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> 5daf8a457SStefano Zampini #include <petsc/private/sfimpl.h> 6674ae819SStefano Zampini 7a3df083aSStefano Zampini #undef __FUNCT__ 81f4df5f7SStefano Zampini #define __FUNCT__ "PCBDDCComputeLocalTopologyInfo" 91f4df5f7SStefano Zampini PetscErrorCode PCBDDCComputeLocalTopologyInfo(PC pc) 101f4df5f7SStefano Zampini { 111f4df5f7SStefano Zampini PetscErrorCode ierr; 121f4df5f7SStefano Zampini Vec local,global; 131f4df5f7SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 141f4df5f7SStefano Zampini Mat_IS *matis = (Mat_IS*)pc->pmat->data; 151f4df5f7SStefano Zampini 161f4df5f7SStefano Zampini PetscFunctionBegin; 171f4df5f7SStefano Zampini ierr = MatCreateVecs(pc->pmat,&global,NULL);CHKERRQ(ierr); 181f4df5f7SStefano Zampini /* need to convert from global to local topology information and remove references to information in global ordering */ 191f4df5f7SStefano Zampini ierr = MatCreateVecs(matis->A,&local,NULL);CHKERRQ(ierr); 201f4df5f7SStefano Zampini if (pcbddc->user_provided_isfordofs) { 211f4df5f7SStefano Zampini if (pcbddc->n_ISForDofs) { 221f4df5f7SStefano Zampini PetscInt i; 231f4df5f7SStefano Zampini ierr = PetscMalloc1(pcbddc->n_ISForDofs,&pcbddc->ISForDofsLocal);CHKERRQ(ierr); 241f4df5f7SStefano Zampini for (i=0;i<pcbddc->n_ISForDofs;i++) { 251f4df5f7SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->rctx,global,local,pcbddc->ISForDofs[i],&pcbddc->ISForDofsLocal[i]);CHKERRQ(ierr); 261f4df5f7SStefano Zampini ierr = ISDestroy(&pcbddc->ISForDofs[i]);CHKERRQ(ierr); 271f4df5f7SStefano Zampini } 281f4df5f7SStefano Zampini pcbddc->n_ISForDofsLocal = pcbddc->n_ISForDofs; 291f4df5f7SStefano Zampini pcbddc->n_ISForDofs = 0; 301f4df5f7SStefano Zampini ierr = PetscFree(pcbddc->ISForDofs);CHKERRQ(ierr); 311f4df5f7SStefano Zampini } 321f4df5f7SStefano Zampini } else { 331f4df5f7SStefano Zampini if (!pcbddc->n_ISForDofsLocal) { /* field split not present, create it in local ordering */ 341f4df5f7SStefano Zampini PetscInt i, n = matis->A->rmap->n; 351f4df5f7SStefano Zampini ierr = MatGetBlockSize(pc->pmat,&pcbddc->n_ISForDofsLocal);CHKERRQ(ierr); 361f4df5f7SStefano Zampini ierr = PetscMalloc1(pcbddc->n_ISForDofsLocal,&pcbddc->ISForDofsLocal);CHKERRQ(ierr); 371f4df5f7SStefano Zampini for (i=0;i<pcbddc->n_ISForDofsLocal;i++) { 381f4df5f7SStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)pc),n/pcbddc->n_ISForDofsLocal,i,pcbddc->n_ISForDofsLocal,&pcbddc->ISForDofsLocal[i]);CHKERRQ(ierr); 391f4df5f7SStefano Zampini } 401f4df5f7SStefano Zampini } 411f4df5f7SStefano Zampini } 421f4df5f7SStefano Zampini 431f4df5f7SStefano Zampini if (!pcbddc->DirichletBoundariesLocal && pcbddc->DirichletBoundaries) { 441f4df5f7SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->rctx,global,local,pcbddc->DirichletBoundaries,&pcbddc->DirichletBoundariesLocal);CHKERRQ(ierr); 451f4df5f7SStefano Zampini } 461f4df5f7SStefano Zampini if (!pcbddc->NeumannBoundariesLocal && pcbddc->NeumannBoundaries) { 471f4df5f7SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->rctx,global,local,pcbddc->NeumannBoundaries,&pcbddc->NeumannBoundariesLocal);CHKERRQ(ierr); 481f4df5f7SStefano Zampini } 491f4df5f7SStefano Zampini if (!pcbddc->user_primal_vertices_local && pcbddc->user_primal_vertices) { 501f4df5f7SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->rctx,global,local,pcbddc->user_primal_vertices,&pcbddc->user_primal_vertices_local);CHKERRQ(ierr); 511f4df5f7SStefano Zampini } 521f4df5f7SStefano Zampini ierr = VecDestroy(&global);CHKERRQ(ierr); 531f4df5f7SStefano Zampini ierr = VecDestroy(&local);CHKERRQ(ierr); 541f4df5f7SStefano Zampini PetscFunctionReturn(0); 551f4df5f7SStefano Zampini } 561f4df5f7SStefano Zampini 571f4df5f7SStefano Zampini #undef __FUNCT__ 583e589ea0SStefano Zampini #define __FUNCT__ "PCBDDCBenignRemoveInterior" 593e589ea0SStefano Zampini PetscErrorCode PCBDDCBenignRemoveInterior(PC pc,Vec r,Vec z) 603e589ea0SStefano Zampini { 613e589ea0SStefano Zampini PC_IS *pcis = (PC_IS*)(pc->data); 623e589ea0SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)(pc->data); 633e589ea0SStefano Zampini PetscErrorCode ierr; 643e589ea0SStefano Zampini 653e589ea0SStefano Zampini PetscFunctionBegin; 663e589ea0SStefano Zampini if (!pcbddc->benign_have_null) { 673e589ea0SStefano Zampini PetscFunctionReturn(0); 683e589ea0SStefano Zampini } 693e589ea0SStefano Zampini if (pcbddc->ChangeOfBasisMatrix) { 703e589ea0SStefano Zampini Vec swap; 713e589ea0SStefano Zampini 723e589ea0SStefano Zampini ierr = MatMultTranspose(pcbddc->ChangeOfBasisMatrix,r,pcbddc->work_change);CHKERRQ(ierr); 733e589ea0SStefano Zampini swap = pcbddc->work_change; 743e589ea0SStefano Zampini pcbddc->work_change = r; 753e589ea0SStefano Zampini r = swap; 763e589ea0SStefano Zampini } 773e589ea0SStefano Zampini ierr = VecScatterBegin(pcis->global_to_D,r,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 783e589ea0SStefano Zampini ierr = VecScatterEnd(pcis->global_to_D,r,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 793e589ea0SStefano Zampini if (pcbddc->benign_n) { 803e589ea0SStefano Zampini ierr = KSPSolve(pcbddc->ksp_D,pcis->vec1_D,pcis->vec2_D);CHKERRQ(ierr); 813e589ea0SStefano Zampini } else { 823e589ea0SStefano Zampini ierr = VecSet(pcis->vec2_D,0.);CHKERRQ(ierr); 833e589ea0SStefano Zampini } 843e589ea0SStefano Zampini ierr = VecSet(z,0.);CHKERRQ(ierr); 853e589ea0SStefano Zampini ierr = VecScatterBegin(pcis->global_to_D,pcis->vec2_D,z,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 863e589ea0SStefano Zampini ierr = VecScatterEnd(pcis->global_to_D,pcis->vec2_D,z,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 873e589ea0SStefano Zampini if (pcbddc->ChangeOfBasisMatrix) { 883e589ea0SStefano Zampini Vec swap; 893e589ea0SStefano Zampini 903e589ea0SStefano Zampini swap = r; 913e589ea0SStefano Zampini r = pcbddc->work_change; 923e589ea0SStefano Zampini pcbddc->work_change = swap; 933e589ea0SStefano Zampini ierr = VecCopy(z,pcbddc->work_change);CHKERRQ(ierr); 943e589ea0SStefano Zampini ierr = MatMult(pcbddc->ChangeOfBasisMatrix,pcbddc->work_change,z);CHKERRQ(ierr); 953e589ea0SStefano Zampini } 963e589ea0SStefano Zampini PetscFunctionReturn(0); 973e589ea0SStefano Zampini } 983e589ea0SStefano Zampini 993e589ea0SStefano Zampini #undef __FUNCT__ 100a3df083aSStefano Zampini #define __FUNCT__ "PCBDDCBenignMatMult_Private_Private" 101a3df083aSStefano Zampini PetscErrorCode PCBDDCBenignMatMult_Private_Private(Mat A, Vec x, Vec y, PetscBool transpose) 102a3df083aSStefano Zampini { 103a3df083aSStefano Zampini PCBDDCBenignMatMult_ctx ctx; 104a3df083aSStefano Zampini PetscErrorCode ierr; 105a3df083aSStefano Zampini PetscBool apply_right,apply_left,reset_x; 106a3df083aSStefano Zampini 107a3df083aSStefano Zampini PetscFunctionBegin; 108a3df083aSStefano Zampini ierr = MatShellGetContext(A,&ctx);CHKERRQ(ierr); 109a3df083aSStefano Zampini if (transpose) { 110a3df083aSStefano Zampini apply_right = ctx->apply_left; 111a3df083aSStefano Zampini apply_left = ctx->apply_right; 112a3df083aSStefano Zampini } else { 113a3df083aSStefano Zampini apply_right = ctx->apply_right; 114a3df083aSStefano Zampini apply_left = ctx->apply_left; 115a3df083aSStefano Zampini } 116a3df083aSStefano Zampini reset_x = PETSC_FALSE; 117a3df083aSStefano Zampini if (apply_right) { 118a3df083aSStefano Zampini const PetscScalar *ax; 119a3df083aSStefano Zampini PetscInt nl,i; 120a3df083aSStefano Zampini 121a3df083aSStefano Zampini ierr = VecGetLocalSize(x,&nl);CHKERRQ(ierr); 122a3df083aSStefano Zampini ierr = VecGetArrayRead(x,&ax);CHKERRQ(ierr); 123a3df083aSStefano Zampini ierr = PetscMemcpy(ctx->work,ax,nl*sizeof(PetscScalar));CHKERRQ(ierr); 124a3df083aSStefano Zampini ierr = VecRestoreArrayRead(x,&ax);CHKERRQ(ierr); 125a3df083aSStefano Zampini for (i=0;i<ctx->benign_n;i++) { 126a3df083aSStefano Zampini PetscScalar sum,val; 127a3df083aSStefano Zampini const PetscInt *idxs; 128a3df083aSStefano Zampini PetscInt nz,j; 129a3df083aSStefano Zampini ierr = ISGetLocalSize(ctx->benign_zerodiag_subs[i],&nz);CHKERRQ(ierr); 130a3df083aSStefano Zampini ierr = ISGetIndices(ctx->benign_zerodiag_subs[i],&idxs);CHKERRQ(ierr); 131a3df083aSStefano Zampini sum = 0.; 132a3df083aSStefano Zampini if (ctx->apply_p0) { 133a3df083aSStefano Zampini val = ctx->work[idxs[nz-1]]; 134a3df083aSStefano Zampini for (j=0;j<nz-1;j++) { 135a3df083aSStefano Zampini sum += ctx->work[idxs[j]]; 136a3df083aSStefano Zampini ctx->work[idxs[j]] += val; 137a3df083aSStefano Zampini } 138a3df083aSStefano Zampini } else { 139a3df083aSStefano Zampini for (j=0;j<nz-1;j++) { 140a3df083aSStefano Zampini sum += ctx->work[idxs[j]]; 141a3df083aSStefano Zampini } 142a3df083aSStefano Zampini } 143a3df083aSStefano Zampini ctx->work[idxs[nz-1]] -= sum; 144a3df083aSStefano Zampini ierr = ISRestoreIndices(ctx->benign_zerodiag_subs[i],&idxs);CHKERRQ(ierr); 145a3df083aSStefano Zampini } 146a3df083aSStefano Zampini ierr = VecPlaceArray(x,ctx->work);CHKERRQ(ierr); 147a3df083aSStefano Zampini reset_x = PETSC_TRUE; 148a3df083aSStefano Zampini } 149a3df083aSStefano Zampini if (transpose) { 150a3df083aSStefano Zampini ierr = MatMultTranspose(ctx->A,x,y);CHKERRQ(ierr); 151a3df083aSStefano Zampini } else { 152a3df083aSStefano Zampini ierr = MatMult(ctx->A,x,y);CHKERRQ(ierr); 153a3df083aSStefano Zampini } 154a3df083aSStefano Zampini if (reset_x) { 155a3df083aSStefano Zampini ierr = VecResetArray(x);CHKERRQ(ierr); 156a3df083aSStefano Zampini } 157a3df083aSStefano Zampini if (apply_left) { 158a3df083aSStefano Zampini PetscScalar *ay; 159a3df083aSStefano Zampini PetscInt i; 160a3df083aSStefano Zampini 161a3df083aSStefano Zampini ierr = VecGetArray(y,&ay);CHKERRQ(ierr); 162a3df083aSStefano Zampini for (i=0;i<ctx->benign_n;i++) { 163a3df083aSStefano Zampini PetscScalar sum,val; 164a3df083aSStefano Zampini const PetscInt *idxs; 165a3df083aSStefano Zampini PetscInt nz,j; 166a3df083aSStefano Zampini ierr = ISGetLocalSize(ctx->benign_zerodiag_subs[i],&nz);CHKERRQ(ierr); 167a3df083aSStefano Zampini ierr = ISGetIndices(ctx->benign_zerodiag_subs[i],&idxs);CHKERRQ(ierr); 168a3df083aSStefano Zampini val = -ay[idxs[nz-1]]; 169a3df083aSStefano Zampini if (ctx->apply_p0) { 170a3df083aSStefano Zampini sum = 0.; 171a3df083aSStefano Zampini for (j=0;j<nz-1;j++) { 172a3df083aSStefano Zampini sum += ay[idxs[j]]; 173a3df083aSStefano Zampini ay[idxs[j]] += val; 174a3df083aSStefano Zampini } 175a3df083aSStefano Zampini ay[idxs[nz-1]] += sum; 176a3df083aSStefano Zampini } else { 177a3df083aSStefano Zampini for (j=0;j<nz-1;j++) { 178a3df083aSStefano Zampini ay[idxs[j]] += val; 179a3df083aSStefano Zampini } 180a3df083aSStefano Zampini ay[idxs[nz-1]] = 0.; 181a3df083aSStefano Zampini } 182a3df083aSStefano Zampini ierr = ISRestoreIndices(ctx->benign_zerodiag_subs[i],&idxs);CHKERRQ(ierr); 183a3df083aSStefano Zampini } 184a3df083aSStefano Zampini ierr = VecRestoreArray(y,&ay);CHKERRQ(ierr); 185a3df083aSStefano Zampini } 186a3df083aSStefano Zampini PetscFunctionReturn(0); 187a3df083aSStefano Zampini } 188a3df083aSStefano Zampini 189a3df083aSStefano Zampini #undef __FUNCT__ 190a3df083aSStefano Zampini #define __FUNCT__ "PCBDDCBenignMatMultTranspose_Private" 191a3df083aSStefano Zampini PetscErrorCode PCBDDCBenignMatMultTranspose_Private(Mat A, Vec x, Vec y) 192a3df083aSStefano Zampini { 193a3df083aSStefano Zampini PetscErrorCode ierr; 194a3df083aSStefano Zampini 195a3df083aSStefano Zampini PetscFunctionBegin; 196a3df083aSStefano Zampini ierr = PCBDDCBenignMatMult_Private_Private(A,x,y,PETSC_TRUE);CHKERRQ(ierr); 197a3df083aSStefano Zampini PetscFunctionReturn(0); 198a3df083aSStefano Zampini } 199a3df083aSStefano Zampini 200a3df083aSStefano Zampini #undef __FUNCT__ 201a3df083aSStefano Zampini #define __FUNCT__ "PCBDDCBenignMatMult_Private" 202a3df083aSStefano Zampini PetscErrorCode PCBDDCBenignMatMult_Private(Mat A, Vec x, Vec y) 203a3df083aSStefano Zampini { 204a3df083aSStefano Zampini PetscErrorCode ierr; 205a3df083aSStefano Zampini 206a3df083aSStefano Zampini PetscFunctionBegin; 207a3df083aSStefano Zampini ierr = PCBDDCBenignMatMult_Private_Private(A,x,y,PETSC_FALSE);CHKERRQ(ierr); 208a3df083aSStefano Zampini PetscFunctionReturn(0); 209a3df083aSStefano Zampini } 210a3df083aSStefano Zampini 211a3df083aSStefano Zampini #undef __FUNCT__ 212a3df083aSStefano Zampini #define __FUNCT__ "PCBDDCBenignShellMat" 213a3df083aSStefano Zampini PetscErrorCode PCBDDCBenignShellMat(PC pc, PetscBool restore) 214a3df083aSStefano Zampini { 215a3df083aSStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 216a3df083aSStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 217a3df083aSStefano Zampini PCBDDCBenignMatMult_ctx ctx; 218a3df083aSStefano Zampini PetscErrorCode ierr; 219a3df083aSStefano Zampini 220a3df083aSStefano Zampini PetscFunctionBegin; 221a3df083aSStefano Zampini if (!restore) { 2221dd7afcfSStefano Zampini Mat A_IB,A_BI; 223a3df083aSStefano Zampini PetscScalar *work; 224df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse = pcbddc->sub_schurs->reuse_solver; 225a3df083aSStefano Zampini 2261dd7afcfSStefano Zampini if (pcbddc->benign_original_mat) { 2271dd7afcfSStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Benign original mat has not been restored"); 2281dd7afcfSStefano Zampini } 2291dd7afcfSStefano Zampini if (!pcbddc->benign_change || !pcbddc->benign_n || pcbddc->benign_change_explicit) { 2301dd7afcfSStefano Zampini PetscFunctionReturn(0); 2311dd7afcfSStefano Zampini } 232a3df083aSStefano Zampini ierr = PetscMalloc1(pcis->n,&work);CHKERRQ(ierr); 233a3df083aSStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&A_IB);CHKERRQ(ierr); 234a3df083aSStefano Zampini ierr = MatSetSizes(A_IB,pcis->n-pcis->n_B,pcis->n_B,PETSC_DECIDE,PETSC_DECIDE);CHKERRQ(ierr); 235a3df083aSStefano Zampini ierr = MatSetType(A_IB,MATSHELL);CHKERRQ(ierr); 236a3df083aSStefano Zampini ierr = MatShellSetOperation(A_IB,MATOP_MULT,(void (*)(void))PCBDDCBenignMatMult_Private);CHKERRQ(ierr); 237a3df083aSStefano Zampini ierr = MatShellSetOperation(A_IB,MATOP_MULT_TRANSPOSE,(void (*)(void))PCBDDCBenignMatMultTranspose_Private);CHKERRQ(ierr); 238a3df083aSStefano Zampini ierr = PetscNew(&ctx);CHKERRQ(ierr); 239a3df083aSStefano Zampini ierr = MatShellSetContext(A_IB,ctx);CHKERRQ(ierr); 240a3df083aSStefano Zampini ctx->apply_left = PETSC_TRUE; 241a3df083aSStefano Zampini ctx->apply_right = PETSC_FALSE; 242a3df083aSStefano Zampini ctx->apply_p0 = PETSC_FALSE; 243a3df083aSStefano Zampini ctx->benign_n = pcbddc->benign_n; 244059032f7SStefano Zampini if (reuse) { 245a3df083aSStefano Zampini ctx->benign_zerodiag_subs = reuse->benign_zerodiag_subs; 2461dd7afcfSStefano Zampini ctx->free = PETSC_FALSE; 247059032f7SStefano Zampini } else { /* TODO: could be optimized for successive solves */ 248059032f7SStefano Zampini ISLocalToGlobalMapping N_to_D; 249059032f7SStefano Zampini PetscInt i; 250059032f7SStefano Zampini 251059032f7SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(pcis->is_I_local,&N_to_D);CHKERRQ(ierr); 252059032f7SStefano Zampini ierr = PetscMalloc1(pcbddc->benign_n,&ctx->benign_zerodiag_subs);CHKERRQ(ierr); 253059032f7SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 254059032f7SStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(N_to_D,IS_GTOLM_DROP,pcbddc->benign_zerodiag_subs[i],&ctx->benign_zerodiag_subs[i]);CHKERRQ(ierr); 255059032f7SStefano Zampini } 256059032f7SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&N_to_D);CHKERRQ(ierr); 2571dd7afcfSStefano Zampini ctx->free = PETSC_TRUE; 258059032f7SStefano Zampini } 259a3df083aSStefano Zampini ctx->A = pcis->A_IB; 260a3df083aSStefano Zampini ctx->work = work; 261a3df083aSStefano Zampini ierr = MatSetUp(A_IB);CHKERRQ(ierr); 262a3df083aSStefano Zampini ierr = MatAssemblyBegin(A_IB,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 263a3df083aSStefano Zampini ierr = MatAssemblyEnd(A_IB,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 264a3df083aSStefano Zampini pcis->A_IB = A_IB; 265a3df083aSStefano Zampini 266a3df083aSStefano Zampini /* A_BI as A_IB^T */ 267a3df083aSStefano Zampini ierr = MatCreateTranspose(A_IB,&A_BI);CHKERRQ(ierr); 268a3df083aSStefano Zampini pcbddc->benign_original_mat = pcis->A_BI; 269a3df083aSStefano Zampini pcis->A_BI = A_BI; 270a3df083aSStefano Zampini } else { 2711dd7afcfSStefano Zampini if (!pcbddc->benign_original_mat) { 2721dd7afcfSStefano Zampini PetscFunctionReturn(0); 2731dd7afcfSStefano Zampini } 274a3df083aSStefano Zampini ierr = MatShellGetContext(pcis->A_IB,&ctx);CHKERRQ(ierr); 275a3df083aSStefano Zampini ierr = MatDestroy(&pcis->A_IB);CHKERRQ(ierr); 276a3df083aSStefano Zampini pcis->A_IB = ctx->A; 2771dd7afcfSStefano Zampini ctx->A = NULL; 2781dd7afcfSStefano Zampini ierr = MatDestroy(&pcis->A_BI);CHKERRQ(ierr); 2791dd7afcfSStefano Zampini pcis->A_BI = pcbddc->benign_original_mat; 2801dd7afcfSStefano Zampini pcbddc->benign_original_mat = NULL; 2811dd7afcfSStefano Zampini if (ctx->free) { 282059032f7SStefano Zampini PetscInt i; 2831dd7afcfSStefano Zampini for (i=0;i<ctx->benign_n;i++) { 284059032f7SStefano Zampini ierr = ISDestroy(&ctx->benign_zerodiag_subs[i]);CHKERRQ(ierr); 285059032f7SStefano Zampini } 286059032f7SStefano Zampini ierr = PetscFree(ctx->benign_zerodiag_subs);CHKERRQ(ierr); 287059032f7SStefano Zampini } 288a3df083aSStefano Zampini ierr = PetscFree(ctx->work);CHKERRQ(ierr); 289a3df083aSStefano Zampini ierr = PetscFree(ctx);CHKERRQ(ierr); 290a3df083aSStefano Zampini } 291a3df083aSStefano Zampini PetscFunctionReturn(0); 292a3df083aSStefano Zampini } 293a3df083aSStefano Zampini 294a3df083aSStefano Zampini /* used just in bddc debug mode */ 295a3df083aSStefano Zampini #undef __FUNCT__ 296a3df083aSStefano Zampini #define __FUNCT__ "PCBDDCBenignProject" 297a3df083aSStefano Zampini PetscErrorCode PCBDDCBenignProject(PC pc, IS is1, IS is2, Mat *B) 298a3df083aSStefano Zampini { 299a3df083aSStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 300a3df083aSStefano Zampini Mat_IS *matis = (Mat_IS*)pc->pmat->data; 301a3df083aSStefano Zampini Mat An; 302a3df083aSStefano Zampini PetscErrorCode ierr; 303a3df083aSStefano Zampini 304a3df083aSStefano Zampini PetscFunctionBegin; 305a3df083aSStefano Zampini ierr = MatPtAP(matis->A,pcbddc->benign_change,MAT_INITIAL_MATRIX,2.0,&An);CHKERRQ(ierr); 306a3df083aSStefano Zampini ierr = MatZeroRowsColumns(An,pcbddc->benign_n,pcbddc->benign_p0_lidx,1.0,NULL,NULL);CHKERRQ(ierr); 307a3df083aSStefano Zampini if (is1) { 308a3df083aSStefano Zampini ierr = MatGetSubMatrix(An,is1,is2,MAT_INITIAL_MATRIX,B);CHKERRQ(ierr); 309a3df083aSStefano Zampini ierr = MatDestroy(&An);CHKERRQ(ierr); 310a3df083aSStefano Zampini } else { 311a3df083aSStefano Zampini *B = An; 312a3df083aSStefano Zampini } 313a3df083aSStefano Zampini PetscFunctionReturn(0); 314a3df083aSStefano Zampini } 315a3df083aSStefano Zampini 3161cf9b237SStefano Zampini /* TODO: add reuse flag */ 3171cf9b237SStefano Zampini #undef __FUNCT__ 3181cf9b237SStefano Zampini #define __FUNCT__ "MatSeqAIJCompress" 3191cf9b237SStefano Zampini PetscErrorCode MatSeqAIJCompress(Mat A, Mat *B) 3201cf9b237SStefano Zampini { 3211cf9b237SStefano Zampini Mat Bt; 3221cf9b237SStefano Zampini PetscScalar *a,*bdata; 3231cf9b237SStefano Zampini const PetscInt *ii,*ij; 3241cf9b237SStefano Zampini PetscInt m,n,i,nnz,*bii,*bij; 3251cf9b237SStefano Zampini PetscBool flg_row; 3261cf9b237SStefano Zampini PetscErrorCode ierr; 3271cf9b237SStefano Zampini 3281cf9b237SStefano Zampini PetscFunctionBegin; 3291cf9b237SStefano Zampini ierr = MatGetSize(A,&n,&m);CHKERRQ(ierr); 3301cf9b237SStefano Zampini ierr = MatGetRowIJ(A,0,PETSC_FALSE,PETSC_FALSE,&n,&ii,&ij,&flg_row);CHKERRQ(ierr); 3311cf9b237SStefano Zampini ierr = MatSeqAIJGetArray(A,&a);CHKERRQ(ierr); 3321cf9b237SStefano Zampini nnz = n; 3331cf9b237SStefano Zampini for (i=0;i<ii[n];i++) { 3341cf9b237SStefano Zampini if (PetscLikely(PetscAbsScalar(a[i]) > PETSC_SMALL)) nnz++; 3351cf9b237SStefano Zampini } 3361cf9b237SStefano Zampini ierr = PetscMalloc1(n+1,&bii);CHKERRQ(ierr); 3371cf9b237SStefano Zampini ierr = PetscMalloc1(nnz,&bij);CHKERRQ(ierr); 3381cf9b237SStefano Zampini ierr = PetscMalloc1(nnz,&bdata);CHKERRQ(ierr); 3391cf9b237SStefano Zampini nnz = 0; 3401cf9b237SStefano Zampini bii[0] = 0; 3411cf9b237SStefano Zampini for (i=0;i<n;i++) { 3421cf9b237SStefano Zampini PetscInt j; 3431cf9b237SStefano Zampini for (j=ii[i];j<ii[i+1];j++) { 3441cf9b237SStefano Zampini PetscScalar entry = a[j]; 3451cf9b237SStefano Zampini if (PetscLikely(PetscAbsScalar(entry) > PETSC_SMALL) || ij[j] == i) { 3461cf9b237SStefano Zampini bij[nnz] = ij[j]; 3471cf9b237SStefano Zampini bdata[nnz] = entry; 3481cf9b237SStefano Zampini nnz++; 3491cf9b237SStefano Zampini } 3501cf9b237SStefano Zampini } 3511cf9b237SStefano Zampini bii[i+1] = nnz; 3521cf9b237SStefano Zampini } 3531cf9b237SStefano Zampini ierr = MatSeqAIJRestoreArray(A,&a);CHKERRQ(ierr); 3541cf9b237SStefano Zampini ierr = MatCreateSeqAIJWithArrays(PetscObjectComm((PetscObject)A),n,m,bii,bij,bdata,&Bt);CHKERRQ(ierr); 3551cf9b237SStefano Zampini ierr = MatRestoreRowIJ(A,0,PETSC_FALSE,PETSC_FALSE,&n,&ii,&ij,&flg_row);CHKERRQ(ierr); 3561cf9b237SStefano Zampini { 3571cf9b237SStefano Zampini Mat_SeqAIJ *b = (Mat_SeqAIJ*)(Bt->data); 3581cf9b237SStefano Zampini b->free_a = PETSC_TRUE; 3591cf9b237SStefano Zampini b->free_ij = PETSC_TRUE; 3601cf9b237SStefano Zampini } 3611cf9b237SStefano Zampini *B = Bt; 3621cf9b237SStefano Zampini PetscFunctionReturn(0); 3631cf9b237SStefano Zampini } 3641cf9b237SStefano Zampini 365674ae819SStefano Zampini #undef __FUNCT__ 3664f1b2e48SStefano Zampini #define __FUNCT__ "MatDetectDisconnectedComponents" 3674f1b2e48SStefano Zampini PetscErrorCode MatDetectDisconnectedComponents(Mat A, PetscBool filter, PetscInt *ncc, IS* cc[]) 3684f1b2e48SStefano Zampini { 3694f1b2e48SStefano Zampini Mat B; 3704f1b2e48SStefano Zampini IS is_dummy,*cc_n; 3714f1b2e48SStefano Zampini ISLocalToGlobalMapping l2gmap_dummy; 3724f1b2e48SStefano Zampini PCBDDCGraph graph; 3734f1b2e48SStefano Zampini PetscInt i,n; 3744f1b2e48SStefano Zampini PetscInt *xadj,*adjncy; 3754f1b2e48SStefano Zampini PetscInt *xadj_filtered,*adjncy_filtered; 3764f1b2e48SStefano Zampini PetscBool flg_row,isseqaij; 3774f1b2e48SStefano Zampini PetscErrorCode ierr; 3784f1b2e48SStefano Zampini 3794f1b2e48SStefano Zampini PetscFunctionBegin; 38063c961adSStefano Zampini if (!A->rmap->N || !A->cmap->N) { 38163c961adSStefano Zampini *ncc = 0; 38263c961adSStefano Zampini *cc = NULL; 38363c961adSStefano Zampini PetscFunctionReturn(0); 38463c961adSStefano Zampini } 3854f1b2e48SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)A,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 3864f1b2e48SStefano Zampini if (!isseqaij && filter) { 3871cf9b237SStefano Zampini PetscBool isseqdense; 3881cf9b237SStefano Zampini 3891cf9b237SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)A,MATSEQDENSE,&isseqdense);CHKERRQ(ierr); 3901cf9b237SStefano Zampini if (!isseqdense) { 3914f1b2e48SStefano Zampini ierr = MatConvert(A,MATSEQAIJ,MAT_INITIAL_MATRIX,&B);CHKERRQ(ierr); 3921cf9b237SStefano Zampini } else { /* TODO: rectangular case and LDA */ 3931cf9b237SStefano Zampini PetscScalar *array; 3941cf9b237SStefano Zampini PetscReal chop=1.e-6; 3951cf9b237SStefano Zampini 3961cf9b237SStefano Zampini ierr = MatDuplicate(A,MAT_COPY_VALUES,&B);CHKERRQ(ierr); 3971cf9b237SStefano Zampini ierr = MatDenseGetArray(B,&array);CHKERRQ(ierr); 3981cf9b237SStefano Zampini ierr = MatGetSize(B,&n,NULL);CHKERRQ(ierr); 3991cf9b237SStefano Zampini for (i=0;i<n;i++) { 4001cf9b237SStefano Zampini PetscInt j; 4011cf9b237SStefano Zampini for (j=i+1;j<n;j++) { 4021cf9b237SStefano Zampini PetscReal thresh = chop*(PetscAbsScalar(array[i*(n+1)])+PetscAbsScalar(array[j*(n+1)])); 4031cf9b237SStefano Zampini if (PetscAbsScalar(array[i*n+j]) < thresh) array[i*n+j] = 0.; 4041cf9b237SStefano Zampini if (PetscAbsScalar(array[j*n+i]) < thresh) array[j*n+i] = 0.; 4051cf9b237SStefano Zampini } 4061cf9b237SStefano Zampini } 4071cf9b237SStefano Zampini ierr = MatDenseRestoreArray(B,&array);CHKERRQ(ierr); 4089d54b7f4SStefano Zampini ierr = MatConvert(B,MATSEQAIJ,MAT_INPLACE_MATRIX,&B);CHKERRQ(ierr); 4091cf9b237SStefano Zampini } 4104f1b2e48SStefano Zampini } else { 4114f1b2e48SStefano Zampini B = A; 4124f1b2e48SStefano Zampini } 4134f1b2e48SStefano Zampini ierr = MatGetRowIJ(B,0,PETSC_TRUE,PETSC_FALSE,&n,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 4144f1b2e48SStefano Zampini 4154f1b2e48SStefano Zampini /* if filter is true, then removes entries lower than PETSC_SMALL in magnitude */ 4164f1b2e48SStefano Zampini if (filter) { 4174f1b2e48SStefano Zampini PetscScalar *data; 4184f1b2e48SStefano Zampini PetscInt j,cum; 4194f1b2e48SStefano Zampini 4204f1b2e48SStefano Zampini ierr = PetscCalloc2(n+1,&xadj_filtered,xadj[n],&adjncy_filtered);CHKERRQ(ierr); 4214f1b2e48SStefano Zampini ierr = MatSeqAIJGetArray(B,&data);CHKERRQ(ierr); 4224f1b2e48SStefano Zampini cum = 0; 4234f1b2e48SStefano Zampini for (i=0;i<n;i++) { 4244f1b2e48SStefano Zampini PetscInt t; 4254f1b2e48SStefano Zampini 4264f1b2e48SStefano Zampini for (j=xadj[i];j<xadj[i+1];j++) { 4274f1b2e48SStefano Zampini if (PetscUnlikely(PetscAbsScalar(data[j]) < PETSC_SMALL)) { 4284f1b2e48SStefano Zampini continue; 4294f1b2e48SStefano Zampini } 4304f1b2e48SStefano Zampini adjncy_filtered[cum+xadj_filtered[i]++] = adjncy[j]; 4314f1b2e48SStefano Zampini } 4324f1b2e48SStefano Zampini t = xadj_filtered[i]; 4334f1b2e48SStefano Zampini xadj_filtered[i] = cum; 4344f1b2e48SStefano Zampini cum += t; 4354f1b2e48SStefano Zampini } 4364f1b2e48SStefano Zampini ierr = MatSeqAIJRestoreArray(B,&data);CHKERRQ(ierr); 4374f1b2e48SStefano Zampini } else { 4384f1b2e48SStefano Zampini xadj_filtered = NULL; 4394f1b2e48SStefano Zampini adjncy_filtered = NULL; 4404f1b2e48SStefano Zampini } 4414f1b2e48SStefano Zampini 4424f1b2e48SStefano Zampini /* compute local connected components using PCBDDCGraph */ 4434f1b2e48SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,n,0,1,&is_dummy);CHKERRQ(ierr); 4444f1b2e48SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(is_dummy,&l2gmap_dummy);CHKERRQ(ierr); 4454f1b2e48SStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 4464f1b2e48SStefano Zampini ierr = PCBDDCGraphCreate(&graph);CHKERRQ(ierr); 4474f1b2e48SStefano Zampini ierr = PCBDDCGraphInit(graph,l2gmap_dummy,n);CHKERRQ(ierr); 4484f1b2e48SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&l2gmap_dummy);CHKERRQ(ierr); 4494f1b2e48SStefano Zampini if (xadj_filtered) { 4504f1b2e48SStefano Zampini graph->xadj = xadj_filtered; 4514f1b2e48SStefano Zampini graph->adjncy = adjncy_filtered; 4524f1b2e48SStefano Zampini } else { 4534f1b2e48SStefano Zampini graph->xadj = xadj; 4544f1b2e48SStefano Zampini graph->adjncy = adjncy; 4554f1b2e48SStefano Zampini } 4564f1b2e48SStefano Zampini ierr = PCBDDCGraphSetUp(graph,1,NULL,NULL,0,NULL,NULL);CHKERRQ(ierr); 4574f1b2e48SStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(graph);CHKERRQ(ierr); 4584f1b2e48SStefano Zampini /* partial clean up */ 4594f1b2e48SStefano Zampini ierr = PetscFree2(xadj_filtered,adjncy_filtered);CHKERRQ(ierr); 4604f1b2e48SStefano Zampini ierr = MatRestoreRowIJ(B,0,PETSC_TRUE,PETSC_FALSE,&n,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 4611cf9b237SStefano Zampini if (A != B) { 4624f1b2e48SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 4634f1b2e48SStefano