,&Blas_K,qr_basis,&Blas_LDA,qr_tau,gqr_work,&lgqr_work,&lierr)); 3951984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in UNGQR Lapack routine %d",(int)lierr); 3952984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 3953984c4197SStefano Zampini 3954984c4197SStefano Zampini /* first primal_dofs columns of Q need to be re-scaled in order to be unitary w.r.t constraints 3955984c4197SStefano Zampini i.e. C_{pxn}*Q_{nxn} should be equal to [I_pxp | 0_pxd] (see check below) 3956984c4197SStefano Zampini where n=size_of_constraint, p=primal_dofs, d=dual_dofs (n=p+d), I and 0 identity and null matrix resp. */ 39573f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 39583f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 39593f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_K);CHKERRQ(ierr); 39603f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 39613f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDB);CHKERRQ(ierr); 39623f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDC);CHKERRQ(ierr); 3963984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 39649162d606SStefano 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)); 3965984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 39669162d606SStefano Zampini ierr = PetscMemcpy(qr_basis,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 3967984c4197SStefano Zampini 3968984c4197SStefano Zampini /* insert values in change of basis matrix respecting global ordering of new primal dofs */ 39699162d606SStefano Zampini start_rows = &constraints_idxs[constraints_idxs_ptr[total_counts]]; 3970984c4197SStefano Zampini /* insert cols for primal dofs */ 3971984c4197SStefano Zampini for (j=0;j<primal_dofs;j++) { 3972984c4197SStefano Zampini start_vals = &qr_basis[j*size_of_constraint]; 39739162d606SStefano Zampini start_cols = &constraints_idxs[constraints_idxs_ptr[total_counts]+j]; 3974906d46d4SStefano Zampini ierr = MatSetValues(localChangeOfBasisMatrix,size_of_constraint,start_rows,1,start_cols,start_vals,INSERT_VALUES);CHKERRQ(ierr); 3975984c4197SStefano Zampini } 3976984c4197SStefano Zampini /* insert cols for dual dofs */ 3977984c4197SStefano Zampini for (j=0,k=0;j<dual_dofs;k++) { 39789162d606SStefano Zampini if (!PetscBTLookup(is_primal,constraints_idxs_B[constraints_idxs_ptr[total_counts]+k])) { 3979984c4197SStefano Zampini start_vals = &qr_basis[(primal_dofs+j)*size_of_constraint]; 39809162d606SStefano Zampini start_cols = &constraints_idxs[constraints_idxs_ptr[total_counts]+k]; 3981906d46d4SStefano Zampini ierr = MatSetValues(localChangeOfBasisMatrix,size_of_constraint,start_rows,1,start_cols,start_vals,INSERT_VALUES);CHKERRQ(ierr); 3982984c4197SStefano Zampini j++; 3983674ae819SStefano Zampini } 3984674ae819SStefano Zampini } 3985984c4197SStefano Zampini 3986984c4197SStefano Zampini /* check change of basis */ 3987984c4197SStefano Zampini if (pcbddc->dbg_flag) { 3988984c4197SStefano Zampini PetscInt ii,jj; 3989984c4197SStefano Zampini PetscBool valid_qr=PETSC_TRUE; 3990c4303822SStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_M);CHKERRQ(ierr); 3991c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 3992c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_K);CHKERRQ(ierr); 3993c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 3994c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDB);CHKERRQ(ierr); 3995c4303822SStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDC);CHKERRQ(ierr); 3996984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 3997cf5a6209SStefano 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)); 3998984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 3999984c4197SStefano Zampini for (jj=0;jj<size_of_constraint;jj++) { 4000984c4197SStefano Zampini for (ii=0;ii<primal_dofs;ii++) { 4001cf5a6209SStefano Zampini if (ii != jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]) > 1.e-12) valid_qr = PETSC_FALSE; 4002cf5a6209SStefano 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; 4003674ae819SStefano Zampini } 4004674ae819SStefano Zampini } 4005984c4197SStefano Zampini if (!valid_qr) { 400622d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> wrong change of basis!\n");CHKERRQ(ierr); 4007984c4197SStefano Zampini for (jj=0;jj<size_of_constraint;jj++) { 4008984c4197SStefano Zampini for (ii=0;ii<primal_dofs;ii++) { 4009cf5a6209SStefano Zampini if (ii != jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]) > 1.e-12) { 4010cf5a6209SStefano 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])); 4011674ae819SStefano Zampini } 4012cf5a6209SStefano Zampini if (ii == jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]-1.0) > 1.e-12) { 4013cf5a6209SStefano 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])); 4014984c4197SStefano Zampini } 4015984c4197SStefano Zampini } 4016984c4197SStefano Zampini } 4017674ae819SStefano Zampini } else { 401822d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> right change of basis!\n");CHKERRQ(ierr); 4019674ae819SStefano Zampini } 4020674ae819SStefano Zampini } 4021a717540cSStefano Zampini } else { /* simple transformation block */ 4022a717540cSStefano Zampini PetscInt row,col; 4023a6b551f4SStefano Zampini PetscScalar val,norm; 4024a6b551f4SStefano Zampini 4025a6b551f4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 40269162d606SStefano 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)); 4027a717540cSStefano Zampini for (j=0;j<size_of_constraint;j++) { 40289162d606SStefano Zampini PetscInt row_B = constraints_idxs_B[constraints_idxs_ptr[total_counts]+j]; 40299162d606SStefano Zampini row = constraints_idxs[constraints_idxs_ptr[total_counts]+j]; 4030bbb9e6c6SStefano Zampini if (!PetscBTLookup(is_primal,row_B)) { 40319162d606SStefano Zampini col = constraints_idxs[constraints_idxs_ptr[total_counts]]; 4032906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,row,1.0,INSERT_VALUES);CHKERRQ(ierr); 40339162d606SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,col,constraints_data[constraints_data_ptr[total_counts]+j]/norm,INSERT_VALUES);CHKERRQ(ierr); 4034a717540cSStefano Zampini } else { 4035a717540cSStefano Zampini for (k=0;k<size_of_constraint;k++) { 40369162d606SStefano Zampini col = constraints_idxs[constraints_idxs_ptr[total_counts]+k]; 4037a717540cSStefano Zampini if (row != col) { 40389162d606SStefano Zampini val = -constraints_data[constraints_data_ptr[total_counts]+k]/constraints_data[constraints_data_ptr[total_counts]]; 4039a717540cSStefano Zampini } else { 40409162d606SStefano Zampini val = constraints_data[constraints_data_ptr[total_counts]]/norm; 4041a717540cSStefano Zampini } 4042906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,col,val,INSERT_VALUES);CHKERRQ(ierr); 4043a717540cSStefano Zampini } 4044a717540cSStefano Zampini } 4045a717540cSStefano Zampini } 404698a51de6SStefano Zampini if (pcbddc->dbg_flag) { 404722d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> using standard change of basis\n");CHKERRQ(ierr); 4048a717540cSStefano Zampini } 4049674ae819SStefano Zampini } 4050984c4197SStefano Zampini } else { 4051984c4197SStefano Zampini if (pcbddc->dbg_flag) { 40529162d606SStefano 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); 4053674ae819SStefano Zampini } 4054674ae819SStefano Zampini } 4055674ae819SStefano Zampini } 4056a717540cSStefano Zampini 4057a717540cSStefano Zampini /* free workspace */ 4058a717540cSStefano Zampini if (qr_needed) { 4059984c4197SStefano Zampini if (pcbddc->dbg_flag) { 4060cf5a6209SStefano Zampini ierr = PetscFree(dbg_work);CHKERRQ(ierr); 4061984c4197SStefano Zampini } 4062984c4197SStefano Zampini ierr = PetscFree(trs_rhs);CHKERRQ(ierr); 4063984c4197SStefano Zampini ierr = PetscFree(qr_tau);CHKERRQ(ierr); 4064984c4197SStefano Zampini ierr = PetscFree(qr_work);CHKERRQ(ierr); 4065984c4197SStefano Zampini ierr = PetscFree(gqr_work);CHKERRQ(ierr); 4066984c4197SStefano Zampini ierr = PetscFree(qr_basis);CHKERRQ(ierr); 4067674ae819SStefano Zampini } 4068a717540cSStefano Zampini ierr = PetscBTDestroy(&is_primal);CHKERRQ(ierr); 4069906d46d4SStefano Zampini ierr = MatAssemblyBegin(localChangeOfBasisMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4070906d46d4SStefano Zampini ierr = MatAssemblyEnd(localChangeOfBasisMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4071906d46d4SStefano Zampini 4072906d46d4SStefano Zampini /* assembling of global change of variable */ 4073bbb9e6c6SStefano Zampini { 4074bbb9e6c6SStefano Zampini Mat tmat; 407516f15bc4SStefano Zampini PetscInt bs; 407616f15bc4SStefano Zampini 4077906d46d4SStefano Zampini ierr = VecGetSize(pcis->vec1_global,&global_size);CHKERRQ(ierr); 4078906d46d4SStefano Zampini ierr = VecGetLocalSize(pcis->vec1_global,&local_size);CHKERRQ(ierr); 4079bbb9e6c6SStefano Zampini ierr = MatDuplicate(pc->pmat,MAT_DO_NOT_COPY_VALUES,&tmat);CHKERRQ(ierr); 4080bbb9e6c6SStefano Zampini ierr = MatISSetLocalMat(tmat,localChangeOfBasisMatrix);CHKERRQ(ierr); 4081bbb9e6c6SStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)pc),&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 4082bbb9e6c6SStefano Zampini ierr = MatSetType(pcbddc->ChangeOfBasisMatrix,MATAIJ);CHKERRQ(ierr); 408316f15bc4SStefano Zampini ierr = MatGetBlockSize(pc->pmat,&bs);CHKERRQ(ierr); 408416f15bc4SStefano Zampini ierr = MatSetBlockSize(pcbddc->ChangeOfBasisMatrix,bs);CHKERRQ(ierr); 4085906d46d4SStefano Zampini ierr = MatSetSizes(pcbddc->ChangeOfBasisMatrix,local_size,local_size,global_size,global_size);CHKERRQ(ierr); 4086bbb9e6c6SStefano Zampini ierr = MatISSetMPIXAIJPreallocation_Private(tmat,pcbddc->ChangeOfBasisMatrix,PETSC_TRUE);CHKERRQ(ierr); 4087bbb9e6c6SStefano Zampini ierr = MatISGetMPIXAIJ(tmat,MAT_REUSE_MATRIX,&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 4088bbb9e6c6SStefano Zampini ierr = MatDestroy(&tmat);CHKERRQ(ierr); 4089bbb9e6c6SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 4090bbb9e6c6SStefano Zampini ierr = VecSet(pcis->vec1_N,1.0);CHKERRQ(ierr); 4091e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4092e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4093bbb9e6c6SStefano Zampini ierr = VecReciprocal(pcis->vec1_global);CHKERRQ(ierr); 4094bbb9e6c6SStefano Zampini ierr = MatDiagonalScale(pcbddc->ChangeOfBasisMatrix,pcis->vec1_global,NULL);CHKERRQ(ierr); 4095906d46d4SStefano Zampini } 4096906d46d4SStefano Zampini /* check */ 4097906d46d4SStefano Zampini if (pcbddc->dbg_flag) { 4098906d46d4SStefano Zampini PetscReal error; 4099906d46d4SStefano Zampini Vec x,x_change; 4100906d46d4SStefano Zampini 4101906d46d4SStefano Zampini ierr = VecDuplicate(pcis->vec1_global,&x);CHKERRQ(ierr); 4102906d46d4SStefano Zampini ierr = VecDuplicate(pcis->vec1_global,&x_change);CHKERRQ(ierr); 4103906d46d4SStefano Zampini ierr = VecSetRandom(x,NULL);CHKERRQ(ierr); 4104906d46d4SStefano Zampini ierr = VecCopy(x,pcis->vec1_global);CHKERRQ(ierr); 4105e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,x,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4106e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,x,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4107bbb9e6c6SStefano Zampini ierr = MatMult(localChangeOfBasisMatrix,pcis->vec1_N,pcis->vec2_N);CHKERRQ(ierr); 4108e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec2_N,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4109e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec2_N,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4110906d46d4SStefano Zampini ierr = MatMult(pcbddc->ChangeOfBasisMatrix,pcis->vec1_global,x_change);CHKERRQ(ierr); 4111906d46d4SStefano Zampini ierr = VecAXPY(x,-1.0,x_change);CHKERRQ(ierr); 4112906d46d4SStefano Zampini ierr = VecNorm(x,NORM_INFINITY,&error);CHKERRQ(ierr); 4113906d46d4SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4114bbb9e6c6SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Error global vs local change: %1.6e\n",error);CHKERRQ(ierr); 4115906d46d4SStefano Zampini ierr = VecDestroy(&x);CHKERRQ(ierr); 4116906d46d4SStefano Zampini ierr = VecDestroy(&x_change);CHKERRQ(ierr); 4117906d46d4SStefano Zampini } 4118b96c3477SStefano Zampini 4119b96c3477SStefano Zampini /* adapt sub_schurs computed (if any) */ 4120b96c3477SStefano Zampini if (pcbddc->use_deluxe_scaling) { 4121b96c3477SStefano Zampini PCBDDCSubSchurs sub_schurs=pcbddc->sub_schurs; 4122b96c3477SStefano Zampini if (sub_schurs->S_Ej_all) { 4123ac632422SStefano Zampini Mat S_new,tmat; 4124b087196eSStefano Zampini ISLocalToGlobalMapping NtoSall; 4125b087196eSStefano Zampini IS is_all_N,is_V,is_V_Sall; 4126b087196eSStefano Zampini const PetscScalar *array; 4127b087196eSStefano Zampini const PetscInt *idxs_V,*idxs_all; 4128b087196eSStefano Zampini PetscInt i,n_V; 4129bbb9e6c6SStefano Zampini 4130bbb9e6c6SStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(pcis->BtoNmap,sub_schurs->is_Ej_all,&is_all_N);CHKERRQ(ierr); 41316816873aSStefano Zampini ierr = MatGetSubMatrix(localChangeOfBasisMatrix,is_all_N,is_all_N,MAT_INITIAL_MATRIX,&tmat);CHKERRQ(ierr); 4132b087196eSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,pcbddc->n_vertices,pcbddc->local_primal_ref_node,PETSC_COPY_VALUES,&is_V);CHKERRQ(ierr); 4133b087196eSStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(is_all_N,&NtoSall);CHKERRQ(ierr); 4134b087196eSStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(NtoSall,IS_GTOLM_DROP,is_V,&is_V_Sall);CHKERRQ(ierr); 4135b087196eSStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&NtoSall);CHKERRQ(ierr); 4136bbb9e6c6SStefano Zampini ierr = ISDestroy(&is_all_N);CHKERRQ(ierr); 4137b087196eSStefano Zampini ierr = ISDestroy(&is_V);CHKERRQ(ierr); 4138ac632422SStefano Zampini ierr = MatPtAP(sub_schurs->S_Ej_all,tmat,MAT_INITIAL_MATRIX,1.0,&S_new);CHKERRQ(ierr); 4139b96c3477SStefano Zampini ierr = MatDestroy(&sub_schurs->S_Ej_all);CHKERRQ(ierr); 4140ac632422SStefano Zampini ierr = PetscObjectReference((PetscObject)S_new);CHKERRQ(ierr); 4141b087196eSStefano Zampini ierr = MatZeroRowsColumnsIS(S_new,is_V_Sall,1.,NULL,NULL);CHKERRQ(ierr); 4142b087196eSStefano Zampini ierr = ISGetLocalSize(is_V_Sall,&n_V);CHKERRQ(ierr); 4143b087196eSStefano Zampini ierr = ISGetIndices(is_V_Sall,&idxs_V);CHKERRQ(ierr); 4144b087196eSStefano Zampini ierr = ISGetIndices(sub_schurs->is_Ej_all,&idxs_all);CHKERRQ(ierr); 4145b087196eSStefano Zampini ierr = VecGetArrayRead(pcis->D,&array);CHKERRQ(ierr); 4146b087196eSStefano Zampini for (i=0;i<n_V;i++) { 4147b087196eSStefano Zampini PetscScalar val; 4148b087196eSStefano Zampini PetscInt idx; 4149b087196eSStefano Zampini 4150b087196eSStefano Zampini idx = idxs_V[i]; 4151b087196eSStefano Zampini val = array[idxs_all[idxs_V[i]]]; 4152b087196eSStefano Zampini ierr = MatSetValue(S_new,idx,idx,val,INSERT_VALUES);CHKERRQ(ierr); 4153b087196eSStefano Zampini } 4154b087196eSStefano Zampini ierr = MatAssemblyBegin(S_new,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4155b087196eSStefano Zampini ierr = MatAssemblyEnd(S_new,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4156ac632422SStefano Zampini sub_schurs->S_Ej_all = S_new; 4157ac632422SStefano Zampini ierr = MatDestroy(&S_new);CHKERRQ(ierr); 4158ac632422SStefano Zampini if (sub_schurs->sum_S_Ej_all) { 4159ac632422SStefano Zampini ierr = MatPtAP(sub_schurs->sum_S_Ej_all,tmat,MAT_INITIAL_MATRIX,1.0,&S_new);CHKERRQ(ierr); 4160b96c3477SStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_all);CHKERRQ(ierr); 4161ac632422SStefano Zampini ierr = PetscObjectReference((PetscObject)S_new);CHKERRQ(ierr); 4162b087196eSStefano Zampini ierr = MatZeroRowsColumnsIS(S_new,is_V_Sall,1.,NULL,NULL);CHKERRQ(ierr); 4163ac632422SStefano Zampini sub_schurs->sum_S_Ej_all = S_new; 4164ac632422SStefano Zampini ierr = MatDestroy(&S_new);CHKERRQ(ierr); 4165ac632422SStefano Zampini } 4166b087196eSStefano Zampini ierr = VecRestoreArrayRead(pcis->D,&array);CHKERRQ(ierr); 4167b087196eSStefano Zampini ierr = ISRestoreIndices(sub_schurs->is_Ej_all,&idxs_all);CHKERRQ(ierr); 4168b087196eSStefano Zampini ierr = ISRestoreIndices(is_V_Sall,&idxs_V);CHKERRQ(ierr); 4169b96c3477SStefano Zampini ierr = MatDestroy(&tmat);CHKERRQ(ierr); 4170b087196eSStefano Zampini ierr = ISDestroy(&is_V_Sall);CHKERRQ(ierr); 4171b96c3477SStefano Zampini } 4172b96c3477SStefano Zampini } 4173906d46d4SStefano Zampini ierr = MatDestroy(&localChangeOfBasisMatrix);CHKERRQ(ierr); 4174906d46d4SStefano Zampini } else if (pcbddc->user_ChangeOfBasisMatrix) { 4175b9b85e73SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->user_ChangeOfBasisMatrix);CHKERRQ(ierr); 4176b9b85e73SStefano Zampini pcbddc->ChangeOfBasisMatrix = pcbddc->user_ChangeOfBasisMatrix; 4177b9b85e73SStefano Zampini } 4178906d46d4SStefano Zampini 4179906d46d4SStefano Zampini /* set up change of basis context */ 4180906d46d4SStefano Zampini if (pcbddc->ChangeOfBasisMatrix) { 4181906d46d4SStefano Zampini PCBDDCChange_ctx change_ctx; 4182906d46d4SStefano Zampini 4183906d46d4SStefano Zampini if (!pcbddc->new_global_mat) { 4184906d46d4SStefano Zampini PetscInt global_size,local_size; 4185906d46d4SStefano Zampini 4186906d46d4SStefano Zampini ierr = VecGetSize(pcis->vec1_global,&global_size);CHKERRQ(ierr); 4187906d46d4SStefano Zampini ierr = VecGetLocalSize(pcis->vec1_global,&local_size);CHKERRQ(ierr); 4188906d46d4SStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)pc),&pcbddc->new_global_mat);CHKERRQ(ierr); 4189906d46d4SStefano Zampini ierr = MatSetSizes(pcbddc->new_global_mat,local_size,local_size,global_size,global_size);CHKERRQ(ierr); 4190906d46d4SStefano Zampini ierr = MatSetType(pcbddc->new_global_mat,MATSHELL);CHKERRQ(ierr); 4191906d46d4SStefano Zampini ierr = MatShellSetOperation(pcbddc->new_global_mat,MATOP_MULT,(void (*)(void))PCBDDCMatMult_Private);CHKERRQ(ierr); 4192906d46d4SStefano Zampini ierr = MatShellSetOperation(pcbddc->new_global_mat,MATOP_MULT_TRANSPOSE,(void (*)(void))PCBDDCMatMultTranspose_Private);CHKERRQ(ierr); 4193906d46d4SStefano Zampini ierr = PetscNew(&change_ctx);CHKERRQ(ierr); 4194906d46d4SStefano Zampini ierr = MatShellSetContext(pcbddc->new_global_mat,change_ctx);CHKERRQ(ierr); 4195906d46d4SStefano Zampini } else { 4196906d46d4SStefano Zampini ierr = MatShellGetContext(pcbddc->new_global_mat,&change_ctx);CHKERRQ(ierr); 4197906d46d4SStefano Zampini ierr = MatDestroy(&change_ctx->global_change);CHKERRQ(ierr); 4198906d46d4SStefano Zampini ierr = VecDestroyVecs(2,&change_ctx->work);CHKERRQ(ierr); 4199906d46d4SStefano Zampini } 4200906d46d4SStefano Zampini if (!pcbddc->user_ChangeOfBasisMatrix) { 4201906d46d4SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 4202906d46d4SStefano Zampini change_ctx->global_change = pcbddc->ChangeOfBasisMatrix; 4203906d46d4SStefano Zampini } else { 4204906d46d4SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->user_ChangeOfBasisMatrix);CHKERRQ(ierr); 4205906d46d4SStefano Zampini change_ctx->global_change = pcbddc->user_ChangeOfBasisMatrix; 4206906d46d4SStefano Zampini } 4207906d46d4SStefano Zampini ierr = VecDuplicateVecs(pcis->vec1_global,2,&change_ctx->work);CHKERRQ(ierr); 4208906d46d4SStefano Zampini ierr = MatSetUp(pcbddc->new_global_mat);CHKERRQ(ierr); 4209906d46d4SStefano Zampini ierr = MatAssemblyBegin(pcbddc->new_global_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4210906d46d4SStefano Zampini ierr = MatAssemblyEnd(pcbddc->new_global_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4211a8661815SStefano Zampini } else { 4212a8661815SStefano Zampini ierr = MatDestroy(&pcbddc->new_global_mat);CHKERRQ(ierr); 4213b9b85e73SStefano Zampini } 4214a717540cSStefano Zampini 42154f1b2e48SStefano Zampini /* add pressure dofs to set of primal nodes for numbering purposes */ 42164f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 42174f1b2e48SStefano Zampini pcbddc->local_primal_ref_node[pcbddc->local_primal_size_cc] = pcbddc->benign_p0_lidx[i]; 42184f1b2e48SStefano Zampini pcbddc->primal_indices_local_idxs[pcbddc->local_primal_size] = pcbddc->benign_p0_lidx[i]; 4219019a44ceSStefano Zampini pcbddc->local_primal_ref_mult[pcbddc->local_primal_size_cc] = 1; 4220019a44ceSStefano Zampini pcbddc->local_primal_size_cc++; 4221019a44ceSStefano Zampini pcbddc->local_primal_size++; 4222019a44ceSStefano Zampini } 4223019a44ceSStefano Zampini 4224019a44ceSStefano Zampini /* check if a new primal space has been introduced (also take into account benign trick) */ 4225727cdba6SStefano Zampini pcbddc->new_primal_space_local = PETSC_TRUE; 4226727cdba6SStefano Zampini if (olocal_primal_size == pcbddc->local_primal_size) { 42279f47a83aSStefano 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); 4228c1c8e736SStefano Zampini pcbddc->new_primal_space_local = (PetscBool)(!pcbddc->new_primal_space_local); 