Zampini } 4644f1b2e48SStefano Zampini 4654f1b2e48SStefano Zampini /* get back data */ 4661cf9b237SStefano Zampini if (ncc) *ncc = graph->ncc; 4671cf9b237SStefano Zampini if (cc) { 4684f1b2e48SStefano Zampini ierr = PetscMalloc1(graph->ncc,&cc_n);CHKERRQ(ierr); 4694f1b2e48SStefano Zampini for (i=0;i<graph->ncc;i++) { 4704f1b2e48SStefano 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); 4714f1b2e48SStefano Zampini } 4724f1b2e48SStefano Zampini *cc = cc_n; 4731cf9b237SStefano Zampini } 4744f1b2e48SStefano Zampini /* clean up graph */ 4754f1b2e48SStefano Zampini graph->xadj = 0; 4764f1b2e48SStefano Zampini graph->adjncy = 0; 4774f1b2e48SStefano Zampini ierr = PCBDDCGraphDestroy(&graph);CHKERRQ(ierr); 4784f1b2e48SStefano Zampini PetscFunctionReturn(0); 4794f1b2e48SStefano Zampini } 4804f1b2e48SStefano Zampini 4814f1b2e48SStefano Zampini #undef __FUNCT__ 4825408967cSStefano Zampini #define __FUNCT__ "PCBDDCBenignCheck" 4835408967cSStefano Zampini PetscErrorCode PCBDDCBenignCheck(PC pc, IS zerodiag) 4845408967cSStefano Zampini { 4855408967cSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 4865408967cSStefano Zampini PC_IS* pcis = (PC_IS*)(pc->data); 487dee84bffSStefano Zampini IS dirIS = NULL; 4884f1b2e48SStefano Zampini PetscInt i; 4895408967cSStefano Zampini PetscErrorCode ierr; 4905408967cSStefano Zampini 4915408967cSStefano Zampini PetscFunctionBegin; 492dee84bffSStefano Zampini ierr = PCBDDCGraphGetDirichletDofs(pcbddc->mat_graph,&dirIS);CHKERRQ(ierr); 4935408967cSStefano Zampini if (zerodiag) { 4945408967cSStefano Zampini Mat A; 4955408967cSStefano Zampini Vec vec3_N; 4965408967cSStefano Zampini PetscScalar *vals; 4975408967cSStefano Zampini const PetscInt *idxs; 498d12d3064SStefano Zampini PetscInt nz,*count; 4995408967cSStefano Zampini 5005408967cSStefano Zampini /* p0 */ 5015408967cSStefano Zampini ierr = VecSet(pcis->vec1_N,0.);CHKERRQ(ierr); 5025408967cSStefano Zampini ierr = PetscMalloc1(pcis->n,&vals);CHKERRQ(ierr); 5035408967cSStefano Zampini ierr = ISGetLocalSize(zerodiag,&nz);CHKERRQ(ierr); 5045408967cSStefano Zampini ierr = ISGetIndices(zerodiag,&idxs);CHKERRQ(ierr); 5054f1b2e48SStefano Zampini for (i=0;i<nz;i++) vals[i] = 1.; 5065408967cSStefano Zampini ierr = VecSetValues(pcis->vec1_N,nz,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 5075408967cSStefano Zampini ierr = VecAssemblyBegin(pcis->vec1_N);CHKERRQ(ierr); 5085408967cSStefano Zampini ierr = VecAssemblyEnd(pcis->vec1_N);CHKERRQ(ierr); 5095408967cSStefano Zampini /* v_I */ 5105408967cSStefano Zampini ierr = VecSetRandom(pcis->vec2_N,NULL);CHKERRQ(ierr); 5115408967cSStefano Zampini for (i=0;i<nz;i++) vals[i] = 0.; 5125408967cSStefano Zampini ierr = VecSetValues(pcis->vec2_N,nz,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 5135408967cSStefano Zampini ierr = ISRestoreIndices(zerodiag,&idxs);CHKERRQ(ierr); 5145408967cSStefano Zampini ierr = ISGetIndices(pcis->is_B_local,&idxs);CHKERRQ(ierr); 5155408967cSStefano Zampini for (i=0;i<pcis->n_B;i++) vals[i] = 0.; 5165408967cSStefano Zampini ierr = VecSetValues(pcis->vec2_N,pcis->n_B,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 5175408967cSStefano Zampini ierr = ISRestoreIndices(pcis->is_B_local,&idxs);CHKERRQ(ierr); 5185408967cSStefano Zampini if (dirIS) { 5195408967cSStefano Zampini PetscInt n; 5205408967cSStefano Zampini 5215408967cSStefano Zampini ierr = ISGetLocalSize(dirIS,&n);CHKERRQ(ierr); 5225408967cSStefano Zampini ierr = ISGetIndices(dirIS,&idxs);CHKERRQ(ierr); 5235408967cSStefano Zampini for (i=0;i<n;i++) vals[i] = 0.; 5245408967cSStefano Zampini ierr = VecSetValues(pcis->vec2_N,n,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 5255408967cSStefano Zampini ierr = ISRestoreIndices(dirIS,&idxs);CHKERRQ(ierr); 5265408967cSStefano Zampini } 5275408967cSStefano Zampini ierr = VecAssemblyBegin(pcis->vec2_N);CHKERRQ(ierr); 5285408967cSStefano Zampini ierr = VecAssemblyEnd(pcis->vec2_N);CHKERRQ(ierr); 5295408967cSStefano Zampini ierr = VecDuplicate(pcis->vec1_N,&vec3_N);CHKERRQ(ierr); 5305408967cSStefano Zampini ierr = VecSet(vec3_N,0.);CHKERRQ(ierr); 5315408967cSStefano Zampini ierr = MatISGetLocalMat(pc->mat,&A);CHKERRQ(ierr); 5325408967cSStefano Zampini ierr = MatMult(A,pcis->vec1_N,vec3_N);CHKERRQ(ierr); 5335408967cSStefano Zampini ierr = VecDot(vec3_N,pcis->vec2_N,&vals[0]);CHKERRQ(ierr); 534fbfcb133SStefano Zampini if (PetscAbsScalar(vals[0]) > 1.e-1) { 535b9b0e38cSStefano Zampini SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_SUP,"Benign trick can not be applied! b(v_I,p_0) = %1.6e (should be numerically 0.)",PetscAbsScalar(vals[0])); 5365408967cSStefano Zampini } 5375408967cSStefano Zampini ierr = PetscFree(vals);CHKERRQ(ierr); 5385408967cSStefano Zampini ierr = VecDestroy(&vec3_N);CHKERRQ(ierr); 539d12d3064SStefano Zampini 540d12d3064SStefano Zampini /* there should not be any pressure dofs lying on the interface */ 541d12d3064SStefano Zampini ierr = PetscCalloc1(pcis->n,&count);CHKERRQ(ierr); 542d12d3064SStefano Zampini ierr = ISGetIndices(pcis->is_B_local,&idxs);CHKERRQ(ierr); 543d12d3064SStefano Zampini for (i=0;i<pcis->n_B;i++) count[idxs[i]]++; 544d12d3064SStefano Zampini ierr = ISRestoreIndices(pcis->is_B_local,&idxs);CHKERRQ(ierr); 545d12d3064SStefano Zampini ierr = ISGetIndices(zerodiag,&idxs);CHKERRQ(ierr); 546d4d8cf7bSStefano Zampini for (i=0;i<nz;i++) { 547d12d3064SStefano Zampini if (count[idxs[i]]) { 548d12d3064SStefano Zampini SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_SUP,"Benign trick can not be applied! pressure dof %d is an interface dof",idxs[i]); 549d12d3064SStefano Zampini } 550d4d8cf7bSStefano Zampini } 551d12d3064SStefano Zampini ierr = ISRestoreIndices(zerodiag,&idxs);CHKERRQ(ierr); 552d12d3064SStefano Zampini ierr = PetscFree(count);CHKERRQ(ierr); 5535408967cSStefano Zampini } 554dee84bffSStefano Zampini ierr = ISDestroy(&dirIS);CHKERRQ(ierr); 5555408967cSStefano Zampini 5565408967cSStefano Zampini /* check PCBDDCBenignGetOrSetP0 */ 5575408967cSStefano Zampini ierr = VecSetRandom(pcis->vec1_global,NULL);CHKERRQ(ierr); 5584f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) pcbddc->benign_p0[i] = -PetscGlobalRank-i; 5595408967cSStefano Zampini ierr = PCBDDCBenignGetOrSetP0(pc,pcis->vec1_global,PETSC_FALSE);CHKERRQ(ierr); 5604f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) pcbddc->benign_p0[i] = 1; 5615408967cSStefano Zampini ierr = PCBDDCBenignGetOrSetP0(pc,pcis->vec1_global,PETSC_TRUE);CHKERRQ(ierr); 5624f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 5634f1b2e48SStefano Zampini if ((PetscInt)PetscRealPart(pcbddc->benign_p0[i]) != -PetscGlobalRank-i) { 5644f1b2e48SStefano 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); 5654f1b2e48SStefano Zampini } 5665408967cSStefano Zampini } 5675408967cSStefano Zampini PetscFunctionReturn(0); 5685408967cSStefano Zampini } 5695408967cSStefano Zampini 5705408967cSStefano Zampini #undef __FUNCT__ 571339f8db1SStefano Zampini #define __FUNCT__ "PCBDDCBenignDetectSaddlePoint" 572339f8db1SStefano Zampini PetscErrorCode PCBDDCBenignDetectSaddlePoint(PC pc, IS *zerodiaglocal) 573339f8db1SStefano Zampini { 574339f8db1SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 5754f1b2e48SStefano Zampini IS pressures,zerodiag,*zerodiag_subs; 576b0f5fe93SStefano Zampini PetscInt nz,n; 5771f4df5f7SStefano Zampini PetscInt *interior_dofs,n_interior_dofs; 5784f1b2e48SStefano Zampini PetscBool sorted,have_null,has_null_pressures,recompute_zerodiag; 579339f8db1SStefano Zampini PetscErrorCode ierr; 580339f8db1SStefano Zampini 581339f8db1SStefano Zampini PetscFunctionBegin; 5829f47a83aSStefano Zampini ierr = PetscSFDestroy(&pcbddc->benign_sf);CHKERRQ(ierr); 5839f47a83aSStefano Zampini ierr = MatDestroy(&pcbddc->benign_B0);CHKERRQ(ierr); 584a3df083aSStefano Zampini for (n=0;n<pcbddc->benign_n;n++) { 585a3df083aSStefano Zampini ierr = ISDestroy(&pcbddc->benign_zerodiag_subs[n]);CHKERRQ(ierr); 586a3df083aSStefano Zampini } 587a3df083aSStefano Zampini ierr = PetscFree(pcbddc->benign_zerodiag_subs);CHKERRQ(ierr); 588a3df083aSStefano Zampini pcbddc->benign_n = 0; 5894f1b2e48SStefano Zampini /* if a local info on dofs is present, assumes the last field is represented by "pressures" 5904f1b2e48SStefano Zampini otherwise, it uses only zerodiagonal dofs (ok if the pressure block is all zero; it could fail if it is not) 5914f1b2e48SStefano Zampini Checks if all the pressure dofs in each subdomain have a zero diagonal 5924f1b2e48SStefano Zampini If not, a change of basis on pressures is not needed 5934f1b2e48SStefano Zampini since the local Schur complements are SPD 5944f1b2e48SStefano Zampini */ 5954f1b2e48SStefano Zampini has_null_pressures = PETSC_TRUE; 5964f1b2e48SStefano Zampini have_null = PETSC_TRUE; 59740fa8d13SStefano Zampini if (pcbddc->n_ISForDofsLocal) { 5984f1b2e48SStefano Zampini PetscInt npl,*idxs,p = pcbddc->n_ISForDofsLocal-1; 5994f1b2e48SStefano Zampini 6004f1b2e48SStefano Zampini /* Dofs splitting for BDDC cannot have PETSC_COMM_SELF, so create a sequential IS */ 6014f1b2e48SStefano Zampini ierr = ISGetLocalSize(pcbddc->ISForDofsLocal[p],&npl);CHKERRQ(ierr); 6024f1b2e48SStefano Zampini ierr = ISGetIndices(pcbddc->ISForDofsLocal[p],(const PetscInt**)&idxs);CHKERRQ(ierr); 6034f1b2e48SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,npl,idxs,PETSC_COPY_VALUES,&pressures);CHKERRQ(ierr); 604ba14f8e3SStefano Zampini ierr = ISRestoreIndices(pcbddc->ISForDofsLocal[p],(const PetscInt**)&idxs);CHKERRQ(ierr); 60540fa8d13SStefano Zampini ierr = ISSorted(pressures,&sorted);CHKERRQ(ierr); 60640fa8d13SStefano Zampini if (!sorted) { 60740fa8d13SStefano Zampini ierr = ISSort(pressures);CHKERRQ(ierr); 60840fa8d13SStefano Zampini } 60940fa8d13SStefano Zampini } else { 61040fa8d13SStefano Zampini pressures = NULL; 61140fa8d13SStefano Zampini } 61297d764eeSStefano Zampini /* pcis has not been setup yet, so get the local size from the subdomain matrix */ 61397d764eeSStefano Zampini ierr = MatGetLocalSize(pcbddc->local_mat,&n,NULL);CHKERRQ(ierr); 61427b6a85dSStefano Zampini if (!n) pcbddc->benign_change_explicit = PETSC_TRUE; 61597d764eeSStefano Zampini ierr = MatFindZeroDiagonals(pcbddc->local_mat,&zerodiag);CHKERRQ(ierr); 616339f8db1SStefano Zampini ierr = ISSorted(zerodiag,&sorted);CHKERRQ(ierr); 617339f8db1SStefano Zampini if (!sorted) { 618339f8db1SStefano Zampini ierr = ISSort(zerodiag);CHKERRQ(ierr); 619339f8db1SStefano Zampini } 620339f8db1SStefano Zampini ierr = ISGetLocalSize(zerodiag,&nz);CHKERRQ(ierr); 6214f1b2e48SStefano Zampini if (!nz) { 6224f1b2e48SStefano Zampini if (n) have_null = PETSC_FALSE; 6234f1b2e48SStefano Zampini has_null_pressures = PETSC_FALSE; 62440fa8d13SStefano Zampini ierr = ISDestroy(&zerodiag);CHKERRQ(ierr); 62540fa8d13SStefano Zampini } 6264f1b2e48SStefano Zampini recompute_zerodiag = PETSC_FALSE; 6274f1b2e48SStefano Zampini /* in case disconnected subdomains info is present, split the pressures accordingly (otherwise the benign trick could fail) */ 6284f1b2e48SStefano Zampini zerodiag_subs = NULL; 6294f1b2e48SStefano Zampini pcbddc->benign_n = 0; 6301f4df5f7SStefano Zampini n_interior_dofs = 0; 6311f4df5f7SStefano Zampini interior_dofs = NULL; 6321f4df5f7SStefano Zampini if (pcbddc->current_level) { /* need to compute interior nodes */ 6331f4df5f7SStefano Zampini PetscInt n,i,j; 6341f4df5f7SStefano Zampini PetscInt n_neigh,*neigh,*n_shared,**shared; 6351f4df5f7SStefano Zampini PetscInt *iwork; 6361f4df5f7SStefano Zampini 6371f4df5f7SStefano Zampini ierr = ISLocalToGlobalMappingGetSize(pc->pmat->rmap->mapping,&n);CHKERRQ(ierr); 6381f4df5f7SStefano Zampini ierr = ISLocalToGlobalMappingGetInfo(pc->pmat->rmap->mapping,&n_neigh,&neigh,&n_shared,&shared);CHKERRQ(ierr); 6391f4df5f7SStefano Zampini ierr = PetscCalloc1(n,&iwork);CHKERRQ(ierr); 6401f4df5f7SStefano Zampini ierr = PetscMalloc1(n,&interior_dofs);CHKERRQ(ierr); 6411f4df5f7SStefano Zampini for (i=0;i<n_neigh;i++) 6421f4df5f7SStefano Zampini for (j=0;j<n_shared[i];j++) 6431f4df5f7SStefano Zampini iwork[shared[i][j]] += 1; 6441f4df5f7SStefano Zampini for (i=0;i<n;i++) 6451f4df5f7SStefano Zampini if (!iwork[i]) 6461f4df5f7SStefano Zampini interior_dofs[n_interior_dofs++] = i; 6471f4df5f7SStefano Zampini ierr = PetscFree(iwork);CHKERRQ(ierr); 6481f4df5f7SStefano Zampini ierr = ISLocalToGlobalMappingRestoreInfo(pc->pmat->rmap->mapping,&n_neigh,&neigh,&n_shared,&shared);CHKERRQ(ierr); 6491f4df5f7SStefano Zampini } 6504f1b2e48SStefano Zampini if (has_null_pressures) { 6514f1b2e48SStefano Zampini IS *subs; 6521f4df5f7SStefano Zampini PetscInt nsubs,i,j,nl; 6531f4df5f7SStefano Zampini const PetscInt *idxs; 6541f4df5f7SStefano Zampini PetscScalar *array; 6551f4df5f7SStefano Zampini Vec *work; 6561f4df5f7SStefano Zampini Mat_IS* matis = (Mat_IS*)(pc->pmat->data); 6574f1b2e48SStefano Zampini 6584f1b2e48SStefano Zampini subs = pcbddc->local_subs; 6594f1b2e48SStefano Zampini nsubs = pcbddc->n_local_subs; 6601f4df5f7SStefano Zampini /* these vectors are needed to check if the constant on pressures is in the kernel of the local operator B (i.e. B(v_I,p0) should be zero) */ 6611f4df5f7SStefano Zampini if (pcbddc->current_level) { 6621f4df5f7SStefano Zampini ierr = VecDuplicateVecs(matis->y,2,&work);CHKERRQ(ierr); 6631f4df5f7SStefano Zampini ierr = ISGetLocalSize(zerodiag,&nl);CHKERRQ(ierr); 6641f4df5f7SStefano Zampini ierr = ISGetIndices(zerodiag,&idxs);CHKERRQ(ierr); 6651f4df5f7SStefano Zampini /* work[0] = 1_p */ 6661f4df5f7SStefano Zampini ierr = VecSet(work[0],0.);CHKERRQ(ierr); 6671f4df5f7SStefano Zampini ierr = VecGetArray(work[0],&array);CHKERRQ(ierr); 6681f4df5f7SStefano Zampini for (j=0;j<nl;j++) array[idxs[j]] = 1.; 6691f4df5f7SStefano Zampini ierr = VecRestoreArray(work[0],&array);CHKERRQ(ierr); 6701f4df5f7SStefano Zampini /* work[0] = 1_v */ 6711f4df5f7SStefano Zampini ierr = VecSet(work[1],1.);CHKERRQ(ierr); 6721f4df5f7SStefano Zampini ierr = VecGetArray(work[1],&array);CHKERRQ(ierr); 6731f4df5f7SStefano Zampini for (j=0;j<nl;j++) array[idxs[j]] = 0.; 6741f4df5f7SStefano Zampini ierr = VecRestoreArray(work[1],&array);CHKERRQ(ierr); 6751f4df5f7SStefano Zampini ierr = ISRestoreIndices(zerodiag,&idxs);CHKERRQ(ierr); 6761f4df5f7SStefano Zampini } 6774f1b2e48SStefano Zampini if (nsubs > 1) { 6784f1b2e48SStefano Zampini ierr = PetscCalloc1(nsubs,&zerodiag_subs);CHKERRQ(ierr); 6794f1b2e48SStefano Zampini for (i=0;i<nsubs;i++) { 6804f1b2e48SStefano Zampini ISLocalToGlobalMapping l2g; 6814f1b2e48SStefano Zampini IS t_zerodiag_subs; 6824f1b2e48SStefano Zampini PetscInt nl; 6834f1b2e48SStefano Zampini 6844f1b2e48SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(subs[i],&l2g);CHKERRQ(ierr); 6854f1b2e48SStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(l2g,IS_GTOLM_DROP,zerodiag,&t_zerodiag_subs);CHKERRQ(ierr); 6864f1b2e48SStefano Zampini ierr = ISGetLocalSize(t_zerodiag_subs,&nl);CHKERRQ(ierr); 6874f1b2e48SStefano Zampini if (nl) { 6884f1b2e48SStefano Zampini PetscBool valid = PETSC_TRUE; 6894f1b2e48SStefano Zampini 6901f4df5f7SStefano Zampini if (pcbddc->current_level) { 6911f4df5f7SStefano Zampini ierr = VecSet(matis->x,0);CHKERRQ(ierr); 6921f4df5f7SStefano Zampini ierr = ISGetLocalSize(subs[i],&nl);CHKERRQ(ierr); 6931f4df5f7SStefano Zampini ierr = ISGetIndices(subs[i],&idxs);CHKERRQ(ierr); 6941f4df5f7SStefano Zampini ierr = VecGetArray(matis->x,&array);CHKERRQ(ierr); 6951f4df5f7SStefano Zampini for (j=0;j<nl;j++) array[idxs[j]] = 1.; 6961f4df5f7SStefano Zampini ierr = VecRestoreArray(matis->x,&array);CHKERRQ(ierr); 6971f4df5f7SStefano Zampini ierr = ISRestoreIndices(subs[i],&idxs);CHKERRQ(ierr); 6981f4df5f7SStefano Zampini ierr = VecPointwiseMult(matis->x,work[0],matis->x);CHKERRQ(ierr); 6991f4df5f7SStefano Zampini ierr = MatMult(matis->A,matis->x,matis->y);CHKERRQ(ierr); 7001f4df5f7SStefano Zampini ierr = VecPointwiseMult(matis->y,work[1],matis->y);CHKERRQ(ierr); 7011f4df5f7SStefano Zampini ierr = VecGetArray(matis->y,&array);CHKERRQ(ierr); 7021f4df5f7SStefano Zampini for (j=0;j<n_interior_dofs;j++) { 7031f4df5f7SStefano Zampini if (PetscAbsScalar(array[interior_dofs[j]]) > PETSC_SMALL) { 7041f4df5f7SStefano Zampini valid = PETSC_FALSE; 7051f4df5f7SStefano Zampini break; 7061f4df5f7SStefano Zampini } 7071f4df5f7SStefano Zampini } 7081f4df5f7SStefano Zampini ierr = VecRestoreArray(matis->y,&array);CHKERRQ(ierr); 7091f4df5f7SStefano Zampini } 7101f4df5f7SStefano Zampini if (valid && pcbddc->NeumannBoundariesLocal) { 7111f4df5f7SStefano Zampini IS t_bc; 7121f4df5f7SStefano Zampini PetscInt nzb; 7131f4df5f7SStefano Zampini 7141f4df5f7SStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(l2g,IS_GTOLM_DROP,pcbddc->NeumannBoundariesLocal,&t_bc);CHKERRQ(ierr); 7151f4df5f7SStefano Zampini ierr = ISGetLocalSize(t_bc,&nzb);CHKERRQ(ierr); 7161f4df5f7SStefano Zampini ierr = ISDestroy(&t_bc);CHKERRQ(ierr); 7171f4df5f7SStefano Zampini if (nzb) valid = PETSC_FALSE; 7181f4df5f7SStefano Zampini } 7191f4df5f7SStefano Zampini if (valid && pressures) { 7204f1b2e48SStefano Zampini IS t_pressure_subs; 7214f1b2e48SStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(l2g,IS_GTOLM_DROP,pressures,&t_pressure_subs);CHKERRQ(ierr); 7224f1b2e48SStefano Zampini ierr = ISEqual(t_pressure_subs,t_zerodiag_subs,&valid);CHKERRQ(ierr); 7234f1b2e48SStefano Zampini ierr = ISDestroy(&t_pressure_subs);CHKERRQ(ierr); 7244f1b2e48SStefano Zampini } 7254f1b2e48SStefano Zampini if (valid) { 7264f1b2e48SStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(l2g,t_zerodiag_subs,&zerodiag_subs[pcbddc->benign_n]);CHKERRQ(ierr); 7274f1b2e48SStefano Zampini pcbddc->benign_n++; 7284f1b2e48SStefano Zampini } else { 7294f1b2e48SStefano Zampini recompute_zerodiag = PETSC_TRUE; 7304f1b2e48SStefano Zampini } 7314f1b2e48SStefano Zampini } 7324f1b2e48SStefano Zampini ierr = ISDestroy(&t_zerodiag_subs);CHKERRQ(ierr); 7334f1b2e48SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&l2g);CHKERRQ(ierr); 7344f1b2e48SStefano Zampini } 7354f1b2e48SStefano Zampini } else { /* there's just one subdomain (or zero if they have not been detected */ 7364f1b2e48SStefano Zampini PetscBool valid = PETSC_TRUE; 7371f4df5f7SStefano Zampini 7381f4df5f7SStefano Zampini if (pcbddc->NeumannBoundariesLocal) { 7391f4df5f7SStefano Zampini PetscInt nzb; 7401f4df5f7SStefano Zampini ierr = ISGetLocalSize(pcbddc->NeumannBoundariesLocal,&nzb);CHKERRQ(ierr); 7411f4df5f7SStefano Zampini if (nzb) valid = PETSC_FALSE; 7421f4df5f7SStefano Zampini } 7431f4df5f7SStefano Zampini if (valid && pressures) { 7444f1b2e48SStefano Zampini ierr = ISEqual(pressures,zerodiag,&valid);CHKERRQ(ierr); 7454f1b2e48SStefano Zampini } 7461f4df5f7SStefano Zampini if (valid && pcbddc->current_level) { 7471f4df5f7SStefano Zampini ierr = MatMult(matis->A,work[0],matis->x);CHKERRQ(ierr); 7481f4df5f7SStefano Zampini ierr = VecPointwiseMult(matis->x,work[1],matis->x);CHKERRQ(ierr); 7491f4df5f7SStefano Zampini ierr = VecGetArray(matis->x,&array);CHKERRQ(ierr); 7501f4df5f7SStefano Zampini for (j=0;j<n_interior_dofs;j++) { 7511f4df5f7SStefano Zampini if (PetscAbsScalar(array[interior_dofs[j]]) > PETSC_SMALL) { 7521f4df5f7SStefano Zampini valid = PETSC_FALSE; 7531f4df5f7SStefano Zampini break; 7541f4df5f7SStefano Zampini } 7551f4df5f7SStefano Zampini } 7561f4df5f7SStefano Zampini ierr = VecRestoreArray(matis->x,&array);CHKERRQ(ierr); 7571f4df5f7SStefano Zampini } 7584f1b2e48SStefano Zampini if (valid) { 7594f1b2e48SStefano Zampini pcbddc->benign_n = 1; 760ca92afb2SStefano Zampini ierr = PetscMalloc1(pcbddc->benign_n,&zerodiag_subs);CHKERRQ(ierr); 7614f1b2e48SStefano Zampini ierr = PetscObjectReference((PetscObject)zerodiag);CHKERRQ(ierr); 7624f1b2e48SStefano Zampini zerodiag_subs[0] = zerodiag; 7634f1b2e48SStefano Zampini } 7644f1b2e48SStefano Zampini } 7651f4df5f7SStefano Zampini if (pcbddc->current_level) { 7661f4df5f7SStefano Zampini ierr = VecDestroyVecs(2,&work);CHKERRQ(ierr); 7674f1b2e48SStefano Zampini } 7681f4df5f7SStefano Zampini } 7691f4df5f7SStefano Zampini ierr = PetscFree(interior_dofs);CHKERRQ(ierr); 7704f1b2e48SStefano Zampini 7714f1b2e48SStefano Zampini if (!pcbddc->benign_n) { 772b9b0e38cSStefano Zampini PetscInt n; 773b9b0e38cSStefano Zampini 7744f1b2e48SStefano Zampini ierr = ISDestroy(&zerodiag);CHKERRQ(ierr); 7754f1b2e48SStefano Zampini recompute_zerodiag = PETSC_FALSE; 776b9b0e38cSStefano Zampini ierr = MatGetLocalSize(pcbddc->local_mat,&n,NULL);CHKERRQ(ierr); 777b9b0e38cSStefano Zampini if (n) { 7784f1b2e48SStefano Zampini has_null_pressures = PETSC_FALSE; 7794f1b2e48SStefano Zampini have_null = PETSC_FALSE; 7804f1b2e48SStefano Zampini } 781b9b0e38cSStefano Zampini } 7824f1b2e48SStefano Zampini 7834f1b2e48SStefano Zampini /* final check for null pressures */ 7844f1b2e48SStefano Zampini if (zerodiag && pressures) { 7854f1b2e48SStefano Zampini PetscInt nz,np; 7864f1b2e48SStefano Zampini ierr = ISGetLocalSize(zerodiag,&nz);CHKERRQ(ierr); 7874f1b2e48SStefano Zampini ierr = ISGetLocalSize(pressures,&np);CHKERRQ(ierr); 7884f1b2e48SStefano Zampini if (nz != np) have_null = PETSC_FALSE; 7894f1b2e48SStefano Zampini } 7904f1b2e48SStefano Zampini 7914f1b2e48SStefano Zampini if (recompute_zerodiag) { 7924f1b2e48SStefano Zampini ierr = ISDestroy(&zerodiag);CHKERRQ(ierr); 7934f1b2e48SStefano Zampini if (pcbddc->benign_n == 1) { 7944f1b2e48SStefano Zampini ierr = PetscObjectReference((PetscObject)zerodiag_subs[0]);CHKERRQ(ierr); 7954f1b2e48SStefano Zampini zerodiag = zerodiag_subs[0]; 7964f1b2e48SStefano Zampini } else { 7974f1b2e48SStefano Zampini PetscInt i,nzn,*new_idxs; 7984f1b2e48SStefano Zampini 7994f1b2e48SStefano Zampini nzn = 0; 8004f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 8014f1b2e48SStefano Zampini PetscInt ns; 8024f1b2e48SStefano Zampini ierr = ISGetLocalSize(zerodiag_subs[i],&ns);CHKERRQ(ierr); 8034f1b2e48SStefano Zampini nzn += ns; 8044f1b2e48SStefano Zampini } 8054f1b2e48SStefano Zampini ierr = PetscMalloc1(nzn,&new_idxs);CHKERRQ(ierr); 8064f1b2e48SStefano Zampini nzn = 0; 8074f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 8084f1b2e48SStefano Zampini PetscInt ns,*idxs; 8094f1b2e48SStefano Zampini ierr = ISGetLocalSize(zerodiag_subs[i],&ns);CHKERRQ(ierr); 8104f1b2e48SStefano Zampini ierr = ISGetIndices(zerodiag_subs[i],(const PetscInt**)&idxs);CHKERRQ(ierr); 8114f1b2e48SStefano Zampini ierr = PetscMemcpy(new_idxs+nzn,idxs,ns*sizeof(PetscInt));CHKERRQ(ierr); 8124f1b2e48SStefano Zampini ierr = ISRestoreIndices(zerodiag_subs[i],(const PetscInt**)&idxs);CHKERRQ(ierr); 8134f1b2e48SStefano Zampini nzn += ns; 8144f1b2e48SStefano Zampini } 8154f1b2e48SStefano Zampini ierr = PetscSortInt(nzn,new_idxs);CHKERRQ(ierr); 8164f1b2e48SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,nzn,new_idxs,PETSC_OWN_POINTER,&zerodiag);CHKERRQ(ierr); 8174f1b2e48SStefano Zampini } 8184f1b2e48SStefano Zampini have_null = PETSC_FALSE; 8194f1b2e48SStefano Zampini } 8204f1b2e48SStefano Zampini 821b3afcdbeSStefano Zampini /* no-net-flux */ 822b3afcdbeSStefano Zampini if (pcbddc->benign_compute_nonetflux) { 823b3afcdbeSStefano Zampini Mat_IS* matis = (Mat_IS*)(pc->pmat->data); 824b3afcdbeSStefano Zampini MatNullSpace near_null_space; 825b3afcdbeSStefano Zampini ISLocalToGlobalMapping rmap; 826b3afcdbeSStefano Zampini Vec quad_vec; 827b3afcdbeSStefano Zampini PetscScalar *pvals; 828b3afcdbeSStefano Zampini PetscInt i,np,*dummyins; 8291f4df5f7SStefano Zampini IS isused = NULL; 8301f4df5f7SStefano Zampini PetscBool participate = PETSC_TRUE; 831b3afcdbeSStefano Zampini 832b3afcdbeSStefano Zampini /* create vector to hold quadrature weights */ 833b3afcdbeSStefano Zampini ierr = MatCreateVecs(pc->pmat,&quad_vec,NULL);CHKERRQ(ierr); 834b3afcdbeSStefano Zampini ierr = VecSet(quad_vec,0.0);CHKERRQ(ierr); 835b3afcdbeSStefano Zampini ierr = MatGetLocalToGlobalMapping(pc->pmat,&rmap,NULL);CHKERRQ(ierr); 836b3afcdbeSStefano Zampini ierr = VecSetLocalToGlobalMapping(quad_vec,rmap);CHKERRQ(ierr); 837b3afcdbeSStefano Zampini 838b3afcdbeSStefano Zampini /* compute B^{(i)T} * 1_p */ 839b3afcdbeSStefano Zampini np = 0; 8401f4df5f7SStefano Zampini if (pressures) { 8411f4df5f7SStefano Zampini isused = pressures; 8421f4df5f7SStefano Zampini } else { 8431f4df5f7SStefano Zampini isused = zerodiag; 8441f4df5f7SStefano Zampini } 8451f4df5f7SStefano Zampini if (isused) { 8461f4df5f7SStefano Zampini ierr = ISGetLocalSize(isused,&np);CHKERRQ(ierr); 847b3afcdbeSStefano Zampini } 848b3afcdbeSStefano Zampini ierr = PetscMalloc1(np,&pvals);CHKERRQ(ierr); 849b3afcdbeSStefano Zampini for (i=0;i<np;i++) pvals[i] = 1.; 850b3afcdbeSStefano Zampini ierr = VecSet(matis->x,0.);CHKERRQ(ierr); 851b3afcdbeSStefano Zampini if (np) { 852b3afcdbeSStefano Zampini const PetscInt *pidxs; 853b3afcdbeSStefano Zampini 8541f4df5f7SStefano Zampini if (isused) { 8551f4df5f7SStefano Zampini ierr = ISGetIndices(isused,&pidxs);CHKERRQ(ierr); 856b3afcdbeSStefano Zampini } 857b3afcdbeSStefano Zampini ierr = VecSetValues(matis->x,np,pidxs,pvals,INSERT_VALUES);CHKERRQ(ierr); 8581f4df5f7SStefano Zampini if (isused) { 8591f4df5f7SStefano Zampini ierr = ISRestoreIndices(isused,&pidxs);CHKERRQ(ierr); 860b3afcdbeSStefano Zampini } 861b3afcdbeSStefano Zampini } 862b3afcdbeSStefano Zampini ierr = VecAssemblyBegin(matis->x);CHKERRQ(ierr); 863b3afcdbeSStefano Zampini ierr = VecAssemblyEnd(matis->x);CHKERRQ(ierr); 864b3afcdbeSStefano Zampini ierr = MatMult(matis->A,matis->x,matis->y);CHKERRQ(ierr); 865b3afcdbeSStefano Zampini ierr = PetscFree(pvals);CHKERRQ(ierr); 8661f4df5f7SStefano Zampini if (!isused) participate = PETSC_FALSE; 867b3afcdbeSStefano Zampini /* decide which of the sharing ranks (per dof) has to insert the values (should just be a matter of having a different orientation) */ 868b3afcdbeSStefano Zampini ierr = MatISSetUpSF(pc->pmat);CHKERRQ(ierr); 869daf8a457SStefano Zampini for (i=0;i<matis->sf->nroots;i++) matis->sf_rootdata[i] = -1; 8701f4df5f7SStefano Zampini for (i=0;i<matis->sf->nleaves;i++) 8711f4df5f7SStefano Zampini if (participate) matis->sf_leafdata[i] = PetscGlobalRank; 8721f4df5f7SStefano Zampini else matis->sf_leafdata[i] = -1; 8731f4df5f7SStefano Zampini 874b3afcdbeSStefano Zampini ierr = PetscSFReduceBegin(matis->sf,MPIU_INT,matis->sf_leafdata,matis->sf_rootdata,MPI_MAX);CHKERRQ(ierr); 875b3afcdbeSStefano Zampini ierr = PetscSFReduceEnd(matis->sf,MPIU_INT,matis->sf_leafdata,matis->sf_rootdata,MPI_MAX);CHKERRQ(ierr); 876b3afcdbeSStefano Zampini ierr = PetscSFBcastBegin(matis->sf,MPIU_INT,matis->sf_rootdata,matis->sf_leafdata);CHKERRQ(ierr); 877b3afcdbeSStefano Zampini ierr = PetscSFBcastEnd(matis->sf,MPIU_INT,matis->sf_rootdata,matis->sf_leafdata);CHKERRQ(ierr); 878b3afcdbeSStefano Zampini ierr = VecGetArray(matis->y,&pvals);CHKERRQ(ierr); 879daf8a457SStefano Zampini for (i=0;i<matis->sf->nleaves;i++) { 880b3afcdbeSStefano Zampini if (PetscGlobalRank != matis->sf_leafdata[i]) { 881b3afcdbeSStefano Zampini pvals[i] = 0.; 882b3afcdbeSStefano Zampini } 883b3afcdbeSStefano Zampini } 884daf8a457SStefano Zampini ierr = PetscMalloc1(matis->sf->nleaves,&dummyins);CHKERRQ(ierr); 885daf8a457SStefano Zampini for (i=0;i<matis->sf->nleaves;i++) dummyins[i] = i; 886daf8a457SStefano Zampini ierr = VecSetValuesLocal(quad_vec,matis->sf->nleaves,dummyins,pvals,ADD_VALUES);CHKERRQ(ierr); 887b3afcdbeSStefano Zampini ierr = VecRestoreArray(matis->y,&pvals);CHKERRQ(ierr); 888b3afcdbeSStefano Zampini ierr = PetscFree(dummyins);CHKERRQ(ierr); 889b3afcdbeSStefano