42290e6343abSStefano Zampini if (!pcbddc->new_primal_space_local) { 42309f47a83aSStefano 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); 4231727cdba6SStefano Zampini pcbddc->new_primal_space_local = (PetscBool)(!pcbddc->new_primal_space_local); 4232727cdba6SStefano Zampini } 42330e6343abSStefano Zampini } 42340e6343abSStefano Zampini ierr = PetscFree2(olocal_primal_ref_node,olocal_primal_ref_mult);CHKERRQ(ierr); 4235727cdba6SStefano Zampini /* new_primal_space will be used for numbering of coarse dofs, so it should be the same across all subdomains */ 4236727cdba6SStefano Zampini ierr = MPI_Allreduce(&pcbddc->new_primal_space_local,&pcbddc->new_primal_space,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 4237727cdba6SStefano Zampini 4238a717540cSStefano Zampini /* flush dbg viewer */ 4239b8ffe317SStefano Zampini if (pcbddc->dbg_flag) { 4240b8ffe317SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4241b8ffe317SStefano Zampini } 4242a717540cSStefano Zampini 4243e310c8b4SStefano Zampini /* free workspace */ 4244a717540cSStefano Zampini ierr = PetscBTDestroy(&qr_needed_idx);CHKERRQ(ierr); 42454641a718SStefano Zampini ierr = PetscBTDestroy(&change_basis);CHKERRQ(ierr); 424608122e43SStefano Zampini if (!pcbddc->adaptive_selection) { 42479162d606SStefano Zampini ierr = PetscFree3(constraints_idxs_ptr,constraints_data_ptr,constraints_n);CHKERRQ(ierr); 42489162d606SStefano Zampini ierr = PetscFree3(constraints_data,constraints_idxs,constraints_idxs_B);CHKERRQ(ierr); 424908122e43SStefano Zampini } else { 42509162d606SStefano Zampini ierr = PetscFree5(pcbddc->adaptive_constraints_n, 42519162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr, 42529162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr, 425308122e43SStefano Zampini pcbddc->adaptive_constraints_idxs, 425408122e43SStefano Zampini pcbddc->adaptive_constraints_data);CHKERRQ(ierr); 42559162d606SStefano Zampini ierr = PetscFree(constraints_n);CHKERRQ(ierr); 42569162d606SStefano Zampini ierr = PetscFree(constraints_idxs_B);CHKERRQ(ierr); 425708122e43SStefano Zampini } 4258674ae819SStefano Zampini PetscFunctionReturn(0); 4259674ae819SStefano Zampini } 4260674ae819SStefano Zampini 4261674ae819SStefano Zampini #undef __FUNCT__ 4262674ae819SStefano Zampini #define __FUNCT__ "PCBDDCAnalyzeInterface" 4263674ae819SStefano Zampini PetscErrorCode PCBDDCAnalyzeInterface(PC pc) 4264674ae819SStefano Zampini { 4265674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 4266674ae819SStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 4267674ae819SStefano Zampini Mat_IS *matis = (Mat_IS*)pc->pmat->data; 42687fb0e2dbSStefano Zampini PetscInt ierr,i,vertex_size,N; 4269674ae819SStefano Zampini PetscViewer viewer=pcbddc->dbg_viewer; 4270674ae819SStefano Zampini 4271674ae819SStefano Zampini PetscFunctionBegin; 42728e61c736SStefano Zampini /* Reset previously computed graph */ 42738e61c736SStefano Zampini ierr = PCBDDCGraphReset(pcbddc->mat_graph);CHKERRQ(ierr); 4274674ae819SStefano Zampini /* Init local Graph struct */ 42757fb0e2dbSStefano Zampini ierr = MatGetSize(pc->pmat,&N,NULL);CHKERRQ(ierr); 42763bbff08aSStefano Zampini ierr = PCBDDCGraphInit(pcbddc->mat_graph,pcis->mapping,N);CHKERRQ(ierr); 4277674ae819SStefano Zampini 4278575ad6abSStefano Zampini /* Check validity of the csr graph passed in by the user */ 42795099eff2SStefano Zampini if (pcbddc->mat_graph->nvtxs_csr && pcbddc->mat_graph->nvtxs_csr != pcbddc->mat_graph->nvtxs) { 42805099eff2SStefano 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); 4281575ad6abSStefano Zampini } 42829577ea80SStefano Zampini 4283674ae819SStefano Zampini /* Set default CSR adjacency of local dofs if not provided by the user with PCBDDCSetLocalAdjacencyGraph */ 4284d4d8cf7bSStefano Zampini if ( (!pcbddc->mat_graph->xadj || !pcbddc->mat_graph->adjncy) && pcbddc->use_local_adj) { 42854d379d7bSStefano Zampini PetscInt *xadj,*adjncy; 42864d379d7bSStefano Zampini PetscInt nvtxs; 4287e496cd5dSStefano Zampini PetscBool flg_row=PETSC_FALSE; 4288674ae819SStefano Zampini 42892fffb893SStefano Zampini ierr = MatGetRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 42902fffb893SStefano Zampini if (flg_row) { 42914d379d7bSStefano Zampini ierr = PCBDDCSetLocalAdjacencyGraph(pc,nvtxs,xadj,adjncy,PETSC_COPY_VALUES);CHKERRQ(ierr); 4292b96c3477SStefano Zampini pcbddc->computed_rowadj = PETSC_TRUE; 42932fffb893SStefano Zampini } 42942fffb893SStefano Zampini ierr = MatRestoreRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 4295674ae819SStefano Zampini } 42969b28b941SStefano Zampini if (pcbddc->dbg_flag) { 42979b28b941SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4298674ae819SStefano Zampini } 4299674ae819SStefano Zampini 430063602bcaSStefano Zampini /* Set default dofs' splitting if no information has been provided by the user with PCBDDCSetDofsSplitting or PCBDDCSetDofsSplittingLocal */ 4301674ae819SStefano Zampini vertex_size = 1; 430263602bcaSStefano Zampini if (pcbddc->user_provided_isfordofs) { 430363602bcaSStefano Zampini if (pcbddc->n_ISForDofs) { /* need to convert from global to local and remove references to global dofs splitting */ 430495ecbf38SStefano Zampini ierr = PetscMalloc1(pcbddc->n_ISForDofs,&pcbddc->ISForDofsLocal);CHKERRQ(ierr); 430563602bcaSStefano Zampini for (i=0;i<pcbddc->n_ISForDofs;i++) { 4306e176bc59SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->rctx,pcis->vec1_global,pcis->vec1_N,pcbddc->ISForDofs[i],&pcbddc->ISForDofsLocal[i]);CHKERRQ(ierr); 430763602bcaSStefano Zampini ierr = ISDestroy(&pcbddc->ISForDofs[i]);CHKERRQ(ierr); 4308674ae819SStefano Zampini } 430963602bcaSStefano Zampini pcbddc->n_ISForDofsLocal = pcbddc->n_ISForDofs; 431063602bcaSStefano Zampini pcbddc->n_ISForDofs = 0; 431163602bcaSStefano Zampini ierr = PetscFree(pcbddc->ISForDofs);CHKERRQ(ierr); 4312674ae819SStefano Zampini } 431363602bcaSStefano Zampini /* mat block size as vertex size (used for elasticity with rigid body modes as nearnullspace) */ 4314674ae819SStefano Zampini ierr = MatGetBlockSize(matis->A,&vertex_size);CHKERRQ(ierr); 431563602bcaSStefano Zampini } else { 431663602bcaSStefano Zampini if (!pcbddc->n_ISForDofsLocal) { /* field split not present, create it in local ordering */ 431763602bcaSStefano Zampini ierr = MatGetBlockSize(pc->pmat,&pcbddc->n_ISForDofsLocal);CHKERRQ(ierr); 4318854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->n_ISForDofsLocal,&pcbddc->ISForDofsLocal);CHKERRQ(ierr); 431963602bcaSStefano Zampini for (i=0;i<pcbddc->n_ISForDofsLocal;i++) { 432063602bcaSStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)pc),pcis->n/pcbddc->n_ISForDofsLocal,i,pcbddc->n_ISForDofsLocal,&pcbddc->ISForDofsLocal[i]);CHKERRQ(ierr); 432163602bcaSStefano Zampini } 432263602bcaSStefano Zampini } 4323674ae819SStefano Zampini } 4324674ae819SStefano Zampini 4325674ae819SStefano Zampini /* Setup of Graph */ 4326785d1243SStefano Zampini if (!pcbddc->DirichletBoundariesLocal && pcbddc->DirichletBoundaries) { /* need to convert from global to local */ 4327e176bc59SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->rctx,pcis->vec1_global,pcis->vec1_N,pcbddc->DirichletBoundaries,&pcbddc->DirichletBoundariesLocal);CHKERRQ(ierr); 4328785d1243SStefano Zampini } 4329785d1243SStefano Zampini if (!pcbddc->NeumannBoundariesLocal && pcbddc->NeumannBoundaries) { /* need to convert from global to local */ 4330e176bc59SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->rctx,pcis->vec1_global,pcis->vec1_N,pcbddc->NeumannBoundaries,&pcbddc->NeumannBoundariesLocal);CHKERRQ(ierr); 4331785d1243SStefano Zampini } 433230368db7SStefano Zampini if (!pcbddc->user_primal_vertices_local && pcbddc->user_primal_vertices) { /* need to convert from global to local */ 433330368db7SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->rctx,pcis->vec1_global,pcis->vec1_N,pcbddc->user_primal_vertices,&pcbddc->user_primal_vertices_local);CHKERRQ(ierr); 433430368db7SStefano Zampini } 433530368db7SStefano Zampini ierr = PCBDDCGraphSetUp(pcbddc->mat_graph,vertex_size,pcbddc->NeumannBoundariesLocal,pcbddc->DirichletBoundariesLocal,pcbddc->n_ISForDofsLocal,pcbddc->ISForDofsLocal,pcbddc->user_primal_vertices_local);CHKERRQ(ierr); 4336674ae819SStefano Zampini 43374f1b2e48SStefano Zampini /* attach info on disconnected subdomains if present */ 43384f1b2e48SStefano Zampini if (pcbddc->n_local_subs) { 43394f1b2e48SStefano Zampini PetscInt *local_subs; 43404f1b2e48SStefano Zampini 43414f1b2e48SStefano Zampini ierr = PetscMalloc1(N,&local_subs);CHKERRQ(ierr); 43424f1b2e48SStefano Zampini for (i=0;i<pcbddc->n_local_subs;i++) { 43434f1b2e48SStefano Zampini const PetscInt *idxs; 43444f1b2e48SStefano Zampini PetscInt nl,j; 43454f1b2e48SStefano Zampini 43464f1b2e48SStefano Zampini ierr = ISGetLocalSize(pcbddc->local_subs[i],&nl);CHKERRQ(ierr); 43474f1b2e48SStefano Zampini ierr = ISGetIndices(pcbddc->local_subs[i],&idxs);CHKERRQ(ierr); 43484f1b2e48SStefano Zampini for (j=0;j<nl;j++) { 43494f1b2e48SStefano Zampini local_subs[idxs[j]] = i; 43504f1b2e48SStefano Zampini } 43514f1b2e48SStefano Zampini ierr = ISRestoreIndices(pcbddc->local_subs[i],&idxs);CHKERRQ(ierr); 43524f1b2e48SStefano Zampini } 43534f1b2e48SStefano Zampini pcbddc->mat_graph->n_local_subs = pcbddc->n_local_subs; 43544f1b2e48SStefano Zampini pcbddc->mat_graph->local_subs = local_subs; 43554f1b2e48SStefano Zampini } 43564f1b2e48SStefano Zampini 4357674ae819SStefano Zampini /* Graph's connected components analysis */ 4358674ae819SStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(pcbddc->mat_graph);CHKERRQ(ierr); 4359674ae819SStefano Zampini 4360674ae819SStefano Zampini /* print some info to stdout */ 4361674ae819SStefano Zampini if (pcbddc->dbg_flag) { 4362302440fdSBarry Smith ierr = PCBDDCGraphASCIIView(pcbddc->mat_graph,pcbddc->dbg_flag,viewer);CHKERRQ(ierr); 4363674ae819SStefano Zampini } 4364fb180af4SStefano Zampini 4365fb180af4SStefano Zampini /* mark topography has done */ 4366fb180af4SStefano Zampini pcbddc->recompute_topography = PETSC_FALSE; 4367674ae819SStefano Zampini PetscFunctionReturn(0); 4368674ae819SStefano Zampini } 4369674ae819SStefano Zampini 4370dc456d91SStefano Zampini /* given an index sets possibly with holes, renumbers the indexes removing the holes */ 4371674ae819SStefano Zampini #undef __FUNCT__ 4372674ae819SStefano Zampini #define __FUNCT__ "PCBDDCSubsetNumbering" 4373dc456d91SStefano Zampini PetscErrorCode PCBDDCSubsetNumbering(IS subset, IS subset_mult, PetscInt *N_n, IS *subset_n) 4374674ae819SStefano Zampini { 4375dc456d91SStefano Zampini PetscSF sf; 4376dc456d91SStefano Zampini PetscLayout map; 4377dc456d91SStefano Zampini const PetscInt *idxs; 4378dc456d91SStefano Zampini PetscInt *leaf_data,*root_data,*gidxs; 4379dc456d91SStefano Zampini PetscInt N,n,i,lbounds[2],gbounds[2],Nl; 4380dc456d91SStefano Zampini PetscInt n_n,nlocals,start,first_index; 4381dc456d91SStefano Zampini PetscMPIInt commsize; 4382674ae819SStefano Zampini PetscBool first_found; 4383674ae819SStefano Zampini PetscErrorCode ierr; 4384674ae819SStefano Zampini 4385674ae819SStefano Zampini PetscFunctionBegin; 4386dc456d91SStefano Zampini ierr = ISGetLocalSize(subset,&n);CHKERRQ(ierr); 4387dc456d91SStefano Zampini if (subset_mult) { 4388dc456d91SStefano Zampini PetscCheckSameComm(subset,1,subset_mult,2); 4389dc456d91SStefano Zampini ierr = ISGetLocalSize(subset,&i);CHKERRQ(ierr); 4390dc456d91SStefano Zampini if (i != n) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Local subset and multiplicity sizes don't match! %d != %d",n,i); 4391674ae819SStefano Zampini } 4392dc456d91SStefano Zampini /* create workspace layout for computing global indices of subset */ 4393dc456d91SStefano Zampini ierr = ISGetIndices(subset,&idxs);CHKERRQ(ierr); 4394dc456d91SStefano Zampini lbounds[0] = lbounds[1] = 0; 4395dc456d91SStefano Zampini for (i=0;i<n;i++) { 4396dc456d91SStefano Zampini if (idxs[i] < lbounds[0]) lbounds[0] = idxs[i]; 4397dc456d91SStefano Zampini else if (idxs[i] > lbounds[1]) lbounds[1] = idxs[i]; 4398674ae819SStefano Zampini } 4399dc456d91SStefano Zampini lbounds[0] = -lbounds[0]; 4400dc456d91SStefano Zampini ierr = MPI_Allreduce(lbounds,gbounds,2,MPIU_INT,MPI_MAX,PetscObjectComm((PetscObject)subset));CHKERRQ(ierr); 4401dc456d91SStefano Zampini gbounds[0] = -gbounds[0]; 4402dc456d91SStefano Zampini N = gbounds[1] - gbounds[0] + 1; 4403dc456d91SStefano Zampini ierr = PetscLayoutCreate(PetscObjectComm((PetscObject)subset),&map);CHKERRQ(ierr); 4404dc456d91SStefano Zampini ierr = PetscLayoutSetBlockSize(map,1);CHKERRQ(ierr); 4405dc456d91SStefano Zampini ierr = PetscLayoutSetSize(map,N);CHKERRQ(ierr); 4406dc456d91SStefano Zampini ierr = PetscLayoutSetUp(map);CHKERRQ(ierr); 4407dc456d91SStefano Zampini ierr = PetscLayoutGetLocalSize(map,&Nl);CHKERRQ(ierr); 4408dc456d91SStefano Zampini 4409dc456d91SStefano Zampini /* create sf : leaf_data == multiplicity of indexes, root data == global index in layout */ 4410dc456d91SStefano Zampini ierr = PetscMalloc2(n,&leaf_data,Nl,&root_data);CHKERRQ(ierr); 4411dc456d91SStefano Zampini if (subset_mult) { 4412dc456d91SStefano Zampini const PetscInt* idxs_mult; 4413dc456d91SStefano Zampini 4414dc456d91SStefano Zampini ierr = ISGetIndices(subset_mult,&idxs_mult);CHKERRQ(ierr); 4415dc456d91SStefano Zampini ierr = PetscMemcpy(leaf_data,idxs_mult,n*sizeof(PetscInt));CHKERRQ(ierr); 4416dc456d91SStefano Zampini ierr = ISRestoreIndices(subset_mult,&idxs_mult);CHKERRQ(ierr); 4417674ae819SStefano Zampini } else { 4418dc456d91SStefano Zampini for (i=0;i<n;i++) leaf_data[i] = 1; 4419674ae819SStefano Zampini } 4420dc456d91SStefano Zampini /* local size of new subset */ 4421dc456d91SStefano Zampini n_n = 0; 4422dc456d91SStefano Zampini for (i=0;i<n;i++) n_n += leaf_data[i]; 4423dc456d91SStefano Zampini 4424dc456d91SStefano Zampini /* global indexes in layout */ 4425dc456d91SStefano Zampini ierr = PetscMalloc1(n_n,&gidxs);CHKERRQ(ierr); /* allocating possibly extra space in gidxs which will be used later */ 4426dc456d91SStefano Zampini for (i=0;i<n;i++) gidxs[i] = idxs[i] - gbounds[0]; 4427dc456d91SStefano Zampini ierr = ISRestoreIndices(subset,&idxs);CHKERRQ(ierr); 4428dc456d91SStefano Zampini ierr = PetscSFCreate(PetscObjectComm((PetscObject)subset),&sf);CHKERRQ(ierr); 4429dc456d91SStefano Zampini ierr = PetscSFSetGraphLayout(sf,map,n,NULL,PETSC_COPY_VALUES,gidxs);CHKERRQ(ierr); 4430dc456d91SStefano Zampini ierr = PetscLayoutDestroy(&map);CHKERRQ(ierr); 4431dc456d91SStefano Zampini 4432dc456d91SStefano Zampini /* reduce from leaves to roots */ 4433dc456d91SStefano Zampini ierr = PetscMemzero(root_data,Nl*sizeof(PetscInt));CHKERRQ(ierr); 443464a8e5bcSStefano Zampini ierr = PetscSFReduceBegin(sf,MPIU_INT,leaf_data,root_data,MPI_MAX);CHKERRQ(ierr); 443564a8e5bcSStefano Zampini ierr = PetscSFReduceEnd(sf,MPIU_INT,leaf_data,root_data,MPI_MAX);CHKERRQ(ierr); 4436dc456d91SStefano Zampini 4437dc456d91SStefano Zampini /* count indexes in local part of layout */ 4438674ae819SStefano Zampini nlocals = 0; 4439674ae819SStefano Zampini first_index = -1; 4440674ae819SStefano Zampini first_found = PETSC_FALSE; 4441dc456d91SStefano Zampini for (i=0;i<Nl;i++) { 4442dc456d91SStefano Zampini if (!first_found && root_data[i]) { 4443674ae819SStefano Zampini first_found = PETSC_TRUE; 4444674ae819SStefano Zampini first_index = i; 4445674ae819SStefano Zampini } 4446dc456d91SStefano Zampini nlocals += root_data[i]; 4447674ae819SStefano Zampini } 4448dc456d91SStefano Zampini 4449dc456d91SStefano Zampini /* cumulative of number of indexes and size of subset without holes */ 44505fa240b1SStefano Zampini #if defined(PETSC_HAVE_MPI_EXSCAN) 4451dc456d91SStefano Zampini start = 0; 445264a8e5bcSStefano Zampini ierr = MPI_Exscan(&nlocals,&start,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)subset));CHKERRQ(ierr); 44535fa240b1SStefano Zampini #else 445464a8e5bcSStefano Zampini ierr = MPI_Scan(&nlocals,&start,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)subset));CHKERRQ(ierr); 44555fa240b1SStefano Zampini start = start-nlocals; 44565fa240b1SStefano Zampini #endif 44575fa240b1SStefano Zampini 4458dc456d91SStefano Zampini if (N_n) { /* compute total size of new subset if requested */ 4459dc456d91SStefano Zampini *N_n = start + nlocals; 4460dc456d91SStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)subset),&commsize);CHKERRQ(ierr); 4461dc456d91SStefano Zampini ierr = MPI_Bcast(N_n,1,MPIU_INT,commsize-1,PetscObjectComm((PetscObject)subset));CHKERRQ(ierr); 4462674ae819SStefano Zampini } 44635fa240b1SStefano Zampini 44645fa240b1SStefano Zampini /* adapt root data with cumulative */ 4465674ae819SStefano Zampini if (first_found) { 4466dc456d91SStefano Zampini PetscInt old_index; 4467dc456d91SStefano Zampini 4468dc456d91SStefano Zampini root_data[first_index] += start; 4469674ae819SStefano Zampini old_index = first_index; 4470dc456d91SStefano Zampini for (i=first_index+1;i<Nl;i++) { 4471dc456d91SStefano Zampini if (root_data[i]) { 4472dc456d91SStefano Zampini root_data[i] += root_data[old_index]; 4473674ae819SStefano Zampini old_index = i; 4474674ae819SStefano Zampini } 4475674ae819SStefano Zampini } 4476674ae819SStefano Zampini } 4477dc456d91SStefano Zampini 4478dc456d91SStefano Zampini /* from roots to leaves */ 4479dc456d91SStefano Zampini ierr = PetscSFBcastBegin(sf,MPIU_INT,root_data,leaf_data);CHKERRQ(ierr); 4480dc456d91SStefano Zampini ierr = PetscSFBcastEnd(sf,MPIU_INT,root_data,leaf_data);CHKERRQ(ierr); 4481dc456d91SStefano Zampini ierr = PetscSFDestroy(&sf);CHKERRQ(ierr); 4482dc456d91SStefano Zampini 4483dc456d91SStefano Zampini /* create new IS with global indexes without holes */ 4484dc456d91SStefano Zampini if (subset_mult) { 4485dc456d91SStefano Zampini const PetscInt* idxs_mult; 4486dc456d91SStefano Zampini PetscInt cum; 4487dc456d91SStefano Zampini 4488dc456d91SStefano Zampini cum = 0; 4489dc456d91SStefano Zampini ierr = ISGetIndices(subset_mult,&idxs_mult);CHKERRQ(ierr); 4490dc456d91SStefano Zampini for (i=0;i<n;i++) { 4491dc456d91SStefano Zampini PetscInt j; 4492dc456d91SStefano Zampini for (j=0;j<idxs_mult[i];j++) gidxs[cum++] = leaf_data[i] - idxs_mult[i] + j; 4493674ae819SStefano Zampini } 4494dc456d91SStefano Zampini ierr = ISRestoreIndices(subset_mult,&idxs_mult);CHKERRQ(ierr); 4495674ae819SStefano Zampini } else { 4496dc456d91SStefano Zampini for (i=0;i<n;i++) { 4497dc456d91SStefano Zampini gidxs[i] = leaf_data[i]-1; 4498674ae819SStefano Zampini } 4499674ae819SStefano Zampini } 4500dc456d91SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)subset),n_n,gidxs,PETSC_OWN_POINTER,subset_n);CHKERRQ(ierr); 4501dc456d91SStefano Zampini ierr = PetscFree2(leaf_data,root_data);CHKERRQ(ierr); 4502674ae819SStefano Zampini PetscFunctionReturn(0); 4503674ae819SStefano Zampini } 45049a7d3425SStefano Zampini 45059a7d3425SStefano Zampini #undef __FUNCT__ 45069a7d3425SStefano Zampini #define __FUNCT__ "PCBDDCOrthonormalizeVecs" 45079a7d3425SStefano Zampini PetscErrorCode PCBDDCOrthonormalizeVecs(PetscInt n, Vec vecs[]) 45089a7d3425SStefano Zampini { 45099a7d3425SStefano Zampini PetscInt i,j; 45109a7d3425SStefano Zampini PetscScalar *alphas; 45119a7d3425SStefano Zampini PetscErrorCode ierr; 45129a7d3425SStefano Zampini 45139a7d3425SStefano Zampini PetscFunctionBegin; 45149a7d3425SStefano Zampini /* this implements stabilized Gram-Schmidt */ 4515785e854fSJed Brown ierr = PetscMalloc1(n,&alphas);CHKERRQ(ierr); 45169a7d3425SStefano Zampini for (i=0;i<n;i++) { 45179a7d3425SStefano Zampini ierr = VecNormalize(vecs[i],NULL);CHKERRQ(ierr); 45189a7d3425SStefano Zampini if (i<n) { ierr = VecMDot(vecs[i],n-i-1,&vecs[i+1],&alphas[i+1]);CHKERRQ(ierr); } 45199a7d3425SStefano Zampini for (j=i+1;j<n;j++) { ierr = VecAXPY(vecs[j],PetscConj(-alphas[j]),vecs[i]);CHKERRQ(ierr); } 45209a7d3425SStefano Zampini } 45219a7d3425SStefano Zampini ierr = PetscFree(alphas);CHKERRQ(ierr); 45229a7d3425SStefano Zampini PetscFunctionReturn(0); 45239a7d3425SStefano Zampini } 45249a7d3425SStefano Zampini 4525e7931f94SStefano Zampini #undef __FUNCT__ 452670cf5478SStefano Zampini #define __FUNCT__ "MatISGetSubassemblingPattern" 4527b0c7d250SStefano Zampini PetscErrorCode MatISGetSubassemblingPattern(Mat mat, PetscInt n_subdomains, PetscInt redprocs, IS* is_sends) 4528e7931f94SStefano Zampini { 452952e5ac9dSStefano Zampini IS ranks_send_to; 4530e7931f94SStefano Zampini PetscInt n_neighs,*neighs,*n_shared,**shared; 4531e7931f94SStefano Zampini PetscMPIInt size,rank,color; 453252e5ac9dSStefano Zampini PetscInt *xadj,*adjncy; 453352e5ac9dSStefano Zampini PetscInt *adjncy_wgt,*v_wgt,*ranks_send_to_idx; 45343837a79fSStefano Zampini PetscInt i,local_size,threshold=0; 45352b510759SStefano Zampini PetscBool use_vwgt=PETSC_FALSE,use_square=PETSC_FALSE; 4536e7931f94SStefano Zampini PetscSubcomm subcomm; 453752e5ac9dSStefano Zampini PetscErrorCode ierr; 4538a57a6d2fSStefano Zampini 4539e7931f94SStefano Zampini PetscFunctionBegin; 45402b510759SStefano Zampini ierr = PetscOptionsGetBool(NULL,"-matis_partitioning_use_square",&use_square,NULL);CHKERRQ(ierr); 45412b510759SStefano Zampini ierr = PetscOptionsGetBool(NULL,"-matis_partitioning_use_vwgt",&use_vwgt,NULL);CHKERRQ(ierr); 45422b510759SStefano Zampini ierr = PetscOptionsGetInt(NULL,"-matis_partitioning_threshold",&threshold,NULL);CHKERRQ(ierr); 4543e7931f94SStefano Zampini 4544e7931f94SStefano Zampini /* Get info on mapping */ 45453bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetSize(mat->rmap->mapping,&local_size);CHKERRQ(ierr); 45463bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetInfo(mat->rmap->mapping,&n_neighs,&neighs,&n_shared,&shared);CHKERRQ(ierr); 4547e7931f94SStefano Zampini 4548e7931f94SStefano Zampini /* build local CSR graph of subdomains' connectivity */ 4549785e854fSJed Brown ierr = PetscMalloc1(2,&xadj);CHKERRQ(ierr); 4550e7931f94SStefano Zampini xadj[0] = 0; 4551e7931f94SStefano Zampini xadj[1] = PetscMax(n_neighs-1,0); 4552785e854fSJed Brown ierr = PetscMalloc1(xadj[1],&adjncy);CHKERRQ(ierr); 4553785e854fSJed Brown ierr = PetscMalloc1(xadj[1],&adjncy_wgt);CHKERRQ(ierr); 4554e7931f94SStefano Zampini 45552b510759SStefano Zampini if (threshold) { 4556d023bfaeSStefano Zampini PetscInt xadj_count = 0; 45572b510759SStefano Zampini for (i=1;i<n_neighs;i++) { 4558d023bfaeSStefano Zampini if (n_shared[i] > threshold) { 4559d023bfaeSStefano Zampini adjncy[xadj_count] = neighs[i]; 4560d023bfaeSStefano Zampini adjncy_wgt[xadj_count] = n_shared[i]; 4561d023bfaeSStefano Zampini xadj_count++; 4562e7931f94SStefano Zampini } 4563e7931f94SStefano Zampini } 4564d023bfaeSStefano Zampini xadj[1] = xadj_count; 4565c8587f34SStefano Zampini } else { 4566e7931f94SStefano Zampini if (xadj[1]) { 4567e7931f94SStefano Zampini ierr = PetscMemcpy(adjncy,&neighs[1],xadj[1]*sizeof(*adjncy));CHKERRQ(ierr); 4568e7931f94SStefano Zampini ierr = PetscMemcpy(adjncy_wgt,&n_shared[1],xadj[1]*sizeof(*adjncy_wgt));CHKERRQ(ierr); 4569c8587f34SStefano Zampini } 4570e7931f94SStefano Zampini } 45713bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingRestoreInfo(mat->rmap->mapping,&n_neighs,&neighs,&n_shared,&shared);CHKERRQ(ierr); 4572e7931f94SStefano Zampini if (use_square) { 4573e7931f94SStefano Zampini for (i=0;i<xadj[1];i++) { 4574e7931f94SStefano Zampini adjncy_wgt[i] = adjncy_wgt[i]*adjncy_wgt[i]; 4575e7931f94SStefano Zampini } 4576e7931f94SStefano Zampini } 4577e7931f94SStefano Zampini ierr = PetscSortIntWithArray(xadj[1],adjncy,adjncy_wgt);CHKERRQ(ierr); 4578e7931f94SStefano Zampini 45793837a79fSStefano Zampini ierr = PetscMalloc1(1,&ranks_send_to_idx);CHKERRQ(ierr); 4580e7931f94SStefano Zampini 4581e7931f94SStefano Zampini /* 4582e7931f94SStefano Zampini Restrict work on active processes only. 4583e7931f94SStefano Zampini */ 4584e7931f94SStefano Zampini ierr = PetscSubcommCreate(PetscObjectComm((PetscObject)mat),&subcomm);CHKERRQ(ierr); 4585e7931f94SStefano Zampini ierr = PetscSubcommSetNumber(subcomm,2);CHKERRQ(ierr); /* 2 groups, active process and not active processes */ 4586e7931f94SStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)mat),&rank);CHKERRQ(ierr); 45872b510759SStefano Zampini ierr = PetscMPIIntCast(!local_size,&color);CHKERRQ(ierr); 4588d3531aaaSJed Brown ierr = PetscSubcommSetTypeGeneral(subcomm,color,rank);CHKERRQ(ierr); 4589e7931f94SStefano Zampini if (color) { 4590e7931f94SStefano Zampini ierr = PetscFree(xadj);CHKERRQ(ierr); 4591e7931f94SStefano Zampini ierr = PetscFree(adjncy);CHKERRQ(ierr); 4592e7931f94SStefano Zampini ierr = PetscFree(adjncy_wgt);CHKERRQ(ierr); 4593c8587f34SStefano Zampini } else { 459452e5ac9dSStefano Zampini Mat subdomain_adj; 459552e5ac9dSStefano Zampini IS new_ranks,new_ranks_contig; 459652e5ac9dSStefano Zampini MatPartitioning partitioner; 459752e5ac9dSStefano Zampini PetscInt prank,rstart=0,rend=0; 459852e5ac9dSStefano Zampini PetscInt *is_indices,*oldranks; 4599b0c7d250SStefano Zampini PetscBool aggregate; 4600b0c7d250SStefano Zampini 4601306c2d5bSBarry Smith ierr = MPI_Comm_size(PetscSubcommChild(subcomm),&size);CHKERRQ(ierr); 4602785e854fSJed Brown ierr = PetscMalloc1(size,&oldranks);CHKERRQ(ierr); 4603e7931f94SStefano Zampini prank = rank; 4604306c2d5bSBarry Smith ierr = MPI_Allgather(&prank,1,MPIU_INT,oldranks,1,MPIU_INT,PetscSubcommChild(subcomm));CHKERRQ(ierr); 46058002ef2cSStefano Zampini /* 4606e7931f94SStefano Zampini for (i=0;i<size;i++) { 4607e7931f94SStefano Zampini PetscPrintf(subcomm->comm,"oldranks[%d] = %d\n",i,oldranks[i]); 4608c8587f34SStefano Zampini } 46098002ef2cSStefano Zampini */ 4610e7931f94SStefano Zampini for (i=0;i<xadj[1];i++) { 4611e7931f94SStefano Zampini ierr = PetscFindInt(adjncy[i],size,oldranks,&adjncy[i]);CHKERRQ(ierr); 4612c8587f34SStefano Zampini } 4613e7931f94SStefano Zampini ierr = PetscSortIntWithArray(xadj[1],adjncy,adjncy_wgt);CHKERRQ(ierr); 4614b0c7d250SStefano Zampini aggregate = ((redprocs > 0 && redprocs < size) ? PETSC_TRUE : PETSC_FALSE); 4615b0c7d250SStefano Zampini if (aggregate) { 4616b0c7d250SStefano Zampini PetscInt lrows,row,ncols,*cols; 4617b0c7d250SStefano Zampini PetscMPIInt nrank; 4618b0c7d250SStefano Zampini PetscScalar *vals; 4619b0c7d250SStefano Zampini 4620b0c7d250SStefano Zampini ierr = MPI_Comm_rank(PetscSubcommChild(subcomm),&nrank);CHKERRQ(ierr); 4621b0c7d250SStefano Zampini lrows = 0; 4622b0c7d250SStefano Zampini if (nrank<redprocs) { 4623b0c7d250SStefano Zampini lrows = size/redprocs; 4624b0c7d250SStefano Zampini if (nrank<size%redprocs) lrows++; 4625b0c7d250SStefano Zampini } 46265fa240b1SStefano Zampini ierr = MatCreateAIJ(PetscSubcommChild(subcomm),lrows,lrows,size,size,50,NULL,50,NULL,&subdomain_adj);CHKERRQ(ierr); 4627b0c7d250SStefano Zampini ierr = MatGetOwnershipRange(subdomain_adj,&rstart,&rend);CHKERRQ(ierr); 4628b0c7d250SStefano Zampini ierr = MatSetOption(subdomain_adj,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_FALSE);CHKERRQ(ierr); 4629b0c7d250SStefano Zampini ierr = MatSetOption(subdomain_adj,MAT_NEW_NONZERO_ALLOCATION_ERR,PETSC_FALSE);CHKERRQ(ierr); 4630b0c7d250SStefano Zampini row = nrank; 4631b0c7d250SStefano Zampini ncols = xadj[1]-xadj[0]; 4632b0c7d250SStefano Zampini cols = adjncy; 4633b0c7d250SStefano Zampini ierr = PetscMalloc1(ncols,&vals);CHKERRQ(ierr); 4634b0c7d250SStefano Zampini for (i=0;i<ncols;i++) vals[i] = adjncy_wgt[i]; 4635b0c7d250SStefano Zampini ierr = MatSetValues(subdomain_adj,1,&row,ncols,cols,vals,INSERT_VALUES);CHKERRQ(ierr); 4636b0c7d250SStefano Zampini ierr = MatAssemblyBegin(subdomain_adj,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4637b0c7d250SStefano Zampini ierr = MatAssemblyEnd(subdomain_adj,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 463852e5ac9dSStefano Zampini ierr = PetscFree(xadj);CHKERRQ(ierr); 463952e5ac9dSStefano Zampini ierr = PetscFree(adjncy);CHKERRQ(ierr); 464052e5ac9dSStefano Zampini ierr = PetscFree(adjncy_wgt);CHKERRQ(ierr); 4641b0c7d250SStefano Zampini ierr = PetscFree(vals);CHKERRQ(ierr); 4642b0c7d250SStefano Zampini } else { 4643306c2d5bSBarry Smith ierr = MatCreateMPIAdj(PetscSubcommChild(subcomm),1,(PetscInt)size,xadj,adjncy,adjncy_wgt,&subdomain_adj);CHKERRQ(ierr); 4644b0c7d250SStefano Zampini } 464522b6e8a2SStefano Zampini /* ierr = MatView(subdomain_adj,0);CHKERRQ(ierr); */ 4646e7931f94SStefano Zampini 4647e7931f94SStefano Zampini /* Partition */ 4648306c2d5bSBarry Smith ierr = MatPartitioningCreate(PetscSubcommChild(subcomm),&partitioner);CHKERRQ(ierr); 4649e7931f94SStefano Zampini ierr = MatPartitioningSetAdjacency(partitioner,subdomain_adj);CHKERRQ(ierr); 4650e7931f94SStefano Zampini if (use_vwgt) { 46513837a79fSStefano Zampini ierr = PetscMalloc1(1,&v_wgt);CHKERRQ(ierr); 4652e7931f94SStefano Zampini v_wgt[0] = local_size; 4653e7931f94SStefano Zampini ierr = MatPartitioningSetVertexWeights(partitioner,v_wgt);CHKERRQ(ierr); 4654c8587f34SStefano Zampini } 465528143c3dSStefano Zampini n_subdomains = PetscMin((PetscInt)size,n_subdomains); 465628143c3dSStefano Zampini ierr = MatPartitioningSetNParts(partitioner,n_subdomains);CHKERRQ(ierr); 4657e7931f94SStefano Zampini ierr = MatPartitioningSetFromOptions(partitioner);CHKERRQ(ierr); 4658e7931f94SStefano Zampini ierr = MatPartitioningApply(partitioner,&new_ranks);CHKERRQ(ierr); 465922b6e8a2SStefano Zampini /* ierr = MatPartitioningView(partitioner,0);CHKERRQ(ierr); */ 4660e7931f94SStefano Zampini 466152e5ac9dSStefano Zampini /* renumber new_ranks to avoid "holes" in new set of processors */ 466252e5ac9dSStefano Zampini ierr = PCBDDCSubsetNumbering(new_ranks,NULL,NULL,&new_ranks_contig);CHKERRQ(ierr); 466352e5ac9dSStefano Zampini ierr = ISDestroy(&new_ranks);CHKERRQ(ierr); 466452e5ac9dSStefano Zampini ierr = ISGetIndices(new_ranks_contig,(const PetscInt**)&is_indices);CHKERRQ(ierr); 4665b0c7d250SStefano Zampini if (!redprocs) { 4666b0c7d250SStefano Zampini ranks_send_to_idx[0] = oldranks[is_indices[0]]; 466728143c3dSStefano Zampini } else { 4668b0c7d250SStefano Zampini PetscInt idxs[1]; 4669b0c7d250SStefano Zampini PetscMPIInt tag; 4670b0c7d250SStefano Zampini MPI_Request *reqs; 4671b0c7d250SStefano Zampini 4672b0c7d250SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)subdomain_adj,&tag);CHKERRQ(ierr); 4673b0c7d250SStefano Zampini ierr = PetscMalloc1(rend-rstart,&reqs);CHKERRQ(ierr); 4674b0c7d250SStefano Zampini for (i=rstart;i<rend;i++) { 4675b0c7d250SStefano Zampini ierr = MPI_Isend(is_indices+i-rstart,1,MPIU_INT,i,tag,PetscSubcommChild(subcomm),&reqs[i-rstart]);CHKERRQ(ierr); 467628143c3dSStefano Zampini } 4677b0c7d250SStefano Zampini ierr = MPI_Recv(idxs,1,MPIU_INT,MPI_ANY_SOURCE,tag,PetscSubcommChild(subcomm),MPI_STATUS_IGNORE);CHKERRQ(ierr); 4678b0c7d250SStefano Zampini ierr = MPI_Waitall(rend-rstart,reqs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 4679b0c7d250SStefano Zampini ierr = PetscFree(reqs);CHKERRQ(ierr); 4680b0c7d250SStefano Zampini ranks_send_to_idx[0] = oldranks[idxs[0]]; 4681e7931f94SStefano Zampini } 468252e5ac9dSStefano Zampini ierr = ISRestoreIndices(new_ranks_contig,(const PetscInt**)&is_indices);CHKERRQ(ierr); 4683e7931f94SStefano Zampini /* clean up */ 4684e7931f94SStefano Zampini ierr = PetscFree(oldranks);CHKERRQ(ierr); 468552e5ac9dSStefano Zampini ierr = ISDestroy(&new_ranks_contig);CHKERRQ(ierr); 4686e7931f94SStefano Zampini ierr = MatDestroy(&subdomain_adj);CHKERRQ(ierr); 4687e7931f94SStefano Zampini ierr = MatPartitioningDestroy(&partitioner);CHKERRQ(ierr); 4688e7931f94SStefano Zampini } 4689e7931f94SStefano Zampini ierr = PetscSubcommDestroy(&subcomm);CHKERRQ(ierr); 4690e7931f94SStefano Zampini 4691e7931f94SStefano Zampini /* assemble parallel IS for sends */ 4692e7931f94SStefano Zampini i = 1; 4693e7931f94SStefano Zampini if (color) i=0; 4694e7931f94SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)mat),i,ranks_send_to_idx,PETSC_OWN_POINTER,&ranks_send_to);CHKERRQ(ierr); 4695e7931f94SStefano Zampini /* get back IS */ 4696e7931f94SStefano Zampini *is_sends = ranks_send_to; 4697e7931f94SStefano Zampini PetscFunctionReturn(0); 4698e7931f94SStefano Zampini } 4699e7931f94SStefano Zampini 4700e7931f94SStefano Zampini typedef enum {MATDENSE_PRIVATE=0,MATAIJ_PRIVATE,MATBAIJ_PRIVATE,MATSBAIJ_PRIVATE}MatTypePrivate; 4701e7931f94SStefano Zampini 4702e7931f94SStefano Zampini #undef __FUNCT__ 4703e7931f94SStefano Zampini #define __FUNCT__ "MatISSubassemble" 470453a05cb3SStefano 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[]) 4705e7931f94SStefano Zampini { 470670cf5478SStefano Zampini Mat local_mat; 4707e7931f94SStefano Zampini IS is_sends_internal; 47089d30be91SStefano Zampini PetscInt rows,cols,new_local_rows; 470928143c3dSStefano Zampini PetscInt i,bs,buf_size_idxs,buf_size_idxs_is,buf_size_vals; 47109d30be91SStefano Zampini PetscBool ismatis,isdense,newisdense,destroy_mat; 4711e7931f94SStefano Zampini ISLocalToGlobalMapping l2gmap; 4712e7931f94SStefano Zampini PetscInt* l2gmap_indices; 4713e7931f94SStefano Zampini const PetscInt* is_indices; 4714e7931f94SStefano Zampini MatType new_local_type; 4715e7931f94SStefano Zampini /* buffers */ 4716e7931f94SStefano Zampini PetscInt *ptr_idxs,*send_buffer_idxs,*recv_buffer_idxs; 471728143c3dSStefano Zampini PetscInt *ptr_idxs_is,*send_buffer_idxs_is,*recv_buffer_idxs_is; 47189d30be91SStefano Zampini PetscInt *recv_buffer_idxs_local; 4719e7931f94SStefano Zampini PetscScalar *ptr_vals,*send_buffer_vals,*recv_buffer_vals; 4720e7931f94SStefano Zampini /* MPI */ 472128143c3dSStefano Zampini MPI_Comm comm,comm_n; 472228143c3dSStefano Zampini PetscSubcomm subcomm; 4723e7931f94SStefano Zampini PetscMPIInt n_sends,n_recvs,commsize; 472428143c3dSStefano Zampini PetscMPIInt *iflags,*ilengths_idxs,*ilengths_vals,*ilengths_idxs_is; 472528143c3dSStefano Zampini PetscMPIInt *onodes,*onodes_is,*olengths_idxs,*olengths_idxs_is,*olengths_vals; 472628143c3dSStefano Zampini PetscMPIInt len,tag_idxs,tag_idxs_is,tag_vals,source_dest; 472728143c3dSStefano Zampini MPI_Request *send_req_idxs,*send_req_idxs_is,*send_req_vals; 472828143c3dSStefano Zampini MPI_Request *recv_req_idxs,*recv_req_idxs_is,*recv_req_vals; 4729e7931f94SStefano Zampini PetscErrorCode ierr; 4730e7931f94SStefano Zampini 4731e7931f94SStefano Zampini PetscFunctionBegin; 473228143c3dSStefano Zampini /* TODO: add missing checks */ 473328143c3dSStefano Zampini PetscValidLogicalCollectiveInt(mat,n_subdomains,3); 473428143c3dSStefano Zampini PetscValidLogicalCollectiveBool(mat,restrict_comm,4); 473528143c3dSStefano Zampini PetscValidLogicalCollectiveEnum(mat,reuse,5); 473628143c3dSStefano Zampini PetscValidLogicalCollectiveInt(mat,nis,7); 4737e7931f94SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)mat,MATIS,&ismatis);CHKERRQ(ierr); 473828143c3dSStefano Zampini if (!ismatis) SETERRQ1(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot use %s on a matrix object which is not of type MATIS",__FUNCT__); 4739e7931f94SStefano Zampini ierr = MatISGetLocalMat(mat,&local_mat);CHKERRQ(ierr); 4740e7931f94SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)local_mat,MATSEQDENSE,&isdense);CHKERRQ(ierr); 4741e7931f94SStefano Zampini if (!isdense) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Currently cannot subassemble MATIS when local matrix type is not of type SEQDENSE"); 4742e7931f94SStefano Zampini ierr = MatGetSize(local_mat,&rows,&cols);CHKERRQ(ierr); 4743e7931f94SStefano Zampini if (rows != cols) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Local MATIS matrices should be square"); 474428143c3dSStefano Zampini if (reuse == MAT_REUSE_MATRIX && *mat_n) { 474570cf5478SStefano Zampini PetscInt mrows,mcols,mnrows,mncols; 474670cf5478SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)*mat_n,MATIS,&ismatis);CHKERRQ(ierr); 474728143c3dSStefano Zampini if (!ismatis) SETERRQ(PetscObjectComm((PetscObject)*mat_n),PETSC_ERR_SUP,"Cannot reuse a matrix which is not of type MATIS"); 474870cf5478SStefano Zampini ierr = MatGetSize(mat,&mrows,&mcols);CHKERRQ(ierr); 474970cf5478SStefano Zampini ierr = MatGetSize(*mat_n,&mnrows,&mncols);CHKERRQ(ierr); 475070cf5478SStefano Zampini if (mrows != mnrows) SETERRQ2(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot reuse matrix! Wrong number of rows %D != %D",mrows,mnrows); 475170cf5478SStefano Zampini if (mcols != mncols) SETERRQ2(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot reuse matrix! Wrong number of cols %D != %D",mcols,mncols); 475270cf5478SStefano Zampini } 4753e7931f94SStefano Zampini ierr = MatGetBlockSize(local_mat,&bs);CHKERRQ(ierr); 4754e7931f94SStefano Zampini PetscValidLogicalCollectiveInt(mat,bs,0); 4755e7931f94SStefano Zampini /* prepare IS for sending if not provided */ 4756e7931f94SStefano Zampini if (!is_sends) { 475728143c3dSStefano Zampini if (!n_subdomains) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"You should specify either an IS or a target number of subdomains"); 4758b0c7d250SStefano Zampini ierr = MatISGetSubassemblingPattern(mat,n_subdomains,0,&is_sends_internal);CHKERRQ(ierr); 4759c8587f34SStefano Zampini } else { 4760e7931f94SStefano Zampini ierr = PetscObjectReference((PetscObject)is_sends);CHKERRQ(ierr); 4761e7931f94SStefano Zampini is_sends_internal = is_sends; 4762c8587f34SStefano Zampini } 4763e7931f94SStefano Zampini 4764e7931f94SStefano Zampini /* get comm */ 4765a316fed8SStefano Zampini ierr = PetscObjectGetComm((PetscObject)mat,&comm);CHKERRQ(ierr); 4766e7931f94SStefano Zampini 4767e7931f94SStefano Zampini /* compute number of sends */ 4768e7931f94SStefano Zampini ierr = ISGetLocalSize(is_sends_internal,&i);CHKERRQ(ierr); 4769e7931f94SStefano Zampini ierr = PetscMPIIntCast(i,&n_sends);CHKERRQ(ierr); 4770e7931f94SStefano Zampini 4771e7931f94SStefano Zampini /* compute number of receives */ 4772e7931f94SStefano Zampini ierr = MPI_Comm_size(comm,&commsize);CHKERRQ(ierr); 4773785e854fSJed Brown ierr = PetscMalloc1(commsize,&iflags);CHKERRQ(ierr); 4774e7931f94SStefano Zampini ierr = PetscMemzero(iflags,commsize*sizeof(*iflags));CHKERRQ(ierr); 4775e7931f94SStefano Zampini ierr = ISGetIndices(is_sends_internal,&is_indices);CHKERRQ(ierr); 4776e7931f94SStefano Zampini for (i=0;i<n_sends;i++) iflags[is_indices[i]] = 1; 4777e7931f94SStefano Zampini ierr = PetscGatherNumberOfMessages(comm,iflags,NULL,&n_recvs);CHKERRQ(ierr); 4778e7931f94SStefano Zampini ierr = PetscFree(iflags);CHKERRQ(ierr); 4779e7931f94SStefano Zampini 478028143c3dSStefano Zampini /* restrict comm if requested */ 478128143c3dSStefano Zampini subcomm = 0; 478228143c3dSStefano Zampini destroy_mat = PETSC_FALSE; 478328143c3dSStefano Zampini if (restrict_comm) { 4784779c1cceSStefano Zampini PetscMPIInt color,subcommsize; 4785779c1cceSStefano Zampini 478628143c3dSStefano Zampini color = 0; 478753a05cb3SStefano Zampini if (restrict_full) { 478853a05cb3SStefano Zampini if (!n_recvs) color = 1; /* processes not receiving anything will not partecipate in new comm (full restriction) */ 478953a05cb3SStefano Zampini } else { 479053a05cb3SStefano Zampini if (!n_recvs && n_sends) color = 1; /* just those processes that are sending but not receiving anything will not partecipate in new comm */ 479153a05cb3SStefano Zampini } 479228143c3dSStefano Zampini ierr = MPI_Allreduce(&color,&subcommsize,1,MPI_INT,MPI_SUM,comm);CHKERRQ(ierr); 479328143c3dSStefano Zampini subcommsize = commsize - subcommsize; 479428143c3dSStefano Zampini /* check if reuse has been requested */ 479528143c3dSStefano Zampini if (reuse == MAT_REUSE_MATRIX) { 479628143c3dSStefano Zampini if (*mat_n) { 479728143c3dSStefano Zampini PetscMPIInt subcommsize2; 479828143c3dSStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)*mat_n),&subcommsize2);CHKERRQ(ierr); 479928143c3dSStefano Zampini if (subcommsize != subcommsize2) SETERRQ2(PetscObjectComm((PetscObject)*mat_n),PETSC_ERR_PLIB,"Cannot reuse matrix! wrong subcomm size %d != %d",subcommsize,subcommsize2); 480028143c3dSStefano Zampini comm_n = PetscObjectComm((PetscObject)*mat_n); 480128143c3dSStefano Zampini } else { 480228143c3dSStefano Zampini comm_n = PETSC_COMM_SELF; 480328143c3dSStefano Zampini } 480428143c3dSStefano Zampini } else { /* MAT_INITIAL_MATRIX */ 4805779c1cceSStefano Zampini PetscMPIInt rank; 4806779c1cceSStefano Zampini 4807779c1cceSStefano Zampini ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 480828143c3dSStefano Zampini ierr = PetscSubcommCreate(comm,&subcomm);CHKERRQ(ierr); 480928143c3dSStefano Zampini ierr = PetscSubcommSetNumber(subcomm,2);CHKERRQ(ierr); 481028143c3dSStefano Zampini ierr = PetscSubcommSetTypeGeneral(subcomm,color,rank);CHKERRQ(ierr); 4811306c2d5bSBarry Smith comm_n = PetscSubcommChild(subcomm); 481228143c3dSStefano Zampini } 481328143c3dSStefano Zampini /* flag to destroy *mat_n if not significative */ 481428143c3dSStefano Zampini if (color) destroy_mat = PETSC_TRUE; 481528143c3dSStefano Zampini } else { 481628143c3dSStefano Zampini comm_n = comm; 481728143c3dSStefano Zampini } 481828143c3dSStefano Zampini 4819e7931f94SStefano Zampini /* prepare send/receive buffers */ 4820785e854fSJed Brown ierr = PetscMalloc1(commsize,&ilengths_idxs);CHKERRQ(ierr); 4821e7931f94SStefano Zampini ierr = PetscMemzero(ilengths_idxs,commsize*sizeof(*ilengths_idxs));CHKERRQ(ierr); 4822785e854fSJed Brown ierr = PetscMalloc1(commsize,&ilengths_vals);CHKERRQ(ierr); 4823e7931f94SStefano Zampini ierr = PetscMemzero(ilengths_vals,commsize*sizeof(*ilengths_vals));CHKERRQ(ierr); 482428143c3dSStefano Zampini if (nis) { 4825854ce69bSBarry Smith ierr = PetscCalloc1(commsize,&ilengths_idxs_is);CHKERRQ(ierr); 482628143c3dSStefano Zampini } 4827e7931f94SStefano Zampini 482828143c3dSStefano Zampini /* Get data from local matrices */ 4829e7931f94SStefano Zampini if (!isdense) { 4830a26c9d0eSStefano Zampini SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Subassembling of AIJ local matrices not yet implemented"); 4831e7931f94SStefano Zampini /* TODO: See below some guidelines on how to prepare the local buffers */ 4832e7931f94SStefano Zampini /* 4833e7931f94SStefano Zampini send_buffer_vals should contain the raw values of the local matrix 4834e7931f94SStefano Zampini send_buffer_idxs should contain: 4835e7931f94SStefano Zampini - MatType_PRIVATE type 4836e7931f94SStefano Zampini - PetscInt size_of_l2gmap 4837e7931f94SStefano Zampini - PetscInt global_row_indices[size_of_l2gmap] 4838e7931f94SStefano Zampini - PetscInt all_other_info_which_is_needed_to_compute_preallocation_and_set_values 4839e7931f94SStefano Zampini */ 4840e7931f94SStefano Zampini } else { 4841e7931f94SStefano Zampini ierr = MatDenseGetArray(local_mat,&send_buffer_vals);CHKERRQ(ierr); 48423bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetSize(mat->rmap->mapping,&i);CHKERRQ(ierr); 4843854ce69bSBarry Smith ierr = PetscMalloc1(i+2,&send_buffer_idxs);CHKERRQ(ierr); 4844e7931f94SStefano Zampini send_buffer_idxs[0] = (PetscInt)MATDENSE_PRIVATE; 4845e7931f94SStefano Zampini send_buffer_idxs[1] = i; 48463bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetIndices(mat->rmap->mapping,(const PetscInt**)&ptr_idxs);CHKERRQ(ierr); 4847e7931f94SStefano Zampini ierr = PetscMemcpy(&send_buffer_idxs[2],ptr_idxs,i*sizeof(PetscInt));CHKERRQ(ierr); 48483bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingRestoreIndices(mat->rmap->mapping,(const PetscInt**)&ptr_idxs);CHKERRQ(ierr); 4849e7931f94SStefano Zampini ierr = PetscMPIIntCast(i,&len);CHKERRQ(ierr); 4850e7931f94SStefano Zampini for (i=0;i<n_sends;i++) { 4851e7931f94SStefano Zampini ilengths_vals[is_indices[i]] = len*len; 4852e7931f94SStefano Zampini ilengths_idxs[is_indices[i]] = len+2; 4853c8587f34SStefano Zampini } 4854c8587f34SStefano Zampini } 4855e7931f94SStefano Zampini ierr = PetscGatherMessageLengths2(comm,n_sends,n_recvs,ilengths_idxs,ilengths_vals,&onodes,&olengths_idxs,&olengths_vals);CHKERRQ(ierr); 485628143c3dSStefano Zampini /* additional is (if any) */ 485728143c3dSStefano Zampini if (nis) { 485828143c3dSStefano Zampini PetscMPIInt psum; 485928143c3dSStefano Zampini PetscInt j; 486028143c3dSStefano Zampini for (j=0,psum=0;j<nis;j++) { 486128143c3dSStefano Zampini PetscInt plen; 486228143c3dSStefano Zampini ierr = ISGetLocalSize(isarray[j],&plen);CHKERRQ(ierr); 486328143c3dSStefano Zampini ierr = PetscMPIIntCast(plen,&len);CHKERRQ(ierr); 486428143c3dSStefano Zampini psum += len+1; /* indices + lenght */ 486528143c3dSStefano Zampini } 4866854ce69bSBarry Smith ierr = PetscMalloc1(psum,&send_buffer_idxs_is);CHKERRQ(ierr); 486728143c3dSStefano Zampini for (j=0,psum=0;j<nis;j++) { 486828143c3dSStefano Zampini PetscInt plen; 486928143c3dSStefano Zampini const PetscInt *is_array_idxs; 487028143c3dSStefano Zampini ierr = ISGetLocalSize(isarray[j],&plen);CHKERRQ(ierr); 487128143c3dSStefano Zampini send_buffer_idxs_is[psum] = plen; 487228143c3dSStefano Zampini ierr = ISGetIndices(isarray[j],&is_array_idxs);CHKERRQ(ierr); 487328143c3dSStefano Zampini ierr = PetscMemcpy(&send_buffer_idxs_is[psum+1],is_array_idxs,plen*sizeof(PetscInt));CHKERRQ(ierr); 487428143c3dSStefano Zampini ierr = ISRestoreIndices(isarray[j],&is_array_idxs);CHKERRQ(ierr); 487528143c3dSStefano Zampini psum += plen+1; /* indices + lenght */ 487628143c3dSStefano Zampini } 487728143c3dSStefano Zampini for (i=0;i<n_sends;i++) { 487828143c3dSStefano Zampini ilengths_idxs_is[is_indices[i]] = psum; 487928143c3dSStefano Zampini } 488028143c3dSStefano Zampini ierr = PetscGatherMessageLengths(comm,n_sends,n_recvs,ilengths_idxs_is,&onodes_is,&olengths_idxs_is);CHKERRQ(ierr); 488128143c3dSStefano Zampini } 488228143c3dSStefano Zampini 4883e7931f94SStefano Zampini buf_size_idxs = 0; 4884e7931f94SStefano Zampini buf_size_vals = 0; 488528143c3dSStefano Zampini buf_size_idxs_is = 0; 4886e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 4887e7931f94SStefano Zampini buf_size_idxs += (PetscInt)olengths_idxs[i]; 4888e7931f94SStefano Zampini buf_size_vals += (PetscInt)olengths_vals[i]; 488928143c3dSStefano Zampini if (nis) buf_size_idxs_is += (PetscInt)olengths_idxs_is[i]; 4890e7931f94SStefano Zampini } 4891785e854fSJed Brown ierr = PetscMalloc1(buf_size_idxs,&recv_buffer_idxs);CHKERRQ(ierr); 4892785e854fSJed Brown ierr = PetscMalloc1(buf_size_vals,&recv_buffer_vals);CHKERRQ(ierr); 489395ecbf38SStefano Zampini ierr = PetscMalloc1(buf_size_idxs_is,&recv_buffer_idxs_is);CHKERRQ(ierr); 4894e7931f94SStefano Zampini 4895e7931f94SStefano Zampini /* get new tags for clean communications */ 4896e7931f94SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_idxs);CHKERRQ(ierr); 4897e7931f94SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_vals);CHKERRQ(ierr); 489828143c3dSStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_idxs_is);CHKERRQ(ierr); 4899e7931f94SStefano Zampini 4900e7931f94SStefano Zampini /* allocate for requests */ 4901785e854fSJed Brown ierr = PetscMalloc1(n_sends,&send_req_idxs);CHKERRQ(ierr); 4902785e854fSJed Brown ierr = PetscMalloc1(n_sends,&send_req_vals);CHKERRQ(ierr); 490395ecbf38SStefano Zampini ierr = PetscMalloc1(n_sends,&send_req_idxs_is);CHKERRQ(ierr); 4904785e854fSJed Brown ierr = PetscMalloc1(n_recvs,&recv_req_idxs);CHKERRQ(ierr); 4905785e854fSJed Brown ierr = PetscMalloc1(n_recvs,&recv_req_vals);CHKERRQ(ierr); 490695ecbf38SStefano Zampini ierr = PetscMalloc1(n_recvs,&recv_req_idxs_is);CHKERRQ(ierr); 4907e7931f94SStefano Zampini 4908e7931f94SStefano Zampini /* communications */ 4909e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 4910e7931f94SStefano Zampini ptr_vals = recv_buffer_vals; 491128143c3dSStefano Zampini ptr_idxs_is = recv_buffer_idxs_is; 4912e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 4913e7931f94SStefano Zampini source_dest = onodes[i]; 4914e7931f94SStefano Zampini ierr = MPI_Irecv(ptr_idxs,olengths_idxs[i],MPIU_INT,source_dest,tag_idxs,comm,&recv_req_idxs[i]);CHKERRQ(ierr); 4915e7931f94SStefano Zampini ierr = MPI_Irecv(ptr_vals,olengths_vals[i],MPIU_SCALAR,source_dest,tag_vals,comm,&recv_req_vals[i]);CHKERRQ(ierr); 4916e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 4917e7931f94SStefano Zampini ptr_vals += olengths_vals[i]; 491828143c3dSStefano Zampini if (nis) { 491928143c3dSStefano 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); 492028143c3dSStefano Zampini ptr_idxs_is += olengths_idxs_is[i]; 492128143c3dSStefano Zampini } 4922e7931f94SStefano Zampini } 4923e7931f94SStefano Zampini for (i=0;i<n_sends;i++) { 4924e7931f94SStefano Zampini ierr = PetscMPIIntCast(is_indices[i],&source_dest);CHKERRQ(ierr); 4925e7931f94SStefano Zampini ierr = MPI_Isend(send_buffer_idxs,ilengths_idxs[source_dest],MPIU_INT,source_dest,tag_idxs,comm,&send_req_idxs[i]);CHKERRQ(ierr); 4926e7931f94SStefano Zampini ierr = MPI_Isend(send_buffer_vals,ilengths_vals[source_dest],MPIU_SCALAR,source_dest,tag_vals,comm,&send_req_vals[i]);CHKERRQ(ierr); 492728143c3dSStefano Zampini if (nis) { 492828143c3dSStefano 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); 492928143c3dSStefano Zampini } 4930e7931f94SStefano Zampini } 4931e7931f94SStefano Zampini ierr = ISRestoreIndices(is_sends_internal,&is_indices);CHKERRQ(ierr); 4932e7931f94SStefano Zampini ierr = ISDestroy(&is_sends_internal);CHKERRQ(ierr); 4933e7931f94SStefano Zampini 4934e7931f94SStefano Zampini /* assemble new l2g map */ 4935e7931f94SStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_idxs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 4936e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 49379d30be91SStefano Zampini new_local_rows = 0; 4938e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 49399d30be91SStefano Zampini new_local_rows += *(ptr_idxs+1); /* second element is the local size of the l2gmap */ 4940e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 4941e7931f94SStefano Zampini } 49429d30be91SStefano Zampini ierr = PetscMalloc1(new_local_rows,&l2gmap_indices);CHKERRQ(ierr); 4943e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 49449d30be91SStefano Zampini new_local_rows = 0; 4945e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 49469d30be91SStefano Zampini ierr = PetscMemcpy(&l2gmap_indices[new_local_rows],ptr_idxs+2,(*(ptr_idxs+1))*sizeof(PetscInt));CHKERRQ(ierr); 49479d30be91SStefano Zampini new_local_rows += *(ptr_idxs+1); /* second element is the local size of the l2gmap */ 4948e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 4949e7931f94SStefano Zampini } 49509d30be91SStefano Zampini ierr = PetscSortRemoveDupsInt(&new_local_rows,l2gmap_indices);CHKERRQ(ierr); 49519d30be91SStefano Zampini ierr = ISLocalToGlobalMappingCreate(comm_n,1,new_local_rows,l2gmap_indices,PETSC_COPY_VALUES,&l2gmap);CHKERRQ(ierr); 4952e7931f94SStefano Zampini ierr = PetscFree(l2gmap_indices);CHKERRQ(ierr); 4953e7931f94SStefano Zampini 4954e7931f94SStefano Zampini /* infer new local matrix type from received local matrices type */ 4955e7931f94SStefano Zampini /* currently if all local matrices are of type X, then the resulting matrix will be of type X, except for the dense case */ 4956e7931f94SStefano 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) */ 4957e7931f94SStefano Zampini if (n_recvs) { 495828143c3dSStefano Zampini MatTypePrivate new_local_type_private = (MatTypePrivate)send_buffer_idxs[0]; 4959e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 4960e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 4961e7931f94SStefano Zampini if ((PetscInt)new_local_type_private != *ptr_idxs) { 4962e7931f94SStefano Zampini new_local_type_private = MATAIJ_PRIVATE; 4963e7931f94SStefano Zampini break; 4964e7931f94SStefano Zampini } 4965e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 4966e7931f94SStefano Zampini } 4967e7931f94SStefano Zampini switch (new_local_type_private) { 496828143c3dSStefano Zampini case MATDENSE_PRIVATE: 496928143c3dSStefano Zampini if (n_recvs>1) { /* subassembling of dense matrices does not give a dense matrix! */ 4970e7931f94SStefano Zampini new_local_type = MATSEQAIJ; 4971e7931f94SStefano Zampini bs = 1; 497228143c3dSStefano Zampini } else { /* if I receive only 1 dense matrix */ 497328143c3dSStefano Zampini new_local_type = MATSEQDENSE; 497428143c3dSStefano Zampini bs = 1; 497528143c3dSStefano Zampini } 4976e7931f94SStefano Zampini break; 4977e7931f94SStefano Zampini case MATAIJ_PRIVATE: 4978e7931f94SStefano Zampini new_local_type = MATSEQAIJ; 4979e7931f94SStefano Zampini bs = 1; 4980e7931f94SStefano Zampini break; 4981e7931f94SStefano Zampini case MATBAIJ_PRIVATE: 4982e7931f94SStefano Zampini new_local_type = MATSEQBAIJ; 4983e7931f94SStefano Zampini break; 4984e7931f94SStefano Zampini case MATSBAIJ_PRIVATE: 4985e7931f94SStefano Zampini new_local_type = MATSEQSBAIJ; 4986e7931f94SStefano Zampini break; 4987e7931f94SStefano Zampini default: 49889d30be91SStefano Zampini SETERRQ2(comm,PETSC_ERR_SUP,"Unsupported private type %d in %s",new_local_type_private,__FUNCT__); 4989e7931f94SStefano Zampini break; 4990e7931f94SStefano Zampini } 499128143c3dSStefano Zampini } else { /* by default, new_local_type is seqdense */ 499228143c3dSStefano Zampini new_local_type = MATSEQDENSE; 499328143c3dSStefano Zampini bs = 1; 4994e7931f94SStefano Zampini } 4995e7931f94SStefano Zampini 499670cf5478SStefano Zampini /* create MATIS object if needed */ 499770cf5478SStefano Zampini if (reuse == MAT_INITIAL_MATRIX) { 4998e7931f94SStefano Zampini ierr = MatGetSize(mat,&rows,&cols);CHKERRQ(ierr); 4999e176bc59SStefano Zampini ierr = MatCreateIS(comm_n,bs,PETSC_DECIDE,PETSC_DECIDE,rows,cols,l2gmap,NULL,mat_n);CHKERRQ(ierr); 500070cf5478SStefano Zampini } else { 500170cf5478SStefano Zampini /* it also destroys the local matrices */ 500270cf5478SStefano Zampini ierr = MatSetLocalToGlobalMapping(*mat_n,l2gmap,l2gmap);CHKERRQ(ierr); 500370cf5478SStefano Zampini } 500470cf5478SStefano Zampini ierr = MatISGetLocalMat(*mat_n,&local_mat);CHKERRQ(ierr); 5005e7931f94SStefano Zampini ierr = MatSetType(local_mat,new_local_type);CHKERRQ(ierr); 50069d30be91SStefano Zampini 50079d30be91SStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_vals,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 50089d30be91SStefano Zampini 50099d30be91SStefano Zampini /* Global to local map of received indices */ 50109d30be91SStefano Zampini ierr = PetscMalloc1(buf_size_idxs,&recv_buffer_idxs_local);CHKERRQ(ierr); /* needed for values insertion */ 50119d30be91SStefano Zampini ierr = ISGlobalToLocalMappingApply(l2gmap,IS_GTOLM_MASK,buf_size_idxs,recv_buffer_idxs,&i,recv_buffer_idxs_local);CHKERRQ(ierr); 50129d30be91SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&l2gmap);CHKERRQ(ierr); 50139d30be91SStefano Zampini 50149d30be91SStefano Zampini /* restore attributes -> type of incoming data and its size */ 50159d30be91SStefano Zampini buf_size_idxs = 0; 50169d30be91SStefano Zampini for (i=0;i<n_recvs;i++) { 50179d30be91SStefano Zampini recv_buffer_idxs_local[buf_size_idxs] = recv_buffer_idxs[buf_size_idxs]; 50189d30be91SStefano Zampini recv_buffer_idxs_local[buf_size_idxs+1] = recv_buffer_idxs[buf_size_idxs+1]; 50199d30be91SStefano Zampini buf_size_idxs += (PetscInt)olengths_idxs[i]; 50209d30be91SStefano Zampini } 50219d30be91SStefano Zampini ierr = PetscFree(recv_buffer_idxs);CHKERRQ(ierr); 50229d30be91SStefano Zampini 50239d30be91SStefano Zampini /* set preallocation */ 50249d30be91SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)local_mat,MATSEQDENSE,&newisdense);CHKERRQ(ierr); 50259d30be91SStefano Zampini if (!newisdense) { 50269d30be91SStefano Zampini PetscInt *new_local_nnz=0; 50279d30be91SStefano Zampini 50289d30be91SStefano Zampini ptr_vals = recv_buffer_vals; 50299d30be91SStefano Zampini ptr_idxs = recv_buffer_idxs_local; 50309d30be91SStefano Zampini if (n_recvs) { 50319d30be91SStefano Zampini ierr = PetscCalloc1(new_local_rows,&new_local_nnz);CHKERRQ(ierr); 50329d30be91SStefano Zampini } 50339d30be91SStefano Zampini for (i=0;i<n_recvs;i++) { 50349d30be91SStefano Zampini PetscInt j; 50359d30be91SStefano Zampini if (*ptr_idxs == (PetscInt)MATDENSE_PRIVATE) { /* preallocation provided for dense case only */ 50369d30be91SStefano Zampini for (j=0;j<*(ptr_idxs+1);j++) { 50379d30be91SStefano Zampini new_local_nnz[*(ptr_idxs+2+j)] += *(ptr_idxs+1); 50389d30be91SStefano Zampini } 50399d30be91SStefano Zampini } else { 50409d30be91SStefano Zampini /* TODO */ 50419d30be91SStefano Zampini } 50429d30be91SStefano Zampini ptr_idxs += olengths_idxs[i]; 50439d30be91SStefano Zampini } 50449d30be91SStefano Zampini if (new_local_nnz) { 50459d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] = PetscMin(new_local_nnz[i],new_local_rows); 50469d30be91SStefano Zampini ierr = MatSeqAIJSetPreallocation(local_mat,0,new_local_nnz);CHKERRQ(ierr); 50479d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] /= bs; 50489d30be91SStefano Zampini ierr = MatSeqBAIJSetPreallocation(local_mat,bs,0,new_local_nnz);CHKERRQ(ierr); 50499d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] = PetscMax(new_local_nnz[i]-i,0); 50509d30be91SStefano Zampini ierr = MatSeqSBAIJSetPreallocation(local_mat,bs,0,new_local_nnz);CHKERRQ(ierr); 50519d30be91SStefano Zampini } else { 50529d30be91SStefano Zampini ierr = MatSetUp(local_mat);CHKERRQ(ierr); 50539d30be91SStefano Zampini } 50549d30be91SStefano Zampini ierr = PetscFree(new_local_nnz);CHKERRQ(ierr); 50559d30be91SStefano Zampini } else { 50569d30be91SStefano Zampini ierr = MatSetUp(local_mat);CHKERRQ(ierr); 50579d30be91SStefano Zampini } 5058e7931f94SStefano Zampini 5059e7931f94SStefano Zampini /* set values */ 5060e7931f94SStefano Zampini ptr_vals = recv_buffer_vals; 50619d30be91SStefano Zampini ptr_idxs = recv_buffer_idxs_local; 5062e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 5063e7931f94SStefano Zampini if (*ptr_idxs == (PetscInt)MATDENSE_PRIVATE) { /* values insertion provided for dense case only */ 5064e7931f94SStefano Zampini ierr = MatSetOption(local_mat,MAT_ROW_ORIENTED,PETSC_FALSE);CHKERRQ(ierr); 50659d30be91SStefano Zampini ierr = MatSetValues(local_mat,*(ptr_idxs+1),ptr_idxs+2,*(ptr_idxs+1),ptr_idxs+2,ptr_vals,ADD_VALUES);CHKERRQ(ierr); 5066e7931f94SStefano Zampini ierr = MatAssemblyBegin(local_mat,MAT_FLUSH_ASSEMBLY);CHKERRQ(ierr); 5067e7931f94SStefano Zampini ierr = MatAssemblyEnd(local_mat,MAT_FLUSH_ASSEMBLY);CHKERRQ(ierr); 5068e7931f94SStefano Zampini ierr = MatSetOption(local_mat,MAT_ROW_ORIENTED,PETSC_TRUE);CHKERRQ(ierr); 506928143c3dSStefano Zampini } else { 507028143c3dSStefano Zampini /* TODO */ 5071e7931f94SStefano Zampini } 5072e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 5073e7931f94SStefano Zampini ptr_vals += olengths_vals[i]; 5074e7931f94SStefano Zampini } 5075e7931f94SStefano Zampini ierr = MatAssemblyBegin(local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5076e7931f94SStefano Zampini ierr = MatAssemblyEnd(local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 507770cf5478SStefano Zampini ierr = MatAssemblyBegin(*mat_n,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 507870cf5478SStefano Zampini ierr = MatAssemblyEnd(*mat_n,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 50799d30be91SStefano Zampini ierr = PetscFree(recv_buffer_vals);CHKERRQ(ierr); 50809d30be91SStefano Zampini ierr = PetscFree(recv_buffer_idxs_local);CHKERRQ(ierr); 5081e7931f94SStefano Zampini 5082dfd14d43SStefano Zampini #if 0 508328143c3dSStefano Zampini if (!restrict_comm) { /* check */ 5084e7931f94SStefano Zampini Vec lvec,rvec; 5085e7931f94SStefano Zampini PetscReal infty_error; 5086e7931f94SStefano Zampini 50872a7a6963SBarry Smith ierr = MatCreateVecs(mat,&rvec,&lvec);CHKERRQ(ierr); 5088e7931f94SStefano Zampini ierr = VecSetRandom(rvec,NULL);CHKERRQ(ierr); 5089e7931f94SStefano Zampini ierr = MatMult(mat,rvec,lvec);CHKERRQ(ierr); 5090e7931f94SStefano Zampini ierr = VecScale(lvec,-1.0);CHKERRQ(ierr); 509170cf5478SStefano Zampini ierr = MatMultAdd(*mat_n,rvec,lvec,lvec);CHKERRQ(ierr); 5092e7931f94SStefano Zampini ierr = VecNorm(lvec,NORM_INFINITY,&infty_error);CHKERRQ(ierr); 5093e7931f94SStefano Zampini ierr = PetscPrintf(PetscObjectComm((PetscObject)mat),"Infinity error subassembling %1.6e\n",infty_error); 5094e7931f94SStefano Zampini ierr = VecDestroy(&rvec);CHKERRQ(ierr); 5095e7931f94SStefano Zampini ierr = VecDestroy(&lvec);CHKERRQ(ierr); 5096e7931f94SStefano Zampini } 509728143c3dSStefano Zampini #endif 5098e7931f94SStefano Zampini 509928143c3dSStefano Zampini /* assemble new additional is (if any) */ 510028143c3dSStefano Zampini if (nis) { 510128143c3dSStefano Zampini PetscInt **temp_idxs,*count_is,j,psum; 510228143c3dSStefano Zampini 510328143c3dSStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_idxs_is,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 5104854ce69bSBarry Smith ierr = PetscCalloc1(nis,&count_is);CHKERRQ(ierr); 510528143c3dSStefano Zampini ptr_idxs = recv_buffer_idxs_is; 510628143c3dSStefano Zampini psum = 0; 510728143c3dSStefano Zampini for (i=0;i<n_recvs;i++) { 510828143c3dSStefano Zampini for (j=0;j<nis;j++) { 510928143c3dSStefano Zampini PetscInt plen = *(ptr_idxs); /* first element is the local size of IS's indices */ 511028143c3dSStefano Zampini count_is[j] += plen; /* increment counting of buffer for j-th IS */ 511128143c3dSStefano Zampini psum += plen; 511228143c3dSStefano Zampini ptr_idxs += plen+1; /* shift pointer to received data */ 511328143c3dSStefano Zampini } 511428143c3dSStefano Zampini } 5115854ce69bSBarry Smith ierr = PetscMalloc1(nis,&temp_idxs);CHKERRQ(ierr); 5116854ce69bSBarry Smith ierr = PetscMalloc1(psum,&temp_idxs[0]);CHKERRQ(ierr); 511728143c3dSStefano Zampini for (i=1;i<nis;i++) { 511828143c3dSStefano Zampini temp_idxs[i] = temp_idxs[i-1]+count_is[i-1]; 511928143c3dSStefano Zampini } 512028143c3dSStefano Zampini ierr = PetscMemzero(count_is,nis*sizeof(PetscInt));CHKERRQ(ierr); 512128143c3dSStefano Zampini ptr_idxs = recv_buffer_idxs_is; 512228143c3dSStefano Zampini for (i=0;i<n_recvs;i++) { 512328143c3dSStefano Zampini for (j=0;j<nis;j++) { 512428143c3dSStefano Zampini PetscInt plen = *(ptr_idxs); /* first element is the local size of IS's indices */ 512528143c3dSStefano Zampini ierr = PetscMemcpy(&temp_idxs[j][count_is[j]],ptr_idxs+1,plen*sizeof(PetscInt));CHKERRQ(ierr); 512628143c3dSStefano Zampini count_is[j] += plen; /* increment starting point