Zampini 890b3afcdbeSStefano Zampini /* assembly quadrature vec and attach near null space to pmat */ 891b3afcdbeSStefano Zampini ierr = VecAssemblyBegin(quad_vec);CHKERRQ(ierr); 892b3afcdbeSStefano Zampini ierr = VecAssemblyEnd(quad_vec);CHKERRQ(ierr); 893b3afcdbeSStefano Zampini ierr = VecNormalize(quad_vec,NULL);CHKERRQ(ierr); 894b3afcdbeSStefano Zampini ierr = MatNullSpaceCreate(PetscObjectComm((PetscObject)pc),PETSC_FALSE,1,&quad_vec,&near_null_space);CHKERRQ(ierr); 895b3afcdbeSStefano Zampini ierr = VecDestroy(&quad_vec);CHKERRQ(ierr); 896b3afcdbeSStefano Zampini ierr = MatSetNearNullSpace(pc->pmat,near_null_space);CHKERRQ(ierr); 897b3afcdbeSStefano Zampini ierr = MatNullSpaceDestroy(&near_null_space);CHKERRQ(ierr); 898b3afcdbeSStefano Zampini } 899b3afcdbeSStefano Zampini 900b3afcdbeSStefano Zampini /* change of basis and p0 dofs */ 9014f1b2e48SStefano Zampini if (has_null_pressures) { 9024f1b2e48SStefano Zampini IS zerodiagc; 9034f1b2e48SStefano Zampini const PetscInt *idxs,*idxsc; 9044f1b2e48SStefano Zampini PetscInt i,s,*nnz; 9054f1b2e48SStefano Zampini 9064f1b2e48SStefano Zampini ierr = ISGetLocalSize(zerodiag,&nz);CHKERRQ(ierr); 907339f8db1SStefano Zampini ierr = ISComplement(zerodiag,0,n,&zerodiagc);CHKERRQ(ierr); 908339f8db1SStefano Zampini ierr = ISGetIndices(zerodiagc,&idxsc);CHKERRQ(ierr); 909339f8db1SStefano Zampini /* local change of basis for pressures */ 910339f8db1SStefano Zampini ierr = MatDestroy(&pcbddc->benign_change);CHKERRQ(ierr); 91197d764eeSStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)pcbddc->local_mat),&pcbddc->benign_change);CHKERRQ(ierr); 912339f8db1SStefano Zampini ierr = MatSetType(pcbddc->benign_change,MATAIJ);CHKERRQ(ierr); 913339f8db1SStefano Zampini ierr = MatSetSizes(pcbddc->benign_change,n,n,PETSC_DECIDE,PETSC_DECIDE);CHKERRQ(ierr); 914339f8db1SStefano Zampini ierr = PetscMalloc1(n,&nnz);CHKERRQ(ierr); 9154f1b2e48SStefano Zampini for (i=0;i<n-nz;i++) nnz[idxsc[i]] = 1; /* identity on velocities plus pressure dofs for non-singular subdomains */ 9164f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 9174f1b2e48SStefano Zampini PetscInt nzs,j; 9184f1b2e48SStefano Zampini 9194f1b2e48SStefano Zampini ierr = ISGetLocalSize(zerodiag_subs[i],&nzs);CHKERRQ(ierr); 9204f1b2e48SStefano Zampini ierr = ISGetIndices(zerodiag_subs[i],&idxs);CHKERRQ(ierr); 9214f1b2e48SStefano Zampini for (j=0;j<nzs-1;j++) nnz[idxs[j]] = 2; /* change on pressures */ 9224f1b2e48SStefano Zampini nnz[idxs[nzs-1]] = nzs; /* last local pressure dof in subdomain */ 9234f1b2e48SStefano Zampini ierr = ISRestoreIndices(zerodiag_subs[i],&idxs);CHKERRQ(ierr); 9244f1b2e48SStefano Zampini } 925339f8db1SStefano Zampini ierr = MatSeqAIJSetPreallocation(pcbddc->benign_change,0,nnz);CHKERRQ(ierr); 926339f8db1SStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 927339f8db1SStefano Zampini /* set identity on velocities */ 928339f8db1SStefano Zampini for (i=0;i<n-nz;i++) { 929339f8db1SStefano Zampini ierr = MatSetValue(pcbddc->benign_change,idxsc[i],idxsc[i],1.,INSERT_VALUES);CHKERRQ(ierr); 930339f8db1SStefano Zampini } 9314f1b2e48SStefano Zampini ierr = ISRestoreIndices(zerodiagc,&idxsc);CHKERRQ(ierr); 9324f1b2e48SStefano Zampini ierr = ISDestroy(&zerodiagc);CHKERRQ(ierr); 9339f47a83aSStefano Zampini ierr = PetscFree3(pcbddc->benign_p0_lidx,pcbddc->benign_p0_gidx,pcbddc->benign_p0);CHKERRQ(ierr); 9344f1b2e48SStefano 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); 935339f8db1SStefano Zampini /* set change on pressures */ 9364f1b2e48SStefano Zampini for (s=0;s<pcbddc->benign_n;s++) { 9374f1b2e48SStefano Zampini PetscScalar *array; 9384f1b2e48SStefano Zampini PetscInt nzs; 9394f1b2e48SStefano Zampini 9404f1b2e48SStefano Zampini ierr = ISGetLocalSize(zerodiag_subs[s],&nzs);CHKERRQ(ierr); 9414f1b2e48SStefano Zampini ierr = ISGetIndices(zerodiag_subs[s],&idxs);CHKERRQ(ierr); 9424f1b2e48SStefano Zampini for (i=0;i<nzs-1;i++) { 943339f8db1SStefano Zampini PetscScalar vals[2]; 944339f8db1SStefano Zampini PetscInt cols[2]; 945339f8db1SStefano Zampini 946339f8db1SStefano Zampini cols[0] = idxs[i]; 9474f1b2e48SStefano Zampini cols[1] = idxs[nzs-1]; 948339f8db1SStefano Zampini vals[0] = 1.; 949b0f5fe93SStefano Zampini vals[1] = 1.; 9504f1b2e48SStefano Zampini ierr = MatSetValues(pcbddc->benign_change,1,cols,2,cols,vals,INSERT_VALUES);CHKERRQ(ierr); 951339f8db1SStefano Zampini } 9524f1b2e48SStefano Zampini ierr = PetscMalloc1(nzs,&array);CHKERRQ(ierr); 9534f1b2e48SStefano Zampini for (i=0;i<nzs-1;i++) array[i] = -1.; 9544f1b2e48SStefano Zampini array[nzs-1] = 1.; 9554f1b2e48SStefano Zampini ierr = MatSetValues(pcbddc->benign_change,1,idxs+nzs-1,nzs,idxs,array,INSERT_VALUES);CHKERRQ(ierr); 9564f1b2e48SStefano Zampini /* store local idxs for p0 */ 9574f1b2e48SStefano Zampini pcbddc->benign_p0_lidx[s] = idxs[nzs-1]; 9584f1b2e48SStefano Zampini ierr = ISRestoreIndices(zerodiag_subs[s],&idxs);CHKERRQ(ierr); 959339f8db1SStefano Zampini ierr = PetscFree(array);CHKERRQ(ierr); 9604f1b2e48SStefano Zampini } 961339f8db1SStefano Zampini ierr = MatAssemblyBegin(pcbddc->benign_change,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 962339f8db1SStefano Zampini ierr = MatAssemblyEnd(pcbddc->benign_change,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 963a3df083aSStefano Zampini /* project if needed */ 964a3df083aSStefano Zampini if (pcbddc->benign_change_explicit) { 9651dd7afcfSStefano Zampini Mat M; 9661dd7afcfSStefano Zampini 9671dd7afcfSStefano Zampini ierr = MatPtAP(pcbddc->local_mat,pcbddc->benign_change,MAT_INITIAL_MATRIX,2.0,&M);CHKERRQ(ierr); 968339f8db1SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 9691dd7afcfSStefano Zampini ierr = MatSeqAIJCompress(M,&pcbddc->local_mat);CHKERRQ(ierr); 9701dd7afcfSStefano Zampini ierr = MatDestroy(&M);CHKERRQ(ierr); 971a3df083aSStefano Zampini } 9724f1b2e48SStefano Zampini /* store global idxs for p0 */ 9734f1b2e48SStefano Zampini ierr = ISLocalToGlobalMappingApply(pc->pmat->rmap->mapping,pcbddc->benign_n,pcbddc->benign_p0_lidx,pcbddc->benign_p0_gidx);CHKERRQ(ierr); 974339f8db1SStefano Zampini } 975ca92afb2SStefano Zampini pcbddc->benign_zerodiag_subs = zerodiag_subs; 9764f1b2e48SStefano Zampini ierr = ISDestroy(&pressures);CHKERRQ(ierr); 977b0f5fe93SStefano Zampini 978b0f5fe93SStefano Zampini /* determines if the coarse solver will be singular or not */ 979b0f5fe93SStefano Zampini ierr = MPI_Allreduce(&have_null,&pcbddc->benign_null,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 98027b6a85dSStefano Zampini /* determines if the problem has subdomains with 0 pressure block */ 98127b6a85dSStefano Zampini ierr = MPI_Allreduce(&have_null,&pcbddc->benign_have_null,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 982339f8db1SStefano Zampini *zerodiaglocal = zerodiag; 983339f8db1SStefano Zampini PetscFunctionReturn(0); 984339f8db1SStefano Zampini } 985339f8db1SStefano Zampini 986339f8db1SStefano Zampini #undef __FUNCT__ 987015636ebSStefano Zampini #define __FUNCT__ "PCBDDCBenignGetOrSetP0" 988015636ebSStefano Zampini PetscErrorCode PCBDDCBenignGetOrSetP0(PC pc, Vec v, PetscBool get) 989efc2fbd9SStefano Zampini { 990efc2fbd9SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 991de9d7bd0SStefano Zampini PetscScalar *array; 992efc2fbd9SStefano Zampini PetscErrorCode ierr; 993efc2fbd9SStefano Zampini 994efc2fbd9SStefano Zampini PetscFunctionBegin; 995efc2fbd9SStefano Zampini if (!pcbddc->benign_sf) { 996efc2fbd9SStefano Zampini ierr = PetscSFCreate(PetscObjectComm((PetscObject)pc),&pcbddc->benign_sf);CHKERRQ(ierr); 9974f1b2e48SStefano Zampini ierr = PetscSFSetGraphLayout(pcbddc->benign_sf,pc->pmat->rmap,pcbddc->benign_n,NULL,PETSC_OWN_POINTER,pcbddc->benign_p0_gidx);CHKERRQ(ierr); 998efc2fbd9SStefano Zampini } 999de9d7bd0SStefano Zampini if (get) { 1000efc2fbd9SStefano Zampini ierr = VecGetArrayRead(v,(const PetscScalar**)&array);CHKERRQ(ierr); 10014f1b2e48SStefano Zampini ierr = PetscSFBcastBegin(pcbddc->benign_sf,MPIU_SCALAR,array,pcbddc->benign_p0);CHKERRQ(ierr); 10024f1b2e48SStefano Zampini ierr = PetscSFBcastEnd(pcbddc->benign_sf,MPIU_SCALAR,array,pcbddc->benign_p0);CHKERRQ(ierr); 1003efc2fbd9SStefano Zampini ierr = VecRestoreArrayRead(v,(const PetscScalar**)&array);CHKERRQ(ierr); 1004de9d7bd0SStefano Zampini } else { 1005de9d7bd0SStefano Zampini ierr = VecGetArray(v,&array);CHKERRQ(ierr); 1006de9d7bd0SStefano Zampini ierr = PetscSFReduceBegin(pcbddc->benign_sf,MPIU_SCALAR,pcbddc->benign_p0,array,MPIU_REPLACE);CHKERRQ(ierr); 1007de9d7bd0SStefano Zampini ierr = PetscSFReduceEnd(pcbddc->benign_sf,MPIU_SCALAR,pcbddc->benign_p0,array,MPIU_REPLACE);CHKERRQ(ierr); 1008de9d7bd0SStefano Zampini ierr = VecRestoreArray(v,&array);CHKERRQ(ierr); 1009efc2fbd9SStefano Zampini } 1010efc2fbd9SStefano Zampini PetscFunctionReturn(0); 1011efc2fbd9SStefano Zampini } 1012efc2fbd9SStefano Zampini 1013efc2fbd9SStefano Zampini #undef __FUNCT__ 1014c263805aSStefano Zampini #define __FUNCT__ "PCBDDCBenignPopOrPushB0" 1015c263805aSStefano Zampini PetscErrorCode PCBDDCBenignPopOrPushB0(PC pc, PetscBool pop) 1016c263805aSStefano Zampini { 1017c263805aSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 1018c263805aSStefano Zampini PetscErrorCode ierr; 1019c263805aSStefano Zampini 1020c263805aSStefano Zampini PetscFunctionBegin; 1021c263805aSStefano Zampini /* TODO: add error checking 1022c263805aSStefano Zampini - avoid nested pop (or push) calls. 1023c263805aSStefano Zampini - cannot push before pop. 10241c604dc7SStefano Zampini - cannot call this if pcbddc->local_mat is NULL 1025c263805aSStefano Zampini */ 10264f1b2e48SStefano Zampini if (!pcbddc->benign_n) { 1027efc2fbd9SStefano Zampini PetscFunctionReturn(0); 1028efc2fbd9SStefano Zampini } 1029c263805aSStefano Zampini if (pop) { 1030a3df083aSStefano Zampini if (pcbddc->benign_change_explicit) { 10314f1b2e48SStefano Zampini IS is_p0; 10324f1b2e48SStefano Zampini MatReuse reuse; 1033c263805aSStefano Zampini 1034c263805aSStefano Zampini /* extract B_0 */ 10354f1b2e48SStefano Zampini reuse = MAT_INITIAL_MATRIX; 10364f1b2e48SStefano Zampini if (pcbddc->benign_B0) { 10374f1b2e48SStefano Zampini reuse = MAT_REUSE_MATRIX; 10384f1b2e48SStefano Zampini } 10394f1b2e48SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,pcbddc->benign_n,pcbddc->benign_p0_lidx,PETSC_COPY_VALUES,&is_p0);CHKERRQ(ierr); 10404f1b2e48SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,is_p0,NULL,reuse,&pcbddc->benign_B0);CHKERRQ(ierr); 1041c263805aSStefano Zampini /* remove rows and cols from local problem */ 1042c263805aSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_KEEP_NONZERO_PATTERN,PETSC_TRUE);CHKERRQ(ierr); 104397d764eeSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_FALSE);CHKERRQ(ierr); 10444f1b2e48SStefano Zampini ierr = MatZeroRowsColumnsIS(pcbddc->local_mat,is_p0,1.0,NULL,NULL);CHKERRQ(ierr); 10454f1b2e48SStefano Zampini ierr = ISDestroy(&is_p0);CHKERRQ(ierr); 1046a3df083aSStefano Zampini } else { 1047a3df083aSStefano Zampini Mat_IS *matis = (Mat_IS*)pc->pmat->data; 1048a3df083aSStefano Zampini PetscScalar *vals; 1049a3df083aSStefano Zampini PetscInt i,n,*idxs_ins; 1050a3df083aSStefano Zampini 1051a3df083aSStefano Zampini ierr = VecGetLocalSize(matis->y,&n);CHKERRQ(ierr); 1052a3df083aSStefano Zampini ierr = PetscMalloc2(n,&idxs_ins,n,&vals);CHKERRQ(ierr); 1053a3df083aSStefano Zampini if (!pcbddc->benign_B0) { 10540b5adadeSStefano Zampini PetscInt *nnz; 1055a3df083aSStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)pcbddc->local_mat),&pcbddc->benign_B0);CHKERRQ(ierr); 1056a3df083aSStefano Zampini ierr = MatSetType(pcbddc->benign_B0,MATAIJ);CHKERRQ(ierr); 1057a3df083aSStefano Zampini ierr = MatSetSizes(pcbddc->benign_B0,pcbddc->benign_n,n,PETSC_DECIDE,PETSC_DECIDE);CHKERRQ(ierr); 1058331e053bSStefano Zampini ierr = PetscMalloc1(pcbddc->benign_n,&nnz);CHKERRQ(ierr); 1059331e053bSStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 1060331e053bSStefano Zampini ierr = ISGetLocalSize(pcbddc->benign_zerodiag_subs[i],&nnz[i]);CHKERRQ(ierr); 1061331e053bSStefano Zampini nnz[i] = n - nnz[i]; 1062331e053bSStefano Zampini } 1063331e053bSStefano Zampini ierr = MatSeqAIJSetPreallocation(pcbddc->benign_B0,0,nnz);CHKERRQ(ierr); 1064331e053bSStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 1065331e053bSStefano Zampini } 1066a3df083aSStefano Zampini 1067a3df083aSStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 1068a3df083aSStefano Zampini PetscScalar *array; 1069a3df083aSStefano Zampini PetscInt *idxs,j,nz,cum; 1070a3df083aSStefano Zampini 1071a3df083aSStefano Zampini ierr = VecSet(matis->x,0.);CHKERRQ(ierr); 1072a3df083aSStefano Zampini ierr = ISGetLocalSize(pcbddc->benign_zerodiag_subs[i],&nz);CHKERRQ(ierr); 1073a3df083aSStefano Zampini ierr = ISGetIndices(pcbddc->benign_zerodiag_subs[i],(const PetscInt**)&idxs);CHKERRQ(ierr); 1074a3df083aSStefano Zampini for (j=0;j<nz;j++) vals[j] = 1.; 1075a3df083aSStefano Zampini ierr = VecSetValues(matis->x,nz,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 1076a3df083aSStefano Zampini ierr = VecAssemblyBegin(matis->x);CHKERRQ(ierr); 1077a3df083aSStefano Zampini ierr = VecAssemblyEnd(matis->x);CHKERRQ(ierr); 1078a3df083aSStefano Zampini ierr = VecSet(matis->y,0.);CHKERRQ(ierr); 1079a3df083aSStefano Zampini ierr = MatMult(matis->A,matis->x,matis->y);CHKERRQ(ierr); 1080a3df083aSStefano Zampini ierr = VecGetArray(matis->y,&array);CHKERRQ(ierr); 1081a3df083aSStefano Zampini cum = 0; 1082a3df083aSStefano Zampini for (j=0;j<n;j++) { 108322db5ddcSStefano Zampini if (PetscUnlikely(PetscAbsScalar(array[j]) > PETSC_SMALL)) { 1084a3df083aSStefano Zampini vals[cum] = array[j]; 1085a3df083aSStefano Zampini idxs_ins[cum] = j; 1086a3df083aSStefano Zampini cum++; 1087a3df083aSStefano Zampini } 1088a3df083aSStefano Zampini } 1089a3df083aSStefano Zampini ierr = MatSetValues(pcbddc->benign_B0,1,&i,cum,idxs_ins,vals,INSERT_VALUES);CHKERRQ(ierr); 1090a3df083aSStefano Zampini ierr = VecRestoreArray(matis->y,&array);CHKERRQ(ierr); 1091a3df083aSStefano Zampini ierr = ISRestoreIndices(pcbddc->benign_zerodiag_subs[i],(const PetscInt**)&idxs);CHKERRQ(ierr); 1092a3df083aSStefano Zampini } 1093a3df083aSStefano Zampini ierr = MatAssemblyBegin(pcbddc->benign_B0,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1094a3df083aSStefano Zampini ierr = MatAssemblyEnd(pcbddc->benign_B0,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1095a3df083aSStefano Zampini ierr = PetscFree2(idxs_ins,vals);CHKERRQ(ierr); 1096a3df083aSStefano Zampini } 1097c263805aSStefano Zampini } else { /* push */ 1098a3df083aSStefano Zampini if (pcbddc->benign_change_explicit) { 10994f1b2e48SStefano Zampini PetscInt i; 11004f1b2e48SStefano Zampini 11014f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 11024f1b2e48SStefano Zampini PetscScalar *B0_vals; 11034f1b2e48SStefano Zampini PetscInt *B0_cols,B0_ncol; 11044f1b2e48SStefano Zampini 11054f1b2e48SStefano Zampini ierr = MatGetRow(pcbddc->benign_B0,i,&B0_ncol,(const PetscInt**)&B0_cols,(const PetscScalar**)&B0_vals);CHKERRQ(ierr); 11064f1b2e48SStefano Zampini ierr = MatSetValues(pcbddc->local_mat,1,pcbddc->benign_p0_lidx+i,B0_ncol,B0_cols,B0_vals,INSERT_VALUES);CHKERRQ(ierr); 11077b034428SStefano Zampini ierr = MatSetValues(pcbddc->local_mat,B0_ncol,B0_cols,1,pcbddc->benign_p0_lidx+i,B0_vals,INSERT_VALUES);CHKERRQ(ierr); 11084f1b2e48SStefano Zampini ierr = MatSetValue(pcbddc->local_mat,pcbddc->benign_p0_lidx[i],pcbddc->benign_p0_lidx[i],0.0,INSERT_VALUES);CHKERRQ(ierr); 11094f1b2e48SStefano Zampini ierr = MatRestoreRow(pcbddc->benign_B0,i,&B0_ncol,(const PetscInt**)&B0_cols,(const PetscScalar**)&B0_vals);CHKERRQ(ierr); 11104f1b2e48SStefano Zampini } 1111c263805aSStefano Zampini ierr = MatAssemblyBegin(pcbddc->local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1112c263805aSStefano Zampini ierr = MatAssemblyEnd(pcbddc->local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1113a3df083aSStefano Zampini } else { 1114a3df083aSStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Cannot push B0!\n"); 1115a3df083aSStefano Zampini } 1116c263805aSStefano Zampini } 1117c263805aSStefano Zampini PetscFunctionReturn(0); 1118c263805aSStefano Zampini } 1119c263805aSStefano Zampini 1120c263805aSStefano Zampini #undef __FUNCT__ 1121b1b3d7a2SStefano Zampini #define __FUNCT__ "PCBDDCAdaptiveSelection" 112208122e43SStefano Zampini PetscErrorCode PCBDDCAdaptiveSelection(PC pc) 1123b1b3d7a2SStefano Zampini { 1124b1b3d7a2SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 112508122e43SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 112608122e43SStefano Zampini PetscBLASInt B_dummyint,B_neigs,B_ierr,B_lwork; 112708122e43SStefano Zampini PetscBLASInt *B_iwork,*B_ifail; 112808122e43SStefano Zampini PetscScalar *work,lwork; 112908122e43SStefano Zampini PetscScalar *St,*S,*eigv; 113008122e43SStefano Zampini PetscScalar *Sarray,*Starray; 113108122e43SStefano Zampini PetscReal *eigs,thresh; 11321b968477SStefano Zampini PetscInt i,nmax,nmin,nv,cum,mss,cum2,cumarray,maxneigs; 1133f6f667cfSStefano Zampini PetscBool allocated_S_St; 113408122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 113508122e43SStefano Zampini PetscReal *rwork; 113608122e43SStefano Zampini #endif 1137b1b3d7a2SStefano Zampini PetscErrorCode ierr; 1138b1b3d7a2SStefano Zampini 1139b1b3d7a2SStefano Zampini PetscFunctionBegin; 1140af25d912SStefano Zampini if (!sub_schurs->schur_explicit) SETERRQ(PetscObjectComm((PetscObject)pc),PETSC_ERR_SUP,"Adaptive selection of constraints requires MUMPS and/or MKL_CPARDISO"); 1141af25d912SStefano Zampini if (sub_schurs->n_subs && (!sub_schurs->is_hermitian || !sub_schurs->is_posdef)) 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); 114206a4e24aSStefano Zampini 1143fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 1144fd14bc51SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1145fd14bc51SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 1146fd14bc51SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Check adaptive selection of constraints\n");CHKERRQ(ierr); 11471575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 1148fd14bc51SStefano Zampini } 1149fd14bc51SStefano Zampini 1150e496cd5dSStefano Zampini if (pcbddc->dbg_flag) { 1151e496cd5dSStefano 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); 1152e496cd5dSStefano Zampini } 1153e496cd5dSStefano Zampini 115408122e43SStefano Zampini /* max size of subsets */ 115508122e43SStefano Zampini mss = 0; 115608122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 115708122e43SStefano Zampini PetscInt subset_size; 1158862806e4SStefano Zampini 115908122e43SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 116008122e43SStefano Zampini mss = PetscMax(mss,subset_size); 116108122e43SStefano Zampini } 116208122e43SStefano Zampini 116308122e43SStefano Zampini /* min/max and threshold */ 116408122e43SStefano Zampini nmax = pcbddc->adaptive_nmax > 0 ? pcbddc->adaptive_nmax : mss; 1165f6f667cfSStefano Zampini nmin = pcbddc->adaptive_nmin > 0 ? pcbddc->adaptive_nmin : 0; 116608122e43SStefano Zampini nmax = PetscMax(nmin,nmax); 1167f6f667cfSStefano Zampini allocated_S_St = PETSC_FALSE; 1168f6f667cfSStefano Zampini if (nmin) { 1169f6f667cfSStefano Zampini allocated_S_St = PETSC_TRUE; 1170f6f667cfSStefano Zampini } 117108122e43SStefano Zampini 117208122e43SStefano Zampini /* allocate lapack workspace */ 117308122e43SStefano Zampini cum = cum2 = 0; 117408122e43SStefano Zampini maxneigs = 0; 117508122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 117608122e43SStefano Zampini PetscInt n,subset_size; 1177f6f667cfSStefano Zampini 117808122e43SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 117908122e43SStefano Zampini n = PetscMin(subset_size,nmax); 11809162d606SStefano Zampini cum += subset_size; 11819162d606SStefano Zampini cum2 += subset_size*n; 118208122e43SStefano Zampini maxneigs = PetscMax(maxneigs,n); 118308122e43SStefano Zampini } 118408122e43SStefano Zampini if (mss) { 11859ab7bb16SStefano Zampini if (sub_schurs->is_hermitian && sub_schurs->is_posdef) { 118608122e43SStefano Zampini PetscBLASInt B_itype = 1; 118708122e43SStefano Zampini PetscBLASInt B_N = mss; 11884c6709b3SStefano Zampini PetscReal zero = 0.0; 11894c6709b3SStefano Zampini PetscReal eps = 0.0; /* dlamch? */ 119008122e43SStefano Zampini 119108122e43SStefano Zampini B_lwork = -1; 119208122e43SStefano Zampini S = NULL; 119308122e43SStefano Zampini St = NULL; 1194a58a30b4SStefano Zampini eigs = NULL; 1195a58a30b4SStefano Zampini eigv = NULL; 1196a58a30b4SStefano Zampini B_iwork = NULL; 1197a58a30b4SStefano Zampini B_ifail = NULL; 1198d1710679SStefano Zampini #if defined(PETSC_USE_COMPLEX) 1199d1710679SStefano Zampini rwork = NULL; 1200d1710679SStefano Zampini #endif 12018bec7fa6SStefano Zampini thresh = 1.0; 120208122e43SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 120308122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 120408122e43SStefano 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)); 120508122e43SStefano Zampini #else 120608122e43SStefano 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)); 120708122e43SStefano Zampini #endif 120808122e43SStefano Zampini if (B_ierr != 0) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to SYGVX Lapack routine %d",(int)B_ierr); 120908122e43SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 121008122e43SStefano Zampini } else { 121108122e43SStefano Zampini /* TODO */ 121208122e43SStefano Zampini } 121308122e43SStefano Zampini } else { 121408122e43SStefano Zampini lwork = 0; 121508122e43SStefano Zampini } 121608122e43SStefano Zampini 121708122e43SStefano Zampini nv = 0; 1218d62866d3SStefano 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) */ 1219d62866d3SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_vertices,&nv);CHKERRQ(ierr); 122008122e43SStefano Zampini } 12214c6709b3SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lwork),&B_lwork);CHKERRQ(ierr); 1222f6f667cfSStefano Zampini if (allocated_S_St) { 1223f6f667cfSStefano Zampini ierr = PetscMalloc2(mss*mss,&S,mss*mss,&St);CHKERRQ(ierr); 1224f6f667cfSStefano Zampini } 1225f6f667cfSStefano Zampini ierr = PetscMalloc5(mss*mss,&eigv,mss,&eigs,B_lwork,&work,5*mss,&B_iwork,mss,&B_ifail);CHKERRQ(ierr); 122608122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 122708122e43SStefano Zampini ierr = PetscMalloc1(7*mss,&rwork);CHKERRQ(ierr); 122808122e43SStefano Zampini #endif 12299162d606SStefano Zampini ierr = PetscMalloc5(nv+sub_schurs->n_subs,&pcbddc->adaptive_constraints_n, 12309162d606SStefano Zampini nv+sub_schurs->n_subs+1,&pcbddc->adaptive_constraints_idxs_ptr, 12319162d606SStefano Zampini nv+sub_schurs->n_subs+1,&pcbddc->adaptive_constraints_data_ptr, 123208122e43SStefano Zampini nv+cum,&pcbddc->adaptive_constraints_idxs, 12339162d606SStefano Zampini nv+cum2,&pcbddc->adaptive_constraints_data);CHKERRQ(ierr); 123408122e43SStefano Zampini ierr = PetscMemzero(pcbddc->adaptive_constraints_n,(nv+sub_schurs->n_subs)*sizeof(PetscInt));CHKERRQ(ierr); 123508122e43SStefano Zampini 123608122e43SStefano Zampini maxneigs = 0; 123772b8c272SStefano Zampini cum = cumarray = 0; 12389162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr[0] = 0; 12399162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr[0] = 0; 1240d62866d3SStefano Zampini if (sub_schurs->is_vertices && pcbddc->use_vertices) { 124108122e43SStefano Zampini const PetscInt *idxs; 124208122e43SStefano Zampini 1243d62866d3SStefano Zampini ierr = ISGetIndices(sub_schurs->is_vertices,&idxs);CHKERRQ(ierr); 124408122e43SStefano Zampini for (cum=0;cum<nv;cum++) { 124508122e43SStefano Zampini pcbddc->adaptive_constraints_n[cum] = 1; 124608122e43SStefano Zampini pcbddc->adaptive_constraints_idxs[cum] = idxs[cum]; 124708122e43SStefano Zampini pcbddc->adaptive_constraints_data[cum] = 1.0; 12489162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr[cum+1] = pcbddc->adaptive_constraints_idxs_ptr[cum]+1; 12499162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr[cum+1] = pcbddc->adaptive_constraints_data_ptr[cum]+1; 125008122e43SStefano Zampini } 1251d62866d3SStefano Zampini ierr = ISRestoreIndices(sub_schurs->is_vertices,&idxs);CHKERRQ(ierr); 125208122e43SStefano Zampini } 125308122e43SStefano Zampini 125408122e43SStefano Zampini if (mss) { /* multilevel */ 125508122e43SStefano Zampini ierr = MatSeqAIJGetArray(sub_schurs->sum_S_Ej_inv_all,&Sarray);CHKERRQ(ierr); 125608122e43SStefano Zampini ierr = MatSeqAIJGetArray(sub_schurs->sum_S_Ej_tilda_all,&Starray);CHKERRQ(ierr); 125708122e43SStefano Zampini } 125808122e43SStefano Zampini 1259ffd830a3SStefano Zampini thresh = pcbddc->adaptive_threshold; 126008122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 126108122e43SStefano Zampini const PetscInt *idxs; 12629d54b7f4SStefano Zampini PetscReal upper,lower; 1263862806e4SStefano Zampini PetscInt j,subset_size,eigs_start = 0; 126408122e43SStefano Zampini PetscBLASInt B_N; 1265aff50787SStefano Zampini PetscBool same_data = PETSC_FALSE; 126608122e43SStefano Zampini 12679d54b7f4SStefano Zampini if (pcbddc->use_deluxe_scaling) { 12689d54b7f4SStefano Zampini upper = PETSC_MAX_REAL; 12699d54b7f4SStefano Zampini lower = thresh; 12709d54b7f4SStefano Zampini } else { 12719d54b7f4SStefano Zampini upper = 1./thresh; 12729d54b7f4SStefano Zampini lower = 0.; 12739d54b7f4SStefano Zampini } 1274862806e4SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 1275ffd830a3SStefano Zampini ierr = ISGetIndices(sub_schurs->is_subs[i],&idxs);CHKERRQ(ierr); 1276f6f667cfSStefano Zampini ierr = PetscBLASIntCast(subset_size,&B_N);CHKERRQ(ierr); 1277f6f667cfSStefano Zampini if (allocated_S_St) { /* S and S_t should be copied since we could need them later */ 12789ab7bb16SStefano Zampini if (sub_schurs->is_hermitian) { 1279aff50787SStefano Zampini PetscInt j,k; 1280aff50787SStefano Zampini if (sub_schurs->n_subs == 1) { /* zeroing memory to use PetscMemcmp later */ 1281aff50787SStefano Zampini ierr = PetscMemzero(S,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 1282aff50787SStefano Zampini ierr = PetscMemzero(St,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 128308122e43SStefano Zampini } 128408122e43SStefano Zampini for (j=0;j<subset_size;j++) { 1285aff50787SStefano Zampini for (k=j;k<subset_size;k++) { 1286aff50787SStefano Zampini S [j*subset_size+k] = Sarray [cumarray+j*subset_size+k]; 1287aff50787SStefano Zampini St[j*subset_size+k] = Starray[cumarray+j*subset_size+k]; 1288aff50787SStefano Zampini } 128908122e43SStefano Zampini } 129008122e43SStefano Zampini } else { 129108122e43SStefano Zampini ierr = PetscMemcpy(S,Sarray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 129208122e43SStefano Zampini ierr = PetscMemcpy(St,Starray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 129308122e43SStefano Zampini } 12948bec7fa6SStefano Zampini } else { 1295f6f667cfSStefano Zampini S = Sarray + cumarray; 1296f6f667cfSStefano Zampini St = Starray + cumarray; 12978bec7fa6SStefano Zampini } 1298aff50787SStefano Zampini /* see if we can save some work */ 1299b7ab4a40SStefano Zampini if (sub_schurs->n_subs == 1 && pcbddc->use_deluxe_scaling) { 1300aff50787SStefano Zampini ierr = PetscMemcmp(S,St,subset_size*subset_size*sizeof(PetscScalar),&same_data);CHKERRQ(ierr); 1301aff50787SStefano Zampini } 1302aff50787SStefano Zampini 1303b7ab4a40SStefano Zampini if (same_data && !sub_schurs->change) { /* there's no need of constraints here */ 1304aff50787SStefano Zampini B_neigs = 0; 1305aff50787SStefano Zampini } else { 13069ab7bb16SStefano Zampini if (sub_schurs->is_hermitian && sub_schurs->is_posdef) { 130708122e43SStefano Zampini PetscBLASInt B_itype = 1; 1308f6f667cfSStefano Zampini PetscBLASInt B_IL, B_IU; 13094c6709b3SStefano Zampini PetscReal eps = -1.0; /* dlamch? */ 13109552c7c7SStefano Zampini PetscInt nmin_s; 1311b7ab4a40SStefano Zampini PetscBool compute_range = PETSC_FALSE; 131208122e43SStefano Zampini 1313fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 13148bec7fa6SStefano 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]]); 1315fd14bc51SStefano Zampini } 1316d16cbb6bSStefano Zampini 1317b7ab4a40SStefano Zampini compute_range = PETSC_FALSE; 1318b7ab4a40SStefano Zampini if (thresh > 1.