of buffer for j-th IS */ 512728143c3dSStefano Zampini ptr_idxs += plen+1; /* shift pointer to received data */ 512828143c3dSStefano Zampini } 512928143c3dSStefano Zampini } 513028143c3dSStefano Zampini for (i=0;i<nis;i++) { 513128143c3dSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 513228143c3dSStefano Zampini ierr = PetscSortRemoveDupsInt(&count_is[i],temp_idxs[i]);CHKERRQ(ierr);CHKERRQ(ierr); 513328143c3dSStefano Zampini ierr = ISCreateGeneral(comm_n,count_is[i],temp_idxs[i],PETSC_COPY_VALUES,&isarray[i]);CHKERRQ(ierr); 513428143c3dSStefano Zampini } 513528143c3dSStefano Zampini ierr = PetscFree(count_is);CHKERRQ(ierr); 513628143c3dSStefano Zampini ierr = PetscFree(temp_idxs[0]);CHKERRQ(ierr); 513728143c3dSStefano Zampini ierr = PetscFree(temp_idxs);CHKERRQ(ierr); 513828143c3dSStefano Zampini } 5139e7931f94SStefano Zampini /* free workspace */ 514028143c3dSStefano Zampini ierr = PetscFree(recv_buffer_idxs_is);CHKERRQ(ierr); 5141e7931f94SStefano Zampini ierr = MPI_Waitall(n_sends,send_req_idxs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 5142e7931f94SStefano Zampini ierr = PetscFree(send_buffer_idxs);CHKERRQ(ierr); 5143e7931f94SStefano Zampini ierr = MPI_Waitall(n_sends,send_req_vals,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 5144e7931f94SStefano Zampini if (isdense) { 5145e7931f94SStefano Zampini ierr = MatISGetLocalMat(mat,&local_mat);CHKERRQ(ierr); 5146e7931f94SStefano Zampini ierr = MatDenseRestoreArray(local_mat,&send_buffer_vals);CHKERRQ(ierr); 5147e7931f94SStefano Zampini } else { 5148e7931f94SStefano Zampini /* ierr = PetscFree(send_buffer_vals);CHKERRQ(ierr); */ 5149e7931f94SStefano Zampini } 515028143c3dSStefano Zampini if (nis) { 515128143c3dSStefano Zampini ierr = MPI_Waitall(n_sends,send_req_idxs_is,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 515228143c3dSStefano Zampini ierr = PetscFree(send_buffer_idxs_is);CHKERRQ(ierr); 515328143c3dSStefano Zampini } 5154e7931f94SStefano Zampini ierr = PetscFree(recv_req_idxs);CHKERRQ(ierr); 5155e7931f94SStefano Zampini ierr = PetscFree(recv_req_vals);CHKERRQ(ierr); 515628143c3dSStefano Zampini ierr = PetscFree(recv_req_idxs_is);CHKERRQ(ierr); 5157e7931f94SStefano Zampini ierr = PetscFree(send_req_idxs);CHKERRQ(ierr); 5158e7931f94SStefano Zampini ierr = PetscFree(send_req_vals);CHKERRQ(ierr); 515928143c3dSStefano Zampini ierr = PetscFree(send_req_idxs_is);CHKERRQ(ierr); 5160e7931f94SStefano Zampini ierr = PetscFree(ilengths_vals);CHKERRQ(ierr); 5161e7931f94SStefano Zampini ierr = PetscFree(ilengths_idxs);CHKERRQ(ierr); 5162e7931f94SStefano Zampini ierr = PetscFree(olengths_vals);CHKERRQ(ierr); 5163e7931f94SStefano Zampini ierr = PetscFree(olengths_idxs);CHKERRQ(ierr); 5164e7931f94SStefano Zampini ierr = PetscFree(onodes);CHKERRQ(ierr); 516528143c3dSStefano Zampini if (nis) { 516628143c3dSStefano Zampini ierr = PetscFree(ilengths_idxs_is);CHKERRQ(ierr); 516728143c3dSStefano Zampini ierr = PetscFree(olengths_idxs_is);CHKERRQ(ierr); 516828143c3dSStefano Zampini ierr = PetscFree(onodes_is);CHKERRQ(ierr); 516928143c3dSStefano Zampini } 517028143c3dSStefano Zampini ierr = PetscSubcommDestroy(&subcomm);CHKERRQ(ierr); 517128143c3dSStefano Zampini if (destroy_mat) { /* destroy mat is true only if restrict comm is true and process will not partecipate */ 517228143c3dSStefano Zampini ierr = MatDestroy(mat_n);CHKERRQ(ierr); 517328143c3dSStefano Zampini for (i=0;i<nis;i++) { 517428143c3dSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 517528143c3dSStefano Zampini } 517653a05cb3SStefano Zampini *mat_n = NULL; 517728143c3dSStefano Zampini } 5178e7931f94SStefano Zampini PetscFunctionReturn(0); 5179e7931f94SStefano Zampini } 5180a57a6d2fSStefano Zampini 518112edc857SStefano Zampini /* temporary hack into ksp private data structure */ 5182af0996ceSBarry Smith #include <petsc/private/kspimpl.h> 518312edc857SStefano Zampini 5184c8587f34SStefano Zampini #undef __FUNCT__ 5185c8587f34SStefano Zampini #define __FUNCT__ "PCBDDCSetUpCoarseSolver" 5186c8587f34SStefano Zampini PetscErrorCode PCBDDCSetUpCoarseSolver(PC pc,PetscScalar* coarse_submat_vals) 5187c8587f34SStefano Zampini { 5188c8587f34SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 5189c8587f34SStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 519020a2ab83SStefano Zampini Mat coarse_mat,coarse_mat_is,coarse_submat_dense; 51919881197aSStefano Zampini MatNullSpace CoarseNullSpace=NULL; 519220a2ab83SStefano Zampini ISLocalToGlobalMapping coarse_islg; 51936e683305SStefano Zampini IS coarse_is,*isarray; 51946e683305SStefano Zampini PetscInt i,im_active=-1,active_procs=-1; 519530368db7SStefano Zampini PetscInt nis,nisdofs,nisneu,nisvert; 5196f9eb5b7dSStefano Zampini PC pc_temp; 5197c8587f34SStefano Zampini PCType coarse_pc_type; 5198c8587f34SStefano Zampini KSPType coarse_ksp_type; 5199f9eb5b7dSStefano Zampini PetscBool multilevel_requested,multilevel_allowed; 52004f3a063dSStefano Zampini PetscBool isredundant,isbddc,isnn,coarse_reuse; 52016e683305SStefano Zampini Mat t_coarse_mat_is; 52026e683305SStefano Zampini PetscInt void_procs,ncoarse_ml,ncoarse_ds,ncoarse; 52036e683305SStefano Zampini PetscMPIInt all_procs; 520474e2c79eSStefano Zampini PetscBool csin_ml,csin_ds,csin,csin_type_simple,redist; 520568457ee5SStefano Zampini PetscBool compute_vecs = PETSC_FALSE; 520622bc73bbSStefano Zampini PetscScalar *array; 52079881197aSStefano Zampini PetscErrorCode ierr; 5208fdc09c96SStefano Zampini 5209c8587f34SStefano Zampini PetscFunctionBegin; 5210c8587f34SStefano Zampini /* Assign global numbering to coarse dofs */ 521168457ee5SStefano 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 */ 5212fa7f1dd8SStefano Zampini PetscInt ocoarse_size; 52135a75c04eSSatish Balay compute_vecs = PETSC_TRUE; 5214fa7f1dd8SStefano Zampini ocoarse_size = pcbddc->coarse_size; 5215f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->global_primal_indices);CHKERRQ(ierr); 5216f4ddd8eeSStefano Zampini ierr = PCBDDCComputePrimalNumbering(pc,&pcbddc->coarse_size,&pcbddc->global_primal_indices);CHKERRQ(ierr); 5217f4ddd8eeSStefano Zampini /* see if we can avoid some work */ 5218fa7f1dd8SStefano Zampini if (pcbddc->coarse_ksp) { /* coarse ksp has already been created */ 521951bea450SStefano Zampini /* if the coarse size is different or we are using adaptive selection, better to not reuse the coarse matrix */ 522051bea450SStefano Zampini if (ocoarse_size != pcbddc->coarse_size || pcbddc->adaptive_selection) { 5221dc4bcba2SStefano Zampini PC pc; 5222dc4bcba2SStefano Zampini PetscBool isbddc; 5223dc4bcba2SStefano Zampini 5224dc4bcba2SStefano Zampini /* temporary workaround since PCBDDC does not have a reset method so far */ 5225dc4bcba2SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&pc);CHKERRQ(ierr); 5226dc4bcba2SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc,PCBDDC,&isbddc);CHKERRQ(ierr); 5227dc4bcba2SStefano Zampini if (isbddc) { 522863c961adSStefano Zampini ierr = KSPDestroy(&pcbddc->coarse_ksp);CHKERRQ(ierr); 522963c961adSStefano Zampini } else { 5230727cdba6SStefano Zampini ierr = KSPReset(pcbddc->coarse_ksp);CHKERRQ(ierr); 523163c961adSStefano Zampini } 5232fa7f1dd8SStefano Zampini coarse_reuse = PETSC_FALSE; 5233fa7f1dd8SStefano Zampini } else { /* we can safely reuse already computed coarse matrix */ 5234fa7f1dd8SStefano Zampini coarse_reuse = PETSC_TRUE; 5235f4ddd8eeSStefano Zampini } 5236fa7f1dd8SStefano Zampini } else { /* there's no coarse ksp, so we need to create the coarse matrix too */ 5237fa7f1dd8SStefano Zampini coarse_reuse = PETSC_FALSE; 5238f4ddd8eeSStefano Zampini } 523970cf5478SStefano Zampini /* reset any subassembling information */ 524070cf5478SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling);CHKERRQ(ierr); 52416e683305SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 52426e683305SStefano Zampini } else { /* primal space is unchanged, so we can reuse coarse matrix */ 5243fa7f1dd8SStefano Zampini coarse_reuse = PETSC_TRUE; 5244f4ddd8eeSStefano Zampini } 5245c8587f34SStefano Zampini 52466e683305SStefano Zampini /* count "active" (i.e. with positive local size) and "void" processes */ 52472b510759SStefano Zampini im_active = !!(pcis->n); 52482b510759SStefano Zampini ierr = MPI_Allreduce(&im_active,&active_procs,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 52496e683305SStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)pc),&all_procs);CHKERRQ(ierr); 52506e683305SStefano Zampini void_procs = all_procs-active_procs; 52516e683305SStefano Zampini csin_type_simple = PETSC_TRUE; 525274e2c79eSStefano Zampini redist = PETSC_FALSE; 525322bc73bbSStefano Zampini if (pcbddc->current_level && void_procs) { 52546e683305SStefano Zampini csin_ml = PETSC_TRUE; 52556e683305SStefano Zampini ncoarse_ml = void_procs; 5256779c1cceSStefano Zampini /* it has no sense to redistribute on a set of processors larger than the number of active processes */ 5257779c1cceSStefano Zampini if (pcbddc->redistribute_coarse > 0 && pcbddc->redistribute_coarse < active_procs) { 52586e683305SStefano Zampini csin_ds = PETSC_TRUE; 525918a45a71SStefano Zampini ncoarse_ds = pcbddc->redistribute_coarse; 526018a45a71SStefano Zampini redist = PETSC_TRUE; 526118a45a71SStefano Zampini } else { 52626e683305SStefano Zampini csin_ds = PETSC_TRUE; 5263779c1cceSStefano Zampini ncoarse_ds = active_procs; 5264779c1cceSStefano Zampini redist = PETSC_TRUE; 526518a45a71SStefano Zampini } 52666e683305SStefano Zampini } else { 52676e683305SStefano Zampini csin_ml = PETSC_FALSE; 52686e683305SStefano Zampini ncoarse_ml = all_procs; 52696e683305SStefano Zampini if (void_procs) { 52706e683305SStefano Zampini csin_ds = PETSC_TRUE; 52716e683305SStefano Zampini ncoarse_ds = void_procs; 52726e683305SStefano Zampini csin_type_simple = PETSC_FALSE; 52736e683305SStefano Zampini } else { 5274779c1cceSStefano Zampini if (pcbddc->redistribute_coarse > 0 && pcbddc->redistribute_coarse < all_procs) { 527574e2c79eSStefano Zampini csin_ds = PETSC_TRUE; 527674e2c79eSStefano Zampini ncoarse_ds = pcbddc->redistribute_coarse; 527774e2c79eSStefano Zampini redist = PETSC_TRUE; 527874e2c79eSStefano Zampini } else { 52796e683305SStefano Zampini csin_ds = PETSC_FALSE; 52806e683305SStefano Zampini ncoarse_ds = all_procs; 52816e683305SStefano Zampini } 52826e683305SStefano Zampini } 528374e2c79eSStefano Zampini } 52846e683305SStefano Zampini 52856e683305SStefano Zampini /* 52866e683305SStefano Zampini test if we can go multilevel: three conditions must be satisfied: 52876e683305SStefano Zampini - we have not exceeded the number of levels requested 52886e683305SStefano Zampini - we can actually subassemble the active processes 52896e683305SStefano Zampini - we can find a suitable number of MPI processes where we can place the subassembled problem 52906e683305SStefano Zampini */ 52916e683305SStefano Zampini multilevel_allowed = PETSC_FALSE; 52926e683305SStefano Zampini multilevel_requested = PETSC_FALSE; 52936e683305SStefano Zampini if (pcbddc->current_level < pcbddc->max_levels) { 52946e683305SStefano Zampini multilevel_requested = PETSC_TRUE; 52956e683305SStefano Zampini if (active_procs/pcbddc->coarsening_ratio < 2 || ncoarse_ml/pcbddc->coarsening_ratio < 2) { 5296f9eb5b7dSStefano Zampini multilevel_allowed = PETSC_FALSE; 52972b510759SStefano Zampini } else { 5298f9eb5b7dSStefano Zampini multilevel_allowed = PETSC_TRUE; 5299c8587f34SStefano Zampini } 5300c8587f34SStefano Zampini } 53016e683305SStefano Zampini /* determine number of process partecipating to coarse solver */ 53026e683305SStefano Zampini if (multilevel_allowed) { 53036e683305SStefano Zampini ncoarse = ncoarse_ml; 53046e683305SStefano Zampini csin = csin_ml; 530558da7f69SStefano Zampini redist = PETSC_FALSE; 53066e683305SStefano Zampini } else { 53076e683305SStefano Zampini ncoarse = ncoarse_ds; 53086e683305SStefano Zampini csin = csin_ds; 53096e683305SStefano Zampini } 5310e7931f94SStefano Zampini 5311abbbba34SStefano Zampini /* creates temporary l2gmap and IS for coarse indexes */ 5312abbbba34SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),pcbddc->local_primal_size,pcbddc->global_primal_indices,PETSC_COPY_VALUES,&coarse_is);CHKERRQ(ierr); 5313abbbba34SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(coarse_is,&coarse_islg);CHKERRQ(ierr); 5314abbbba34SStefano Zampini 5315abbbba34SStefano Zampini /* creates temporary MATIS object for coarse matrix */ 531622bc73bbSStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,pcbddc->local_primal_size,pcbddc->local_primal_size,NULL,&coarse_submat_dense);CHKERRQ(ierr); 531722bc73bbSStefano Zampini ierr = MatDenseGetArray(coarse_submat_dense,&array);CHKERRQ(ierr); 531822bc73bbSStefano Zampini ierr = PetscMemcpy(array,coarse_submat_vals,sizeof(*coarse_submat_vals)*pcbddc->local_primal_size*pcbddc->local_primal_size);CHKERRQ(ierr); 531922bc73bbSStefano Zampini ierr = MatDenseRestoreArray(coarse_submat_dense,&array);CHKERRQ(ierr); 5320e176bc59SStefano 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); 53216e683305SStefano Zampini ierr = MatISSetLocalMat(t_coarse_mat_is,coarse_submat_dense);CHKERRQ(ierr); 53226e683305SStefano Zampini ierr = MatAssemblyBegin(t_coarse_mat_is,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 53236e683305SStefano Zampini ierr = MatAssemblyEnd(t_coarse_mat_is,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5324abbbba34SStefano Zampini ierr = MatDestroy(&coarse_submat_dense);CHKERRQ(ierr); 5325abbbba34SStefano Zampini 53266e683305SStefano Zampini /* compute dofs splitting and neumann boundaries for coarse dofs */ 532730368db7SStefano Zampini if (multilevel_allowed && (pcbddc->n_ISForDofsLocal || pcbddc->NeumannBoundariesLocal || (pcbddc->benign_saddle_point && pcbddc->user_primal_vertices_local))) { /* protects from unneded computations */ 53286e683305SStefano Zampini PetscInt *tidxs,*tidxs2,nout,tsize,i; 53296e683305SStefano Zampini const PetscInt *idxs; 53306e683305SStefano Zampini ISLocalToGlobalMapping tmap; 53316e683305SStefano Zampini 53326e683305SStefano Zampini /* create map between primal indices (in local representative ordering) and local primal numbering */ 53330be93d54SStefano Zampini ierr = ISLocalToGlobalMappingCreate(PETSC_COMM_SELF,1,pcbddc->local_primal_size,pcbddc->primal_indices_local_idxs,PETSC_COPY_VALUES,&tmap);CHKERRQ(ierr); 53346e683305SStefano Zampini /* allocate space for temporary storage */ 5335854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->local_primal_size,&tidxs);CHKERRQ(ierr); 5336854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->local_primal_size,&tidxs2);CHKERRQ(ierr); 53376e683305SStefano Zampini /* allocate for IS array */ 53386e683305SStefano Zampini nisdofs = pcbddc->n_ISForDofsLocal; 53396e683305SStefano Zampini nisneu = !!pcbddc->NeumannBoundariesLocal; 534030368db7SStefano Zampini nisvert = 0; 534130368db7SStefano Zampini if (pcbddc->benign_saddle_point && pcbddc->user_primal_vertices_local) { 534230368db7SStefano Zampini nisvert = 1; 534330368db7SStefano Zampini } 534430368db7SStefano Zampini nis = nisdofs + nisneu + nisvert; 5345854ce69bSBarry Smith ierr = PetscMalloc1(nis,&isarray);CHKERRQ(ierr); 53466e683305SStefano Zampini /* dofs splitting */ 53476e683305SStefano Zampini for (i=0;i<nisdofs;i++) { 53486e683305SStefano Zampini /* ierr = ISView(pcbddc->ISForDofsLocal[i],0);CHKERRQ(ierr); */ 53496e683305SStefano Zampini ierr = ISGetLocalSize(pcbddc->ISForDofsLocal[i],&tsize);CHKERRQ(ierr); 53506e683305SStefano Zampini ierr = ISGetIndices(pcbddc->ISForDofsLocal[i],&idxs);CHKERRQ(ierr); 53516e683305SStefano Zampini ierr = ISGlobalToLocalMappingApply(tmap,IS_GTOLM_DROP,tsize,idxs,&nout,tidxs);CHKERRQ(ierr); 53526e683305SStefano Zampini ierr = ISRestoreIndices(pcbddc->ISForDofsLocal[i],&idxs);CHKERRQ(ierr); 53536e683305SStefano Zampini ierr = ISLocalToGlobalMappingApply(coarse_islg,nout,tidxs,tidxs2);CHKERRQ(ierr); 535430368db7SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),nout,tidxs2,PETSC_COPY_VALUES,&isarray[i]);CHKERRQ(ierr); 53556e683305SStefano Zampini /* ierr = ISView(isarray[i],0);CHKERRQ(ierr); */ 53566e683305SStefano Zampini } 53576e683305SStefano Zampini /* neumann boundaries */ 53586e683305SStefano Zampini if (pcbddc->NeumannBoundariesLocal) { 53596e683305SStefano Zampini /* ierr = ISView(pcbddc->NeumannBoundariesLocal,0);CHKERRQ(ierr); */ 53606e683305SStefano Zampini ierr = ISGetLocalSize(pcbddc->NeumannBoundariesLocal,&tsize);CHKERRQ(ierr); 53616e683305SStefano Zampini ierr = ISGetIndices(pcbddc->NeumannBoundariesLocal,&idxs);CHKERRQ(ierr); 53626e683305SStefano Zampini ierr = ISGlobalToLocalMappingApply(tmap,IS_GTOLM_DROP,tsize,idxs,&nout,tidxs);CHKERRQ(ierr); 53636e683305SStefano Zampini ierr = ISRestoreIndices(pcbddc->NeumannBoundariesLocal,&idxs);CHKERRQ(ierr); 53646e683305SStefano Zampini ierr = ISLocalToGlobalMappingApply(coarse_islg,nout,tidxs,tidxs2);CHKERRQ(ierr); 536530368db7SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),nout,tidxs2,PETSC_COPY_VALUES,&isarray[nisdofs]);CHKERRQ(ierr); 53666e683305SStefano Zampini /* ierr = ISView(isarray[nisdofs],0);CHKERRQ(ierr); */ 53676e683305SStefano Zampini } 536830368db7SStefano Zampini /* primal vertices (benign) */ 536930368db7SStefano Zampini if (pcbddc->benign_saddle_point && pcbddc->user_primal_vertices_local) { 537030368db7SStefano Zampini ierr = ISGetLocalSize(pcbddc->user_primal_vertices_local,&tsize);CHKERRQ(ierr); 537130368db7SStefano Zampini ierr = ISGetIndices(pcbddc->user_primal_vertices_local,&idxs);CHKERRQ(ierr); 537230368db7SStefano Zampini ierr = ISGlobalToLocalMappingApply(tmap,IS_GTOLM_DROP,tsize,idxs,&nout,tidxs);CHKERRQ(ierr); 537330368db7SStefano Zampini ierr = ISRestoreIndices(pcbddc->user_primal_vertices_local,&idxs);CHKERRQ(ierr); 537430368db7SStefano Zampini ierr = ISLocalToGlobalMappingApply(coarse_islg,nout,tidxs,tidxs2);CHKERRQ(ierr); 537530368db7SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),nout,tidxs2,PETSC_COPY_VALUES,&isarray[nis-1]);CHKERRQ(ierr); 537630368db7SStefano Zampini } 53776e683305SStefano Zampini /* free memory */ 53786e683305SStefano Zampini ierr = PetscFree(tidxs);CHKERRQ(ierr); 53796e683305SStefano Zampini ierr = PetscFree(tidxs2);CHKERRQ(ierr); 53806e683305SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&tmap);CHKERRQ(ierr); 53816e683305SStefano Zampini } else { 53826e683305SStefano Zampini nis = 0; 53836e683305SStefano Zampini nisdofs = 0; 53846e683305SStefano Zampini nisneu = 0; 538530368db7SStefano Zampini nisvert = 0; 53866e683305SStefano Zampini isarray = NULL; 53876e683305SStefano Zampini } 53886e683305SStefano Zampini /* destroy no longer needed map */ 53896e683305SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&coarse_islg);CHKERRQ(ierr); 53906e683305SStefano Zampini 53916e683305SStefano Zampini /* restrict on coarse candidates (if needed) */ 53926e683305SStefano Zampini coarse_mat_is = NULL; 53936e683305SStefano Zampini if (csin) { 53946e683305SStefano Zampini if (!pcbddc->coarse_subassembling_init ) { /* creates subassembling init pattern if not present */ 539574e2c79eSStefano Zampini if (redist) { 539674e2c79eSStefano Zampini PetscMPIInt rank; 5397779c1cceSStefano Zampini PetscInt spc,n_spc_p1,dest[1],destsize; 539874e2c79eSStefano Zampini 539974e2c79eSStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)pc),&rank);CHKERRQ(ierr); 540058da7f69SStefano Zampini spc = active_procs/ncoarse; 540158da7f69SStefano Zampini n_spc_p1 = active_procs%ncoarse; 5402779c1cceSStefano Zampini if (im_active) { 5403779c1cceSStefano Zampini destsize = 1; 540474e2c79eSStefano Zampini if (rank > n_spc_p1*(spc+1)-1) { 540574e2c79eSStefano Zampini dest[0] = n_spc_p1+(rank-(n_spc_p1*(spc+1)))/spc; 540674e2c79eSStefano Zampini } else { 540774e2c79eSStefano Zampini dest[0] = rank/(spc+1); 540874e2c79eSStefano Zampini } 540974e2c79eSStefano Zampini } else { 5410779c1cceSStefano Zampini destsize = 0; 54116e683305SStefano Zampini } 5412779c1cceSStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),destsize,dest,PETSC_COPY_VALUES,&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 5413779c1cceSStefano Zampini } else if (csin_type_simple) { 54146e683305SStefano Zampini PetscMPIInt rank; 54156e683305SStefano Zampini PetscInt issize,isidx; 5416779c1cceSStefano Zampini 54176e683305SStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)pc),&rank);CHKERRQ(ierr); 54186e683305SStefano Zampini if (im_active) { 54196e683305SStefano Zampini issize = 1; 54206e683305SStefano Zampini isidx = (PetscInt)rank; 54216e683305SStefano Zampini } else { 54226e683305SStefano Zampini issize = 0; 54236e683305SStefano Zampini isidx = -1; 54246e683305SStefano Zampini } 54256e683305SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),issize,&isidx,PETSC_COPY_VALUES,&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 5426779c1cceSStefano Zampini } else { /* get a suitable subassembling pattern from MATIS code */ 5427b0c7d250SStefano Zampini ierr = MatISGetSubassemblingPattern(t_coarse_mat_is,ncoarse,pcbddc->coarse_adj_red,&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 54286e683305SStefano Zampini } 5429779c1cceSStefano Zampini 5430779c1cceSStefano Zampini /* we need to shift on coarse candidates either if we are not redistributing or we are redistributing and we have enough void processes */ 5431779c1cceSStefano Zampini if (!redist || ncoarse <= void_procs) { 5432779c1cceSStefano Zampini PetscInt ncoarse_cand,tissize,*nisindices; 5433779c1cceSStefano Zampini PetscInt *coarse_candidates; 5434779c1cceSStefano Zampini const PetscInt* tisindices; 5435779c1cceSStefano Zampini 5436779c1cceSStefano Zampini /* get coarse candidates' ranks in pc communicator */ 5437779c1cceSStefano Zampini ierr = PetscMalloc1(all_procs,&coarse_candidates);CHKERRQ(ierr); 5438779c1cceSStefano Zampini ierr = MPI_Allgather(&im_active,1,MPIU_INT,coarse_candidates,1,MPIU_INT,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 5439779c1cceSStefano Zampini for (i=0,ncoarse_cand=0;i<all_procs;i++) { 5440779c1cceSStefano Zampini if (!coarse_candidates[i]) { 5441779c1cceSStefano Zampini coarse_candidates[ncoarse_cand++]=i; 5442779c1cceSStefano Zampini } 5443779c1cceSStefano Zampini } 5444779c1cceSStefano Zampini if (ncoarse_cand < ncoarse) SETERRQ2(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"This should not happen! %d < %d",ncoarse_cand,ncoarse); 5445779c1cceSStefano Zampini 5446779c1cceSStefano Zampini 54476e683305SStefano Zampini if (pcbddc->dbg_flag) { 54486e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 54496e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Subassembling pattern init (before shift)\n");CHKERRQ(ierr); 54506e683305SStefano Zampini ierr = ISView(pcbddc->coarse_subassembling_init,pcbddc->dbg_viewer);CHKERRQ(ierr); 54516e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Coarse candidates\n");CHKERRQ(ierr); 5452779c1cceSStefano Zampini for (i=0;i<ncoarse_cand;i++) { 54536e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"%d ",coarse_candidates[i]);CHKERRQ(ierr); 54546e683305SStefano Zampini } 54556e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"\n");CHKERRQ(ierr); 54566e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 54576e683305SStefano Zampini } 54586e683305SStefano Zampini /* shift the pattern on coarse candidates */ 54596e683305SStefano Zampini ierr = ISGetLocalSize(pcbddc->coarse_subassembling_init,&tissize);CHKERRQ(ierr); 54606e683305SStefano Zampini ierr = ISGetIndices(pcbddc->coarse_subassembling_init,&tisindices);CHKERRQ(ierr); 5461854ce69bSBarry Smith ierr = PetscMalloc1(tissize,&nisindices);CHKERRQ(ierr); 54626e683305SStefano Zampini for (i=0;i<tissize;i++) nisindices[i] = coarse_candidates[tisindices[i]]; 54636e683305SStefano Zampini ierr = ISRestoreIndices(pcbddc->coarse_subassembling_init,&tisindices);CHKERRQ(ierr); 54646e683305SStefano Zampini ierr = ISGeneralSetIndices(pcbddc->coarse_subassembling_init,tissize,nisindices,PETSC_OWN_POINTER);CHKERRQ(ierr); 54656e683305SStefano Zampini ierr = PetscFree(coarse_candidates);CHKERRQ(ierr); 54666e683305SStefano Zampini } 54676e683305SStefano Zampini if (pcbddc->dbg_flag) { 54686e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 54696e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Subassembling pattern init\n");CHKERRQ(ierr); 54706e683305SStefano Zampini ierr = ISView(pcbddc->coarse_subassembling_init,pcbddc->dbg_viewer);CHKERRQ(ierr); 54716e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 54726e683305SStefano Zampini } 5473779c1cceSStefano Zampini } 54746e683305SStefano Zampini /* get temporary coarse mat in IS format restricted on coarse procs (plus additional index sets of isarray) */ 547553a05cb3SStefano Zampini if (multilevel_allowed) { /* we need to keep tracking of void processes for future placements */ 547653a05cb3SStefano 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); 547753a05cb3SStefano Zampini } else { /* this is the last level, so use just receiving processes in subcomm */ 547853a05cb3SStefano 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); 547953a05cb3SStefano Zampini } 54806e683305SStefano Zampini } else { 54816e683305SStefano Zampini if (pcbddc->dbg_flag) { 54826e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 54836e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Subassembling pattern init not needed\n");CHKERRQ(ierr); 54846e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 54856e683305SStefano Zampini } 54866e683305SStefano Zampini ierr = PetscObjectReference((PetscObject)t_coarse_mat_is);CHKERRQ(ierr); 54876e683305SStefano Zampini coarse_mat_is = t_coarse_mat_is; 54886e683305SStefano Zampini } 54896e683305SStefano Zampini 54906e683305SStefano Zampini /* create local to global scatters for coarse problem */ 549168457ee5SStefano Zampini if (compute_vecs) { 54926e683305SStefano Zampini PetscInt lrows; 54936e683305SStefano Zampini ierr = VecDestroy(&pcbddc->coarse_vec);CHKERRQ(ierr); 54946e683305SStefano Zampini if (coarse_mat_is) { 54956e683305SStefano Zampini ierr = MatGetLocalSize(coarse_mat_is,&lrows,NULL);CHKERRQ(ierr); 54966e683305SStefano Zampini } else { 54976e683305SStefano Zampini lrows = 0; 54986e683305SStefano Zampini } 54996e683305SStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pc),&pcbddc->coarse_vec);CHKERRQ(ierr); 55006e683305SStefano Zampini ierr = VecSetSizes(pcbddc->coarse_vec,lrows,PETSC_DECIDE);CHKERRQ(ierr); 55016e683305SStefano Zampini ierr = VecSetType(pcbddc->coarse_vec,VECSTANDARD);CHKERRQ(ierr); 55026e683305SStefano Zampini ierr = VecScatterDestroy(&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 55036e683305SStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_P,NULL,pcbddc->coarse_vec,coarse_is,&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 55046e683305SStefano Zampini } 55056e683305SStefano Zampini ierr = ISDestroy(&coarse_is);CHKERRQ(ierr); 55066e683305SStefano Zampini ierr = MatDestroy(&t_coarse_mat_is);CHKERRQ(ierr); 5507c8587f34SStefano Zampini 5508f9eb5b7dSStefano Zampini /* set defaults for coarse KSP and PC */ 5509f9eb5b7dSStefano Zampini if (multilevel_allowed) { 5510f9eb5b7dSStefano Zampini coarse_ksp_type = KSPRICHARDSON; 5511f9eb5b7dSStefano Zampini coarse_pc_type = PCBDDC; 5512f9eb5b7dSStefano Zampini } else { 5513f9eb5b7dSStefano Zampini coarse_ksp_type = KSPPREONLY; 5514f9eb5b7dSStefano Zampini coarse_pc_type = PCREDUNDANT; 5515c8587f34SStefano Zampini } 5516c8587f34SStefano Zampini 55176e683305SStefano Zampini /* print some info if requested */ 55186e683305SStefano Zampini if (pcbddc->dbg_flag) { 55196e683305SStefano Zampini if (!multilevel_allowed) { 55206e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 55216e683305SStefano Zampini if (multilevel_requested) { 55226e683305SStefano 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); 55236e683305SStefano Zampini } else if (pcbddc->max_levels) { 55246e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Maximum number of requested levels reached (%d)\n",pcbddc->max_levels);CHKERRQ(ierr); 55256e683305SStefano Zampini } 55266e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 55276e683305SStefano Zampini } 55286e683305SStefano Zampini } 55296e683305SStefano Zampini 5530f9eb5b7dSStefano Zampini /* create the coarse KSP object only once with defaults */ 55316e683305SStefano Zampini if (coarse_mat_is) { 55326e683305SStefano Zampini MatReuse coarse_mat_reuse; 55336a1308c2SStefano Zampini PetscViewer dbg_viewer = NULL; 55346e683305SStefano Zampini if (pcbddc->dbg_flag) { 55356e683305SStefano Zampini dbg_viewer = PETSC_VIEWER_STDOUT_(PetscObjectComm((PetscObject)coarse_mat_is)); 55366e683305SStefano Zampini ierr = PetscViewerASCIIAddTab(dbg_viewer,2*pcbddc->current_level);CHKERRQ(ierr); 55376e683305SStefano Zampini } 5538f9eb5b7dSStefano Zampini if (!pcbddc->coarse_ksp) { 5539312be037SStefano Zampini char prefix[256],str_level[16]; 5540e604994aSStefano Zampini size_t len; 55416e683305SStefano Zampini ierr = KSPCreate(PetscObjectComm((PetscObject)coarse_mat_is),&pcbddc->coarse_ksp);CHKERRQ(ierr); 5542422a814eSBarry Smith ierr = KSPSetErrorIfNotConverged(pcbddc->coarse_ksp,pc->erroriffailure);CHKERRQ(ierr); 5543c8587f34SStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->coarse_ksp,(PetscObject)pc,1);CHKERRQ(ierr); 5544f9eb5b7dSStefano Zampini ierr = KSPSetTolerances(pcbddc->coarse_ksp,PETSC_DEFAULT,PETSC_DEFAULT,PETSC_DEFAULT,1);CHKERRQ(ierr); 55455f76c7aeSStefano Zampini ierr = KSPSetOperators(pcbddc->coarse_ksp,coarse_mat_is,coarse_mat_is);CHKERRQ(ierr); 5546c8587f34SStefano Zampini ierr = KSPSetType(pcbddc->coarse_ksp,coarse_ksp_type);CHKERRQ(ierr); 55476e683305SStefano Zampini ierr = KSPSetNormType(pcbddc->coarse_ksp,KSP_NORM_NONE);CHKERRQ(ierr); 5548c8587f34SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&pc_temp);CHKERRQ(ierr); 5549c8587f34SStefano Zampini ierr = PCSetType(pc_temp,coarse_pc_type);CHKERRQ(ierr); 5550e604994aSStefano Zampini /* prefix */ 5551e604994aSStefano Zampini ierr = PetscStrcpy(prefix,"");CHKERRQ(ierr); 5552e604994aSStefano Zampini ierr = PetscStrcpy(str_level,"");CHKERRQ(ierr); 5553e604994aSStefano Zampini if (!pcbddc->current_level) { 5554e604994aSStefano Zampini ierr = PetscStrcpy(prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 5555e604994aSStefano Zampini ierr = PetscStrcat(prefix,"pc_bddc_coarse_");CHKERRQ(ierr); 5556c8587f34SStefano Zampini } else { 5557e604994aSStefano Zampini ierr = PetscStrlen(((PetscObject)pc)->prefix,&len);CHKERRQ(ierr); 5558312be037SStefano Zampini if (pcbddc->current_level>1) len -= 3; /* remove "lX_" with X level number */ 5559312be037SStefano Zampini if (pcbddc->current_level>10) len -= 1; /* remove another char from level number */ 556034d6797cSStefano Zampini ierr = PetscStrncpy(prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 5561312be037SStefano Zampini sprintf(str_level,"l%d_",(int)(pcbddc->current_level)); 5562e604994aSStefano Zampini ierr = PetscStrcat(prefix,str_level);CHKERRQ(ierr); 5563e604994aSStefano Zampini } 5564e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->coarse_ksp,prefix);CHKERRQ(ierr); 55653e3c6dadSStefano Zampini /* propagate BDDC info to the next level (these are dummy calls if pc_temp is not of type PCBDDC) */ 55663e3c6dadSStefano Zampini ierr = PCBDDCSetLevel(pc_temp,pcbddc->current_level+1);CHKERRQ(ierr); 55673e3c6dadSStefano Zampini ierr = PCBDDCSetCoarseningRatio(pc_temp,pcbddc->coarsening_ratio);CHKERRQ(ierr); 55683e3c6dadSStefano Zampini ierr = PCBDDCSetLevels(pc_temp,pcbddc->max_levels);CHKERRQ(ierr); 5569f9eb5b7dSStefano Zampini /* allow user customization */ 5570f9eb5b7dSStefano Zampini ierr = KSPSetFromOptions(pcbddc->coarse_ksp);CHKERRQ(ierr); 55713e3c6dadSStefano Zampini } 55723e3c6dadSStefano Zampini /* propagate BDDC info to the next level (these are dummy calls if pc_temp is not of type PCBDDC) */ 557351bea450SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&pc_temp);CHKERRQ(ierr); 55743e3c6dadSStefano Zampini if (nisdofs) { 55753e3c6dadSStefano Zampini ierr = PCBDDCSetDofsSplitting(pc_temp,nisdofs,isarray);CHKERRQ(ierr); 55763e3c6dadSStefano Zampini for (i=0;i<nisdofs;i++) { 55773e3c6dadSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 55783e3c6dadSStefano Zampini } 55793e3c6dadSStefano Zampini } 55803e3c6dadSStefano Zampini if (nisneu) { 55813e3c6dadSStefano Zampini ierr = PCBDDCSetNeumannBoundaries(pc_temp,isarray[nisdofs]);CHKERRQ(ierr); 55823e3c6dadSStefano Zampini ierr = ISDestroy(&isarray[nisdofs]);CHKERRQ(ierr); 5583312be037SStefano Zampini } 558430368db7SStefano Zampini if (nisvert) { 558530368db7SStefano Zampini ierr = PCBDDCSetPrimalVerticesIS(pc_temp,isarray[nis-1]);CHKERRQ(ierr); 558630368db7SStefano Zampini ierr = ISDestroy(&isarray[nis-1]);CHKERRQ(ierr); 558730368db7SStefano Zampini } 5588f9eb5b7dSStefano Zampini 5589f9eb5b7dSStefano Zampini /* get some info after set from options */ 5590f9eb5b7dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCNN,&isnn);CHKERRQ(ierr); 5591f9eb5b7dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCBDDC,&isbddc);CHKERRQ(ierr); 55924f3a063dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCREDUNDANT,&isredundant);CHKERRQ(ierr); 55936e683305SStefano Zampini if (isbddc && !multilevel_allowed) { /* multilevel can only be requested via pc_bddc_set_levels */ 5594f9eb5b7dSStefano Zampini ierr = PCSetType(pc_temp,coarse_pc_type);CHKERRQ(ierr); 5595f9eb5b7dSStefano Zampini isbddc = PETSC_FALSE; 5596f9eb5b7dSStefano Zampini } 559739f0f02cSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 55984f3a063dSStefano Zampini if (isredundant) { 55994f3a063dSStefano Zampini KSP inner_ksp; 56004f3a063dSStefano Zampini PC inner_pc; 56014f3a063dSStefano Zampini ierr = PCRedundantGetKSP(pc_temp,&inner_ksp);CHKERRQ(ierr); 56024f3a063dSStefano Zampini ierr = KSPGetPC(inner_ksp,&inner_pc);CHKERRQ(ierr); 56034f3a063dSStefano Zampini ierr = PCFactorSetReuseFill(inner_pc,PETSC_TRUE);CHKERRQ(ierr); 56044f3a063dSStefano Zampini } 5605f9eb5b7dSStefano Zampini 5606f9eb5b7dSStefano Zampini /* assemble coarse matrix */ 5607fa7f1dd8SStefano Zampini if (coarse_reuse) { 560881d9aea3SBarry Smith ierr = KSPGetOperators(pcbddc->coarse_ksp,&coarse_mat,NULL);CHKERRQ(ierr); 5609fa7f1dd8SStefano Zampini ierr = PetscObjectReference((PetscObject)coarse_mat);CHKERRQ(ierr); 56106e683305SStefano Zampini coarse_mat_reuse = MAT_REUSE_MATRIX; 5611fa7f1dd8SStefano Zampini } else { 56126e683305SStefano Zampini coarse_mat_reuse = MAT_INITIAL_MATRIX; 5613fa7f1dd8SStefano Zampini } 5614c8587f34SStefano Zampini if (isbddc || isnn) { 5615d4d8cf7bSStefano Zampini if (isbddc) { /* currently there's no API for this */ 5616d4d8cf7bSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc_temp->data; 5617d4d8cf7bSStefano Zampini pcbddc->detect_disconnected = PETSC_TRUE; 5618d4d8cf7bSStefano Zampini } 561922bc73bbSStefano Zampini if (pcbddc->coarsening_ratio > 1) { 562070cf5478SStefano Zampini if (!pcbddc->coarse_subassembling) { /* subassembling info is not present */ 5621b0c7d250SStefano Zampini ierr = MatISGetSubassemblingPattern(coarse_mat_is,active_procs/pcbddc->coarsening_ratio,pcbddc->coarse_adj_red,&pcbddc->coarse_subassembling);CHKERRQ(ierr); 562222b6e8a2SStefano Zampini if (pcbddc->dbg_flag) { 56236e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 56246e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Subassembling pattern\n");CHKERRQ(ierr); 56256e683305SStefano Zampini ierr = ISView(pcbddc->coarse_subassembling,dbg_viewer);CHKERRQ(ierr); 56266e683305SStefano Zampini ierr = PetscViewerFlush(dbg_viewer);CHKERRQ(ierr); 562722b6e8a2SStefano Zampini } 562870cf5478SStefano Zampini } 562953a05cb3SStefano Zampini ierr = MatISSubassemble(coarse_mat_is,pcbddc->coarse_subassembling,0,PETSC_FALSE,PETSC_FALSE,coarse_mat_reuse,&coarse_mat,0,NULL);CHKERRQ(ierr); 563070cf5478SStefano Zampini } else { 563122bc73bbSStefano Zampini ierr = PetscObjectReference((PetscObject)coarse_mat_is);CHKERRQ(ierr); 563222bc73bbSStefano Zampini coarse_mat = coarse_mat_is; 563322bc73bbSStefano Zampini } 563422bc73bbSStefano Zampini } else { 56352e1e5fa4SStefano Zampini ierr = MatISGetMPIXAIJ(coarse_mat_is,coarse_mat_reuse,&coarse_mat);CHKERRQ(ierr); 5636c8587f34SStefano Zampini } 5637c8587f34SStefano Zampini ierr = MatDestroy(&coarse_mat_is);CHKERRQ(ierr); 5638c8587f34SStefano Zampini 56393301b35fSStefano Zampini /* propagate symmetry info of coarse matrix */ 56405a16e3a0SStefano Zampini ierr = MatSetOption(coarse_mat,MAT_STRUCTURALLY_SYMMETRIC,PETSC_TRUE);CHKERRQ(ierr); 56413301b35fSStefano Zampini if (pc->pmat->symmetric_set) { 56423301b35fSStefano Zampini ierr = MatSetOption(coarse_mat,MAT_SYMMETRIC,pc->pmat->symmetric);CHKERRQ(ierr); 56433301b35fSStefano Zampini } 56443301b35fSStefano Zampini if (pc->pmat->hermitian_set) { 56453301b35fSStefano Zampini ierr = MatSetOption(coarse_mat,MAT_HERMITIAN,pc->pmat->hermitian);CHKERRQ(ierr); 56463301b35fSStefano Zampini } 56473301b35fSStefano Zampini if (pc->pmat->spd_set) { 56483301b35fSStefano Zampini ierr = MatSetOption(coarse_mat,MAT_SPD,pc->pmat->spd);CHKERRQ(ierr); 56493301b35fSStefano Zampini } 56506e683305SStefano Zampini /* set operators */ 56515f76c7aeSStefano Zampini ierr = KSPSetOperators(pcbddc->coarse_ksp,coarse_mat,coarse_mat);CHKERRQ(ierr); 56526e683305SStefano Zampini if (pcbddc->dbg_flag) { 56536e683305SStefano Zampini ierr = PetscViewerASCIISubtractTab(dbg_viewer,2*pcbddc->current_level);CHKERRQ(ierr); 56546e683305SStefano Zampini } 56556e683305SStefano Zampini } else { /* processes non partecipating to coarse solver (if any) */ 56566e683305SStefano Zampini coarse_mat = 0; 56576e683305SStefano Zampini } 56586e683305SStefano Zampini ierr = PetscFree(isarray);CHKERRQ(ierr); 5659b1ecc7b1SStefano Zampini #if 0 5660b9b85e73SStefano Zampini { 5661b9b85e73SStefano Zampini PetscViewer viewer; 5662b9b85e73SStefano Zampini char filename[256]; 5663b1ecc7b1SStefano Zampini sprintf(filename,"coarse_mat_level%d.m",pcbddc->current_level); 5664b1ecc7b1SStefano Zampini ierr = PetscViewerASCIIOpen(PetscObjectComm((PetscObject)coarse_mat),filename,&viewer);CHKERRQ(ierr); 5665b9b85e73SStefano Zampini ierr = PetscViewerSetFormat(viewer,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 5666b9b85e73SStefano Zampini ierr = MatView(coarse_mat,viewer);CHKERRQ(ierr); 5667b9b85e73SStefano Zampini ierr = PetscViewerDestroy(&viewer);CHKERRQ(ierr); 5668b9b85e73SStefano Zampini } 5669b9b85e73SStefano Zampini #endif 5670c8587f34SStefano Zampini 5671c8587f34SStefano Zampini /* Compute coarse null space (special handling by BDDC only) */ 5672298c0119SStefano Zampini #if 0 5673c8587f34SStefano Zampini if (pcbddc->NullSpace) { 5674c8587f34SStefano Zampini ierr = PCBDDCNullSpaceAssembleCoarse(pc,coarse_mat,&CoarseNullSpace);CHKERRQ(ierr); 567598a51de6SStefano Zampini } 5676298c0119SStefano Zampini #endif 5677b0f5fe93SStefano Zampini /* hack */ 567898a51de6SStefano Zampini if (pcbddc->coarse_ksp) { 567998a51de6SStefano Zampini Vec crhs,csol; 568004708bb6SStefano Zampini 5681f347579bSStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&csol);CHKERRQ(ierr); 5682f347579bSStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&crhs);CHKERRQ(ierr); 5683f347579bSStefano Zampini if (!csol) { 56842a7a6963SBarry Smith ierr = MatCreateVecs(coarse_mat,&((pcbddc->coarse_ksp)->vec_sol),NULL);CHKERRQ(ierr); 5685f9eb5b7dSStefano Zampini } 5686f347579bSStefano Zampini if (!crhs) { 56872a7a6963SBarry Smith ierr = MatCreateVecs(coarse_mat,NULL,&((pcbddc->coarse_ksp)->vec_rhs));CHKERRQ(ierr); 5688f347579bSStefano Zampini } 5689b0f5fe93SStefano Zampini } 5690b0f5fe93SStefano Zampini 5691b0f5fe93SStefano Zampini /* compute null space for coarse solver if the benign trick has been requested */ 5692b0f5fe93SStefano Zampini if (pcbddc->benign_null) { 5693b0f5fe93SStefano Zampini 5694b0f5fe93SStefano Zampini ierr = VecSet(pcbddc->vec1_P,0.);CHKERRQ(ierr); 56954f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 56964f1b2e48SStefano Zampini ierr = VecSetValue(pcbddc->vec1_P,pcbddc->local_primal_size-pcbddc->benign_n+i,1.0,INSERT_VALUES);CHKERRQ(ierr); 56974f1b2e48SStefano Zampini } 5698b0f5fe93SStefano Zampini ierr = VecAssemblyBegin(pcbddc->vec1_P);CHKERRQ(ierr); 5699b0f5fe93SStefano Zampini ierr = VecAssemblyEnd(pcbddc->vec1_P);CHKERRQ(ierr); 5700b0f5fe93SStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 5701b0f5fe93SStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 5702b0f5fe93SStefano Zampini if (coarse_mat) { 5703b0f5fe93SStefano Zampini Vec nullv; 5704b0f5fe93SStefano Zampini PetscScalar *array,*array2; 5705b0f5fe93SStefano Zampini PetscInt nl; 5706b0f5fe93SStefano Zampini 5707b0f5fe93SStefano Zampini ierr = MatCreateVecs(coarse_mat,&nullv,NULL);CHKERRQ(ierr); 5708b0f5fe93SStefano Zampini ierr = VecGetLocalSize(nullv,&nl);CHKERRQ(ierr); 5709b0f5fe93SStefano Zampini ierr = VecGetArrayRead(pcbddc->coarse_vec,(const PetscScalar**)&array);CHKERRQ(ierr); 5710b0f5fe93SStefano Zampini ierr = VecGetArray(nullv,&array2);CHKERRQ(ierr); 5711b0f5fe93SStefano Zampini ierr = PetscMemcpy(array2,array,nl*sizeof(*array));CHKERRQ(ierr); 5712b0f5fe93SStefano Zampini ierr = VecRestoreArray(nullv,&array2);CHKERRQ(ierr); 5713b0f5fe93SStefano Zampini ierr = VecRestoreArrayRead(pcbddc->coarse_vec,(const PetscScalar**)&array);CHKERRQ(ierr); 5714b0f5fe93SStefano Zampini ierr = VecNormalize(nullv,NULL);CHKERRQ(ierr); 5715b0f5fe93SStefano Zampini ierr = MatNullSpaceCreate(PetscObjectComm((PetscObject)coarse_mat),PETSC_FALSE,1,&nullv,&CoarseNullSpace);CHKERRQ(ierr); 5716b0f5fe93SStefano Zampini ierr = VecDestroy(&nullv);CHKERRQ(ierr); 5717b0f5fe93SStefano Zampini } 5718b0f5fe93SStefano Zampini } 5719b0f5fe93SStefano Zampini 5720b0f5fe93SStefano Zampini if (pcbddc->coarse_ksp) { 5721b0f5fe93SStefano Zampini PetscBool ispreonly; 5722b0f5fe93SStefano Zampini 5723b0f5fe93SStefano Zampini if (CoarseNullSpace) { 5724b0f5fe93SStefano Zampini PetscBool isnull; 5725b0f5fe93SStefano Zampini ierr = MatNullSpaceTest(CoarseNullSpace,coarse_mat,&isnull);CHKERRQ(ierr); 5726d4d8cf7bSStefano Zampini if (0) { 572730368db7SStefano Zampini if (isbddc && !pcbddc->benign_saddle_point) { 5728b0f5fe93SStefano Zampini ierr = PCBDDCSetNullSpace(pc_temp,CoarseNullSpace);CHKERRQ(ierr); 5729b0f5fe93SStefano Zampini } else { 5730b0f5fe93SStefano Zampini ierr = MatSetNullSpace(coarse_mat,CoarseNullSpace);CHKERRQ(ierr); 5731b0f5fe93SStefano Zampini } 5732b0f5fe93SStefano Zampini } else { 5733b0f5fe93SStefano Zampini ierr = MatNullSpaceDestroy(&CoarseNullSpace);CHKERRQ(ierr); 5734b0f5fe93SStefano Zampini } 5735b0f5fe93SStefano Zampini } 5736b0f5fe93SStefano Zampini /* setup coarse ksp */ 5737b0f5fe93SStefano Zampini ierr = KSPSetUp(pcbddc->coarse_ksp);CHKERRQ(ierr); 5738cbcc2c2aSStefano Zampini /* Check coarse problem if in debug mode or if solving with an iterative method */ 5739cbcc2c2aSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcbddc->coarse_ksp,KSPPREONLY,&ispreonly);CHKERRQ(ierr); 57406e683305SStefano Zampini if (pcbddc->dbg_flag || (!ispreonly && pcbddc->use_coarse_estimates) ) { 5741c8587f34SStefano Zampini KSP check_ksp; 57422b510759SStefano Zampini KSPType check_ksp_type; 5743c8587f34SStefano Zampini PC check_pc; 57446e683305SStefano Zampini Vec check_vec,coarse_vec; 57456a1308c2SStefano Zampini PetscReal abs_infty_error,infty_error,lambda_min=1.0,lambda_max=1.0; 57462b510759SStefano Zampini PetscInt its; 57476e683305SStefano Zampini PetscBool compute_eigs; 57486e683305SStefano Zampini PetscReal *eigs_r,*eigs_c; 57496e683305SStefano Zampini PetscInt neigs; 57508e185a42SStefano Zampini const char *prefix; 5751c8587f34SStefano Zampini 57522b510759SStefano Zampini /* Create ksp object suitable for estimation of extreme eigenvalues */ 57536e683305SStefano Zampini ierr = KSPCreate(PetscObjectComm((PetscObject)pcbddc->coarse_ksp),&check_ksp);CHKERRQ(ierr); 5754422a814eSBarry Smith ierr = KSPSetErrorIfNotConverged(pcbddc->coarse_ksp,pc->erroriffailure);CHKERRQ(ierr); 575523ee1639SBarry Smith ierr = KSPSetOperators(check_ksp,coarse_mat,coarse_mat);CHKERRQ(ierr); 5756f4ddd8eeSStefano Zampini ierr = KSPSetTolerances(check_ksp,1.e-12,1.e-12,PETSC_DEFAULT,pcbddc->coarse_size);CHKERRQ(ierr); 57572b510759SStefano Zampini if (ispreonly) { 57582b510759SStefano Zampini check_ksp_type = KSPPREONLY; 57596e683305SStefano Zampini compute_eigs = PETSC_FALSE; 57602b510759SStefano Zampini } else { 5761cbcc2c2aSStefano Zampini check_ksp_type = KSPGMRES; 57626e683305SStefano Zampini compute_eigs = PETSC_TRUE; 5763c8587f34SStefano Zampini } 5764c8587f34SStefano Zampini ierr = KSPSetType(check_ksp,check_ksp_type);CHKERRQ(ierr); 57656e683305SStefano Zampini ierr = KSPSetComputeSingularValues(check_ksp,compute_eigs);CHKERRQ(ierr); 57666e683305SStefano Zampini ierr = KSPSetComputeEigenvalues(check_ksp,compute_eigs);CHKERRQ(ierr); 57676e683305SStefano Zampini ierr = KSPGMRESSetRestart(check_ksp,pcbddc->coarse_size+1);CHKERRQ(ierr); 5768a7dc3881SStefano Zampini ierr = KSPGetOptionsPrefix(pcbddc->coarse_ksp,&prefix);CHKERRQ(ierr); 5769a7dc3881SStefano Zampini ierr = KSPSetOptionsPrefix(check_ksp,prefix);CHKERRQ(ierr); 5770a7dc3881SStefano Zampini ierr = KSPAppendOptionsPrefix(check_ksp,"check_");CHKERRQ(ierr); 5771a7dc3881SStefano Zampini ierr = KSPSetFromOptions(check_ksp);CHKERRQ(ierr); 5772c8587f34SStefano Zampini ierr = KSPSetUp(check_ksp);CHKERRQ(ierr); 5773c8587f34SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&check_pc);CHKERRQ(ierr); 5774c8587f34SStefano Zampini ierr = KSPSetPC(check_ksp,check_pc);CHKERRQ(ierr); 5775c8587f34SStefano Zampini /* create random vec */ 57766e683305SStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&coarse_vec);CHKERRQ(ierr); 57776e683305SStefano Zampini ierr = VecDuplicate(coarse_vec,&check_vec);CHKERRQ(ierr); 5778c8587f34SStefano Zampini ierr = VecSetRandom(check_vec,NULL);CHKERRQ(ierr); 5779c8587f34SStefano Zampini if (CoarseNullSpace) { 5780c8587f34SStefano Zampini ierr = MatNullSpaceRemove(CoarseNullSpace,check_vec);CHKERRQ(ierr); 5781c8587f34SStefano Zampini } 57826e683305SStefano Zampini ierr = MatMult(coarse_mat,check_vec,coarse_vec);CHKERRQ(ierr); 5783c8587f34SStefano Zampini /* solve coarse problem */ 57846e683305SStefano Zampini ierr = KSPSolve(check_ksp,coarse_vec,coarse_vec);CHKERRQ(ierr); 5785c8587f34SStefano Zampini if (CoarseNullSpace) { 57866e683305SStefano Zampini ierr = MatNullSpaceRemove(CoarseNullSpace,coarse_vec);CHKERRQ(ierr); 5787c8587f34SStefano Zampini } 5788cbcc2c2aSStefano Zampini /* set eigenvalue estimation if preonly has not been requested */ 57896e683305SStefano Zampini if (compute_eigs) { 5790854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->coarse_size+1,&eigs_r);CHKERRQ(ierr); 5791854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->coarse_size+1,&eigs_c);CHKERRQ(ierr); 57926e683305SStefano Zampini ierr = KSPComputeEigenvalues(check_ksp,pcbddc->coarse_size+1,eigs_r,eigs_c,&neigs);CHKERRQ(ierr); 57936e683305SStefano Zampini lambda_max = eigs_r[neigs-1]; 57946e683305SStefano Zampini lambda_min = eigs_r[0]; 57956e683305SStefano Zampini if (pcbddc->use_coarse_estimates) { 57966e683305SStefano Zampini if (lambda_max>lambda_min) { 5797cbcc2c2aSStefano Zampini ierr = KSPChebyshevSetEigenvalues(pcbddc->coarse_ksp,lambda_max,lambda_min);CHKERRQ(ierr); 5798cbcc2c2aSStefano Zampini ierr = KSPRichardsonSetScale(pcbddc->coarse_ksp,2.0/(lambda_max+lambda_min));CHKERRQ(ierr); 5799cbcc2c2aSStefano Zampini } 5800c8587f34SStefano Zampini } 5801c8587f34SStefano Zampini } 5802cbcc2c2aSStefano Zampini 5803c8587f34SStefano Zampini /* check coarse problem residual error */ 58046e683305SStefano Zampini if (pcbddc->dbg_flag) { 58056e683305SStefano Zampini PetscViewer dbg_viewer = PETSC_VIEWER_STDOUT_(PetscObjectComm((PetscObject)pcbddc->coarse_ksp)); 58066e683305SStefano Zampini ierr = PetscViewerASCIIAddTab(dbg_viewer,2*(pcbddc->current_level+1));CHKERRQ(ierr); 58076e683305SStefano Zampini ierr = VecAXPY(check_vec,-1.0,coarse_vec);CHKERRQ(ierr); 5808c8587f34SStefano Zampini ierr = VecNorm(check_vec,NORM_INFINITY,&infty_error);CHKERRQ(ierr); 58096e683305SStefano Zampini ierr = MatMult(coarse_mat,check_vec,coarse_vec);CHKERRQ(ierr); 58106e683305SStefano Zampini ierr = VecNorm(coarse_vec,NORM_INFINITY,&abs_infty_error);CHKERRQ(ierr); 5811c8587f34SStefano Zampini ierr = VecDestroy(&check_vec);CHKERRQ(ierr); 5812779c1cceSStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem details (use estimates %d)\n",pcbddc->use_coarse_estimates);CHKERRQ(ierr); 58136e683305SStefano Zampini ierr = PetscObjectPrintClassNamePrefixType((PetscObject)(pcbddc->coarse_ksp),dbg_viewer);CHKERRQ(ierr); 58146e683305SStefano Zampini ierr = PetscObjectPrintClassNamePrefixType((PetscObject)(check_pc),dbg_viewer);CHKERRQ(ierr); 58156e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem exact infty_error : %1.6e\n",infty_error);CHKERRQ(ierr); 58166e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem residual infty_error: %1.6e\n",abs_infty_error);CHKERRQ(ierr); 5817b0f5fe93SStefano Zampini if (CoarseNullSpace) { 5818b0f5fe93SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem is singular\n");CHKERRQ(ierr); 5819b0f5fe93SStefano Zampini } 58206e683305SStefano Zampini if (compute_eigs) { 58216e683305SStefano Zampini PetscReal lambda_max_s,lambda_min_s; 5822deec49d1SStefano Zampini ierr = KSPGetType(check_ksp,&check_ksp_type);CHKERRQ(ierr); 5823c8587f34SStefano Zampini ierr = KSPGetIterationNumber(check_ksp,&its);CHKERRQ(ierr); 58246e683305SStefano Zampini ierr = KSPComputeExtremeSingularValues(check_ksp,&lambda_max_s,&lambda_min_s);CHKERRQ(ierr); 58256e683305SStefano 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); 58266e683305SStefano Zampini for (i=0;i<neigs;i++) { 58276e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"%1.6e %1.6ei\n",eigs_r[i],eigs_c[i]);CHKERRQ(ierr); 5828c8587f34SStefano Zampini } 58296e683305SStefano Zampini } 58306e683305SStefano Zampini ierr = PetscViewerFlush(dbg_viewer);CHKERRQ(ierr); 58316e683305SStefano Zampini ierr = PetscViewerASCIISubtractTab(dbg_viewer,2*(pcbddc->current_level+1));CHKERRQ(ierr); 58326e683305SStefano Zampini } 5833c8587f34SStefano Zampini ierr = KSPDestroy(&check_ksp);CHKERRQ(ierr); 58346e683305SStefano Zampini if (compute_eigs) { 58356e683305SStefano Zampini ierr = PetscFree(eigs_r);CHKERRQ(ierr); 58366e683305SStefano Zampini ierr = PetscFree(eigs_c);CHKERRQ(ierr); 5837c8587f34SStefano Zampini } 58386e683305SStefano Zampini } 58396e683305SStefano Zampini } 5840cbcc2c2aSStefano Zampini /* print additional info */ 5841cbcc2c2aSStefano Zampini if (pcbddc->dbg_flag) { 58426e683305SStefano Zampini /* waits until all processes reaches this point */ 58436e683305SStefano Zampini ierr = PetscBarrier((PetscObject)pc);CHKERRQ(ierr); 5844cbcc2c2aSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Coarse solver setup completed at level %d\n",pcbddc->current_level);CHKERRQ(ierr); 5845cbcc2c2aSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 5846cbcc2c2aSStefano Zampini } 5847cbcc2c2aSStefano Zampini 58482b510759SStefano Zampini /* free memory */ 5849c8587f34SStefano Zampini ierr = MatNullSpaceDestroy(&CoarseNullSpace);CHKERRQ(ierr); 5850fdc635d7SStefano Zampini ierr = MatDestroy(&coarse_mat);CHKERRQ(ierr); 5851c8587f34SStefano Zampini PetscFunctionReturn(0); 5852c8587f34SStefano Zampini } 5853674ae819SStefano Zampini 5854f34684f1SStefano Zampini #undef __FUNCT__ 5855f34684f1SStefano Zampini #define __FUNCT__ "PCBDDCComputePrimalNumbering" 5856f34684f1SStefano Zampini PetscErrorCode PCBDDCComputePrimalNumbering(PC pc,PetscInt* coarse_size_n,PetscInt** local_primal_indices_n) 5857f34684f1SStefano Zampini { 5858f34684f1SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 5859f34684f1SStefano Zampini PC_IS* pcis = (PC_IS*)pc->data; 5860f34684f1SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 5861dc456d91SStefano Zampini IS subset,subset_mult,subset_n; 5862dc456d91SStefano Zampini PetscInt local_size,coarse_size=0; 586373be2a3aSStefano Zampini PetscInt *local_primal_indices=NULL; 5864dc456d91SStefano Zampini const PetscInt *t_local_primal_indices; 5865f34684f1SStefano Zampini PetscErrorCode ierr; 5866f34684f1SStefano Zampini 5867f34684f1SStefano Zampini PetscFunctionBegin; 5868f34684f1SStefano Zampini /* Compute global number of coarse dofs */ 58690e6343abSStefano Zampini if (pcbddc->local_primal_size && !pcbddc->local_primal_ref_node) { 58700e6343abSStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"BDDC ConstraintsSetUp should be called first"); 5871727cdba6SStefano Zampini } 5872dc456d91SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)(pc->pmat)),pcbddc->local_primal_size_cc,pcbddc->local_primal_ref_node,PETSC_COPY_VALUES,&subset_n);CHKERRQ(ierr); 58733bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(pcis->mapping,subset_n,&subset);CHKERRQ(ierr); 5874dc456d91SStefano Zampini ierr = ISDestroy(&subset_n);CHKERRQ(ierr); 5875dc456d91SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)(pc->pmat)),pcbddc->local_primal_size_cc,pcbddc->local_primal_ref_mult,PETSC_COPY_VALUES,&subset_mult);CHKERRQ(ierr); 5876dc456d91SStefano Zampini ierr = PCBDDCSubsetNumbering(subset,subset_mult,&coarse_size,&subset_n);CHKERRQ(ierr); 5877dc456d91SStefano Zampini ierr = ISDestroy(&subset);CHKERRQ(ierr); 5878dc456d91SStefano Zampini ierr = ISDestroy(&subset_mult);CHKERRQ(ierr); 5879dc456d91SStefano Zampini ierr = ISGetLocalSize(subset_n,&local_size);CHKERRQ(ierr); 58800e6343abSStefano Zampini if (local_size != pcbddc->local_primal_size) { 58810e6343abSStefano Zampini SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Invalid number of local primal indices computed %d != %d",local_size,pcbddc->local_primal_size); 58820e6343abSStefano Zampini } 5883dc456d91SStefano Zampini ierr = PetscMalloc1(local_size,&local_primal_indices);CHKERRQ(ierr); 5884dc456d91SStefano Zampini ierr = ISGetIndices(subset_n,&t_local_primal_indices);CHKERRQ(ierr); 5885dc456d91SStefano Zampini ierr = PetscMemcpy(local_primal_indices,t_local_primal_indices,local_size*sizeof(PetscInt));CHKERRQ(ierr); 5886dc456d91SStefano Zampini ierr = ISRestoreIndices(subset_n,&t_local_primal_indices);CHKERRQ(ierr); 5887dc456d91SStefano Zampini ierr = ISDestroy(&subset_n);CHKERRQ(ierr); 5888f34684f1SStefano Zampini 5889f34684f1SStefano Zampini /* check numbering */ 5890f34684f1SStefano Zampini if (pcbddc->dbg_flag) { 5891019a44ceSStefano Zampini PetscScalar coarsesum,*array,*array2; 5892dc456d91SStefano Zampini PetscInt i; 5893b9b85e73SStefano Zampini PetscBool set_error = PETSC_FALSE,set_error_reduced = PETSC_FALSE; 5894f34684f1SStefano Zampini 5895f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 5896f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 5897f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Check coarse indices\n");CHKERRQ(ierr); 58981575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 5899019a44ceSStefano Zampini /* counter */ 5900019a44ceSStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 5901019a44ceSStefano Zampini ierr = VecSet(pcis->vec1_N,1.0);CHKERRQ(ierr); 5902019a44ceSStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 5903019a44ceSStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 5904019a44ceSStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_global,pcis->vec2_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 5905019a44ceSStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_global,pcis->vec2_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 5906f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_N,0.0);CHKERRQ(ierr); 5907f34684f1SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 5908727cdba6SStefano Zampini ierr = VecSetValue(pcis->vec1_N,pcbddc->primal_indices_local_idxs[i],1.0,INSERT_VALUES);CHKERRQ(ierr); 5909f34684f1SStefano Zampini } 5910f34684f1SStefano Zampini ierr = VecAssemblyBegin(pcis->vec1_N);CHKERRQ(ierr); 5911f34684f1SStefano Zampini ierr = VecAssemblyEnd(pcis->vec1_N);CHKERRQ(ierr); 5912f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 5913e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 5914e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 5915e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_global,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 5916e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_global,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 5917f34684f1SStefano Zampini ierr = VecGetArray(pcis->vec1_N,&array);CHKERRQ(ierr); 5918019a44ceSStefano Zampini ierr = VecGetArray(pcis->vec2_N,&array2);CHKERRQ(ierr); 5919f34684f1SStefano Zampini for (i=0;i<pcis->n;i++) { 5920019a44ceSStefano Zampini if (array[i] != 0.0 && array[i] != array2[i]) { 592175c01103SStefano Zampini PetscInt owned = (PetscInt)PetscRealPart(array[i]); 592275c01103SStefano Zampini PetscInt neigh = (PetscInt)PetscRealPart(array2[i]); 5923b9b85e73SStefano Zampini set_error = PETSC_TRUE; 5924019a44ceSStefano 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); 5925f34684f1SStefano Zampini } 5926f34684f1SStefano Zampini } 5927019a44ceSStefano Zampini ierr = VecRestoreArray(pcis->vec2_N,&array2);CHKERRQ(ierr); 5928b9b85e73SStefano Zampini ierr = MPI_Allreduce(&set_error,&set_error_reduced,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 5929f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 5930f34684f1SStefano Zampini for (i=0;i<pcis->n;i++) { 5931f34684f1SStefano Zampini if (PetscRealPart(array[i]) > 0.0) array[i] = 1.0/PetscRealPart(array[i]); 5932f34684f1SStefano Zampini } 5933f34684f1SStefano Zampini ierr = VecRestoreArray(pcis->vec1_N,&array);CHKERRQ(ierr); 5934f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 5935e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 5936e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 5937f34684f1SStefano Zampini ierr = VecSum(pcis->vec1_global,&coarsesum);CHKERRQ(ierr); 5938f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Size of coarse problem is %d (%lf)\n",coarse_size,PetscRealPart(coarsesum));CHKERRQ(ierr); 5939b9b85e73SStefano Zampini if (pcbddc->dbg_flag > 1 || set_error_reduced) { 5940ca8b9ea9SStefano Zampini PetscInt *gidxs; 5941ca8b9ea9SStefano Zampini 5942ca8b9ea9SStefano Zampini ierr = PetscMalloc1(pcbddc->local_primal_size,&gidxs);CHKERRQ(ierr); 59433bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingApply(pcis->mapping,pcbddc->local_primal_size,pcbddc->primal_indices_local_idxs,gidxs);CHKERRQ(ierr); 5944f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Distribution of local primal indices\n");CHKERRQ(ierr); 5945f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 5946f34684f1SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d\n",PetscGlobalRank);CHKERRQ(ierr); 5947f34684f1SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 59484bc2dc4bSStefano 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); 5949f34684f1SStefano Zampini } 5950f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 5951ca8b9ea9SStefano Zampini ierr = PetscFree(gidxs);CHKERRQ(ierr); 5952f34684f1SStefano Zampini } 5953f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 59541575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 5955302440fdSBarry Smith if (set_error_reduced) SETERRQ(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"BDDC Numbering of coarse dofs failed"); 5956f34684f1SStefano Zampini } 59578bec7fa6SStefano Zampini /* ierr = PetscPrintf(PetscObjectComm((PetscObject)pc),"Size of coarse problem is %d\n",coarse_size);CHKERRQ(ierr); */ 5958f34684f1SStefano Zampini /* get back data */ 5959f34684f1SStefano Zampini *coarse_size_n = coarse_size; 5960f34684f1SStefano Zampini *local_primal_indices_n = local_primal_indices; 5961674ae819SStefano Zampini PetscFunctionReturn(0); 5962674ae819SStefano Zampini } 5963674ae819SStefano Zampini 5964e456f2a8SStefano Zampini #undef __FUNCT__ 5965e456f2a8SStefano Zampini #define __FUNCT__ "PCBDDCGlobalToLocal" 5966a7dc3881SStefano Zampini PetscErrorCode PCBDDCGlobalToLocal(VecScatter g2l_ctx,Vec gwork, Vec lwork, IS globalis, IS* localis) 5967e456f2a8SStefano Zampini { 5968e456f2a8SStefano Zampini IS localis_t; 5969a7dc3881SStefano Zampini PetscInt i,lsize,*idxs,n; 5970e456f2a8SStefano Zampini PetscScalar *vals; 5971e456f2a8SStefano Zampini PetscErrorCode ierr; 5972e456f2a8SStefano Zampini 5973e456f2a8SStefano Zampini PetscFunctionBegin; 5974a7dc3881SStefano Zampini /* get indices in local ordering exploiting local to global map */ 5975e456f2a8SStefano Zampini ierr = ISGetLocalSize(globalis,&lsize);CHKERRQ(ierr); 5976854ce69bSBarry Smith ierr = PetscMalloc1(lsize,&vals);CHKERRQ(ierr); 5977e456f2a8SStefano Zampini for (i=0;i<lsize;i++) vals[i] = 1.0; 5978e456f2a8SStefano Zampini ierr = ISGetIndices(globalis,(const PetscInt**)&idxs);CHKERRQ(ierr); 5979a7dc3881SStefano Zampini ierr = VecSet(gwork,0.0);CHKERRQ(ierr); 5980a7dc3881SStefano Zampini ierr = VecSet(lwork,0.0);CHKERRQ(ierr); 59811035eff8SStefano Zampini if (idxs) { /* multilevel guard */ 5982a7dc3881SStefano Zampini ierr = VecSetValues(gwork,lsize,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 59831035eff8SStefano Zampini } 5984a7dc3881SStefano Zampini ierr = VecAssemblyBegin(gwork);CHKERRQ(ierr); 5985e456f2a8SStefano Zampini ierr = ISRestoreIndices(globalis,(const PetscInt**)&idxs);CHKERRQ(ierr); 5986e456f2a8SStefano Zampini ierr = PetscFree(vals);CHKERRQ(ierr); 5987a7dc3881SStefano Zampini ierr = VecAssemblyEnd(gwork);CHKERRQ(ierr); 5988a7dc3881SStefano Zampini /* now compute set in local ordering */ 5989a7dc3881SStefano Zampini ierr = VecScatterBegin(g2l_ctx,gwork,lwork,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 5990a7dc3881SStefano Zampini ierr = VecScatterEnd(g2l_ctx,gwork,lwork,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 5991a7dc3881SStefano Zampini ierr = VecGetArrayRead(lwork,(const PetscScalar**)&vals);CHKERRQ(ierr); 5992a7dc3881SStefano Zampini ierr = VecGetSize(lwork,&n);CHKERRQ(ierr); 5993a7dc3881SStefano Zampini for (i=0,lsize=0;i<n;i++) { 5994ff92baa0SMatthew