+PETSC_SMALL && !same_data) { 1319b7ab4a40SStefano Zampini compute_range = PETSC_TRUE; 1320b7ab4a40SStefano Zampini } 1321b7ab4a40SStefano Zampini 132208122e43SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 1323b7ab4a40SStefano Zampini if (compute_range) { 1324d16cbb6bSStefano Zampini 1325d16cbb6bSStefano Zampini /* ask for eigenvalues larger than thresh */ 132608122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 13279d54b7f4SStefano Zampini PetscStackCallBLAS("LAPACKsygvx",LAPACKsygvx_(&B_itype,"V","V","L",&B_N,St,&B_N,S,&B_N,&lower,&upper,&B_IL,&B_IU,&eps,&B_neigs,eigs,eigv,&B_N,work,&B_lwork,rwork,B_iwork,B_ifail,&B_ierr)); 132808122e43SStefano Zampini #else 13299d54b7f4SStefano Zampini PetscStackCallBLAS("LAPACKsygvx",LAPACKsygvx_(&B_itype,"V","V","L",&B_N,St,&B_N,S,&B_N,&lower,&upper,&B_IL,&B_IU,&eps,&B_neigs,eigs,eigv,&B_N,work,&B_lwork,B_iwork,B_ifail,&B_ierr)); 133008122e43SStefano Zampini #endif 1331b7ab4a40SStefano Zampini } else if (!same_data) { 1332d16cbb6bSStefano Zampini B_IU = PetscMax(1,PetscMin(B_N,nmax)); 1333d16cbb6bSStefano Zampini B_IL = 1; 1334d16cbb6bSStefano Zampini #if defined(PETSC_USE_COMPLEX) 13359d54b7f4SStefano Zampini PetscStackCallBLAS("LAPACKsygvx",LAPACKsygvx_(&B_itype,"V","I","L",&B_N,St,&B_N,S,&B_N,&lower,&upper,&B_IL,&B_IU,&eps,&B_neigs,eigs,eigv,&B_N,work,&B_lwork,rwork,B_iwork,B_ifail,&B_ierr)); 1336d16cbb6bSStefano Zampini #else 13379d54b7f4SStefano Zampini PetscStackCallBLAS("LAPACKsygvx",LAPACKsygvx_(&B_itype,"V","I","L",&B_N,St,&B_N,S,&B_N,&lower,&upper,&B_IL,&B_IU,&eps,&B_neigs,eigs,eigv,&B_N,work,&B_lwork,B_iwork,B_ifail,&B_ierr)); 1338d16cbb6bSStefano Zampini #endif 1339b7ab4a40SStefano Zampini } else { /* same_data is true, so get the adaptive function requested by the user */ 1340b7ab4a40SStefano Zampini PetscInt k; 1341b7ab4a40SStefano Zampini if (!sub_schurs->change_primal_sub) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"This should not happen"); 1342b7ab4a40SStefano Zampini ierr = ISGetLocalSize(sub_schurs->change_primal_sub[i],&nmax);CHKERRQ(ierr); 1343b7ab4a40SStefano Zampini ierr = PetscBLASIntCast(nmax,&B_neigs);CHKERRQ(ierr); 1344b7ab4a40SStefano Zampini nmin = nmax; 1345b7ab4a40SStefano Zampini ierr = PetscMemzero(eigv,subset_size*nmax*sizeof(PetscScalar));CHKERRQ(ierr); 1346b7ab4a40SStefano Zampini for (k=0;k<nmax;k++) { 1347b7ab4a40SStefano Zampini eigs[k] = 1./PETSC_SMALL; 1348b7ab4a40SStefano Zampini eigv[k*(subset_size+1)] = 1.0; 1349b7ab4a40SStefano Zampini } 1350d16cbb6bSStefano Zampini } 135108122e43SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 135208122e43SStefano Zampini if (B_ierr) { 13536c4ed002SBarry Smith if (B_ierr < 0 ) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYGVX Lapack routine: illegal value for argument %d",-(int)B_ierr); 13546c4ed002SBarry Smith else if (B_ierr <= B_N) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYGVX Lapack routine: %d eigenvalues failed to converge",(int)B_ierr); 13556c4ed002SBarry Smith else 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); 135608122e43SStefano Zampini } 135708122e43SStefano Zampini 135808122e43SStefano Zampini if (B_neigs > nmax) { 1359fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 1360fd14bc51SStefano Zampini PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," found %d eigs, more than maximum required %d.\n",B_neigs,nmax); 1361fd14bc51SStefano Zampini } 13629d54b7f4SStefano Zampini if (pcbddc->use_deluxe_scaling) eigs_start = B_neigs -nmax; 136308122e43SStefano Zampini B_neigs = nmax; 136408122e43SStefano Zampini } 136508122e43SStefano Zampini 13669552c7c7SStefano Zampini nmin_s = PetscMin(nmin,B_N); 13679552c7c7SStefano Zampini if (B_neigs < nmin_s) { 136808122e43SStefano Zampini PetscBLASInt B_neigs2; 136908122e43SStefano Zampini 13709d54b7f4SStefano Zampini if (pcbddc->use_deluxe_scaling) { 1371f6f667cfSStefano Zampini B_IL = B_N - nmin_s + 1; 13729d54b7f4SStefano Zampini B_IU = B_N - B_neigs; 13739d54b7f4SStefano Zampini } else { 13749d54b7f4SStefano Zampini B_IL = B_neigs + 1; 13759d54b7f4SStefano Zampini B_IU = nmin_s; 13769d54b7f4SStefano Zampini } 1377fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 1378fd14bc51SStefano 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); 1379fd14bc51SStefano Zampini } 13809ab7bb16SStefano Zampini if (sub_schurs->is_hermitian) { 138108122e43SStefano Zampini PetscInt j; 138208122e43SStefano Zampini for (j=0;j<subset_size;j++) { 138308122e43SStefano Zampini ierr = PetscMemcpy(S+j*(subset_size+1),Sarray+cumarray+j*(subset_size+1),(subset_size-j)*sizeof(PetscScalar));CHKERRQ(ierr); 138408122e43SStefano Zampini } 138508122e43SStefano Zampini for (j=0;j<subset_size;j++) { 138608122e43SStefano Zampini ierr = PetscMemcpy(St+j*(subset_size+1),Starray+cumarray+j*(subset_size+1),(subset_size-j)*sizeof(PetscScalar));CHKERRQ(ierr); 138708122e43SStefano Zampini } 138808122e43SStefano Zampini } else { 138908122e43SStefano Zampini ierr = PetscMemcpy(S,Sarray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 139008122e43SStefano Zampini ierr = PetscMemcpy(St,Starray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 139108122e43SStefano Zampini } 139208122e43SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 139308122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 13949d54b7f4SStefano Zampini PetscStackCallBLAS("LAPACKsygvx",LAPACKsygvx_(&B_itype,"V","I","L",&B_N,St,&B_N,S,&B_N,&lower,&upper,&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)); 139508122e43SStefano Zampini #else 13969d54b7f4SStefano Zampini PetscStackCallBLAS("LAPACKsygvx",LAPACKsygvx_(&B_itype,"V","I","L",&B_N,St,&B_N,S,&B_N,&lower,&upper,&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)); 139708122e43SStefano Zampini #endif 139808122e43SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 139908122e43SStefano Zampini B_neigs += B_neigs2; 140008122e43SStefano Zampini } 140108122e43SStefano Zampini if (B_ierr) { 14026c4ed002SBarry Smith if (B_ierr < 0 ) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYGVX Lapack routine: illegal value for argument %d",-(int)B_ierr); 14036c4ed002SBarry Smith else if (B_ierr <= B_N) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYGVX Lapack routine: %d eigenvalues failed to converge",(int)B_ierr); 14046c4ed002SBarry Smith else 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); 140508122e43SStefano Zampini } 1406fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 1407ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," -> Got %d eigs\n",B_neigs);CHKERRQ(ierr); 140808122e43SStefano Zampini for (j=0;j<B_neigs;j++) { 140908122e43SStefano Zampini if (eigs[j] == 0.0) { 1410ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," Inf\n");CHKERRQ(ierr); 141108122e43SStefano Zampini } else { 14129d54b7f4SStefano Zampini if (pcbddc->use_deluxe_scaling) { 1413ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," %1.6e\n",eigs[j+eigs_start]);CHKERRQ(ierr); 14149d54b7f4SStefano Zampini } else { 14159d54b7f4SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," %1.6e\n",1./eigs[j+eigs_start]);CHKERRQ(ierr); 14169d54b7f4SStefano Zampini } 1417fd14bc51SStefano Zampini } 141808122e43SStefano Zampini } 141908122e43SStefano Zampini } 142008122e43SStefano Zampini } else { 142108122e43SStefano Zampini /* TODO */ 142208122e43SStefano Zampini } 1423aff50787SStefano Zampini } 14246c3e6151SStefano Zampini /* change the basis back to the original one */ 14256c3e6151SStefano Zampini if (sub_schurs->change) { 142672b8c272SStefano Zampini Mat change,phi,phit; 14276c3e6151SStefano Zampini 14286c3e6151SStefano Zampini if (pcbddc->dbg_flag > 1) { 14296c3e6151SStefano Zampini PetscInt ii; 14306c3e6151SStefano Zampini for (ii=0;ii<B_neigs;ii++) { 14316c3e6151SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," -> Eigenvector (old basis) %d/%d (%d)\n",ii,B_neigs,B_N);CHKERRQ(ierr); 14326c3e6151SStefano Zampini for (j=0;j<B_N;j++) { 14336c3e6151SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," %1.4e\n",eigv[(ii+eigs_start)*subset_size+j]);CHKERRQ(ierr); 14346c3e6151SStefano Zampini } 14356c3e6151SStefano Zampini } 14366c3e6151SStefano Zampini } 143772b8c272SStefano Zampini ierr = KSPGetOperators(sub_schurs->change[i],&change,NULL);CHKERRQ(ierr); 14386c3e6151SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,subset_size,B_neigs,eigv+eigs_start*subset_size,&phit);CHKERRQ(ierr); 143972b8c272SStefano Zampini ierr = MatMatMult(change,phit,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&phi);CHKERRQ(ierr); 14406c3e6151SStefano Zampini ierr = MatCopy(phi,phit,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 14416c3e6151SStefano Zampini ierr = MatDestroy(&phit);CHKERRQ(ierr); 14426c3e6151SStefano Zampini ierr = MatDestroy(&phi);CHKERRQ(ierr); 14436c3e6151SStefano Zampini } 14448bec7fa6SStefano Zampini maxneigs = PetscMax(B_neigs,maxneigs); 14458bec7fa6SStefano Zampini pcbddc->adaptive_constraints_n[i+nv] = B_neigs; 14469162d606SStefano Zampini if (B_neigs) { 14479162d606SStefano 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); 1448fd14bc51SStefano Zampini 1449fd14bc51SStefano Zampini if (pcbddc->dbg_flag > 1) { 14509552c7c7SStefano Zampini PetscInt ii; 14519552c7c7SStefano Zampini for (ii=0;ii<B_neigs;ii++) { 1452ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," -> Eigenvector %d/%d (%d)\n",ii,B_neigs,B_N);CHKERRQ(ierr); 14539552c7c7SStefano Zampini for (j=0;j<B_N;j++) { 1454ac47001eSStefano Zampini #if defined(PETSC_USE_COMPLEX) 1455ac47001eSStefano Zampini PetscReal r = PetscRealPart(pcbddc->adaptive_constraints_data[ii*subset_size+j+pcbddc->adaptive_constraints_data_ptr[cum]]); 1456ac47001eSStefano Zampini PetscReal c = PetscImaginaryPart(pcbddc->adaptive_constraints_data[ii*subset_size+j+pcbddc->adaptive_constraints_data_ptr[cum]]); 1457ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," %1.4e + %1.4e i\n",r,c);CHKERRQ(ierr); 1458ac47001eSStefano Zampini #else 1459ac47001eSStefano 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); 1460ac47001eSStefano Zampini #endif 14619552c7c7SStefano Zampini } 14629552c7c7SStefano Zampini } 1463fd14bc51SStefano Zampini } 14649162d606SStefano Zampini ierr = PetscMemcpy(pcbddc->adaptive_constraints_idxs+pcbddc->adaptive_constraints_idxs_ptr[cum],idxs,subset_size*sizeof(PetscInt));CHKERRQ(ierr); 14659162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr[cum+1] = pcbddc->adaptive_constraints_idxs_ptr[cum] + subset_size; 14669162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr[cum+1] = pcbddc->adaptive_constraints_data_ptr[cum] + subset_size*B_neigs; 14679162d606SStefano Zampini cum++; 146808122e43SStefano Zampini } 146908122e43SStefano Zampini ierr = ISRestoreIndices(sub_schurs->is_subs[i],&idxs);CHKERRQ(ierr); 147008122e43SStefano Zampini /* shift for next computation */ 147108122e43SStefano Zampini cumarray += subset_size*subset_size; 147208122e43SStefano Zampini } 1473fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 1474fd14bc51SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1475fd14bc51SStefano Zampini } 147608122e43SStefano Zampini 147708122e43SStefano Zampini if (mss) { 147808122e43SStefano Zampini ierr = MatSeqAIJRestoreArray(sub_schurs->sum_S_Ej_inv_all,&Sarray);CHKERRQ(ierr); 147908122e43SStefano Zampini ierr = MatSeqAIJRestoreArray(sub_schurs->sum_S_Ej_tilda_all,&Starray);CHKERRQ(ierr); 1480f6f667cfSStefano Zampini /* destroy matrices (junk) */ 1481f6f667cfSStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_inv_all);CHKERRQ(ierr); 1482f6f667cfSStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_tilda_all);CHKERRQ(ierr); 148308122e43SStefano Zampini } 1484f6f667cfSStefano Zampini if (allocated_S_St) { 1485f6f667cfSStefano Zampini ierr = PetscFree2(S,St);CHKERRQ(ierr); 1486f6f667cfSStefano Zampini } 1487f6f667cfSStefano Zampini ierr = PetscFree5(eigv,eigs,work,B_iwork,B_ifail);CHKERRQ(ierr); 148808122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 148908122e43SStefano Zampini ierr = PetscFree(rwork);CHKERRQ(ierr); 149008122e43SStefano Zampini #endif 149108122e43SStefano Zampini if (pcbddc->dbg_flag) { 14921b968477SStefano Zampini PetscInt maxneigs_r; 1493b2566f29SBarry Smith ierr = MPIU_Allreduce(&maxneigs,&maxneigs_r,1,MPIU_INT,MPI_MAX,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 14949b28b941SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Maximum number of constraints per cc %d\n",maxneigs_r);CHKERRQ(ierr); 149508122e43SStefano Zampini } 149608122e43SStefano Zampini PetscFunctionReturn(0); 149708122e43SStefano Zampini } 1498b1b3d7a2SStefano Zampini 1499674ae819SStefano Zampini #undef __FUNCT__ 1500c8587f34SStefano Zampini #define __FUNCT__ "PCBDDCSetUpSolvers" 1501c8587f34SStefano Zampini PetscErrorCode PCBDDCSetUpSolvers(PC pc) 1502c8587f34SStefano Zampini { 15038629588bSStefano Zampini PetscScalar *coarse_submat_vals; 1504c8587f34SStefano Zampini PetscErrorCode ierr; 1505c8587f34SStefano Zampini 1506c8587f34SStefano Zampini PetscFunctionBegin; 1507f4ddd8eeSStefano Zampini /* Setup local scatters R_to_B and (optionally) R_to_D */ 15085e8657edSStefano Zampini /* PCBDDCSetUpLocalWorkVectors should be called first! */ 1509c8587f34SStefano Zampini ierr = PCBDDCSetUpLocalScatters(pc);CHKERRQ(ierr); 1510c8587f34SStefano Zampini 1511684f6988SStefano Zampini /* Setup local neumann solver ksp_R */ 15120fccc4e9SStefano Zampini /* PCBDDCSetUpLocalScatters should be called first! */ 1513684f6988SStefano Zampini ierr = PCBDDCSetUpLocalSolvers(pc,PETSC_FALSE,PETSC_TRUE);CHKERRQ(ierr); 1514c8587f34SStefano Zampini 15158629588bSStefano Zampini /* 15168629588bSStefano Zampini Setup local correction and local part of coarse basis. 15178629588bSStefano Zampini Gives back the dense local part of the coarse matrix in column major ordering 15188629588bSStefano Zampini */ 151947f4ddc3SStefano Zampini ierr = PCBDDCSetUpCorrection(pc,&coarse_submat_vals);CHKERRQ(ierr); 15208629588bSStefano Zampini 15218629588bSStefano Zampini /* Compute total number of coarse nodes and setup coarse solver */ 15228629588bSStefano Zampini ierr = PCBDDCSetUpCoarseSolver(pc,coarse_submat_vals);CHKERRQ(ierr); 15238629588bSStefano Zampini 15248629588bSStefano Zampini /* free */ 15258629588bSStefano Zampini ierr = PetscFree(coarse_submat_vals);CHKERRQ(ierr); 1526c8587f34SStefano Zampini PetscFunctionReturn(0); 1527c8587f34SStefano Zampini } 1528c8587f34SStefano Zampini 1529c8587f34SStefano Zampini #undef __FUNCT__ 1530674ae819SStefano Zampini #define __FUNCT__ "PCBDDCResetCustomization" 1531674ae819SStefano Zampini PetscErrorCode PCBDDCResetCustomization(PC pc) 1532674ae819SStefano Zampini { 1533674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 1534674ae819SStefano Zampini PetscErrorCode ierr; 1535674ae819SStefano Zampini 1536674ae819SStefano Zampini PetscFunctionBegin; 1537674ae819SStefano Zampini ierr = PCBDDCGraphResetCSR(pcbddc->mat_graph);CHKERRQ(ierr); 1538674ae819SStefano Zampini ierr = ISDestroy(&pcbddc->user_primal_vertices);CHKERRQ(ierr); 153930368db7SStefano Zampini ierr = ISDestroy(&pcbddc->user_primal_vertices_local);CHKERRQ(ierr); 1540674ae819SStefano Zampini ierr = ISDestroy(&pcbddc->NeumannBoundaries);CHKERRQ(ierr); 1541785d1243SStefano Zampini ierr = ISDestroy(&pcbddc->NeumannBoundariesLocal);CHKERRQ(ierr); 1542674ae819SStefano Zampini ierr = ISDestroy(&pcbddc->DirichletBoundaries);CHKERRQ(ierr); 1543f4ddd8eeSStefano Zampini ierr = MatNullSpaceDestroy(&pcbddc->onearnullspace);CHKERRQ(ierr); 1544f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->onearnullvecs_state);CHKERRQ(ierr); 1545785d1243SStefano Zampini ierr = ISDestroy(&pcbddc->DirichletBoundariesLocal);CHKERRQ(ierr); 154663602bcaSStefano Zampini ierr = PCBDDCSetDofsSplitting(pc,0,NULL);CHKERRQ(ierr); 154763602bcaSStefano Zampini ierr = PCBDDCSetDofsSplittingLocal(pc,0,NULL);CHKERRQ(ierr); 1548674ae819SStefano Zampini PetscFunctionReturn(0); 1549674ae819SStefano Zampini } 1550674ae819SStefano Zampini 1551674ae819SStefano Zampini #undef __FUNCT__ 1552674ae819SStefano Zampini #define __FUNCT__ "PCBDDCResetTopography" 1553674ae819SStefano Zampini PetscErrorCode PCBDDCResetTopography(PC pc) 1554674ae819SStefano Zampini { 1555674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 15564f1b2e48SStefano Zampini PetscInt i; 1557674ae819SStefano Zampini PetscErrorCode ierr; 1558674ae819SStefano Zampini 1559674ae819SStefano Zampini PetscFunctionBegin; 1560b9b85e73SStefano Zampini ierr = MatDestroy(&pcbddc->user_ChangeOfBasisMatrix);CHKERRQ(ierr); 1561674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 156216909a7fSStefano Zampini ierr = MatDestroy(&pcbddc->switch_static_change);CHKERRQ(ierr); 15631dd7afcfSStefano Zampini ierr = VecDestroy(&pcbddc->work_change);CHKERRQ(ierr); 1564674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 1565674ae819SStefano Zampini ierr = PCBDDCGraphReset(pcbddc->mat_graph);CHKERRQ(ierr); 15664f1b2e48SStefano Zampini for (i=0;i<pcbddc->n_local_subs;i++) { 15674f1b2e48SStefano Zampini ierr = ISDestroy(&pcbddc->local_subs[i]);CHKERRQ(ierr); 15684f1b2e48SStefano Zampini } 15694f1b2e48SStefano Zampini ierr = PetscFree(pcbddc->local_subs);CHKERRQ(ierr); 1570b96c3477SStefano Zampini ierr = PCBDDCSubSchursReset(pcbddc->sub_schurs);CHKERRQ(ierr); 1571674ae819SStefano Zampini PetscFunctionReturn(0); 1572674ae819SStefano Zampini } 1573674ae819SStefano Zampini 1574674ae819SStefano Zampini #undef __FUNCT__ 1575674ae819SStefano Zampini #define __FUNCT__ "PCBDDCResetSolvers" 1576674ae819SStefano Zampini PetscErrorCode PCBDDCResetSolvers(PC pc) 1577674ae819SStefano Zampini { 1578674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 1579674ae819SStefano Zampini PetscErrorCode ierr; 1580674ae819SStefano Zampini 1581674ae819SStefano Zampini PetscFunctionBegin; 1582674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->coarse_vec);CHKERRQ(ierr); 158358da7f69SStefano Zampini if (pcbddc->coarse_phi_B) { 1584ca92afb2SStefano Zampini PetscScalar *array; 158506656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&array);CHKERRQ(ierr); 158606656605SStefano Zampini ierr = PetscFree(array);CHKERRQ(ierr); 158758da7f69SStefano Zampini } 1588674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_B);CHKERRQ(ierr); 1589674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_D);CHKERRQ(ierr); 159015aaf578SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_B);CHKERRQ(ierr); 159115aaf578SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_D);CHKERRQ(ierr); 1592674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec1_P);CHKERRQ(ierr); 1593674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec1_C);CHKERRQ(ierr); 1594674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat2);CHKERRQ(ierr); 159506656605SStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat1);CHKERRQ(ierr); 1596674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec1_R);CHKERRQ(ierr); 1597674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec2_R);CHKERRQ(ierr); 15988ce42a96SStefano Zampini ierr = ISDestroy(&pcbddc->is_R_local);CHKERRQ(ierr); 1599674ae819SStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_B);CHKERRQ(ierr); 1600674ae819SStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_D);CHKERRQ(ierr); 1601674ae819SStefano Zampini ierr = VecScatterDestroy(&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 1602f4ddd8eeSStefano Zampini ierr = KSPDestroy(&pcbddc->ksp_D);CHKERRQ(ierr); 1603f4ddd8eeSStefano Zampini ierr = KSPDestroy(&pcbddc->ksp_R);CHKERRQ(ierr); 1604f4ddd8eeSStefano Zampini ierr = KSPDestroy(&pcbddc->coarse_ksp);CHKERRQ(ierr); 1605f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 1606727cdba6SStefano Zampini ierr = PetscFree(pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 16070e6343abSStefano Zampini ierr = PetscFree2(pcbddc->local_primal_ref_node,pcbddc->local_primal_ref_mult);CHKERRQ(ierr); 1608f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->global_primal_indices);CHKERRQ(ierr); 160970cf5478SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling);CHKERRQ(ierr); 161081d14e9dSStefano Zampini ierr = MatDestroy(&pcbddc->benign_change);CHKERRQ(ierr); 16110369aaf7SStefano Zampini ierr = VecDestroy(&pcbddc->benign_vec);CHKERRQ(ierr); 16121dd7afcfSStefano Zampini ierr = PCBDDCBenignShellMat(pc,PETSC_TRUE);CHKERRQ(ierr); 16134f1b2e48SStefano Zampini ierr = MatDestroy(&pcbddc->benign_B0);CHKERRQ(ierr); 16148b9f24d4SStefano Zampini ierr = PetscSFDestroy(&pcbddc->benign_sf);CHKERRQ(ierr); 1615ca92afb2SStefano Zampini if (pcbddc->benign_zerodiag_subs) { 1616ca92afb2SStefano Zampini PetscInt i; 1617ca92afb2SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 1618ca92afb2SStefano Zampini ierr = ISDestroy(&pcbddc->benign_zerodiag_subs[i]);CHKERRQ(ierr); 1619ca92afb2SStefano Zampini } 1620ca92afb2SStefano Zampini ierr = PetscFree(pcbddc->benign_zerodiag_subs);CHKERRQ(ierr); 1621ca92afb2SStefano Zampini } 16224f1b2e48SStefano Zampini ierr = PetscFree3(pcbddc->benign_p0_lidx,pcbddc->benign_p0_gidx,pcbddc->benign_p0);CHKERRQ(ierr); 1623674ae819SStefano Zampini PetscFunctionReturn(0); 1624674ae819SStefano Zampini } 1625674ae819SStefano Zampini 1626674ae819SStefano Zampini #undef __FUNCT__ 1627f4ddd8eeSStefano Zampini #define __FUNCT__ "PCBDDCSetUpLocalWorkVectors" 1628f4ddd8eeSStefano Zampini PetscErrorCode PCBDDCSetUpLocalWorkVectors(PC pc) 16296bfb1811SStefano Zampini { 16306bfb1811SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 16316bfb1811SStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 16326bfb1811SStefano Zampini VecType impVecType; 16334f1b2e48SStefano Zampini PetscInt n_constraints,n_R,old_size; 16346bfb1811SStefano Zampini PetscErrorCode ierr; 16356bfb1811SStefano Zampini 16366bfb1811SStefano Zampini PetscFunctionBegin; 16376c4ed002SBarry Smith if (!pcbddc->ConstraintMatrix) SETERRQ(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"BDDC Constraint matrix has not been created"); 1638e7b262bdSStefano Zampini /* get sizes */ 16394f1b2e48SStefano Zampini n_constraints = pcbddc->local_primal_size - pcbddc->benign_n - pcbddc->n_vertices; 1640b371cd4fSStefano Zampini n_R = pcis->n - pcbddc->n_vertices; 16416bfb1811SStefano Zampini ierr = VecGetType(pcis->vec1_N,&impVecType);CHKERRQ(ierr); 1642e7b262bdSStefano Zampini /* local work vectors (try to avoid unneeded work)*/ 1643e7b262bdSStefano Zampini /* R nodes */ 1644e7b262bdSStefano Zampini old_size = -1; 1645e7b262bdSStefano Zampini if (pcbddc->vec1_R) { 1646e7b262bdSStefano Zampini ierr = VecGetSize(pcbddc->vec1_R,&old_size);CHKERRQ(ierr); 1647e7b262bdSStefano Zampini } 1648e7b262bdSStefano Zampini if (n_R != old_size) { 1649e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec1_R);CHKERRQ(ierr); 1650e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec2_R);CHKERRQ(ierr); 16516bfb1811SStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_R);CHKERRQ(ierr); 16526bfb1811SStefano Zampini ierr = VecSetSizes(pcbddc->vec1_R,PETSC_DECIDE,n_R);CHKERRQ(ierr); 16536bfb1811SStefano Zampini ierr = VecSetType(pcbddc->vec1_R,impVecType);CHKERRQ(ierr); 16546bfb1811SStefano Zampini ierr = VecDuplicate(pcbddc->vec1_R,&pcbddc->vec2_R);CHKERRQ(ierr); 1655e7b262bdSStefano Zampini } 1656e7b262bdSStefano Zampini /* local primal dofs */ 1657e7b262bdSStefano Zampini old_size = -1; 1658e7b262bdSStefano Zampini if (pcbddc->vec1_P) { 1659e7b262bdSStefano Zampini ierr = VecGetSize(pcbddc->vec1_P,&old_size);CHKERRQ(ierr); 1660e7b262bdSStefano Zampini } 1661e9189074SStefano Zampini if (pcbddc->local_primal_size != old_size) { 1662e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec1_P);CHKERRQ(ierr); 166383b7ccabSStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_P);CHKERRQ(ierr); 1664e9189074SStefano Zampini ierr = VecSetSizes(pcbddc->vec1_P,PETSC_DECIDE,pcbddc->local_primal_size);CHKERRQ(ierr); 16656bfb1811SStefano Zampini ierr = VecSetType(pcbddc->vec1_P,impVecType);CHKERRQ(ierr); 1666e7b262bdSStefano Zampini } 1667e7b262bdSStefano Zampini /* local explicit constraints */ 1668e7b262bdSStefano Zampini old_size = -1; 1669e7b262bdSStefano Zampini if (pcbddc->vec1_C) { 1670e7b262bdSStefano Zampini ierr = VecGetSize(pcbddc->vec1_C,&old_size);CHKERRQ(ierr); 1671e7b262bdSStefano Zampini } 1672e7b262bdSStefano Zampini if (n_constraints && n_constraints != old_size) { 1673e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec1_C);CHKERRQ(ierr); 167483b7ccabSStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_C);CHKERRQ(ierr); 167583b7ccabSStefano Zampini ierr = VecSetSizes(pcbddc->vec1_C,PETSC_DECIDE,n_constraints);CHKERRQ(ierr); 167683b7ccabSStefano Zampini ierr = VecSetType(pcbddc->vec1_C,impVecType);CHKERRQ(ierr); 167783b7ccabSStefano Zampini } 16786bfb1811SStefano Zampini PetscFunctionReturn(0); 16796bfb1811SStefano Zampini } 16806bfb1811SStefano Zampini 16816bfb1811SStefano Zampini #undef __FUNCT__ 168247f4ddc3SStefano Zampini #define __FUNCT__ "PCBDDCSetUpCorrection" 168347f4ddc3SStefano Zampini PetscErrorCode PCBDDCSetUpCorrection(PC pc, PetscScalar **coarse_submat_vals_n) 168488ebb749SStefano Zampini { 168525084f0cSStefano Zampini PetscErrorCode ierr; 168625084f0cSStefano Zampini /* pointers to pcis and pcbddc */ 168788ebb749SStefano Zampini PC_IS* pcis = (PC_IS*)pc->data; 168888ebb749SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 1689d62866d3SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 169025084f0cSStefano Zampini /* submatrices of local problem */ 169180677318SStefano Zampini Mat A_RV,A_VR,A_VV,local_auxmat2_R; 169206656605SStefano Zampini /* submatrices of local coarse problem */ 169306656605SStefano Zampini Mat S_VV,S_CV,S_VC,S_CC; 169425084f0cSStefano Zampini /* working matrices */ 169506656605SStefano Zampini Mat C_CR; 169625084f0cSStefano Zampini /* additional working stuff */ 169706656605SStefano Zampini PC pc_R; 16984f1b2e48SStefano Zampini Mat F; 16995cbda25cSStefano Zampini Vec dummy_vec; 1700a3df083aSStefano Zampini PetscBool isLU,isCHOL,isILU,need_benign_correction; 170125084f0cSStefano Zampini PetscScalar *coarse_submat_vals; /* TODO: use a PETSc matrix */ 170206656605SStefano Zampini PetscScalar *work; 170306656605SStefano Zampini PetscInt *idx_V_B; 1704ffd830a3SStefano Zampini PetscInt lda_rhs,n,n_vertices,n_constraints,*p0_lidx_I; 170506656605SStefano Zampini PetscInt i,n_R,n_D,n_B; 1706ffd830a3SStefano Zampini 170725084f0cSStefano Zampini /* some shortcuts to scalars */ 170806656605SStefano Zampini PetscScalar one=1.0,m_one=-1.0; 170988ebb749SStefano Zampini 171088ebb749SStefano Zampini PetscFunctionBegin; 1711ffd830a3SStefano Zampini if (!pcbddc->symmetric_primal && pcbddc->benign_n) { 1712ffd830a3SStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Non-symmetric primal basis computation with benign trick not yet implemented"); 1713ffd830a3SStefano Zampini } 1714ffd830a3SStefano Zampini 1715ffd830a3SStefano Zampini /* Set Non-overlapping dimensions */ 1716b371cd4fSStefano Zampini n_vertices = pcbddc->n_vertices; 17174f1b2e48SStefano Zampini n_constraints = pcbddc->local_primal_size - pcbddc->benign_n - n_vertices; 1718b371cd4fSStefano Zampini n_B = pcis->n_B; 1719b371cd4fSStefano Zampini n_D = pcis->n - n_B; 172088ebb749SStefano Zampini n_R = pcis->n - n_vertices; 172188ebb749SStefano Zampini 172288ebb749SStefano Zampini /* vertices in boundary numbering */ 1723785e854fSJed Brown ierr = PetscMalloc1(n_vertices,&idx_V_B);CHKERRQ(ierr); 17240e6343abSStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,n_vertices,pcbddc->local_primal_ref_node,&i,idx_V_B);CHKERRQ(ierr); 17256c4ed002SBarry Smith if (i != n_vertices) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Error in boundary numbering for BDDC vertices! %D != %D\n",n_vertices,i); 172688ebb749SStefano Zampini 172706656605SStefano Zampini /* Subdomain contribution (Non-overlapping) to coarse matrix */ 1728019a44ceSStefano Zampini ierr = PetscCalloc1(pcbddc->local_primal_size*pcbddc->local_primal_size,&coarse_submat_vals);CHKERRQ(ierr); 172906656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_vertices,n_vertices,coarse_submat_vals,&S_VV);CHKERRQ(ierr); 173006656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_VV,pcbddc->local_primal_size);CHKERRQ(ierr); 173106656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_constraints,n_vertices,coarse_submat_vals+n_vertices,&S_CV);CHKERRQ(ierr); 173206656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_CV,pcbddc->local_primal_size);CHKERRQ(ierr); 173306656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_vertices,n_constraints,coarse_submat_vals+pcbddc->local_primal_size*n_vertices,&S_VC);CHKERRQ(ierr); 173406656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_VC,pcbddc->local_primal_size);CHKERRQ(ierr); 173506656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_constraints,n_constraints,coarse_submat_vals+(pcbddc->local_primal_size+1)*n_vertices,&S_CC);CHKERRQ(ierr); 173606656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_CC,pcbddc->local_primal_size);CHKERRQ(ierr); 173706656605SStefano Zampini 173806656605SStefano Zampini /* determine if can use MatSolve routines instead of calling KSPSolve on ksp_R */ 173906656605SStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_R);CHKERRQ(ierr); 174006656605SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_R,PCLU,&isLU);CHKERRQ(ierr); 174106656605SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_R,PCILU,&isILU);CHKERRQ(ierr); 174206656605SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_R,PCCHOLESKY,&isCHOL);CHKERRQ(ierr); 1743ffd830a3SStefano Zampini lda_rhs = n_R; 1744a3df083aSStefano Zampini need_benign_correction = PETSC_FALSE; 174506656605SStefano Zampini if (isLU || isILU || isCHOL) { 174606656605SStefano Zampini ierr = PCFactorGetMatrix(pc_R,&F);CHKERRQ(ierr); 1747df4d28bfSStefano Zampini } else if (sub_schurs->reuse_solver) { 1748df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 1749d62866d3SStefano Zampini MatFactorType type; 1750d62866d3SStefano Zampini 1751df4d28bfSStefano Zampini F = reuse_solver->F; 17526816873aSStefano Zampini ierr = MatGetFactorType(F,&type);CHKERRQ(ierr); 1753d62866d3SStefano Zampini if (type == MAT_FACTOR_CHOLESKY) isCHOL = PETSC_TRUE; 1754ffd830a3SStefano Zampini ierr = MatGetSize(F,&lda_rhs,NULL);CHKERRQ(ierr); 175522db5ddcSStefano Zampini need_benign_correction = (PetscBool)(!!reuse_solver->benign_n); 175606656605SStefano Zampini } else { 175706656605SStefano Zampini F = NULL; 175806656605SStefano Zampini } 175906656605SStefano Zampini 1760ffd830a3SStefano Zampini /* allocate workspace */ 1761ffd830a3SStefano Zampini n = 0; 1762ffd830a3SStefano Zampini if (n_constraints) { 1763ffd830a3SStefano Zampini n += lda_rhs*n_constraints; 1764ffd830a3SStefano Zampini } 1765ffd830a3SStefano Zampini if (n_vertices) { 1766ffd830a3SStefano Zampini n = PetscMax(2*lda_rhs*n_vertices,n); 1767ffd830a3SStefano Zampini n = PetscMax((lda_rhs+n_B)*n_vertices,n); 1768ffd830a3SStefano Zampini } 1769ffd830a3SStefano Zampini ierr = PetscMalloc1(n,&work);CHKERRQ(ierr); 1770ffd830a3SStefano Zampini 17715cbda25cSStefano Zampini /* create dummy vector to modify rhs and sol of MatMatSolve (work array will never be used) */ 17725cbda25cSStefano Zampini dummy_vec = NULL; 17735cbda25cSStefano Zampini if (need_benign_correction && lda_rhs != n_R && F) { 17745cbda25cSStefano Zampini ierr = VecCreateSeqWithArray(PETSC_COMM_SELF,1,lda_rhs,work,&dummy_vec);CHKERRQ(ierr); 17755cbda25cSStefano Zampini } 17765cbda25cSStefano Zampini 177788ebb749SStefano Zampini /* Precompute stuffs needed for preprocessing and application of BDDC*/ 177888ebb749SStefano Zampini if (n_constraints) { 177972b8c272SStefano Zampini Mat M1,M2,M3,C_B; 178006656605SStefano Zampini IS is_aux; 178180677318SStefano Zampini PetscScalar *array,*array2; 178206656605SStefano Zampini 1783f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat1);CHKERRQ(ierr); 178480677318SStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat2);CHKERRQ(ierr); 178588ebb749SStefano Zampini 178625084f0cSStefano Zampini /* Extract constraints on R nodes: C_{CR} */ 178725084f0cSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,n_constraints,n_vertices,1,&is_aux);CHKERRQ(ierr); 17888ce42a96SStefano Zampini ierr = MatGetSubMatrix(pcbddc->ConstraintMatrix,is_aux,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&C_CR);CHKERRQ(ierr); 178972b8c272SStefano Zampini ierr = MatGetSubMatrix(pcbddc->ConstraintMatrix,is_aux,pcis->is_B_local,MAT_INITIAL_MATRIX,&C_B);CHKERRQ(ierr); 179088ebb749SStefano Zampini 179180677318SStefano Zampini /* Assemble local_auxmat2_R = (- A_{RR}^{-1} C^T_{CR}) needed by BDDC setup */ 179280677318SStefano Zampini /* Assemble pcbddc->local_auxmat2 = R_to_B (- A_{RR}^{-1} C^T_{CR}) needed by BDDC application */ 1793ffd830a3SStefano Zampini ierr = PetscMemzero(work,lda_rhs*n_constraints*sizeof(PetscScalar));CHKERRQ(ierr); 179488ebb749SStefano Zampini for (i=0;i<n_constraints;i++) { 179506656605SStefano Zampini const PetscScalar *row_cmat_values; 179606656605SStefano Zampini const PetscInt *row_cmat_indices; 179706656605SStefano Zampini PetscInt size_of_constraint,j; 179888ebb749SStefano Zampini 179906656605SStefano Zampini ierr = MatGetRow(C_CR,i,&size_of_constraint,&row_cmat_indices,&row_cmat_values);CHKERRQ(ierr); 180006656605SStefano Zampini for (j=0;j<size_of_constraint;j++) { 1801ffd830a3SStefano Zampini work[row_cmat_indices[j]+i*lda_rhs] = -row_cmat_values[j]; 180206656605SStefano Zampini } 180306656605SStefano Zampini ierr = MatRestoreRow(C_CR,i,&size_of_constraint,&row_cmat_indices,&row_cmat_values);CHKERRQ(ierr); 180406656605SStefano Zampini } 1805ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_constraints,NULL,&local_auxmat2_R);CHKERRQ(ierr); 180606656605SStefano Zampini if (F) { 180706656605SStefano Zampini Mat B; 180806656605SStefano Zampini 1809ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_constraints,work,&B);CHKERRQ(ierr); 1810a3df083aSStefano Zampini if (need_benign_correction) { 1811df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 1812a3df083aSStefano Zampini 181372b8c272SStefano Zampini /* rhs is already zero on interior dofs, no need to change the rhs */ 181472b8c272SStefano Zampini ierr = PetscMemzero(reuse_solver->benign_save_vals,pcbddc->benign_n*sizeof(PetscScalar));CHKERRQ(ierr); 1815a3df083aSStefano Zampini } 181680677318SStefano Zampini ierr = MatMatSolve(F,B,local_auxmat2_R);CHKERRQ(ierr); 1817a3df083aSStefano Zampini if (need_benign_correction) { 1818a3df083aSStefano Zampini PetscScalar *marr; 1819df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 1820a3df083aSStefano Zampini 1821a3df083aSStefano Zampini ierr = MatDenseGetArray(local_auxmat2_R,&marr);CHKERRQ(ierr); 18225cbda25cSStefano Zampini if (lda_rhs != n_R) { 18235cbda25cSStefano Zampini for (i=0;i<n_constraints;i++) { 18245cbda25cSStefano Zampini ierr = VecPlaceArray(dummy_vec,marr+i*lda_rhs);CHKERRQ(ierr); 18255cbda25cSStefano Zampini ierr = PCBDDCReuseSolversBenignAdapt(reuse_solver,dummy_vec,NULL,PETSC_TRUE,PETSC_TRUE);CHKERRQ(ierr); 18265cbda25cSStefano Zampini ierr = VecResetArray(dummy_vec);CHKERRQ(ierr); 18275cbda25cSStefano Zampini } 18285cbda25cSStefano Zampini } else { 1829a3df083aSStefano Zampini for (i=0;i<n_constraints;i++) { 1830a3df083aSStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,marr+i*lda_rhs);CHKERRQ(ierr); 18315cbda25cSStefano Zampini ierr = PCBDDCReuseSolversBenignAdapt(reuse_solver,pcbddc->vec1_R,NULL,PETSC_TRUE,PETSC_TRUE);CHKERRQ(ierr); 1832a3df083aSStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 1833a3df083aSStefano Zampini } 18345cbda25cSStefano Zampini } 1835a3df083aSStefano Zampini ierr = MatDenseRestoreArray(local_auxmat2_R,&marr);CHKERRQ(ierr); 1836a3df083aSStefano Zampini } 183706656605SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 183806656605SStefano Zampini } else { 183980677318SStefano Zampini PetscScalar *marr; 184080677318SStefano Zampini 184180677318SStefano Zampini ierr = MatDenseGetArray(local_auxmat2_R,&marr);CHKERRQ(ierr); 184206656605SStefano Zampini for (i=0;i<n_constraints;i++) { 1843ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*lda_rhs);CHKERRQ(ierr); 1844ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,marr+i*lda_rhs);CHKERRQ(ierr); 184506656605SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 184606656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 184706656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 184806656605SStefano Zampini } 184980677318SStefano Zampini ierr = MatDenseRestoreArray(local_auxmat2_R,&marr);CHKERRQ(ierr); 185006656605SStefano Zampini } 185180677318SStefano Zampini if (!pcbddc->switch_static) { 185280677318SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,n_constraints,NULL,&pcbddc->local_auxmat2);CHKERRQ(ierr); 185380677318SStefano Zampini ierr = MatDenseGetArray(pcbddc->local_auxmat2,&array);CHKERRQ(ierr); 185480677318SStefano Zampini ierr = MatDenseGetArray(local_auxmat2_R,&array2);CHKERRQ(ierr); 185580677318SStefano Zampini for (i=0;i<n_constraints;i++) { 1856ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,array2+i*lda_rhs);CHKERRQ(ierr); 185780677318SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,array+i*n_B);CHKERRQ(ierr); 185880677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 185980677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 186080677318SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 186180677318SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 186280677318SStefano Zampini } 186380677318SStefano Zampini ierr = MatDenseRestoreArray(local_auxmat2_R,&array2);CHKERRQ(ierr); 186480677318SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->local_auxmat2,&array);CHKERRQ(ierr); 186572b8c272SStefano Zampini ierr = MatMatMult(C_B,pcbddc->local_auxmat2,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&M3);CHKERRQ(ierr); 186680677318SStefano Zampini } else { 1867ffd830a3SStefano Zampini if (lda_rhs != n_R) { 1868ffd830a3SStefano Zampini IS dummy; 1869ffd830a3SStefano Zampini 1870ffd830a3SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,n_R,0,1,&dummy);CHKERRQ(ierr); 187172b8c272SStefano Zampini ierr = MatGetSubMatrix(local_auxmat2_R,dummy,NULL,MAT_INITIAL_MATRIX,&pcbddc->local_auxmat2);CHKERRQ(ierr); 1872ffd830a3SStefano Zampini ierr = ISDestroy(&dummy);CHKERRQ(ierr); 1873ffd830a3SStefano Zampini } else { 187480677318SStefano Zampini ierr = PetscObjectReference((PetscObject)local_auxmat2_R);CHKERRQ(ierr); 187580677318SStefano Zampini pcbddc->local_auxmat2 = local_auxmat2_R; 1876ffd830a3SStefano Zampini } 187725084f0cSStefano Zampini ierr = MatMatMult(C_CR,pcbddc->local_auxmat2,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&M3);CHKERRQ(ierr); 187880677318SStefano Zampini } 187980677318SStefano Zampini ierr = ISDestroy(&is_aux);CHKERRQ(ierr); 188080677318SStefano Zampini /* Assemble explicitly S_CC = ( C_{CR} A_{RR}^{-1} C^T_{CR} )^{-1} */ 188180677318SStefano Zampini ierr = MatScale(M3,m_one);CHKERRQ(ierr); 188206656605SStefano Zampini ierr = MatDuplicate(M3,MAT_DO_NOT_COPY_VALUES,&M1);CHKERRQ(ierr); 188306656605SStefano Zampini ierr = MatDuplicate(M3,MAT_DO_NOT_COPY_VALUES,&M2);CHKERRQ(ierr); 188480677318SStefano Zampini if (isCHOL) { 188580677318SStefano Zampini ierr = MatCholeskyFactor(M3,NULL,NULL);CHKERRQ(ierr); 188680677318SStefano Zampini } else { 188725084f0cSStefano Zampini ierr = MatLUFactor(M3,NULL,NULL,NULL);CHKERRQ(ierr); 188880677318SStefano Zampini } 188980677318SStefano Zampini ierr = VecSet(pcbddc->vec1_C,one);CHKERRQ(ierr); 189006656605SStefano Zampini ierr = MatDiagonalSet(M2,pcbddc->vec1_C,INSERT_VALUES);CHKERRQ(ierr); 189125084f0cSStefano Zampini ierr = MatMatSolve(M3,M2,M1);CHKERRQ(ierr); 189225084f0cSStefano Zampini ierr = MatDestroy(&M2);CHKERRQ(ierr); 189325084f0cSStefano Zampini ierr = MatDestroy(&M3);CHKERRQ(ierr); 189480677318SStefano Zampini /* Assemble local_auxmat1 = S_CC*C_{CB} needed by BDDC application in KSP and in preproc */ 189572b8c272SStefano Zampini ierr = MatMatMult(M1,C_B,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&pcbddc->local_auxmat1);CHKERRQ(ierr); 189672b8c272SStefano Zampini ierr = MatDestroy(&C_B);CHKERRQ(ierr); 189706656605SStefano Zampini ierr = MatCopy(M1,S_CC,SAME_NONZERO_PATTERN);CHKERRQ(ierr); /* S_CC can have a different LDA, MatMatSolve doesn't support it */ 189806656605SStefano Zampini ierr = MatDestroy(&M1);CHKERRQ(ierr); 1899f4ddd8eeSStefano Zampini } 1900fc227af8SStefano Zampini 1901fc227af8SStefano Zampini /* Get submatrices from subdomain matrix */ 190288ebb749SStefano Zampini if (n_vertices) { 190306656605SStefano Zampini IS is_aux; 19043a50541eSStefano Zampini 1905df4d28bfSStefano Zampini if (sub_schurs->reuse_solver) { /* is_R_local is not sorted, ISComplement doesn't like it */ 19066816873aSStefano Zampini IS tis; 19076816873aSStefano Zampini 19086816873aSStefano Zampini ierr = ISDuplicate(pcbddc->is_R_local,&tis);CHKERRQ(ierr); 19096816873aSStefano Zampini ierr = ISSort(tis);CHKERRQ(ierr); 19106816873aSStefano Zampini ierr = ISComplement(tis,0,pcis->n,&is_aux);CHKERRQ(ierr); 19116816873aSStefano Zampini ierr = ISDestroy(&tis);CHKERRQ(ierr); 19126816873aSStefano Zampini } else { 19133a50541eSStefano Zampini ierr = ISComplement(pcbddc->is_R_local,0,pcis->n,&is_aux);CHKERRQ(ierr); 19146816873aSStefano Zampini } 19159577ea80SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,pcbddc->is_R_local,is_aux,MAT_INITIAL_MATRIX,&A_RV);CHKERRQ(ierr); 19169577ea80SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,is_aux,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&A_VR);CHKERRQ(ierr); 191704708bb6SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,is_aux,is_aux,MAT_INITIAL_MATRIX,&A_VV);CHKERRQ(ierr); 191825084f0cSStefano Zampini ierr = ISDestroy(&is_aux);CHKERRQ(ierr); 191988ebb749SStefano Zampini } 192088ebb749SStefano Zampini 192188ebb749SStefano Zampini /* Matrix of coarse basis functions (local) */ 1922f4ddd8eeSStefano Zampini if (pcbddc->coarse_phi_B) { 192306656605SStefano Zampini PetscInt on_B,on_primal,on_D=n_D; 192406656605SStefano Zampini if (pcbddc->coarse_phi_D) { 192506656605SStefano Zampini ierr = MatGetSize(pcbddc->coarse_phi_D,&on_D,NULL);CHKERRQ(ierr); 192606656605SStefano Zampini } 1927f4ddd8eeSStefano Zampini ierr = MatGetSize(pcbddc->coarse_phi_B,&on_B,&on_primal);CHKERRQ(ierr); 192806656605SStefano Zampini if (on_B != n_B || on_primal != pcbddc->local_primal_size || on_D != n_D) { 192906656605SStefano Zampini PetscScalar *marray; 193006656605SStefano Zampini 193106656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&marray);CHKERRQ(ierr); 193206656605SStefano Zampini ierr = PetscFree(marray);CHKERRQ(ierr); 1933f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_B);CHKERRQ(ierr); 1934f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_B);CHKERRQ(ierr); 1935f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_D);CHKERRQ(ierr); 1936f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_D);CHKERRQ(ierr); 1937f4ddd8eeSStefano Zampini } 1938f4ddd8eeSStefano Zampini } 193906656605SStefano Zampini 1940f4ddd8eeSStefano Zampini if (!pcbddc->coarse_phi_B) { 194106656605SStefano Zampini PetscScalar *marray; 194288ebb749SStefano Zampini 194306656605SStefano Zampini n = n_B*pcbddc->local_primal_size; 19448eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 194506656605SStefano Zampini n += n_D*pcbddc->local_primal_size; 194688ebb749SStefano Zampini } 19473301b35fSStefano Zampini if (!pcbddc->symmetric_primal) { 194806656605SStefano Zampini n *= 2; 194988ebb749SStefano Zampini } 195006656605SStefano Zampini ierr = PetscCalloc1(n,&marray);CHKERRQ(ierr); 195106656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,pcbddc->local_primal_size,marray,&pcbddc->coarse_phi_B);CHKERRQ(ierr); 195206656605SStefano Zampini n = n_B*pcbddc->local_primal_size; 19538eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 195406656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_D,pcbddc->local_primal_size,marray+n,&pcbddc->coarse_phi_D);CHKERRQ(ierr); 195506656605SStefano Zampini n += n_D*pcbddc->local_primal_size; 195688ebb749SStefano Zampini } 19573301b35fSStefano Zampini if (!pcbddc->symmetric_primal) { 195806656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,pcbddc->local_primal_size,marray+n,&pcbddc->coarse_psi_B);CHKERRQ(ierr); 19598eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 196006656605SStefano Zampini n = n_B*pcbddc->local_primal_size; 196106656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_D,pcbddc->local_primal_size,marray+n,&pcbddc->coarse_psi_D);CHKERRQ(ierr); 196288ebb749SStefano Zampini } 196388ebb749SStefano Zampini } else { 1964c0553b1fSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->coarse_phi_B);CHKERRQ(ierr); 1965c0553b1fSStefano Zampini pcbddc->coarse_psi_B = pcbddc->coarse_phi_B; 19661b968477SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 1967c0553b1fSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->coarse_phi_D);CHKERRQ(ierr); 1968c0553b1fSStefano Zampini pcbddc->coarse_psi_D = pcbddc->coarse_phi_D; 1969c0553b1fSStefano Zampini } 197088ebb749SStefano Zampini } 197106656605SStefano Zampini } 1972019a44ceSStefano Zampini 197306656605SStefano Zampini /* We are now ready to evaluate coarse basis functions and subdomain contribution to coarse problem */ 19744f1b2e48SStefano Zampini p0_lidx_I = NULL; 19754f1b2e48SStefano Zampini if (pcbddc->benign_n && (pcbddc->switch_static || pcbddc->dbg_flag)) { 1976d12edf2fSStefano Zampini const PetscInt *idxs; 1977d12edf2fSStefano Zampini 1978d12edf2fSStefano Zampini ierr = ISGetIndices(pcis->is_I_local,&idxs);CHKERRQ(ierr); 19794f1b2e48SStefano Zampini ierr = PetscMalloc1(pcbddc->benign_n,&p0_lidx_I);CHKERRQ(ierr); 19804f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 19814f1b2e48SStefano Zampini ierr = PetscFindInt(pcbddc->benign_p0_lidx[i],pcis->n-pcis->n_B,idxs,&p0_lidx_I[i]);CHKERRQ(ierr); 19824f1b2e48SStefano Zampini } 1983d12edf2fSStefano Zampini ierr = ISRestoreIndices(pcis->is_I_local,&idxs);CHKERRQ(ierr); 1984d12edf2fSStefano Zampini } 1985d16cbb6bSStefano Zampini 198606656605SStefano Zampini /* vertices */ 198706656605SStefano Zampini if (n_vertices) { 198816f15bc4SStefano Zampini 1989af25d912SStefano Zampini ierr = MatConvert(A_VV,MATDENSE,MAT_INPLACE_MATRIX,&A_VV);CHKERRQ(ierr); 199004708bb6SStefano Zampini 199116f15bc4SStefano Zampini if (n_R) { 199214393ed6SStefano Zampini Mat A_RRmA_RV,A_RV_bcorr=NULL,S_VVt; /* S_VVt with LDA=N */ 199306656605SStefano Zampini PetscBLASInt B_N,B_one = 1; 199416f15bc4SStefano Zampini PetscScalar *x,*y; 199504708bb6SStefano Zampini PetscBool isseqaij; 199606656605SStefano Zampini 199721eccb56SStefano Zampini ierr = MatScale(A_RV,m_one);CHKERRQ(ierr); 199814393ed6SStefano Zampini if (need_benign_correction) { 199914393ed6SStefano Zampini ISLocalToGlobalMapping RtoN; 200014393ed6SStefano Zampini IS is_p0; 200114393ed6SStefano Zampini PetscInt *idxs_p0,n; 200214393ed6SStefano Zampini 200314393ed6SStefano Zampini ierr = PetscMalloc1(pcbddc->benign_n,&idxs_p0);CHKERRQ(ierr); 200414393ed6SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(pcbddc->is_R_local,&RtoN);CHKERRQ(ierr); 200514393ed6SStefano Zampini ierr = ISGlobalToLocalMappingApply(RtoN,IS_GTOLM_DROP,pcbddc->benign_n,pcbddc->benign_p0_lidx,&n,idxs_p0);CHKERRQ(ierr); 2006af25d912SStefano Zampini if (n != pcbddc->benign_n) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Error in R numbering for benign p0! %d != %d\n",n,pcbddc->benign_n); 200714393ed6SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&RtoN);CHKERRQ(ierr); 200814393ed6SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,n,idxs_p0,PETSC_OWN_POINTER,&is_p0);CHKERRQ(ierr); 200914393ed6SStefano Zampini ierr = MatGetSubMatrix(A_RV,is_p0,NULL,MAT_INITIAL_MATRIX,&A_RV_bcorr);CHKERRQ(ierr); 201014393ed6SStefano Zampini ierr = ISDestroy(&is_p0);CHKERRQ(ierr); 201114393ed6SStefano Zampini } 201214393ed6SStefano Zampini 2013ffd830a3SStefano Zampini if (lda_rhs == n_R) { 2014af25d912SStefano Zampini ierr = MatConvert(A_RV,MATDENSE,MAT_INPLACE_MATRIX,&A_RV);CHKERRQ(ierr); 2015ffd830a3SStefano Zampini } else { 2016ca92afb2SStefano Zampini PetscScalar *av,*array; 2017ca92afb2SStefano Zampini const PetscInt *xadj,*adjncy; 2018ca92afb2SStefano Zampini PetscInt n; 2019ca92afb2SStefano Zampini PetscBool flg_row; 2020ffd830a3SStefano Zampini 2021ca92afb2SStefano Zampini array = work+lda_rhs*n_vertices; 2022ca92afb2SStefano Zampini ierr = PetscMemzero(array,lda_rhs*n_vertices*sizeof(PetscScalar));CHKERRQ(ierr); 20239d54b7f4SStefano Zampini ierr = MatConvert(A_RV,MATSEQAIJ,MAT_INPLACE_MATRIX,&A_RV);CHKERRQ(ierr); 2024ca92afb2SStefano Zampini ierr = MatGetRowIJ(A_RV,0,PETSC_FALSE,PETSC_FALSE,&n,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 2025ca92afb2SStefano Zampini ierr = MatSeqAIJGetArray(A_RV,&av);CHKERRQ(ierr); 2026ca92afb2SStefano Zampini for (i=0;i<n;i++) { 2027ca92afb2SStefano Zampini PetscInt j; 2028ca92afb2SStefano Zampini for (j=xadj[i];j<xadj[i+1];j++) array[lda_rhs*adjncy[j]+i] = av[j]; 2029ffd830a3SStefano Zampini } 2030ca92afb2SStefano Zampini ierr = MatRestoreRowIJ(A_RV,0,PETSC_FALSE,PETSC_FALSE,&n,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 2031ca92afb2SStefano Zampini ierr = MatDestroy(&A_RV);CHKERRQ(ierr); 2032ca92afb2SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_vertices,array,&A_RV);CHKERRQ(ierr); 2033ffd830a3SStefano Zampini } 2034ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_vertices,work,&A_RRmA_RV);CHKERRQ(ierr); 2035a3df083aSStefano Zampini if (need_benign_correction) { 2036df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 2037a3df083aSStefano Zampini PetscScalar *marr; 2038a3df083aSStefano Zampini 2039a3df083aSStefano Zampini ierr = MatDenseGetArray(A_RV,&marr);CHKERRQ(ierr); 204014393ed6SStefano Zampini /* need \Phi^T A_RV = (I+L)A_RV, L given by 204114393ed6SStefano Zampini 204214393ed6SStefano Zampini | 0 0 0 | (V) 204314393ed6SStefano Zampini L = | 0 0 -1 | (P-p0) 204414393ed6SStefano Zampini | 0 0 -1 | (p0) 204514393ed6SStefano Zampini 204614393ed6SStefano Zampini */ 2047df4d28bfSStefano Zampini for (i=0;i<reuse_solver->benign_n;i++) { 204814393ed6SStefano Zampini const PetscScalar *vals; 204914393ed6SStefano Zampini const PetscInt *idxs,*idxs_zero; 205014393ed6SStefano Zampini PetscInt n,j,nz; 205114393ed6SStefano Zampini 2052df4d28bfSStefano Zampini ierr = ISGetLocalSize(reuse_solver->benign_zerodiag_subs[i],&nz);CHKERRQ(ierr); 2053df4d28bfSStefano Zampini ierr = ISGetIndices(reuse_solver->benign_zerodiag_subs[i],&idxs_zero);CHKERRQ(ierr); 205414393ed6SStefano Zampini ierr = MatGetRow(A_RV_bcorr,i,&n,&idxs,&vals);CHKERRQ(ierr); 205514393ed6SStefano Zampini for (j=0;j<n;j++) { 205614393ed6SStefano Zampini PetscScalar val = vals[j]; 205714393ed6SStefano Zampini PetscInt k,col = idxs[j]; 205814393ed6SStefano Zampini for (k=0;k<nz;k++) marr[idxs_zero[k]+lda_rhs*col] -= val; 205914393ed6SStefano Zampini } 206014393ed6SStefano Zampini ierr = MatRestoreRow(A_RV_bcorr,i,&n,&idxs,&vals);CHKERRQ(ierr); 2061df4d28bfSStefano Zampini ierr = ISRestoreIndices(reuse_solver->benign_zerodiag_subs[i],&idxs_zero);CHKERRQ(ierr); 206214393ed6SStefano Zampini } 206372b8c272SStefano Zampini ierr = MatDenseRestoreArray(A_RV,&marr);CHKERRQ(ierr); 206472b8c272SStefano Zampini } 206572b8c272SStefano Zampini if (F) { 206614393ed6SStefano Zampini /* need to correct the rhs */ 206772b8c272SStefano Zampini if (need_benign_correction) { 206872b8c272SStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 206972b8c272SStefano Zampini PetscScalar *marr; 207072b8c272SStefano Zampini 207172b8c272SStefano Zampini ierr = MatDenseGetArray(A_RV,&marr);CHKERRQ(ierr); 20725cbda25cSStefano Zampini if (lda_rhs != n_R) { 20735cbda25cSStefano Zampini for (i=0;i<n_vertices;i++) { 20745cbda25cSStefano Zampini ierr = VecPlaceArray(dummy_vec,marr+i*lda_rhs);CHKERRQ(ierr); 20755cbda25cSStefano Zampini ierr = PCBDDCReuseSolversBenignAdapt(reuse_solver,dummy_vec,NULL,PETSC_FALSE,PETSC_TRUE);CHKERRQ(ierr); 20765cbda25cSStefano Zampini ierr = VecResetArray(dummy_vec);CHKERRQ(ierr); 20775cbda25cSStefano Zampini } 20785cbda25cSStefano Zampini } else { 2079a3df083aSStefano Zampini for (i=0;i<n_vertices;i++) { 2080a3df083aSStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,marr+i*lda_rhs);CHKERRQ(ierr); 20815cbda25cSStefano Zampini ierr = PCBDDCReuseSolversBenignAdapt(reuse_solver,pcbddc->vec1_R,NULL,PETSC_FALSE,PETSC_TRUE);CHKERRQ(ierr); 2082a3df083aSStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 2083a3df083aSStefano Zampini } 20845cbda25cSStefano Zampini } 2085a3df083aSStefano Zampini ierr = MatDenseRestoreArray(A_RV,&marr);CHKERRQ(ierr); 2086a3df083aSStefano Zampini } 208706656605SStefano Zampini ierr = MatMatSolve(F,A_RV,A_RRmA_RV);CHKERRQ(ierr); 208814393ed6SStefano Zampini /* need to correct the solution */ 2089a3df083aSStefano Zampini if (need_benign_correction) { 2090df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 2091a3df083aSStefano Zampini PetscScalar *marr; 2092a3df083aSStefano Zampini 2093a3df083aSStefano Zampini ierr = MatDenseGetArray(A_RRmA_RV,&marr);CHKERRQ(ierr); 20945cbda25cSStefano Zampini if (lda_rhs != n_R) { 20955cbda25cSStefano Zampini for (i=0;i<n_vertices;i++) { 20965cbda25cSStefano Zampini ierr = VecPlaceArray(dummy_vec,marr+i*lda_rhs);CHKERRQ(ierr); 20975cbda25cSStefano Zampini ierr = PCBDDCReuseSolversBenignAdapt(reuse_solver,dummy_vec,NULL,PETSC_TRUE,PETSC_TRUE);CHKERRQ(ierr); 20985cbda25cSStefano Zampini ierr = VecResetArray(dummy_vec);CHKERRQ(ierr); 20995cbda25cSStefano Zampini } 21005cbda25cSStefano Zampini } else { 2101a3df083aSStefano Zampini for (i=0;i<n_vertices;i++) { 2102a3df083aSStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,marr+i*lda_rhs);CHKERRQ(ierr); 21035cbda25cSStefano Zampini ierr = PCBDDCReuseSolversBenignAdapt(reuse_solver,pcbddc->vec1_R,NULL,PETSC_TRUE,PETSC_TRUE);CHKERRQ(ierr); 2104a3df083aSStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 2105a3df083aSStefano Zampini } 21065cbda25cSStefano Zampini } 2107a3df083aSStefano Zampini ierr = MatDenseRestoreArray(A_RRmA_RV,&marr);CHKERRQ(ierr); 2108a3df083aSStefano Zampini } 210906656605SStefano Zampini } else { 211006656605SStefano Zampini ierr = MatDenseGetArray(A_RV,&y);CHKERRQ(ierr); 211106656605SStefano Zampini for (i=0;i<n_vertices;i++) { 2112ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,y+i*lda_rhs);CHKERRQ(ierr); 2113ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,work+i*lda_rhs);CHKERRQ(ierr); 211406656605SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 211506656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 211606656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 211706656605SStefano Zampini } 211806656605SStefano Zampini ierr = MatDenseRestoreArray(A_RV,&y);CHKERRQ(ierr); 211906656605SStefano Zampini } 212080677318SStefano Zampini ierr = MatDestroy(&A_RV);CHKERRQ(ierr); 2121ffd830a3SStefano Zampini /* S_VV and S_CV */ 212206656605SStefano Zampini if (n_constraints) { 212306656605SStefano Zampini Mat B; 212480677318SStefano Zampini 2125ffd830a3SStefano Zampini ierr = PetscMemzero(work+lda_rhs*n_vertices,n_B*n_vertices*sizeof(PetscScalar));CHKERRQ(ierr); 212680677318SStefano Zampini for (i=0;i<n_vertices;i++) { 2127ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*lda_rhs);CHKERRQ(ierr); 2128ffd830a3SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,work+lda_rhs*n_vertices+i*n_B);CHKERRQ(ierr); 212980677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 213080677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 213180677318SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 213280677318SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 213380677318SStefano Zampini } 2134ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,n_vertices,work+lda_rhs*n_vertices,&B);CHKERRQ(ierr); 213580677318SStefano Zampini ierr = MatMatMult(pcbddc->local_auxmat1,B,MAT_REUSE_MATRIX,PETSC_DEFAULT,&S_CV);CHKERRQ(ierr); 213680677318SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 2137ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_vertices,work+lda_rhs*n_vertices,&B);CHKERRQ(ierr); 213880677318SStefano Zampini ierr = MatMatMult(local_auxmat2_R,S_CV,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B);CHKERRQ(ierr); 213906656605SStefano Zampini ierr = MatScale(S_CV,m_one);CHKERRQ(ierr); 2140ffd830a3SStefano Zampini ierr = PetscBLASIntCast(lda_rhs*n_vertices,&B_N);CHKERRQ(ierr); 2141ffd830a3SStefano Zampini PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&B_N,&one,work+lda_rhs*n_vertices,&B_one,work,&B_one)); 214206656605SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 214306656605SStefano Zampini } 214404708bb6SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)A_VR,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 214504708bb6SStefano Zampini if (!isseqaij) { /* MatMatMult with SEQ(S)BAIJ below will raise an error */ 2146511c6705SHong Zhang ierr = MatConvert(A_VR,MATSEQAIJ,MAT_INPLACE_MATRIX,&A_VR);CHKERRQ(ierr); 214704708bb6SStefano Zampini } 2148ffd830a3SStefano Zampini if (lda_rhs != n_R) { 2149ffd830a3SStefano Zampini ierr = MatDestroy(&A_RRmA_RV);CHKERRQ(ierr); 2150ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_vertices,work,&A_RRmA_RV);CHKERRQ(ierr); 2151ffd830a3SStefano Zampini ierr = MatSeqDenseSetLDA(A_RRmA_RV,lda_rhs);CHKERRQ(ierr); 2152ffd830a3SStefano Zampini } 215306656605SStefano Zampini ierr = MatMatMult(A_VR,A_RRmA_RV,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&S_VVt);CHKERRQ(ierr); 215414393ed6SStefano Zampini /* need A_VR * \Phi * A_RRmA_RV = A_VR * (I+L)^T * A_RRmA_RV, L given as before */ 215514393ed6SStefano Zampini if (need_benign_correction) { 2156df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 215714393ed6SStefano Zampini PetscScalar *marr,*sums; 215814393ed6SStefano Zampini 215914393ed6SStefano Zampini ierr = PetscMalloc1(n_vertices,&sums);CHKERRQ(ierr); 216014393ed6SStefano Zampini ierr = MatDenseGetArray(S_VVt,&marr); 2161df4d28bfSStefano Zampini for (i=0;i<reuse_solver->benign_n;i++) { 216214393ed6SStefano Zampini const PetscScalar *vals; 216314393ed6SStefano Zampini const PetscInt *idxs,*idxs_zero; 216414393ed6SStefano Zampini PetscInt n,j,nz; 216514393ed6SStefano Zampini 2166df4d28bfSStefano Zampini ierr = ISGetLocalSize(reuse_solver->benign_zerodiag_subs[i],&nz);CHKERRQ(ierr); 2167df4d28bfSStefano Zampini ierr = ISGetIndices(reuse_solver->benign_zerodiag_subs[i],&idxs_zero);CHKERRQ(ierr); 216814393ed6SStefano Zampini for (j=0;j<n_vertices;j++) { 216914393ed6SStefano Zampini PetscInt k; 217014393ed6SStefano Zampini sums[j] = 0.; 217114393ed6SStefano Zampini for (k=0;k<nz;k++) sums[j] += work[idxs_zero[k]+j*lda_rhs]; 217214393ed6SStefano Zampini } 217314393ed6SStefano Zampini ierr = MatGetRow(A_RV_bcorr,i,&n,&idxs,&vals);CHKERRQ(ierr); 217414393ed6SStefano Zampini for (j=0;j<n;j++) { 217514393ed6SStefano Zampini PetscScalar val = vals[j]; 217614393ed6SStefano Zampini PetscInt k; 217714393ed6SStefano Zampini for (k=0;k<n_vertices;k++) { 217814393ed6SStefano Zampini marr[idxs[j]+k*n_vertices] += val*sums[k]; 217914393ed6SStefano Zampini } 218014393ed6SStefano Zampini } 218114393ed6SStefano Zampini ierr = MatRestoreRow(A_RV_bcorr,i,&n,&idxs,&vals);CHKERRQ(ierr); 2182df4d28bfSStefano Zampini ierr = ISRestoreIndices(reuse_solver->benign_zerodiag_subs[i],&idxs_zero);CHKERRQ(ierr); 218314393ed6SStefano Zampini } 218414393ed6SStefano Zampini ierr = PetscFree(sums);CHKERRQ(ierr); 218514393ed6SStefano Zampini ierr = MatDenseRestoreArray(S_VVt,&marr); 218614393ed6SStefano Zampini ierr = MatDestroy(&A_RV_bcorr);CHKERRQ(ierr); 218714393ed6SStefano Zampini } 218880677318SStefano Zampini ierr = MatDestroy(&A_RRmA_RV);CHKERRQ(ierr); 218906656605SStefano Zampini ierr = PetscBLASIntCast(n_vertices*n_vertices,&B_N);CHKERRQ(ierr); 219006656605SStefano Zampini ierr = MatDenseGetArray(A_VV,&x);CHKERRQ(ierr); 219106656605SStefano Zampini ierr = MatDenseGetArray(S_VVt,&y);CHKERRQ(ierr); 219206656605SStefano Zampini PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&B_N,&one,x,&B_one,y,&B_one)); 219306656605SStefano Zampini ierr = MatDenseRestoreArray(A_VV,&x);CHKERRQ(ierr); 219406656605SStefano Zampini ierr = MatDenseRestoreArray(S_VVt,&y);CHKERRQ(ierr); 219506656605SStefano Zampini ierr = MatCopy(S_VVt,S_VV,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 2196d16cbb6bSStefano Zampini ierr = MatDestroy(&S_VVt);CHKERRQ(ierr); 2197019a44ceSStefano Zampini } else { 2198d16cbb6bSStefano Zampini ierr = MatCopy(A_VV,S_VV,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 2199d16cbb6bSStefano Zampini } 220021eccb56SStefano Zampini ierr = MatDestroy(&A_VV);CHKERRQ(ierr); 2201d16cbb6bSStefano Zampini 220206656605SStefano Zampini /* coarse basis functions */ 220306656605SStefano Zampini for (i=0;i<n_vertices;i++) { 220416f15bc4SStefano Zampini PetscScalar *y; 220516f15bc4SStefano Zampini 2206ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+lda_rhs*i);CHKERRQ(ierr); 220706656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 220806656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*i);CHKERRQ(ierr); 220906656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 221006656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 221106656605SStefano Zampini y[n_B*i+idx_V_B[i]] = 1.0; 221206656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 221306656605SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 221406656605SStefano Zampini 221506656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 22164f1b2e48SStefano Zampini PetscInt j; 22174f1b2e48SStefano Zampini 221806656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 221906656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*i);CHKERRQ(ierr); 222006656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 222106656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 222206656605SStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 22234f1b2e48SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) y[n_D*i+p0_lidx_I[j]] = 0.0; 222406656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 222506656605SStefano Zampini } 222606656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 222706656605SStefano Zampini } 222804708bb6SStefano Zampini /* if n_R == 0 the object is not destroyed */ 222904708bb6SStefano Zampini ierr = MatDestroy(&A_RV);CHKERRQ(ierr); 223006656605SStefano Zampini } 22315cbda25cSStefano Zampini ierr = VecDestroy(&dummy_vec);CHKERRQ(ierr); 223206656605SStefano Zampini 223306656605SStefano Zampini if (n_constraints) { 223406656605SStefano Zampini Mat B; 223506656605SStefano Zampini 2236ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_constraints,work,&B);CHKERRQ(ierr); 223706656605SStefano Zampini ierr = MatScale(S_CC,m_one);CHKERRQ(ierr); 223880677318SStefano Zampini ierr = MatMatMult(local_auxmat2_R,S_CC,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B);CHKERRQ(ierr); 223906656605SStefano Zampini ierr = MatScale(S_CC,m_one);CHKERRQ(ierr); 224006656605SStefano Zampini if (n_vertices) { 224180677318SStefano Zampini if (isCHOL) { /* if we can solve the interior problem with cholesky, we should also be fine with transposing here */ 224280677318SStefano Zampini ierr = MatTranspose(S_CV,MAT_REUSE_MATRIX,&S_VC);CHKERRQ(ierr); 224380677318SStefano Zampini } else { 224480677318SStefano Zampini Mat S_VCt; 224580677318SStefano Zampini 2246ffd830a3SStefano Zampini if (lda_rhs != n_R) { 2247ffd830a3SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 224872b8c272SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_constraints,work,&B);CHKERRQ(ierr); 2249ffd830a3SStefano Zampini ierr = MatSeqDenseSetLDA(B,lda_rhs);CHKERRQ(ierr); 2250ffd830a3SStefano Zampini } 225180677318SStefano Zampini ierr = MatMatMult(A_VR,B,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&S_VCt);CHKERRQ(ierr); 225280677318SStefano Zampini ierr = MatCopy(S_VCt,S_VC,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 225380677318SStefano Zampini ierr = MatDestroy(&S_VCt);CHKERRQ(ierr); 225480677318SStefano Zampini } 225506656605SStefano Zampini } 225606656605SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 225706656605SStefano Zampini /* coarse basis functions */ 225806656605SStefano Zampini for (i=0;i<n_constraints;i++) { 225906656605SStefano Zampini PetscScalar *y; 226006656605SStefano Zampini 2261ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+lda_rhs*i);CHKERRQ(ierr); 226206656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 226306656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*(i+n_vertices));CHKERRQ(ierr); 226406656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 226506656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 226606656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 226706656605SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 226806656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 22694f1b2e48SStefano Zampini PetscInt j; 22704f1b2e48SStefano Zampini 227106656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 227206656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*(i+n_vertices));CHKERRQ(ierr); 227306656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 227406656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 227506656605SStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 22764f1b2e48SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) y[n_D*i+p0_lidx_I[j]] = 0.0; 227706656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 227806656605SStefano Zampini } 227906656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 228006656605SStefano Zampini } 228106656605SStefano Zampini } 228280677318SStefano Zampini if (n_constraints) { 228380677318SStefano Zampini ierr = MatDestroy(&local_auxmat2_R);CHKERRQ(ierr); 228480677318SStefano Zampini } 22854f1b2e48SStefano Zampini ierr = PetscFree(p0_lidx_I);CHKERRQ(ierr); 228672b8c272SStefano Zampini 228772b8c272SStefano Zampini /* coarse matrix entries relative to B_0 */ 228872b8c272SStefano Zampini if (pcbddc->benign_n) { 228972b8c272SStefano Zampini Mat B0_B,B0_BPHI; 229072b8c272SStefano Zampini IS is_dummy; 229172b8c272SStefano Zampini PetscScalar *data; 229272b8c272SStefano Zampini PetscInt j; 229372b8c272SStefano Zampini 229472b8c272SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcbddc->benign_n,0,1,&is_dummy);CHKERRQ(ierr); 229572b8c272SStefano Zampini ierr = MatGetSubMatrix(pcbddc->benign_B0,is_dummy,pcis->is_B_local,MAT_INITIAL_MATRIX,&B0_B);CHKERRQ(ierr); 229672b8c272SStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 229772b8c272SStefano Zampini ierr = MatMatMult(B0_B,pcbddc->coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&B0_BPHI);CHKERRQ(ierr); 229886c38910SStefano Zampini ierr = MatConvert(B0_BPHI,MATSEQDENSE,MAT_INPLACE_MATRIX,&B0_BPHI);CHKERRQ(ierr); 229972b8c272SStefano Zampini ierr = MatDenseGetArray(B0_BPHI,&data);CHKERRQ(ierr); 230072b8c272SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) { 230172b8c272SStefano Zampini PetscInt primal_idx = pcbddc->local_primal_size - pcbddc->benign_n + j; 230272b8c272SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 230372b8c272SStefano Zampini coarse_submat_vals[primal_idx*pcbddc->local_primal_size+i] = data[i*pcbddc->benign_n+j]; 230472b8c272SStefano Zampini coarse_submat_vals[i*pcbddc->local_primal_size+primal_idx] = data[i*pcbddc->benign_n+j]; 230572b8c272SStefano Zampini } 230672b8c272SStefano Zampini } 230772b8c272SStefano Zampini ierr = MatDenseRestoreArray(B0_BPHI,&data);CHKERRQ(ierr); 230872b8c272SStefano Zampini ierr = MatDestroy(&B0_B);CHKERRQ(ierr); 230972b8c272SStefano Zampini ierr = MatDestroy(&B0_BPHI);CHKERRQ(ierr); 231072b8c272SStefano Zampini } 2311019a44ceSStefano Zampini 231206656605SStefano Zampini /* compute other basis functions for non-symmetric problems */ 23133301b35fSStefano Zampini if (!pcbddc->symmetric_primal) { 2314ffd830a3SStefano Zampini Mat B_V=NULL,B_C=NULL; 2315ffd830a3SStefano Zampini PetscScalar *marray; 231606656605SStefano Zampini 231706656605SStefano Zampini if (n_constraints) { 2318ffd830a3SStefano Zampini Mat S_CCT,C_CRT; 231906656605SStefano Zampini 2320af25d912SStefano Zampini ierr = MatTranspose(C_CR,MAT_INPLACE_MATRIX,&C_CRT);CHKERRQ(ierr); 232106656605SStefano Zampini ierr = MatTranspose(S_CC,MAT_INITIAL_MATRIX,&S_CCT);CHKERRQ(ierr); 2322ffd830a3SStefano Zampini ierr = MatMatMult(C_CRT,S_CCT,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&B_C);CHKERRQ(ierr); 232316f15bc4SStefano Zampini ierr = MatDestroy(&S_CCT);CHKERRQ(ierr); 232406656605SStefano Zampini if (n_vertices) { 2325ffd830a3SStefano Zampini Mat S_VCT; 232606656605SStefano Zampini 232706656605SStefano Zampini ierr = MatTranspose(S_VC,MAT_INITIAL_MATRIX,&S_VCT);CHKERRQ(ierr); 2328ffd830a3SStefano Zampini ierr = MatMatMult(C_CRT,S_VCT,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&B_V);CHKERRQ(ierr); 232916f15bc4SStefano Zampini ierr = MatDestroy(&S_VCT);CHKERRQ(ierr); 233006656605SStefano Zampini } 2331ffd830a3SStefano Zampini ierr = MatDestroy(&C_CRT);CHKERRQ(ierr); 233206656605SStefano Zampini } 233316f15bc4SStefano Zampini if (n_vertices && n_R) { 2334ffd830a3SStefano Zampini PetscScalar *av,*marray; 2335ffd830a3SStefano Zampini const PetscInt *xadj,*adjncy; 2336ffd830a3SStefano Zampini PetscInt n; 2337ffd830a3SStefano Zampini PetscBool flg_row; 233806656605SStefano Zampini 2339ffd830a3SStefano Zampini /* B_V = B_V - A_VR^T */ 2340af25d912SStefano Zampini ierr = MatConvert(A_VR,MATSEQAIJ,MAT_INPLACE_MATRIX,&A_VR);CHKERRQ(ierr); 2341ffd830a3SStefano Zampini ierr = MatGetRowIJ(A_VR,0,PETSC_FALSE,PETSC_FALSE,&n,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 2342ffd830a3SStefano Zampini ierr = MatSeqAIJGetArray(A_VR,&av);CHKERRQ(ierr); 2343ffd830a3SStefano Zampini ierr = MatDenseGetArray(B_V,&marray);CHKERRQ(ierr); 2344ffd830a3SStefano Zampini for (i=0;i<n;i++) { 2345ffd830a3SStefano Zampini PetscInt j; 2346ffd830a3SStefano Zampini for (j=xadj[i];j<xadj[i+1];j++) marray[i*n_R + adjncy[j]] -= av[j]; 2347ffd830a3SStefano Zampini } 2348ffd830a3SStefano Zampini ierr = MatDenseRestoreArray(B_V,&marray);CHKERRQ(ierr); 2349ffd830a3SStefano Zampini ierr = MatRestoreRowIJ(A_VR,0,PETSC_FALSE,PETSC_FALSE,&n,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 2350ffd830a3SStefano Zampini ierr = MatDestroy(&A_VR);CHKERRQ(ierr); 235106656605SStefano Zampini } 235206656605SStefano Zampini 2353ffd830a3SStefano Zampini /* currently there's no support for MatTransposeMatSolve(F,B,X) */ 2354ffd830a3SStefano Zampini ierr = MatDenseGetArray(B_V,&marray);CHKERRQ(ierr); 2355ffd830a3SStefano Zampini for (i=0;i<n_vertices;i++) { 2356ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,marray+i*n_R);CHKERRQ(ierr); 2357ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,work+i*n_R);CHKERRQ(ierr); 235806656605SStefano Zampini ierr = KSPSolveTranspose(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 235906656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 236006656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 236106656605SStefano Zampini } 2362ffd830a3SStefano Zampini ierr = MatDenseRestoreArray(B_V,&marray);CHKERRQ(ierr); 2363ffd830a3SStefano Zampini ierr = MatDenseGetArray(B_C,&marray);CHKERRQ(ierr); 2364ffd830a3SStefano Zampini for (i=n_vertices;i<n_constraints+n_vertices;i++) { 2365ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,marray+(i-n_vertices)*n_R);CHKERRQ(ierr); 2366ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,work+i*n_R);CHKERRQ(ierr); 2367ffd830a3SStefano Zampini ierr = KSPSolveTranspose(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 2368ffd830a3SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 2369ffd830a3SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 237006656605SStefano Zampini } 2371ffd830a3SStefano Zampini ierr = MatDenseRestoreArray(B_C,&marray);CHKERRQ(ierr); 237206656605SStefano Zampini /* coarse basis functions */ 237306656605SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 237406656605SStefano Zampini PetscScalar *y; 237506656605SStefano Zampini 2376ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*n_R);CHKERRQ(ierr); 237706656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_psi_B,&y);CHKERRQ(ierr); 237806656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*i);CHKERRQ(ierr); 237906656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 238006656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 238106656605SStefano Zampini if (i<n_vertices) { 238206656605SStefano Zampini y[n_B*i+idx_V_B[i]] = 1.0; 238306656605SStefano Zampini } 238406656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_psi_B,&y);CHKERRQ(ierr); 238506656605SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 238606656605SStefano Zampini 238706656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 238806656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_psi_D,&y);CHKERRQ(ierr); 238906656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*i);CHKERRQ(ierr); 239006656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 239106656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 239206656605SStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 239306656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_psi_D,&y);CHKERRQ(ierr); 239406656605SStefano Zampini } 239506656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 239606656605SStefano Zampini } 2397ffd830a3SStefano Zampini ierr = MatDestroy(&B_V);CHKERRQ(ierr); 2398ffd830a3SStefano Zampini ierr = MatDestroy(&B_C);CHKERRQ(ierr); 239906656605SStefano Zampini } 2400d62866d3SStefano Zampini /* free memory */ 240188ebb749SStefano Zampini ierr = PetscFree(idx_V_B);CHKERRQ(ierr); 240206656605SStefano Zampini ierr = MatDestroy(&S_VV);CHKERRQ(ierr); 240306656605SStefano Zampini ierr = MatDestroy(&S_CV);CHKERRQ(ierr); 240406656605SStefano Zampini ierr = MatDestroy(&S_VC);CHKERRQ(ierr); 240506656605SStefano Zampini ierr = MatDestroy(&S_CC);CHKERRQ(ierr); 2406d62866d3SStefano Zampini ierr = PetscFree(work);CHKERRQ(ierr); 2407d62866d3SStefano Zampini if (n_vertices) { 2408d62866d3SStefano Zampini ierr = MatDestroy(&A_VR);CHKERRQ(ierr); 2409d62866d3SStefano Zampini } 2410d62866d3SStefano Zampini if (n_constraints) { 2411d62866d3SStefano Zampini ierr = MatDestroy(&C_CR);CHKERRQ(ierr); 2412d62866d3SStefano Zampini } 241388ebb749SStefano Zampini /* Checking coarse_sub_mat and coarse basis functios */ 241488ebb749SStefano Zampini /* Symmetric case : It should be \Phi^{(j)^T} A^{(j)} \Phi^{(j)}=coarse_sub_mat */ 241588ebb749SStefano Zampini /* Non-symmetric case : It should be \Psi^{(j)^T} A^{(j)} \Phi^{(j)}=coarse_sub_mat */ 2416d12edf2fSStefano Zampini if (pcbddc->dbg_flag) { 241788ebb749SStefano Zampini Mat coarse_sub_mat; 241825084f0cSStefano Zampini Mat AUXMAT,TM1,TM2,TM3,TM4; 241988ebb749SStefano Zampini Mat coarse_phi_D,coarse_phi_B; 242088ebb749SStefano Zampini Mat coarse_psi_D,coarse_psi_B; 242188ebb749SStefano Zampini Mat A_II,A_BB,A_IB,A_BI; 24228bec7fa6SStefano Zampini Mat C_B,CPHI; 24238bec7fa6SStefano Zampini IS is_dummy; 24248bec7fa6SStefano Zampini Vec mones; 242588ebb749SStefano Zampini MatType checkmattype=MATSEQAIJ; 242688ebb749SStefano Zampini PetscReal real_value; 242788ebb749SStefano Zampini 2428a3df083aSStefano Zampini if (pcbddc->benign_n && !pcbddc->benign_change_explicit) { 2429a3df083aSStefano Zampini Mat A; 2430a3df083aSStefano Zampini ierr = PCBDDCBenignProject(pc,NULL,NULL,&A);CHKERRQ(ierr); 2431a3df083aSStefano Zampini ierr = MatGetSubMatrix(A,pcis->is_I_local,pcis->is_I_local,MAT_INITIAL_MATRIX,&A_II);CHKERRQ(ierr); 2432a3df083aSStefano Zampini ierr = MatGetSubMatrix(A,pcis->is_I_local,pcis->is_B_local,MAT_INITIAL_MATRIX,&A_IB);CHKERRQ(ierr); 2433a3df083aSStefano Zampini ierr = MatGetSubMatrix(A,pcis->is_B_local,pcis->is_I_local,MAT_INITIAL_MATRIX,&A_BI);CHKERRQ(ierr); 2434a3df083aSStefano Zampini ierr = MatGetSubMatrix(A,pcis->is_B_local,pcis->is_B_local,MAT_INITIAL_MATRIX,&A_BB);CHKERRQ(ierr); 2435a3df083aSStefano Zampini ierr = MatDestroy(&A);CHKERRQ(ierr); 2436a3df083aSStefano Zampini } else { 243788ebb749SStefano Zampini ierr = MatConvert(pcis->A_II,checkmattype,MAT_INITIAL_MATRIX,&A_II);CHKERRQ(ierr); 243888ebb749SStefano Zampini ierr = MatConvert(pcis->A_IB,checkmattype,MAT_INITIAL_MATRIX,&A_IB);CHKERRQ(ierr); 243988ebb749SStefano Zampini ierr = MatConvert(pcis->A_BI,checkmattype,MAT_INITIAL_MATRIX,&A_BI);CHKERRQ(ierr); 244088ebb749SStefano Zampini ierr = MatConvert(pcis->A_BB,checkmattype,MAT_INITIAL_MATRIX,&A_BB);CHKERRQ(ierr); 2441a3df083aSStefano Zampini } 244288ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_phi_D,checkmattype,MAT_INITIAL_MATRIX,&coarse_phi_D);CHKERRQ(ierr); 244388ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_phi_B,checkmattype,MAT_INITIAL_MATRIX,&coarse_phi_B);CHKERRQ(ierr); 2444ffd830a3SStefano Zampini if (!pcbddc->symmetric_primal) { 244588ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_psi_D,checkmattype,MAT_INITIAL_MATRIX,&coarse_psi_D);CHKERRQ(ierr); 244688ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_psi_B,checkmattype,MAT_INITIAL_MATRIX,&coarse_psi_B);CHKERRQ(ierr); 244788ebb749SStefano Zampini } 244888ebb749SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,pcbddc->local_primal_size,pcbddc->local_primal_size,coarse_submat_vals,&coarse_sub_mat);CHKERRQ(ierr); 244988ebb749SStefano Zampini 245025084f0cSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 24513301b35fSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Check coarse sub mat computation (symmetric %d)\n",pcbddc->symmetric_primal);CHKERRQ(ierr); 245225084f0cSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 2453ffd830a3SStefano Zampini if (!pcbddc->symmetric_primal) { 245488ebb749SStefano Zampini ierr = MatMatMult(A_II,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 245588ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM1);CHKERRQ(ierr); 245688ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 245788ebb749SStefano Zampini ierr = MatMatMult(A_BB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 245888ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM2);CHKERRQ(ierr); 245988ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 246088ebb749SStefano Zampini ierr = MatMatMult(A_IB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 246188ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM3);CHKERRQ(ierr); 246288ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 246388ebb749SStefano Zampini ierr = MatMatMult(A_BI,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 246488ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM4);CHKERRQ(ierr); 246588ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 246688ebb749SStefano Zampini } else { 246788ebb749SStefano Zampini ierr = MatPtAP(A_II,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&TM1);CHKERRQ(ierr); 246888ebb749SStefano Zampini ierr = MatPtAP(A_BB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&TM2);CHKERRQ(ierr); 246988ebb749SStefano Zampini ierr = MatMatMult(A_IB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 247088ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_phi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM3);CHKERRQ(ierr); 247188ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 247288ebb749SStefano Zampini ierr = MatMatMult(A_BI,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 247388ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_phi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM4);CHKERRQ(ierr); 247488ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 247588ebb749SStefano Zampini } 247688ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM2,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 247788ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM3,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 247888ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM4,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 2479511c6705SHong Zhang ierr = MatConvert(TM1,MATSEQDENSE,MAT_INPLACE_MATRIX,&TM1);CHKERRQ(ierr); 24804f1b2e48SStefano Zampini if (pcbddc->benign_n) { 2481fc227af8SStefano Zampini Mat B0_B,B0_BPHI; 2482d12edf2fSStefano Zampini PetscScalar *data,*data2; 24834f1b2e48SStefano Zampini PetscInt j; 2484d12edf2fSStefano Zampini 24854f1b2e48SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcbddc->benign_n,0,1,&is_dummy);CHKERRQ(ierr); 2486fc227af8SStefano Zampini ierr = MatGetSubMatrix(pcbddc->benign_B0,is_dummy,pcis->is_B_local,MAT_INITIAL_MATRIX,&B0_B);CHKERRQ(ierr); 2487d12edf2fSStefano Zampini ierr = MatMatMult(B0_B,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&B0_BPHI);CHKERRQ(ierr); 248886c38910SStefano Zampini ierr = MatConvert(B0_BPHI,MATSEQDENSE,MAT_INPLACE_MATRIX,&B0_BPHI);CHKERRQ(ierr); 2489d12edf2fSStefano Zampini ierr = MatDenseGetArray(TM1,&data);CHKERRQ(ierr); 2490d12edf2fSStefano Zampini ierr = MatDenseGetArray(B0_BPHI,&data2);CHKERRQ(ierr); 24914f1b2e48SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) { 24924f1b2e48SStefano Zampini PetscInt primal_idx = pcbddc->local_primal_size - pcbddc->benign_n + j; 