G. Knepley if (PetscRealPart(vals[i]) > 0.5) { 5995e456f2a8SStefano Zampini lsize++; 5996e456f2a8SStefano Zampini } 5997e456f2a8SStefano Zampini } 5998854ce69bSBarry Smith ierr = PetscMalloc1(lsize,&idxs);CHKERRQ(ierr); 5999a7dc3881SStefano Zampini for (i=0,lsize=0;i<n;i++) { 6000ff92baa0SMatthew G. Knepley if (PetscRealPart(vals[i]) > 0.5) { 6001e456f2a8SStefano Zampini idxs[lsize++] = i; 6002e456f2a8SStefano Zampini } 6003e456f2a8SStefano Zampini } 6004a7dc3881SStefano Zampini ierr = VecRestoreArrayRead(lwork,(const PetscScalar**)&vals);CHKERRQ(ierr); 6005a7dc3881SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)gwork),lsize,idxs,PETSC_OWN_POINTER,&localis_t);CHKERRQ(ierr); 6006e456f2a8SStefano Zampini *localis = localis_t; 6007e456f2a8SStefano Zampini PetscFunctionReturn(0); 6008e456f2a8SStefano Zampini } 6009906d46d4SStefano Zampini 6010906d46d4SStefano Zampini /* the next two functions will be called in KSPMatMult if a change of basis has been requested */ 6011906d46d4SStefano Zampini #undef __FUNCT__ 6012906d46d4SStefano Zampini #define __FUNCT__ "PCBDDCMatMult_Private" 6013906d46d4SStefano Zampini static PetscErrorCode PCBDDCMatMult_Private(Mat A, Vec x, Vec y) 6014906d46d4SStefano Zampini { 6015906d46d4SStefano Zampini PCBDDCChange_ctx change_ctx; 6016906d46d4SStefano Zampini PetscErrorCode ierr; 6017906d46d4SStefano Zampini 6018906d46d4SStefano Zampini PetscFunctionBegin; 6019906d46d4SStefano Zampini ierr = MatShellGetContext(A,&change_ctx);CHKERRQ(ierr); 6020906d46d4SStefano Zampini ierr = MatMult(change_ctx->global_change,x,change_ctx->work[0]);CHKERRQ(ierr); 6021906d46d4SStefano Zampini ierr = MatMult(change_ctx->original_mat,change_ctx->work[0],change_ctx->work[1]);CHKERRQ(ierr); 6022906d46d4SStefano Zampini ierr = MatMultTranspose(change_ctx->global_change,change_ctx->work[1],y);CHKERRQ(ierr); 6023906d46d4SStefano Zampini PetscFunctionReturn(0); 6024906d46d4SStefano Zampini } 6025906d46d4SStefano Zampini 6026906d46d4SStefano Zampini #undef __FUNCT__ 6027906d46d4SStefano Zampini #define __FUNCT__ "PCBDDCMatMultTranspose_Private" 6028906d46d4SStefano Zampini static PetscErrorCode PCBDDCMatMultTranspose_Private(Mat A, Vec x, Vec y) 6029906d46d4SStefano Zampini { 6030906d46d4SStefano Zampini PCBDDCChange_ctx change_ctx; 6031906d46d4SStefano Zampini PetscErrorCode ierr; 6032906d46d4SStefano Zampini 6033906d46d4SStefano Zampini PetscFunctionBegin; 6034906d46d4SStefano Zampini ierr = MatShellGetContext(A,&change_ctx);CHKERRQ(ierr); 6035906d46d4SStefano Zampini ierr = MatMult(change_ctx->global_change,x,change_ctx->work[0]);CHKERRQ(ierr); 6036906d46d4SStefano Zampini ierr = MatMultTranspose(change_ctx->original_mat,change_ctx->work[0],change_ctx->work[1]);CHKERRQ(ierr); 6037906d46d4SStefano Zampini ierr = MatMultTranspose(change_ctx->global_change,change_ctx->work[1],y);CHKERRQ(ierr); 6038906d46d4SStefano Zampini PetscFunctionReturn(0); 6039906d46d4SStefano Zampini } 6040b96c3477SStefano Zampini 6041b96c3477SStefano Zampini #undef __FUNCT__ 6042b96c3477SStefano Zampini #define __FUNCT__ "PCBDDCSetUpSubSchurs" 604308122e43SStefano Zampini PetscErrorCode PCBDDCSetUpSubSchurs(PC pc) 6044b96c3477SStefano Zampini { 6045a64f4aa4SStefano Zampini PC_IS *pcis=(PC_IS*)pc->data; 6046b96c3477SStefano Zampini PC_BDDC *pcbddc=(PC_BDDC*)pc->data; 6047b96c3477SStefano Zampini PCBDDCSubSchurs sub_schurs=pcbddc->sub_schurs; 6048a64f4aa4SStefano Zampini Mat S_j; 6049b96c3477SStefano Zampini PetscInt *used_xadj,*used_adjncy; 6050b96c3477SStefano Zampini PetscBool free_used_adj; 6051b96c3477SStefano Zampini PetscErrorCode ierr; 6052b96c3477SStefano Zampini 6053b96c3477SStefano Zampini PetscFunctionBegin; 6054b96c3477SStefano Zampini /* decide the adjacency to be used for determining internal problems for local schur on subsets */ 6055b96c3477SStefano Zampini free_used_adj = PETSC_FALSE; 605608122e43SStefano Zampini if (pcbddc->sub_schurs_layers == -1) { 6057b96c3477SStefano Zampini used_xadj = NULL; 6058b96c3477SStefano Zampini used_adjncy = NULL; 6059b96c3477SStefano Zampini } else { 606008122e43SStefano Zampini if (pcbddc->sub_schurs_use_useradj && pcbddc->mat_graph->xadj) { 606108122e43SStefano Zampini used_xadj = pcbddc->mat_graph->xadj; 606208122e43SStefano Zampini used_adjncy = pcbddc->mat_graph->adjncy; 606308122e43SStefano Zampini } else if (pcbddc->computed_rowadj) { 6064b96c3477SStefano Zampini used_xadj = pcbddc->mat_graph->xadj; 6065b96c3477SStefano Zampini used_adjncy = pcbddc->mat_graph->adjncy; 6066b96c3477SStefano Zampini } else { 60672fffb893SStefano Zampini PetscBool flg_row=PETSC_FALSE; 6068b96c3477SStefano Zampini const PetscInt *xadj,*adjncy; 6069b96c3477SStefano Zampini PetscInt nvtxs; 6070b96c3477SStefano Zampini 60712fffb893SStefano Zampini ierr = MatGetRowIJ(pcbddc->local_mat,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 60722fffb893SStefano Zampini if (flg_row) { 6073b96c3477SStefano Zampini ierr = PetscMalloc2(nvtxs+1,&used_xadj,xadj[nvtxs],&used_adjncy);CHKERRQ(ierr); 6074b96c3477SStefano Zampini ierr = PetscMemcpy(used_xadj,xadj,(nvtxs+1)*sizeof(*xadj));CHKERRQ(ierr); 6075b96c3477SStefano Zampini ierr = PetscMemcpy(used_adjncy,adjncy,(xadj[nvtxs])*sizeof(*adjncy));CHKERRQ(ierr); 6076b96c3477SStefano Zampini free_used_adj = PETSC_TRUE; 60772fffb893SStefano Zampini } else { 60782fffb893SStefano Zampini pcbddc->sub_schurs_layers = -1; 60792fffb893SStefano Zampini used_xadj = NULL; 60802fffb893SStefano Zampini used_adjncy = NULL; 60812fffb893SStefano Zampini } 60822fffb893SStefano Zampini ierr = MatRestoreRowIJ(pcbddc->local_mat,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 6083b96c3477SStefano Zampini } 6084b96c3477SStefano Zampini } 6085d5574798SStefano Zampini 6086d5574798SStefano Zampini /* setup sub_schurs data */ 6087a64f4aa4SStefano Zampini ierr = MatCreateSchurComplement(pcis->A_II,pcis->A_II,pcis->A_IB,pcis->A_BI,pcis->A_BB,&S_j);CHKERRQ(ierr); 6088df4d28bfSStefano Zampini if (!sub_schurs->schur_explicit) { 6089df4d28bfSStefano Zampini /* pcbddc->ksp_D up to date only if not using MatFactor with Schur complement support */ 6090a64f4aa4SStefano Zampini ierr = MatSchurComplementSetKSP(S_j,pcbddc->ksp_D);CHKERRQ(ierr); 6091ca92afb2SStefano 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); 6092a64f4aa4SStefano Zampini } else { 60936816873aSStefano Zampini PetscBool reuse_solvers = (PetscBool)!pcbddc->use_change_of_basis; 609404708bb6SStefano Zampini PetscBool isseqaij; 6095a3df083aSStefano Zampini PetscInt benign_n; 6096a3df083aSStefano Zampini 60975feab87aSStefano Zampini if (!pcbddc->use_vertices && reuse_solvers) { 60985feab87aSStefano Zampini PetscInt n_vertices; 60995feab87aSStefano Zampini 61005feab87aSStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_vertices,&n_vertices);CHKERRQ(ierr); 61012034aafcSStefano Zampini reuse_solvers = (PetscBool)!n_vertices; 61025feab87aSStefano Zampini } 610304708bb6SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcbddc->local_mat,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 610404708bb6SStefano Zampini if (!isseqaij) { 610504708bb6SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 610604708bb6SStefano Zampini if (matis->A == pcbddc->local_mat) { 610704708bb6SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 610804708bb6SStefano Zampini ierr = MatConvert(matis->A,MATSEQAIJ,MAT_INITIAL_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 610904708bb6SStefano Zampini } else { 611004708bb6SStefano Zampini ierr = MatConvert(pcbddc->local_mat,MATSEQAIJ,MAT_REUSE_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 611104708bb6SStefano Zampini } 611204708bb6SStefano Zampini } 6113a3df083aSStefano Zampini if (!pcbddc->benign_change_explicit) { 6114a3df083aSStefano Zampini benign_n = pcbddc->benign_n; 6115ca92afb2SStefano Zampini } else { 6116a3df083aSStefano Zampini benign_n = 0; 6117ca92afb2SStefano Zampini } 6118a3df083aSStefano 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); 6119ca92afb2SStefano Zampini } 6120d12d3064SStefano Zampini ierr = MatDestroy(&S_j);CHKERRQ(ierr); 6121b96c3477SStefano Zampini 6122b96c3477SStefano Zampini /* free adjacency */ 6123b96c3477SStefano Zampini if (free_used_adj) { 6124b96c3477SStefano Zampini ierr = PetscFree2(used_xadj,used_adjncy);CHKERRQ(ierr); 6125b96c3477SStefano Zampini } 6126b96c3477SStefano Zampini PetscFunctionReturn(0); 6127b96c3477SStefano Zampini } 6128b96c3477SStefano Zampini 6129b96c3477SStefano Zampini #undef __FUNCT__ 6130b96c3477SStefano Zampini #define __FUNCT__ "PCBDDCInitSubSchurs" 613108122e43SStefano Zampini PetscErrorCode PCBDDCInitSubSchurs(PC pc) 6132b96c3477SStefano Zampini { 6133b96c3477SStefano Zampini PC_IS *pcis=(PC_IS*)pc->data; 6134b96c3477SStefano Zampini PC_BDDC *pcbddc=(PC_BDDC*)pc->data; 6135b96c3477SStefano Zampini PCBDDCSubSchurs sub_schurs=pcbddc->sub_schurs; 6136b96c3477SStefano Zampini PCBDDCGraph graph; 6137b96c3477SStefano Zampini PetscErrorCode ierr; 6138b96c3477SStefano Zampini 6139b96c3477SStefano Zampini PetscFunctionBegin; 6140b96c3477SStefano Zampini /* attach interface graph for determining subsets */ 614108122e43SStefano Zampini if (pcbddc->sub_schurs_rebuild) { /* in case rebuild has been requested, it uses a graph generated only by the neighbouring information */ 61423301b35fSStefano Zampini IS verticesIS,verticescomm; 61433301b35fSStefano Zampini PetscInt vsize,*idxs; 6144b96c3477SStefano Zampini 6145b96c3477SStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,NULL,NULL,NULL,NULL,&verticesIS);CHKERRQ(ierr); 61463301b35fSStefano Zampini ierr = ISGetSize(verticesIS,&vsize);CHKERRQ(ierr); 61473301b35fSStefano Zampini ierr = ISGetIndices(verticesIS,(const PetscInt**)&idxs);CHKERRQ(ierr); 61483301b35fSStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),vsize,idxs,PETSC_COPY_VALUES,&verticescomm);CHKERRQ(ierr); 61493301b35fSStefano Zampini ierr = ISRestoreIndices(verticesIS,(const PetscInt**)&idxs);CHKERRQ(ierr); 61503301b35fSStefano Zampini ierr = ISDestroy(&verticesIS);CHKERRQ(ierr); 6151b96c3477SStefano Zampini ierr = PCBDDCGraphCreate(&graph);CHKERRQ(ierr); 61527fb0e2dbSStefano Zampini ierr = PCBDDCGraphInit(graph,pcbddc->mat_graph->l2gmap,pcbddc->mat_graph->nvtxs_global);CHKERRQ(ierr); 61533301b35fSStefano Zampini ierr = PCBDDCGraphSetUp(graph,0,NULL,pcbddc->DirichletBoundariesLocal,0,NULL,verticescomm);CHKERRQ(ierr); 61543301b35fSStefano Zampini ierr = ISDestroy(&verticescomm);CHKERRQ(ierr); 6155b96c3477SStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(graph);CHKERRQ(ierr); 6156b96c3477SStefano Zampini /* 6157b96c3477SStefano Zampini if (pcbddc->dbg_flag) { 6158b96c3477SStefano Zampini ierr = PCBDDCGraphASCIIView(graph,pcbddc->dbg_flag,pcbddc->dbg_viewer);CHKERRQ(ierr); 6159b96c3477SStefano Zampini } 6160b96c3477SStefano Zampini */ 6161b96c3477SStefano Zampini } else { 6162b96c3477SStefano Zampini graph = pcbddc->mat_graph; 6163b96c3477SStefano Zampini } 6164b96c3477SStefano Zampini 6165b96c3477SStefano Zampini /* sub_schurs init */ 6166a64f4aa4SStefano Zampini ierr = PCBDDCSubSchursInit(sub_schurs,pcis->is_I_local,pcis->is_B_local,graph,pcis->BtoNmap);CHKERRQ(ierr); 6167a64f4aa4SStefano Zampini 6168b96c3477SStefano Zampini /* free graph struct */ 616908122e43SStefano Zampini if (pcbddc->sub_schurs_rebuild) { 6170b96c3477SStefano Zampini ierr = PCBDDCGraphDestroy(&graph);CHKERRQ(ierr); 6171b96c3477SStefano Zampini } 6172b96c3477SStefano Zampini PetscFunctionReturn(0); 6173b96c3477SStefano Zampini } 6174fa34dd3eSStefano Zampini 6175fa34dd3eSStefano Zampini #undef __FUNCT__ 6176fa34dd3eSStefano Zampini #define __FUNCT__ "PCBDDCCheckOperator" 6177fa34dd3eSStefano Zampini PetscErrorCode PCBDDCCheckOperator(PC pc) 6178fa34dd3eSStefano Zampini { 6179fa34dd3eSStefano Zampini PC_IS *pcis=(PC_IS*)pc->data; 6180fa34dd3eSStefano Zampini PC_BDDC *pcbddc=(PC_BDDC*)pc->data; 6181fa34dd3eSStefano Zampini PetscErrorCode ierr; 6182fa34dd3eSStefano Zampini 6183fa34dd3eSStefano Zampini PetscFunctionBegin; 6184fa34dd3eSStefano Zampini if (pcbddc->n_vertices == pcbddc->local_primal_size) { 6185fa34dd3eSStefano Zampini IS zerodiag = NULL; 61864f1b2e48SStefano Zampini Mat S_j,B0_B=NULL; 6187fa34dd3eSStefano Zampini Vec dummy_vec=NULL,vec_check_B,vec_scale_P; 61884f1b2e48SStefano Zampini PetscScalar *p0_check,*array,*array2; 618975c01103SStefano Zampini PetscReal norm; 6190fa34dd3eSStefano Zampini PetscInt i; 6191fa34dd3eSStefano Zampini 6192fa34dd3eSStefano Zampini /* B0 and B0_B */ 6193fa34dd3eSStefano Zampini if (zerodiag) { 6194fa34dd3eSStefano Zampini IS dummy; 6195fa34dd3eSStefano Zampini 61964f1b2e48SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcbddc->benign_n,0,1,&dummy);CHKERRQ(ierr); 61974f1b2e48SStefano Zampini ierr = MatGetSubMatrix(pcbddc->benign_B0,dummy,pcis->is_B_local,MAT_INITIAL_MATRIX,&B0_B);CHKERRQ(ierr); 6198fa34dd3eSStefano Zampini ierr = MatCreateVecs(B0_B,NULL,&dummy_vec);CHKERRQ(ierr); 6199fa34dd3eSStefano Zampini ierr = ISDestroy(&dummy);CHKERRQ(ierr); 6200fa34dd3eSStefano Zampini } 6201fa34dd3eSStefano Zampini /* I need a primal vector to scale primal nodes since BDDC sums contibutions */ 6202fa34dd3eSStefano Zampini ierr = VecDuplicate(pcbddc->vec1_P,&vec_scale_P);CHKERRQ(ierr); 6203fa34dd3eSStefano Zampini ierr = VecSet(pcbddc->vec1_P,1.0);CHKERRQ(ierr); 6204fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6205fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6206fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,vec_scale_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6207fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,vec_scale_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6208fa34dd3eSStefano Zampini ierr = VecReciprocal(vec_scale_P);CHKERRQ(ierr); 6209fa34dd3eSStefano Zampini /* S_j */ 6210fa34dd3eSStefano Zampini ierr = MatCreateSchurComplement(pcis->A_II,pcis->A_II,pcis->A_IB,pcis->A_BI,pcis->A_BB,&S_j);CHKERRQ(ierr); 6211fa34dd3eSStefano Zampini ierr = MatSchurComplementSetKSP(S_j,pcbddc->ksp_D);CHKERRQ(ierr); 6212fa34dd3eSStefano Zampini 6213fa34dd3eSStefano Zampini /* mimic vector in \widetilde{W}_\Gamma */ 6214fa34dd3eSStefano Zampini ierr = VecSetRandom(pcis->vec1_N,NULL);CHKERRQ(ierr); 6215fa34dd3eSStefano Zampini /* continuous in primal space */ 6216fa34dd3eSStefano Zampini ierr = VecSetRandom(pcbddc->coarse_vec,NULL);CHKERRQ(ierr); 6217fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,pcbddc->vec1_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6218fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,pcbddc->vec1_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6219fa34dd3eSStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 62204f1b2e48SStefano Zampini ierr = PetscCalloc1(pcbddc->benign_n,&p0_check);CHKERRQ(ierr); 62214f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) p0_check[i] = array[pcbddc->local_primal_size-pcbddc->benign_n+i]; 6222fa34dd3eSStefano Zampini ierr = VecSetValues(pcis->vec1_N,pcbddc->local_primal_size,pcbddc->local_primal_ref_node,array,INSERT_VALUES);CHKERRQ(ierr); 6223fa34dd3eSStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 6224fa34dd3eSStefano Zampini ierr = VecAssemblyBegin(pcis->vec1_N);CHKERRQ(ierr); 6225fa34dd3eSStefano Zampini ierr = VecAssemblyEnd(pcis->vec1_N);CHKERRQ(ierr); 6226fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcis->N_to_B,pcis->vec1_N,pcis->vec2_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6227fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcis->N_to_B,pcis->vec1_N,pcis->vec2_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6228fa34dd3eSStefano Zampini ierr = VecDuplicate(pcis->vec2_B,&vec_check_B);CHKERRQ(ierr); 6229fa34dd3eSStefano Zampini ierr = VecCopy(pcis->vec2_B,vec_check_B);CHKERRQ(ierr); 6230fa34dd3eSStefano Zampini 6231fa34dd3eSStefano Zampini /* assemble rhs for coarse problem */ 6232fa34dd3eSStefano Zampini /* widetilde{S}_\Gamma w_\Gamma + \widetilde{B0}^T_B p0 */ 6233fa34dd3eSStefano Zampini /* local with Schur */ 6234fa34dd3eSStefano Zampini ierr = MatMult(S_j,pcis->vec2_B,pcis->vec1_B);CHKERRQ(ierr); 6235fa34dd3eSStefano Zampini if (zerodiag) { 6236fa34dd3eSStefano Zampini ierr = VecGetArray(dummy_vec,&array);CHKERRQ(ierr); 62374f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) array[i] = p0_check[i]; 6238fa34dd3eSStefano Zampini ierr = VecRestoreArray(dummy_vec,&array);CHKERRQ(ierr); 6239fa34dd3eSStefano Zampini ierr = MatMultTransposeAdd(B0_B,dummy_vec,pcis->vec1_B,pcis->vec1_B);CHKERRQ(ierr); 6240fa34dd3eSStefano Zampini } 6241fa34dd3eSStefano Zampini /* sum on primal nodes the local contributions */ 6242fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcis->N_to_B,pcis->vec1_B,pcis->vec1_N,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6243fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcis->N_to_B,pcis->vec1_B,pcis->vec1_N,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6244fa34dd3eSStefano Zampini ierr = VecGetArray(pcis->vec1_N,&array);CHKERRQ(ierr); 6245fa34dd3eSStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array2);CHKERRQ(ierr); 6246fa34dd3eSStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) array2[i] = array[pcbddc->local_primal_ref_node[i]]; 6247fa34dd3eSStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array2);CHKERRQ(ierr); 6248fa34dd3eSStefano Zampini ierr = VecRestoreArray(pcis->vec1_N,&array);CHKERRQ(ierr); 6249fa34dd3eSStefano Zampini ierr = VecSet(pcbddc->coarse_vec,0.);CHKERRQ(ierr); 6250fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6251fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6252fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,pcbddc->vec1_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6253fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,pcbddc->vec1_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6254fa34dd3eSStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 6255fa34dd3eSStefano Zampini /* scale primal nodes (BDDC sums contibutions) */ 6256fa34dd3eSStefano Zampini ierr = VecPointwiseMult(pcbddc->vec1_P,vec_scale_P,pcbddc->vec1_P);CHKERRQ(ierr); 6257fa34dd3eSStefano Zampini ierr = VecSetValues(pcis->vec1_N,pcbddc->local_primal_size,pcbddc->local_primal_ref_node,array,INSERT_VALUES);CHKERRQ(ierr); 6258fa34dd3eSStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 6259fa34dd3eSStefano Zampini ierr = VecAssemblyBegin(pcis->vec1_N);CHKERRQ(ierr); 6260fa34dd3eSStefano Zampini ierr = VecAssemblyEnd(pcis->vec1_N);CHKERRQ(ierr); 6261fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcis->N_to_B,pcis->vec1_N,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6262fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcis->N_to_B,pcis->vec1_N,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6263fa34dd3eSStefano Zampini /* global: \widetilde{B0}_B w_\Gamma */ 6264fa34dd3eSStefano Zampini if (zerodiag) { 6265fa34dd3eSStefano Zampini ierr = MatMult(B0_B,pcis->vec2_B,dummy_vec);CHKERRQ(ierr); 6266fa34dd3eSStefano Zampini ierr = VecGetArray(dummy_vec,&array);CHKERRQ(ierr); 62674f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) pcbddc->benign_p0[i] = array[i]; 6268fa34dd3eSStefano Zampini ierr = VecRestoreArray(dummy_vec,&array);CHKERRQ(ierr); 6269fa34dd3eSStefano Zampini } 6270fa34dd3eSStefano Zampini /* BDDC */ 6271fa34dd3eSStefano Zampini ierr = VecSet(pcis->vec1_D,0.);CHKERRQ(ierr); 6272fa34dd3eSStefano Zampini ierr = PCBDDCApplyInterfacePreconditioner(pc,PETSC_FALSE);CHKERRQ(ierr); 6273fa34dd3eSStefano Zampini 6274fa34dd3eSStefano Zampini ierr = VecCopy(pcis->vec1_B,pcis->vec2_B);CHKERRQ(ierr); 6275fa34dd3eSStefano Zampini ierr = VecAXPY(pcis->vec1_B,-1.0,vec_check_B);CHKERRQ(ierr); 6276fa34dd3eSStefano Zampini ierr = VecNorm(pcis->vec1_B,NORM_INFINITY,&norm);CHKERRQ(ierr); 6277fa34dd3eSStefano Zampini PetscPrintf(PETSC_COMM_SELF,"[%d] BDDC local error is %1.4e\n",PetscGlobalRank,norm); 62784f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 62794f1b2e48SStefano Zampini PetscPrintf(PETSC_COMM_SELF,"[%d] BDDC p0[%d] error is %1.4e\n",PetscGlobalRank,i,PetscAbsScalar(pcbddc->benign_p0[i]-p0_check[i])); 6280fa34dd3eSStefano Zampini } 62814f1b2e48SStefano Zampini ierr = PetscFree(p0_check);CHKERRQ(ierr); 6282fa34dd3eSStefano Zampini ierr = VecDestroy(&vec_scale_P);CHKERRQ(ierr); 6283fa34dd3eSStefano Zampini ierr = VecDestroy(&vec_check_B);CHKERRQ(ierr); 6284fa34dd3eSStefano Zampini ierr = VecDestroy(&dummy_vec);CHKERRQ(ierr); 6285fa34dd3eSStefano Zampini ierr = MatDestroy(&S_j);CHKERRQ(ierr); 6286fa34dd3eSStefano Zampini ierr = MatDestroy(&B0_B);CHKERRQ(ierr); 6287fa34dd3eSStefano Zampini } 6288fa34dd3eSStefano Zampini PetscFunctionReturn(0); 6289fa34dd3eSStefano Zampini } 6290