2493d12edf2fSStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 24944f1b2e48SStefano Zampini data[primal_idx*pcbddc->local_primal_size+i] += data2[i*pcbddc->benign_n+j]; 24954f1b2e48SStefano Zampini data[i*pcbddc->local_primal_size+primal_idx] += data2[i*pcbddc->benign_n+j]; 24964f1b2e48SStefano Zampini } 2497d12edf2fSStefano Zampini } 2498d12edf2fSStefano Zampini ierr = MatDenseRestoreArray(TM1,&data);CHKERRQ(ierr); 2499d12edf2fSStefano Zampini ierr = MatDenseRestoreArray(B0_BPHI,&data2);CHKERRQ(ierr); 2500d12edf2fSStefano Zampini ierr = MatDestroy(&B0_B);CHKERRQ(ierr); 2501d12edf2fSStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 2502d12edf2fSStefano Zampini ierr = MatDestroy(&B0_BPHI);CHKERRQ(ierr); 2503d12edf2fSStefano Zampini } 2504d12edf2fSStefano Zampini #if 0 2505d12edf2fSStefano Zampini { 2506d12edf2fSStefano Zampini PetscViewer viewer; 2507d12edf2fSStefano Zampini char filename[256]; 2508ffd830a3SStefano Zampini sprintf(filename,"details_local_coarse_mat%d_level%d.m",PetscGlobalRank,pcbddc->current_level); 2509d12edf2fSStefano Zampini ierr = PetscViewerASCIIOpen(PETSC_COMM_SELF,filename,&viewer);CHKERRQ(ierr); 2510d12edf2fSStefano Zampini ierr = PetscViewerSetFormat(viewer,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 2511ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)coarse_sub_mat,"computed");CHKERRQ(ierr); 2512ffd830a3SStefano Zampini ierr = MatView(coarse_sub_mat,viewer);CHKERRQ(ierr); 2513ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)TM1,"projected");CHKERRQ(ierr); 2514d12edf2fSStefano Zampini ierr = MatView(TM1,viewer);CHKERRQ(ierr); 251572b8c272SStefano Zampini if (save_change) { 251672b8c272SStefano Zampini Mat phi_B; 251772b8c272SStefano Zampini ierr = MatMatMult(save_change,pcbddc->coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&phi_B);CHKERRQ(ierr); 251872b8c272SStefano Zampini ierr = PetscObjectSetName((PetscObject)phi_B,"phi_B");CHKERRQ(ierr); 251972b8c272SStefano Zampini ierr = MatView(phi_B,viewer);CHKERRQ(ierr); 252072b8c272SStefano Zampini ierr = MatDestroy(&phi_B);CHKERRQ(ierr); 252172b8c272SStefano Zampini } else { 2522ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)pcbddc->coarse_phi_B,"phi_B");CHKERRQ(ierr); 2523ffd830a3SStefano Zampini ierr = MatView(pcbddc->coarse_phi_B,viewer);CHKERRQ(ierr); 252472b8c272SStefano Zampini } 2525ffd830a3SStefano Zampini if (pcbddc->coarse_phi_D) { 2526ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)pcbddc->coarse_phi_D,"phi_D");CHKERRQ(ierr); 2527ffd830a3SStefano Zampini ierr = MatView(pcbddc->coarse_phi_D,viewer);CHKERRQ(ierr); 2528ffd830a3SStefano Zampini } 2529ffd830a3SStefano Zampini if (pcbddc->coarse_psi_B) { 2530ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)pcbddc->coarse_psi_B,"psi_B");CHKERRQ(ierr); 2531ffd830a3SStefano Zampini ierr = MatView(pcbddc->coarse_psi_B,viewer);CHKERRQ(ierr); 2532ffd830a3SStefano Zampini } 253372b8c272SStefano Zampini if (pcbddc->coarse_psi_D) { 2534ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)pcbddc->coarse_psi_D,"psi_D");CHKERRQ(ierr); 2535ffd830a3SStefano Zampini ierr = MatView(pcbddc->coarse_psi_D,viewer);CHKERRQ(ierr); 2536ffd830a3SStefano Zampini } 2537d12edf2fSStefano Zampini ierr = PetscViewerDestroy(&viewer);CHKERRQ(ierr); 2538d12edf2fSStefano Zampini } 2539d12edf2fSStefano Zampini #endif 254081d9aea3SBarry Smith ierr = MatAXPY(TM1,m_one,coarse_sub_mat,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 25418bec7fa6SStefano Zampini ierr = MatNorm(TM1,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 25421575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 254306656605SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d matrix error % 1.14e\n",PetscGlobalRank,real_value);CHKERRQ(ierr); 25448bec7fa6SStefano Zampini 25458bec7fa6SStefano Zampini /* check constraints */ 2546a00504b5SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcbddc->local_primal_size-pcbddc->benign_n,0,1,&is_dummy);CHKERRQ(ierr); 2547a00504b5SStefano Zampini ierr = MatGetSubMatrix(pcbddc->ConstraintMatrix,is_dummy,pcis->is_B_local,MAT_INITIAL_MATRIX,&C_B);CHKERRQ(ierr); 25484f1b2e48SStefano Zampini if (!pcbddc->benign_n) { /* TODO: add benign case */ 25498bec7fa6SStefano Zampini ierr = MatMatMult(C_B,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&CPHI);CHKERRQ(ierr); 2550a00504b5SStefano Zampini } else { 2551a00504b5SStefano Zampini PetscScalar *data; 2552a00504b5SStefano Zampini Mat tmat; 2553a00504b5SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&data);CHKERRQ(ierr); 2554a00504b5SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,pcis->n_B,pcbddc->local_primal_size-pcbddc->benign_n,data,&tmat);CHKERRQ(ierr); 2555a00504b5SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_B,&data);CHKERRQ(ierr); 2556a00504b5SStefano Zampini ierr = MatMatMult(C_B,tmat,MAT_INITIAL_MATRIX,1.0,&CPHI);CHKERRQ(ierr); 2557a00504b5SStefano Zampini ierr = MatDestroy(&tmat);CHKERRQ(ierr); 2558a00504b5SStefano Zampini } 25598bec7fa6SStefano Zampini ierr = MatCreateVecs(CPHI,&mones,NULL);CHKERRQ(ierr); 25608bec7fa6SStefano Zampini ierr = VecSet(mones,-1.0);CHKERRQ(ierr); 25618bec7fa6SStefano Zampini ierr = MatDiagonalSet(CPHI,mones,ADD_VALUES);CHKERRQ(ierr); 25628bec7fa6SStefano Zampini ierr = MatNorm(CPHI,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 2563bdae7319SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d phi constraints error % 1.14e\n",PetscGlobalRank,real_value);CHKERRQ(ierr); 2564ffd830a3SStefano Zampini if (!pcbddc->symmetric_primal) { 2565bdae7319SStefano Zampini ierr = MatMatMult(C_B,coarse_psi_B,MAT_REUSE_MATRIX,1.0,&CPHI);CHKERRQ(ierr); 2566bdae7319SStefano Zampini ierr = VecSet(mones,-1.0);CHKERRQ(ierr); 2567bdae7319SStefano Zampini ierr = MatDiagonalSet(CPHI,mones,ADD_VALUES);CHKERRQ(ierr); 2568bdae7319SStefano Zampini ierr = MatNorm(CPHI,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 2569bdae7319SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d psi constraints error % 1.14e\n",PetscGlobalRank,real_value);CHKERRQ(ierr); 257088ebb749SStefano Zampini } 25718bec7fa6SStefano Zampini ierr = MatDestroy(&C_B);CHKERRQ(ierr); 25728bec7fa6SStefano Zampini ierr = MatDestroy(&CPHI);CHKERRQ(ierr); 25738bec7fa6SStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 25748bec7fa6SStefano Zampini ierr = VecDestroy(&mones);CHKERRQ(ierr); 257525084f0cSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 257688ebb749SStefano Zampini ierr = MatDestroy(&A_II);CHKERRQ(ierr); 257788ebb749SStefano Zampini ierr = MatDestroy(&A_BB);CHKERRQ(ierr); 257888ebb749SStefano Zampini ierr = MatDestroy(&A_IB);CHKERRQ(ierr); 257988ebb749SStefano Zampini ierr = MatDestroy(&A_BI);CHKERRQ(ierr); 258088ebb749SStefano Zampini ierr = MatDestroy(&TM1);CHKERRQ(ierr); 258188ebb749SStefano Zampini ierr = MatDestroy(&TM2);CHKERRQ(ierr); 258288ebb749SStefano Zampini ierr = MatDestroy(&TM3);CHKERRQ(ierr); 258388ebb749SStefano Zampini ierr = MatDestroy(&TM4);CHKERRQ(ierr); 258488ebb749SStefano Zampini ierr = MatDestroy(&coarse_phi_D);CHKERRQ(ierr); 258588ebb749SStefano Zampini ierr = MatDestroy(&coarse_phi_B);CHKERRQ(ierr); 2586ffd830a3SStefano Zampini if (!pcbddc->symmetric_primal) { 258788ebb749SStefano Zampini ierr = MatDestroy(&coarse_psi_D);CHKERRQ(ierr); 258888ebb749SStefano Zampini ierr = MatDestroy(&coarse_psi_B);CHKERRQ(ierr); 258988ebb749SStefano Zampini } 259088ebb749SStefano Zampini ierr = MatDestroy(&coarse_sub_mat);CHKERRQ(ierr); 259188ebb749SStefano Zampini } 25928629588bSStefano Zampini /* get back data */ 25938629588bSStefano Zampini *coarse_submat_vals_n = coarse_submat_vals; 259488ebb749SStefano Zampini PetscFunctionReturn(0); 259588ebb749SStefano Zampini } 259688ebb749SStefano Zampini 259788ebb749SStefano Zampini #undef __FUNCT__ 2598d65f70fdSStefano Zampini #define __FUNCT__ "MatGetSubMatrixUnsorted" 2599d648f858SStefano Zampini PetscErrorCode MatGetSubMatrixUnsorted(Mat A, IS isrow, IS iscol, Mat* B) 2600aa0d41d4SStefano Zampini { 2601d65f70fdSStefano Zampini Mat *work_mat; 2602d65f70fdSStefano Zampini IS isrow_s,iscol_s; 2603d65f70fdSStefano Zampini PetscBool rsorted,csorted; 2604c43ebad9SStefano Zampini PetscInt rsize,*idxs_perm_r=NULL,csize,*idxs_perm_c=NULL; 2605aa0d41d4SStefano Zampini PetscErrorCode ierr; 2606aa0d41d4SStefano Zampini 2607aa0d41d4SStefano Zampini PetscFunctionBegin; 2608d65f70fdSStefano Zampini ierr = ISSorted(isrow,&rsorted);CHKERRQ(ierr); 2609d65f70fdSStefano Zampini ierr = ISSorted(iscol,&csorted);CHKERRQ(ierr); 2610d65f70fdSStefano Zampini ierr = ISGetLocalSize(isrow,&rsize);CHKERRQ(ierr); 2611d65f70fdSStefano Zampini ierr = ISGetLocalSize(iscol,&csize);CHKERRQ(ierr); 2612aa0d41d4SStefano Zampini 2613d65f70fdSStefano Zampini if (!rsorted) { 2614906d46d4SStefano Zampini const PetscInt *idxs; 2615906d46d4SStefano Zampini PetscInt *idxs_sorted,i; 2616aa0d41d4SStefano Zampini 2617d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_perm_r);CHKERRQ(ierr); 2618d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_sorted);CHKERRQ(ierr); 2619d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 2620d65f70fdSStefano Zampini idxs_perm_r[i] = i; 2621aa0d41d4SStefano Zampini } 2622d65f70fdSStefano Zampini ierr = ISGetIndices(isrow,&idxs);CHKERRQ(ierr); 2623d65f70fdSStefano Zampini ierr = PetscSortIntWithPermutation(rsize,idxs,idxs_perm_r);CHKERRQ(ierr); 2624d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 2625d65f70fdSStefano Zampini idxs_sorted[i] = idxs[idxs_perm_r[i]]; 2626aa0d41d4SStefano Zampini } 2627d65f70fdSStefano Zampini ierr = ISRestoreIndices(isrow,&idxs);CHKERRQ(ierr); 2628d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,rsize,idxs_sorted,PETSC_OWN_POINTER,&isrow_s);CHKERRQ(ierr); 2629d65f70fdSStefano Zampini } else { 2630d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)isrow);CHKERRQ(ierr); 2631d65f70fdSStefano Zampini isrow_s = isrow; 2632aa0d41d4SStefano Zampini } 2633906d46d4SStefano Zampini 2634d65f70fdSStefano Zampini if (!csorted) { 2635d65f70fdSStefano Zampini if (isrow == iscol) { 2636d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)isrow_s);CHKERRQ(ierr); 2637d65f70fdSStefano Zampini iscol_s = isrow_s; 2638d65f70fdSStefano Zampini } else { 2639d65f70fdSStefano Zampini const PetscInt *idxs; 2640d65f70fdSStefano Zampini PetscInt *idxs_sorted,i; 2641906d46d4SStefano Zampini 2642d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_perm_c);CHKERRQ(ierr); 2643d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_sorted);CHKERRQ(ierr); 2644d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 2645d65f70fdSStefano Zampini idxs_perm_c[i] = i; 2646d65f70fdSStefano Zampini } 2647d65f70fdSStefano Zampini ierr = ISGetIndices(iscol,&idxs);CHKERRQ(ierr); 2648d65f70fdSStefano Zampini ierr = PetscSortIntWithPermutation(csize,idxs,idxs_perm_c);CHKERRQ(ierr); 2649d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 2650d65f70fdSStefano Zampini idxs_sorted[i] = idxs[idxs_perm_c[i]]; 2651d65f70fdSStefano Zampini } 2652d65f70fdSStefano Zampini ierr = ISRestoreIndices(iscol,&idxs);CHKERRQ(ierr); 2653d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,csize,idxs_sorted,PETSC_OWN_POINTER,&iscol_s);CHKERRQ(ierr); 2654d65f70fdSStefano Zampini } 2655d65f70fdSStefano Zampini } else { 2656d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)iscol);CHKERRQ(ierr); 2657d65f70fdSStefano Zampini iscol_s = iscol; 2658d65f70fdSStefano Zampini } 2659d65f70fdSStefano Zampini 2660d648f858SStefano Zampini ierr = MatGetSubMatrices(A,1,&isrow_s,&iscol_s,MAT_INITIAL_MATRIX,&work_mat);CHKERRQ(ierr); 2661d65f70fdSStefano Zampini 2662d65f70fdSStefano Zampini if (!rsorted || !csorted) { 2663906d46d4SStefano Zampini Mat new_mat; 2664d65f70fdSStefano Zampini IS is_perm_r,is_perm_c; 2665906d46d4SStefano Zampini 2666d65f70fdSStefano Zampini if (!rsorted) { 2667d65f70fdSStefano Zampini PetscInt *idxs_r,i; 2668d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_r);CHKERRQ(ierr); 2669d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 2670d65f70fdSStefano Zampini idxs_r[idxs_perm_r[i]] = i; 2671906d46d4SStefano Zampini } 2672d65f70fdSStefano Zampini ierr = PetscFree(idxs_perm_r);CHKERRQ(ierr); 2673d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,rsize,idxs_r,PETSC_OWN_POINTER,&is_perm_r);CHKERRQ(ierr); 2674d65f70fdSStefano Zampini } else { 2675d65f70fdSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,rsize,0,1,&is_perm_r);CHKERRQ(ierr); 2676906d46d4SStefano Zampini } 2677d65f70fdSStefano Zampini ierr = ISSetPermutation(is_perm_r);CHKERRQ(ierr); 2678d65f70fdSStefano Zampini 2679d65f70fdSStefano Zampini if (!csorted) { 2680d65f70fdSStefano Zampini if (isrow_s == iscol_s) { 2681d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)is_perm_r);CHKERRQ(ierr); 2682d65f70fdSStefano Zampini is_perm_c = is_perm_r; 2683d65f70fdSStefano Zampini } else { 2684d65f70fdSStefano Zampini PetscInt *idxs_c,i; 2685d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_c);CHKERRQ(ierr); 2686d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 2687d65f70fdSStefano Zampini idxs_c[idxs_perm_c[i]] = i; 2688d65f70fdSStefano Zampini } 2689d65f70fdSStefano Zampini ierr = PetscFree(idxs_perm_c);CHKERRQ(ierr); 2690d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,csize,idxs_c,PETSC_OWN_POINTER,&is_perm_c);CHKERRQ(ierr); 2691d65f70fdSStefano Zampini } 2692d65f70fdSStefano Zampini } else { 2693d65f70fdSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,csize,0,1,&is_perm_c);CHKERRQ(ierr); 2694d65f70fdSStefano Zampini } 2695d65f70fdSStefano Zampini ierr = ISSetPermutation(is_perm_c);CHKERRQ(ierr); 2696d65f70fdSStefano Zampini 2697d65f70fdSStefano Zampini ierr = MatPermute(work_mat[0],is_perm_r,is_perm_c,&new_mat);CHKERRQ(ierr); 2698d65f70fdSStefano Zampini ierr = MatDestroy(&work_mat[0]);CHKERRQ(ierr); 2699d65f70fdSStefano Zampini work_mat[0] = new_mat; 2700d65f70fdSStefano Zampini ierr = ISDestroy(&is_perm_r);CHKERRQ(ierr); 2701d65f70fdSStefano Zampini ierr = ISDestroy(&is_perm_c);CHKERRQ(ierr); 2702d65f70fdSStefano Zampini } 2703d65f70fdSStefano Zampini 2704d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)work_mat[0]);CHKERRQ(ierr); 2705d65f70fdSStefano Zampini *B = work_mat[0]; 2706d65f70fdSStefano Zampini ierr = MatDestroyMatrices(1,&work_mat);CHKERRQ(ierr); 2707d65f70fdSStefano Zampini ierr = ISDestroy(&isrow_s);CHKERRQ(ierr); 2708d65f70fdSStefano Zampini ierr = ISDestroy(&iscol_s);CHKERRQ(ierr); 2709d65f70fdSStefano Zampini PetscFunctionReturn(0); 2710d65f70fdSStefano Zampini } 2711d65f70fdSStefano Zampini 2712d65f70fdSStefano Zampini #undef __FUNCT__ 27135e8657edSStefano Zampini #define __FUNCT__ "PCBDDCComputeLocalMatrix" 27145e8657edSStefano Zampini PetscErrorCode PCBDDCComputeLocalMatrix(PC pc, Mat ChangeOfBasisMatrix) 2715aa0d41d4SStefano Zampini { 2716aa0d41d4SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 27175e8657edSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 2718d65f70fdSStefano Zampini Mat new_mat; 27195e8657edSStefano Zampini IS is_local,is_global; 2720d65f70fdSStefano Zampini PetscInt local_size; 2721d65f70fdSStefano Zampini PetscBool isseqaij; 2722aa0d41d4SStefano Zampini PetscErrorCode ierr; 2723aa0d41d4SStefano Zampini 2724aa0d41d4SStefano Zampini PetscFunctionBegin; 2725aa0d41d4SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 27265e8657edSStefano Zampini ierr = MatGetSize(matis->A,&local_size,NULL);CHKERRQ(ierr); 27275e8657edSStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)matis->A),local_size,0,1,&is_local);CHKERRQ(ierr); 2728b087196eSStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(pc->pmat->rmap->mapping,is_local,&is_global);CHKERRQ(ierr); 2729aa0d41d4SStefano Zampini ierr = ISDestroy(&is_local);CHKERRQ(ierr); 2730d648f858SStefano Zampini ierr = MatGetSubMatrixUnsorted(ChangeOfBasisMatrix,is_global,is_global,&new_mat);CHKERRQ(ierr); 2731aa0d41d4SStefano Zampini ierr = ISDestroy(&is_global);CHKERRQ(ierr); 2732906d46d4SStefano Zampini 2733906d46d4SStefano Zampini /* check */ 2734906d46d4SStefano Zampini if (pcbddc->dbg_flag) { 2735906d46d4SStefano Zampini Vec x,x_change; 2736906d46d4SStefano Zampini PetscReal error; 2737906d46d4SStefano Zampini 27385e8657edSStefano Zampini ierr = MatCreateVecs(ChangeOfBasisMatrix,&x,&x_change);CHKERRQ(ierr); 2739906d46d4SStefano Zampini ierr = VecSetRandom(x,NULL);CHKERRQ(ierr); 27405e8657edSStefano Zampini ierr = MatMult(ChangeOfBasisMatrix,x,x_change);CHKERRQ(ierr); 2741e176bc59SStefano Zampini ierr = VecScatterBegin(matis->cctx,x,matis->x,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2742e176bc59SStefano Zampini ierr = VecScatterEnd(matis->cctx,x,matis->x,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2743d65f70fdSStefano Zampini ierr = MatMult(new_mat,matis->x,matis->y);CHKERRQ(ierr); 2744e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,matis->y,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2745e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,matis->y,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2746906d46d4SStefano Zampini ierr = VecAXPY(x,-1.0,x_change);CHKERRQ(ierr); 2747906d46d4SStefano Zampini ierr = VecNorm(x,NORM_INFINITY,&error);CHKERRQ(ierr); 2748906d46d4SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 2749906d46d4SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Error global vs local change on N: %1.6e\n",error);CHKERRQ(ierr); 2750906d46d4SStefano Zampini ierr = VecDestroy(&x);CHKERRQ(ierr); 2751906d46d4SStefano Zampini ierr = VecDestroy(&x_change);CHKERRQ(ierr); 2752906d46d4SStefano Zampini } 2753906d46d4SStefano Zampini 275422d5777bSStefano Zampini /* TODO: HOW TO WORK WITH BAIJ and SBAIJ and SEQDENSE? */ 27559b28b3ffSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)matis->A,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 275622d5777bSStefano Zampini if (isseqaij) { 2757a00504b5SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 2758a00504b5SStefano Zampini ierr = MatPtAP(matis->A,new_mat,MAT_INITIAL_MATRIX,2.0,&pcbddc->local_mat);CHKERRQ(ierr); 2759aa0d41d4SStefano Zampini } else { 2760a00504b5SStefano Zampini Mat work_mat; 27611cf9b237SStefano Zampini 2762a00504b5SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 2763aa0d41d4SStefano Zampini ierr = MatConvert(matis->A,MATSEQAIJ,MAT_INITIAL_MATRIX,&work_mat);CHKERRQ(ierr); 2764a00504b5SStefano Zampini ierr = MatPtAP(work_mat,new_mat,MAT_INITIAL_MATRIX,2.0,&pcbddc->local_mat);CHKERRQ(ierr); 2765aa0d41d4SStefano Zampini } 27663301b35fSStefano Zampini if (matis->A->symmetric_set) { 27673301b35fSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_SYMMETRIC,matis->A->symmetric);CHKERRQ(ierr); 2768e496cd5dSStefano Zampini #if !defined(PETSC_USE_COMPLEX) 27693301b35fSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_HERMITIAN,matis->A->symmetric);CHKERRQ(ierr); 2770e496cd5dSStefano Zampini #endif 27713301b35fSStefano Zampini } 2772d65f70fdSStefano Zampini ierr = MatDestroy(&new_mat);CHKERRQ(ierr); 2773aa0d41d4SStefano Zampini PetscFunctionReturn(0); 2774aa0d41d4SStefano Zampini } 2775aa0d41d4SStefano Zampini 2776aa0d41d4SStefano Zampini #undef __FUNCT__ 2777a64d13efSStefano Zampini #define __FUNCT__ "PCBDDCSetUpLocalScatters" 27788ce42a96SStefano Zampini PetscErrorCode PCBDDCSetUpLocalScatters(PC pc) 2779a64d13efSStefano Zampini { 2780a64d13efSStefano Zampini PC_IS* pcis = (PC_IS*)(pc->data); 2781a64d13efSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 2782d62866d3SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 278353892102SStefano Zampini PetscInt *idx_R_local=NULL; 27843a50541eSStefano Zampini PetscInt n_vertices,i,j,n_R,n_D,n_B; 27853a50541eSStefano Zampini PetscInt vbs,bs; 27866816873aSStefano Zampini PetscBT bitmask=NULL; 2787a64d13efSStefano Zampini PetscErrorCode ierr; 2788a64d13efSStefano Zampini 2789a64d13efSStefano Zampini PetscFunctionBegin; 2790b23d619eSStefano Zampini /* 2791b23d619eSStefano Zampini No need to setup local scatters if 2792b23d619eSStefano Zampini - primal space is unchanged 2793b23d619eSStefano Zampini AND 2794b23d619eSStefano Zampini - we actually have locally some primal dofs (could not be true in multilevel or for isolated subdomains) 2795b23d619eSStefano Zampini AND 2796b23d619eSStefano Zampini - we are not in debugging mode (this is needed since there are Synchronized prints at the end of the subroutine 2797b23d619eSStefano Zampini */ 2798b23d619eSStefano Zampini if (!pcbddc->new_primal_space_local && pcbddc->local_primal_size && !pcbddc->dbg_flag) { 2799f4ddd8eeSStefano Zampini PetscFunctionReturn(0); 2800f4ddd8eeSStefano Zampini } 2801f4ddd8eeSStefano Zampini /* destroy old objects */ 2802f4ddd8eeSStefano Zampini ierr = ISDestroy(&pcbddc->is_R_local);CHKERRQ(ierr); 2803f4ddd8eeSStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_B);CHKERRQ(ierr); 2804f4ddd8eeSStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_D);CHKERRQ(ierr); 2805a64d13efSStefano Zampini /* Set Non-overlapping dimensions */ 2806b371cd4fSStefano Zampini n_B = pcis->n_B; 2807b371cd4fSStefano Zampini n_D = pcis->n - n_B; 2808b371cd4fSStefano Zampini n_vertices = pcbddc->n_vertices; 28093a50541eSStefano Zampini 2810a64d13efSStefano Zampini /* Dohrmann's notation: dofs splitted in R (Remaining: all dofs but the vertices) and V (Vertices) */ 28116816873aSStefano Zampini 281253892102SStefano Zampini /* create auxiliary bitmask and allocate workspace */ 2813df4d28bfSStefano Zampini if (!sub_schurs->reuse_solver) { 2814854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n-n_vertices,&idx_R_local);CHKERRQ(ierr); 2815a64d13efSStefano Zampini ierr = PetscBTCreate(pcis->n,&bitmask);CHKERRQ(ierr); 2816a64d13efSStefano Zampini for (i=0;i<n_vertices;i++) { 28170e6343abSStefano Zampini ierr = PetscBTSet(bitmask,pcbddc->local_primal_ref_node[i]);CHKERRQ(ierr); 2818a64d13efSStefano Zampini } 2819a64d13efSStefano Zampini 2820a64d13efSStefano Zampini for (i=0, n_R=0; i<pcis->n; i++) { 28214641a718SStefano Zampini if (!PetscBTLookup(bitmask,i)) { 28226816873aSStefano Zampini idx_R_local[n_R++] = i; 2823a64d13efSStefano Zampini } 2824a64d13efSStefano Zampini } 2825df4d28bfSStefano Zampini } else { /* A different ordering (already computed) is present if we are reusing the Schur solver */ 2826df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 28276816873aSStefano Zampini 2828df4d28bfSStefano Zampini ierr = ISGetIndices(reuse_solver->is_R,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 2829df4d28bfSStefano Zampini ierr = ISGetLocalSize(reuse_solver->is_R,&n_R);CHKERRQ(ierr); 28306816873aSStefano Zampini } 28313a50541eSStefano Zampini 28323a50541eSStefano Zampini /* Block code */ 28333a50541eSStefano Zampini vbs = 1; 28343a50541eSStefano Zampini ierr = MatGetBlockSize(pcbddc->local_mat,&bs);CHKERRQ(ierr); 28353a50541eSStefano Zampini if (bs>1 && !(n_vertices%bs)) { 28363a50541eSStefano Zampini PetscBool is_blocked = PETSC_TRUE; 28373a50541eSStefano Zampini PetscInt *vary; 2838df4d28bfSStefano Zampini if (!sub_schurs->reuse_solver) { 2839785e854fSJed Brown ierr = PetscMalloc1(pcis->n/bs,&vary);CHKERRQ(ierr); 28403a50541eSStefano Zampini ierr = PetscMemzero(vary,pcis->n/bs*sizeof(PetscInt));CHKERRQ(ierr); 2841d3df7717SStefano Zampini /* Verify that the vertex indices correspond to each element in a block (code taken from sbaij2.c) */ 2842d3df7717SStefano 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 */ 28430e6343abSStefano Zampini for (i=0; i<n_vertices; i++) vary[pcbddc->local_primal_ref_node[i]/bs]++; 2844d3df7717SStefano Zampini for (i=0; i<pcis->n/bs; i++) { 28453a50541eSStefano Zampini if (vary[i]!=0 && vary[i]!=bs) { 28463a50541eSStefano Zampini is_blocked = PETSC_FALSE; 28473a50541eSStefano Zampini break; 28483a50541eSStefano Zampini } 28493a50541eSStefano Zampini } 2850d3df7717SStefano Zampini ierr = PetscFree(vary);CHKERRQ(ierr); 2851d3df7717SStefano Zampini } else { 2852d3df7717SStefano Zampini /* Verify directly the R set */ 2853d3df7717SStefano Zampini for (i=0; i<n_R/bs; i++) { 2854d3df7717SStefano Zampini PetscInt j,node=idx_R_local[bs*i]; 2855d3df7717SStefano Zampini for (j=1; j<bs; j++) { 2856d3df7717SStefano Zampini if (node != idx_R_local[bs*i+j]-j) { 2857d3df7717SStefano Zampini is_blocked = PETSC_FALSE; 2858d3df7717SStefano Zampini break; 2859d3df7717SStefano Zampini } 2860d3df7717SStefano Zampini } 2861d3df7717SStefano Zampini } 2862d3df7717SStefano Zampini } 28633a50541eSStefano Zampini if (is_blocked) { /* build compressed IS for R nodes (complement of vertices) */ 28643a50541eSStefano Zampini vbs = bs; 28653a50541eSStefano Zampini for (i=0;i<n_R/vbs;i++) { 28663a50541eSStefano Zampini idx_R_local[i] = idx_R_local[vbs*i]/vbs; 28673a50541eSStefano Zampini } 28683a50541eSStefano Zampini } 28693a50541eSStefano Zampini } 28703a50541eSStefano Zampini ierr = ISCreateBlock(PETSC_COMM_SELF,vbs,n_R/vbs,idx_R_local,PETSC_COPY_VALUES,&pcbddc->is_R_local);CHKERRQ(ierr); 2871df4d28bfSStefano Zampini if (sub_schurs->reuse_solver) { 2872df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 287353892102SStefano Zampini 2874df4d28bfSStefano Zampini ierr = ISRestoreIndices(reuse_solver->is_R,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 2875df4d28bfSStefano Zampini ierr = ISDestroy(&reuse_solver->is_R);CHKERRQ(ierr); 287653892102SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->is_R_local);CHKERRQ(ierr); 2877df4d28bfSStefano Zampini reuse_solver->is_R = pcbddc->is_R_local; 287853892102SStefano Zampini } else { 28793a50541eSStefano Zampini ierr = PetscFree(idx_R_local);CHKERRQ(ierr); 288053892102SStefano Zampini } 2881a64d13efSStefano Zampini 2882a64d13efSStefano Zampini /* print some info if requested */ 2883a64d13efSStefano Zampini if (pcbddc->dbg_flag) { 2884a64d13efSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 2885a64d13efSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 28861575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 2887a64d13efSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d local dimensions\n",PetscGlobalRank);CHKERRQ(ierr); 2888a64d13efSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"local_size = %d, dirichlet_size = %d, boundary_size = %d\n",pcis->n,n_D,n_B);CHKERRQ(ierr); 28894f1b2e48SStefano 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); 2890a64d13efSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 2891a64d13efSStefano Zampini } 2892a64d13efSStefano Zampini 2893a64d13efSStefano Zampini /* VecScatters pcbddc->R_to_B and (optionally) pcbddc->R_to_D */ 2894df4d28bfSStefano Zampini if (!sub_schurs->reuse_solver) { 28956816873aSStefano Zampini IS is_aux1,is_aux2; 28966816873aSStefano Zampini PetscInt *aux_array1,*aux_array2,*is_indices,*idx_R_local; 28976816873aSStefano Zampini 28983a50541eSStefano Zampini ierr = ISGetIndices(pcbddc->is_R_local,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 2899854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n_B-n_vertices,&aux_array1);CHKERRQ(ierr); 2900854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n_B-n_vertices,&aux_array2);CHKERRQ(ierr); 2901a64d13efSStefano Zampini ierr = ISGetIndices(pcis->is_I_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 29024641a718SStefano Zampini for (i=0; i<n_D; i++) { 29034641a718SStefano Zampini ierr = PetscBTSet(bitmask,is_indices[i]);CHKERRQ(ierr); 29044641a718SStefano Zampini } 2905a64d13efSStefano Zampini ierr = ISRestoreIndices(pcis->is_I_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 2906a64d13efSStefano Zampini for (i=0, j=0; i<n_R; i++) { 29074641a718SStefano Zampini if (!PetscBTLookup(bitmask,idx_R_local[i])) { 29084641a718SStefano Zampini aux_array1[j++] = i; 2909a64d13efSStefano Zampini } 2910a64d13efSStefano Zampini } 2911a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array1,PETSC_OWN_POINTER,&is_aux1);CHKERRQ(ierr); 2912a64d13efSStefano Zampini ierr = ISGetIndices(pcis->is_B_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 2913a64d13efSStefano Zampini for (i=0, j=0; i<n_B; i++) { 29144641a718SStefano Zampini if (!PetscBTLookup(bitmask,is_indices[i])) { 29154641a718SStefano Zampini aux_array2[j++] = i; 2916a64d13efSStefano Zampini } 2917a64d13efSStefano Zampini } 2918a64d13efSStefano Zampini ierr = ISRestoreIndices(pcis->is_B_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 2919a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array2,PETSC_OWN_POINTER,&is_aux2);CHKERRQ(ierr); 2920a64d13efSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,is_aux1,pcis->vec1_B,is_aux2,&pcbddc->R_to_B);CHKERRQ(ierr); 2921a64d13efSStefano Zampini ierr = ISDestroy(&is_aux1);CHKERRQ(ierr); 2922a64d13efSStefano Zampini ierr = ISDestroy(&is_aux2);CHKERRQ(ierr); 2923a64d13efSStefano Zampini 29248eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 2925785e854fSJed Brown ierr = PetscMalloc1(n_D,&aux_array1);CHKERRQ(ierr); 2926a64d13efSStefano Zampini for (i=0, j=0; i<n_R; i++) { 29274641a718SStefano Zampini if (PetscBTLookup(bitmask,idx_R_local[i])) { 29284641a718SStefano Zampini aux_array1[j++] = i; 2929a64d13efSStefano Zampini } 2930a64d13efSStefano Zampini } 2931a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array1,PETSC_OWN_POINTER,&is_aux1);CHKERRQ(ierr); 2932a64d13efSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,is_aux1,pcis->vec1_D,(IS)0,&pcbddc->R_to_D);CHKERRQ(ierr); 2933a64d13efSStefano Zampini ierr = ISDestroy(&is_aux1);CHKERRQ(ierr); 2934a64d13efSStefano Zampini } 29354641a718SStefano Zampini ierr = PetscBTDestroy(&bitmask);CHKERRQ(ierr); 29363a50541eSStefano Zampini ierr = ISRestoreIndices(pcbddc->is_R_local,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 2937d62866d3SStefano Zampini } else { 2938df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 29396816873aSStefano Zampini IS tis; 29406816873aSStefano Zampini PetscInt schur_size; 29416816873aSStefano Zampini 2942df4d28bfSStefano Zampini ierr = ISGetLocalSize(reuse_solver->is_B,&schur_size);CHKERRQ(ierr); 29436816873aSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,schur_size,n_D,1,&tis);CHKERRQ(ierr); 2944df4d28bfSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,tis,pcis->vec1_B,reuse_solver->is_B,&pcbddc->R_to_B);CHKERRQ(ierr); 29456816873aSStefano Zampini ierr = ISDestroy(&tis);CHKERRQ(ierr); 29466816873aSStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 29476816873aSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,n_D,0,1,&tis);CHKERRQ(ierr); 29486816873aSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,tis,pcis->vec1_D,(IS)0,&pcbddc->R_to_D);CHKERRQ(ierr); 29496816873aSStefano Zampini ierr = ISDestroy(&tis);CHKERRQ(ierr); 2950d62866d3SStefano Zampini } 2951d62866d3SStefano Zampini } 2952a64d13efSStefano Zampini PetscFunctionReturn(0); 2953a64d13efSStefano Zampini } 2954a64d13efSStefano Zampini 2955304d26faSStefano Zampini 2956304d26faSStefano Zampini #undef __FUNCT__ 2957304d26faSStefano Zampini #define __FUNCT__ "PCBDDCSetUpLocalSolvers" 2958684f6988SStefano Zampini PetscErrorCode PCBDDCSetUpLocalSolvers(PC pc, PetscBool dirichlet, PetscBool neumann) 2959304d26faSStefano Zampini { 2960304d26faSStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 2961304d26faSStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 2962304d26faSStefano Zampini PC pc_temp; 2963304d26faSStefano Zampini Mat A_RR; 2964f4ddd8eeSStefano Zampini MatReuse reuse; 2965304d26faSStefano Zampini PetscScalar m_one = -1.0; 2966304d26faSStefano Zampini PetscReal value; 296704708bb6SStefano Zampini PetscInt n_D,n_R; 2968c7017625SStefano Zampini PetscBool check_corr[2],issbaij; 2969304d26faSStefano Zampini PetscErrorCode ierr; 2970e604994aSStefano Zampini /* prefixes stuff */ 2971312be037SStefano Zampini char dir_prefix[256],neu_prefix[256],str_level[16]; 2972e604994aSStefano Zampini size_t len; 2973304d26faSStefano Zampini 2974304d26faSStefano Zampini PetscFunctionBegin; 2975304d26faSStefano Zampini 2976e604994aSStefano Zampini /* compute prefixes */ 2977e604994aSStefano Zampini ierr = PetscStrcpy(dir_prefix,"");CHKERRQ(ierr); 2978e604994aSStefano Zampini ierr = PetscStrcpy(neu_prefix,"");CHKERRQ(ierr); 2979e604994aSStefano Zampini if (!pcbddc->current_level) { 2980e604994aSStefano Zampini ierr = PetscStrcpy(dir_prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 2981e604994aSStefano Zampini ierr = PetscStrcpy(neu_prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 2982e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,"pc_bddc_dirichlet_");CHKERRQ(ierr); 2983e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,"pc_bddc_neumann_");CHKERRQ(ierr); 2984e604994aSStefano Zampini } else { 2985e604994aSStefano Zampini ierr = PetscStrcpy(str_level,"");CHKERRQ(ierr); 2986312be037SStefano Zampini sprintf(str_level,"l%d_",(int)(pcbddc->current_level)); 2987e604994aSStefano Zampini ierr = PetscStrlen(((PetscObject)pc)->prefix,&len);CHKERRQ(ierr); 2988e604994aSStefano Zampini len -= 15; /* remove "pc_bddc_coarse_" */ 2989312be037SStefano Zampini if (pcbddc->current_level>1) len -= 3; /* remove "lX_" with X level number */ 2990312be037SStefano Zampini if (pcbddc->current_level>10) len -= 1; /* remove another char from level number */ 299134d6797cSStefano Zampini ierr = PetscStrncpy(dir_prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 299234d6797cSStefano Zampini ierr = PetscStrncpy(neu_prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 2993e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,"pc_bddc_dirichlet_");CHKERRQ(ierr); 2994e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,"pc_bddc_neumann_");CHKERRQ(ierr); 2995e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,str_level);CHKERRQ(ierr); 2996e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,str_level);CHKERRQ(ierr); 2997e604994aSStefano Zampini } 2998e604994aSStefano Zampini 2999304d26faSStefano Zampini /* DIRICHLET PROBLEM */ 3000684f6988SStefano Zampini if (dirichlet) { 3001d5574798SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 3002450f8f5eSStefano Zampini if (pcbddc->benign_n && !pcbddc->benign_change_explicit) { 3003450f8f5eSStefano Zampini if (!sub_schurs->reuse_solver) { 3004450f8f5eSStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Not yet implemented\n"); 3005450f8f5eSStefano Zampini } 3006450f8f5eSStefano Zampini if (pcbddc->dbg_flag) { 3007a3df083aSStefano Zampini Mat A_IIn; 3008a3df083aSStefano Zampini 3009a3df083aSStefano Zampini ierr = PCBDDCBenignProject(pc,pcis->is_I_local,pcis->is_I_local,&A_IIn);CHKERRQ(ierr); 3010a3df083aSStefano Zampini ierr = MatDestroy(&pcis->A_II);CHKERRQ(ierr); 3011a3df083aSStefano Zampini pcis->A_II = A_IIn; 3012a3df083aSStefano Zampini } 3013450f8f5eSStefano Zampini } 30143301b35fSStefano Zampini if (pcbddc->local_mat->symmetric_set) { 30153301b35fSStefano Zampini ierr = MatSetOption(pcis->A_II,MAT_SYMMETRIC,pcbddc->local_mat->symmetric_set);CHKERRQ(ierr); 3016964fefecSStefano Zampini } 3017ac78edfcSStefano Zampini /* Matrix for Dirichlet problem is pcis->A_II */ 3018964fefecSStefano Zampini n_D = pcis->n - pcis->n_B; 3019304d26faSStefano Zampini if (!pcbddc->ksp_D) { /* create object if not yet build */ 3020304d26faSStefano Zampini ierr = KSPCreate(PETSC_COMM_SELF,&pcbddc->ksp_D);CHKERRQ(ierr); 3021304d26faSStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->ksp_D,(PetscObject)pc,1);CHKERRQ(ierr); 3022304d26faSStefano Zampini /* default */ 3023304d26faSStefano Zampini ierr = KSPSetType(pcbddc->ksp_D,KSPPREONLY);CHKERRQ(ierr); 3024e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->ksp_D,dir_prefix);CHKERRQ(ierr); 30259577ea80SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcis->A_II,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 3026304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_D,&pc_temp);CHKERRQ(ierr); 30279577ea80SStefano Zampini if (issbaij) { 30289577ea80SStefano Zampini ierr = PCSetType(pc_temp,PCCHOLESKY);CHKERRQ(ierr); 30299577ea80SStefano Zampini } else { 3030304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCLU);CHKERRQ(ierr); 30319577ea80SStefano Zampini } 3032304d26faSStefano Zampini /* Allow user's customization */ 3033304d26faSStefano Zampini ierr = KSPSetFromOptions(pcbddc->ksp_D);CHKERRQ(ierr); 3034304d26faSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 3035304d26faSStefano Zampini } 3036d1e9a80fSBarry Smith ierr = KSPSetOperators(pcbddc->ksp_D,pcis->A_II,pcis->A_II);CHKERRQ(ierr); 3037df4d28bfSStefano Zampini if (sub_schurs->reuse_solver) { 3038df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 3039d62866d3SStefano Zampini 3040df4d28bfSStefano Zampini ierr = KSPSetPC(pcbddc->ksp_D,reuse_solver->interior_solver);CHKERRQ(ierr); 3041d5574798SStefano Zampini } 3042304d26faSStefano Zampini /* umfpack interface has a bug when matrix dimension is zero. TODO solve from umfpack interface */ 3043304d26faSStefano Zampini if (!n_D) { 3044304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_D,&pc_temp);CHKERRQ(ierr); 3045304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCNONE);CHKERRQ(ierr); 3046304d26faSStefano Zampini } 3047304d26faSStefano Zampini /* Set Up KSP for Dirichlet problem of BDDC */ 3048304d26faSStefano Zampini ierr = KSPSetUp(pcbddc->ksp_D);CHKERRQ(ierr); 3049304d26faSStefano Zampini /* set ksp_D into pcis data */ 3050304d26faSStefano Zampini ierr = KSPDestroy(&pcis->ksp_D);CHKERRQ(ierr); 3051304d26faSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->ksp_D);CHKERRQ(ierr); 3052304d26faSStefano Zampini pcis->ksp_D = pcbddc->ksp_D; 3053684f6988SStefano Zampini } 3054304d26faSStefano Zampini 3055304d26faSStefano Zampini /* NEUMANN PROBLEM */ 3056684f6988SStefano Zampini A_RR = 0; 3057684f6988SStefano Zampini if (neumann) { 3058d62866d3SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 305904708bb6SStefano Zampini PetscInt ibs,mbs; 306004708bb6SStefano Zampini PetscBool issbaij; 306104708bb6SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 3062f4ddd8eeSStefano Zampini /* Matrix for Neumann problem is A_RR -> we need to create/reuse it at this point */ 30638ce42a96SStefano Zampini ierr = ISGetSize(pcbddc->is_R_local,&n_R);CHKERRQ(ierr); 3064f4ddd8eeSStefano Zampini if (pcbddc->ksp_R) { /* already created ksp */ 3065f4ddd8eeSStefano Zampini PetscInt nn_R; 306681d9aea3SBarry Smith ierr = KSPGetOperators(pcbddc->ksp_R,NULL,&A_RR);CHKERRQ(ierr); 3067f4ddd8eeSStefano Zampini ierr = PetscObjectReference((PetscObject)A_RR);CHKERRQ(ierr); 3068f4ddd8eeSStefano Zampini ierr = MatGetSize(A_RR,&nn_R,NULL);CHKERRQ(ierr); 3069f4ddd8eeSStefano Zampini if (nn_R != n_R) { /* old ksp is not reusable, so reset it */ 3070f4ddd8eeSStefano Zampini ierr = KSPReset(pcbddc->ksp_R);CHKERRQ(ierr); 3071f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 3072f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 3073f4ddd8eeSStefano Zampini } else { /* same sizes, but nonzero pattern depend on primal vertices so it can be changed */ 3074727cdba6SStefano Zampini if (pcbddc->new_primal_space_local) { /* we are not sure the matrix will have the same nonzero pattern */ 3075f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 3076f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 3077f4ddd8eeSStefano Zampini } else { /* safe to reuse the matrix */ 3078f4ddd8eeSStefano Zampini reuse = MAT_REUSE_MATRIX; 3079f4ddd8eeSStefano Zampini } 3080f4ddd8eeSStefano Zampini } 3081f4ddd8eeSStefano Zampini /* last check */ 3082d1e9a80fSBarry Smith if (pc->flag == DIFFERENT_NONZERO_PATTERN) { 3083f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 3084f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 3085f4ddd8eeSStefano Zampini } 3086f4ddd8eeSStefano Zampini } else { /* first time, so we need to create the matrix */ 3087f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 3088f4ddd8eeSStefano Zampini } 3089a00504b5SStefano Zampini /* convert pcbddc->local_mat if needed later in PCBDDCSetUpCorrection */ 3090af732b37SStefano Zampini ierr = MatGetBlockSize(pcbddc->local_mat,&mbs);CHKERRQ(ierr); 3091af732b37SStefano Zampini ierr = ISGetBlockSize(pcbddc->is_R_local,&ibs);CHKERRQ(ierr); 309204708bb6SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcbddc->local_mat,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 309304708bb6SStefano Zampini if (ibs != mbs) { /* need to convert to SEQAIJ to extract any submatrix with is_R_local */ 309404708bb6SStefano Zampini if (matis->A == pcbddc->local_mat) { 309504708bb6SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 309604708bb6SStefano Zampini ierr = MatConvert(matis->A,MATSEQAIJ,MAT_INITIAL_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 3097af732b37SStefano Zampini } else { 3098511c6705SHong Zhang ierr = MatConvert(pcbddc->local_mat,MATSEQAIJ,MAT_INPLACE_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 30996816873aSStefano Zampini } 310004708bb6SStefano Zampini } else if (issbaij) { /* need to convert to BAIJ to get offdiagonal blocks */ 310104708bb6SStefano Zampini if (matis->A == pcbddc->local_mat) { 310204708bb6SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 310304708bb6SStefano Zampini ierr = MatConvert(matis->A,MATSEQBAIJ,MAT_INITIAL_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 310404708bb6SStefano Zampini } else { 3105511c6705SHong Zhang ierr = MatConvert(pcbddc->local_mat,MATSEQBAIJ,MAT_INPLACE_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 310604708bb6SStefano Zampini } 310704708bb6SStefano Zampini } 3108a00504b5SStefano Zampini /* extract A_RR */ 3109a00504b5SStefano Zampini if (sub_schurs->reuse_solver) { 3110a00504b5SStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 3111a00504b5SStefano Zampini 3112a00504b5SStefano Zampini if (pcbddc->dbg_flag) { /* we need A_RR to test the solver later */ 311316e386b8SStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 3114a00504b5SStefano Zampini if (reuse_solver->benign_n) { /* we are not using the explicit change of basis on the pressures */ 311516e386b8SStefano Zampini ierr = PCBDDCBenignProject(pc,pcbddc->is_R_local,pcbddc->is_R_local,&A_RR);CHKERRQ(ierr); 311616e386b8SStefano Zampini } else { 3117a00504b5SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,pcbddc->is_R_local,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&A_RR);CHKERRQ(ierr); 3118a00504b5SStefano Zampini } 3119a00504b5SStefano Zampini } else { 3120a00504b5SStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 3121a00504b5SStefano Zampini ierr = PCGetOperators(reuse_solver->correction_solver,&A_RR,NULL);CHKERRQ(ierr); 3122a00504b5SStefano Zampini ierr = PetscObjectReference((PetscObject)A_RR);CHKERRQ(ierr); 3123a00504b5SStefano Zampini } 3124a00504b5SStefano Zampini } else { /* we have to build the neumann solver, so we need to extract the relevant matrix */ 3125f4ddd8eeSStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,pcbddc->is_R_local,pcbddc->is_R_local,reuse,&A_RR);CHKERRQ(ierr); 312616e386b8SStefano Zampini } 31273301b35fSStefano Zampini if (pcbddc->local_mat->symmetric_set) { 31283301b35fSStefano Zampini ierr = MatSetOption(A_RR,MAT_SYMMETRIC,pcbddc->local_mat->symmetric_set);CHKERRQ(ierr); 31296816873aSStefano Zampini } 3130f4ddd8eeSStefano Zampini if (!pcbddc->ksp_R) { /* create object if not present */ 3131304d26faSStefano Zampini ierr = KSPCreate(PETSC_COMM_SELF,&pcbddc->ksp_R);CHKERRQ(ierr); 3132304d26faSStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->ksp_R,(PetscObject)pc,1);CHKERRQ(ierr); 3133304d26faSStefano Zampini /* default */ 3134304d26faSStefano Zampini ierr = KSPSetType(pcbddc->ksp_R,KSPPREONLY);CHKERRQ(ierr); 3135e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->ksp_R,neu_prefix);CHKERRQ(ierr); 3136304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_temp);CHKERRQ(ierr); 31379577ea80SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)A_RR,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 31389577ea80SStefano Zampini if (issbaij) { 31399577ea80SStefano Zampini ierr = PCSetType(pc_temp,PCCHOLESKY);CHKERRQ(ierr); 31409577ea80SStefano Zampini } else { 3141304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCLU);CHKERRQ(ierr); 31429577ea80SStefano Zampini } 3143304d26faSStefano Zampini /* Allow user's customization */ 3144304d26faSStefano Zampini ierr = KSPSetFromOptions(pcbddc->ksp_R);CHKERRQ(ierr); 3145304d26faSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 3146304d26faSStefano Zampini } 3147304d26faSStefano Zampini /* umfpack interface has a bug when matrix dimension is zero. TODO solve from umfpack interface */ 3148304d26faSStefano Zampini if (!n_R) { 3149304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_temp);CHKERRQ(ierr); 3150304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCNONE);CHKERRQ(ierr); 3151304d26faSStefano Zampini } 31525cbda25cSStefano Zampini ierr = KSPSetOperators(pcbddc->ksp_R,A_RR,A_RR);CHKERRQ(ierr); 3153df4d28bfSStefano Zampini /* Reuse solver if it is present */ 3154df4d28bfSStefano Zampini if (sub_schurs->reuse_solver) { 3155df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 3156d62866d3SStefano Zampini 3157df4d28bfSStefano Zampini ierr = KSPSetPC(pcbddc->ksp_R,reuse_solver->correction_solver);CHKERRQ(ierr); 3158d62866d3SStefano Zampini } 3159304d26faSStefano Zampini /* Set Up KSP for Neumann problem of BDDC */ 3160304d26faSStefano Zampini ierr = KSPSetUp(pcbddc->ksp_R);CHKERRQ(ierr); 3161684f6988SStefano Zampini } 3162304d26faSStefano Zampini 3163684f6988SStefano Zampini if (pcbddc->dbg_flag) { 3164684f6988SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 31651575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 3166684f6988SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 3167684f6988SStefano Zampini } 3168c7017625SStefano Zampini 3169c7017625SStefano Zampini /* adapt Dirichlet and Neumann solvers if a nullspace correction has been requested */ 3170c7017625SStefano Zampini check_corr[0] = check_corr[1] = PETSC_FALSE; 3171c7017625SStefano Zampini if (pcbddc->NullSpace_corr[0]) { 3172c7017625SStefano Zampini ierr = PCBDDCSetUseExactDirichlet(pc,PETSC_FALSE);CHKERRQ(ierr); 3173c7017625SStefano Zampini } 3174c7017625SStefano Zampini if (dirichlet && pcbddc->NullSpace_corr[0] && !pcbddc->switch_static) { 3175c7017625SStefano Zampini check_corr[0] = PETSC_TRUE; 3176c7017625SStefano Zampini ierr = PCBDDCNullSpaceAssembleCorrection(pc,PETSC_TRUE,pcbddc->NullSpace_corr[1]);CHKERRQ(ierr); 3177c7017625SStefano Zampini } 3178c7017625SStefano Zampini if (neumann && pcbddc->NullSpace_corr[2]) { 3179c7017625SStefano Zampini check_corr[1] = PETSC_TRUE; 3180c7017625SStefano Zampini ierr = PCBDDCNullSpaceAssembleCorrection(pc,PETSC_FALSE,pcbddc->NullSpace_corr[3]);CHKERRQ(ierr); 3181c7017625SStefano Zampini } 3182c7017625SStefano Zampini 3183c7017625SStefano Zampini /* check Dirichlet and Neumann solvers */ 3184c7017625SStefano Zampini if (pcbddc->dbg_flag) { 3185684f6988SStefano Zampini if (dirichlet) { /* Dirichlet */ 31860fccc4e9SStefano Zampini ierr = VecSetRandom(pcis->vec1_D,NULL);CHKERRQ(ierr); 31870fccc4e9SStefano Zampini ierr = MatMult(pcis->A_II,pcis->vec1_D,pcis->vec2_D);CHKERRQ(ierr); 31880fccc4e9SStefano Zampini ierr = KSPSolve(pcbddc->ksp_D,pcis->vec2_D,pcis->vec2_D);CHKERRQ(ierr); 31890fccc4e9SStefano Zampini ierr = VecAXPY(pcis->vec1_D,m_one,pcis->vec2_D);CHKERRQ(ierr); 31900fccc4e9SStefano Zampini ierr = VecNorm(pcis->vec1_D,NORM_INFINITY,&value);CHKERRQ(ierr); 3191e604994aSStefano 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); 3192c7017625SStefano Zampini if (check_corr[0]) { 3193c7017625SStefano Zampini ierr = PCBDDCNullSpaceCheckCorrection(pc,PETSC_TRUE);CHKERRQ(ierr); 3194c7017625SStefano Zampini } 3195304d26faSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 3196304d26faSStefano Zampini } 3197684f6988SStefano Zampini if (neumann) { /* Neumann */ 31980fccc4e9SStefano Zampini ierr = VecSetRandom(pcbddc->vec1_R,NULL);CHKERRQ(ierr); 31990fccc4e9SStefano Zampini ierr = MatMult(A_RR,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 32000fccc4e9SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec2_R,pcbddc->vec2_R);CHKERRQ(ierr); 32010fccc4e9SStefano Zampini ierr = VecAXPY(pcbddc->vec1_R,m_one,pcbddc->vec2_R);CHKERRQ(ierr); 32020fccc4e9SStefano Zampini ierr = VecNorm(pcbddc->vec1_R,NORM_INFINITY,&value);CHKERRQ(ierr); 3203e604994aSStefano 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); 3204c7017625SStefano Zampini if (check_corr[1]) { 3205c7017625SStefano Zampini ierr = PCBDDCNullSpaceCheckCorrection(pc,PETSC_FALSE);CHKERRQ(ierr); 3206c7017625SStefano Zampini } 3207304d26faSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 3208304d26faSStefano Zampini } 3209684f6988SStefano Zampini } 32105cbda25cSStefano Zampini /* free Neumann problem's matrix */ 32115cbda25cSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 3212304d26faSStefano Zampini PetscFunctionReturn(0); 3213304d26faSStefano Zampini } 3214304d26faSStefano Zampini 3215304d26faSStefano Zampini #undef __FUNCT__ 3216ba15a52eSStefano Zampini #define __FUNCT__ "PCBDDCSolveSubstructureCorrection" 321780677318SStefano Zampini static PetscErrorCode PCBDDCSolveSubstructureCorrection(PC pc, Vec inout_B, Vec inout_D, PetscBool applytranspose) 3218674ae819SStefano Zampini { 3219674ae819SStefano Zampini PetscErrorCode ierr; 3220674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 3221be83ff47SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 3222674ae819SStefano Zampini 3223674ae819SStefano Zampini PetscFunctionBegin; 3224df4d28bfSStefano Zampini if (!sub_schurs->reuse_solver) { 322580677318SStefano Zampini ierr = VecSet(pcbddc->vec1_R,0.);CHKERRQ(ierr); 322620c7b377SStefano Zampini } 322780677318SStefano Zampini if (!pcbddc->switch_static) { 322880677318SStefano Zampini if (applytranspose && pcbddc->local_auxmat1) { 322980677318SStefano Zampini ierr = MatMultTranspose(pcbddc->local_auxmat2,inout_B,pcbddc->vec1_C);CHKERRQ(ierr); 323080677318SStefano Zampini ierr = MatMultTransposeAdd(pcbddc->local_auxmat1,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr); 323120c7b377SStefano Zampini } 3232df4d28bfSStefano Zampini if (!sub_schurs->reuse_solver) { 323380677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 323480677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 323520c7b377SStefano Zampini } else { 3236df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 3237be83ff47SStefano Zampini 3238df4d28bfSStefano Zampini ierr = VecScatterBegin(reuse_solver->correction_scatter_B,inout_B,reuse_solver->rhs_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 3239df4d28bfSStefano Zampini ierr = VecScatterEnd(reuse_solver->correction_scatter_B,inout_B,reuse_solver->rhs_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 324020c7b377SStefano Zampini } 3241be83ff47SStefano Zampini } else { 324280677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 324380677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 324480677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,inout_D,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 324580677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,inout_D,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 324680677318SStefano Zampini if (applytranspose && pcbddc->local_auxmat1) { 324780677318SStefano Zampini ierr = MatMultTranspose(pcbddc->local_auxmat2,pcbddc->vec1_R,pcbddc->vec1_C);CHKERRQ(ierr); 324880677318SStefano Zampini ierr = MatMultTransposeAdd(pcbddc->local_auxmat1,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr); 324980677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 325080677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 3251674ae819SStefano Zampini } 3252674ae819SStefano Zampini } 325372b8c272SStefano Zampini if (!sub_schurs->reuse_solver || pcbddc->switch_static) { 325480677318SStefano Zampini if (applytranspose) { 325580677318SStefano Zampini ierr = KSPSolveTranspose(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 325680677318SStefano Zampini } else { 325780677318SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 325880677318SStefano Zampini } 3259be83ff47SStefano Zampini } else { 3260df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 3261be83ff47SStefano Zampini 3262be83ff47SStefano Zampini if (applytranspose) { 3263df4d28bfSStefano Zampini ierr = MatFactorSolveSchurComplementTranspose(reuse_solver->F,reuse_solver->rhs_B,reuse_solver->sol_B);CHKERRQ(ierr); 3264be83ff47SStefano Zampini } else { 3265df4d28bfSStefano Zampini ierr = MatFactorSolveSchurComplement(reuse_solver->F,reuse_solver->rhs_B,reuse_solver->sol_B);CHKERRQ(ierr); 3266be83ff47SStefano Zampini } 3267be83ff47SStefano Zampini } 326880677318SStefano Zampini ierr = VecSet(inout_B,0.);CHKERRQ(ierr); 326980677318SStefano Zampini if (!pcbddc->switch_static) { 3270df4d28bfSStefano Zampini if (!sub_schurs->reuse_solver) { 327180677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 327280677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 3273be83ff47SStefano Zampini } else { 3274df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 3275be83ff47SStefano Zampini 3276df4d28bfSStefano Zampini ierr = VecScatterBegin(reuse_solver->correction_scatter_B,reuse_solver->sol_B,inout_B,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 3277df4d28bfSStefano Zampini ierr = VecScatterEnd(reuse_solver->correction_scatter_B,reuse_solver->sol_B,inout_B,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 3278be83ff47SStefano Zampini } 327980677318SStefano Zampini if (!applytranspose && pcbddc->local_auxmat1) { 328080677318SStefano Zampini ierr = MatMult(pcbddc->local_auxmat1,inout_B,pcbddc->vec1_C);CHKERRQ(ierr); 328180677318SStefano Zampini ierr = MatMultAdd(pcbddc->local_auxmat2,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr); 328280677318SStefano Zampini } 328380677318SStefano Zampini } else { 328480677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 328580677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 328680677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 328780677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 328880677318SStefano Zampini if (!applytranspose && pcbddc->local_auxmat1) { 328980677318SStefano Zampini ierr = MatMult(pcbddc->local_auxmat1,inout_B,pcbddc->vec1_C);CHKERRQ(ierr); 329080677318SStefano Zampini ierr = MatMultAdd(pcbddc->local_auxmat2,pcbddc->vec1_C,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 329180677318SStefano Zampini } 329280677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 329380677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 329480677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 329580677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 3296674ae819SStefano Zampini } 3297674ae819SStefano Zampini PetscFunctionReturn(0); 3298674ae819SStefano Zampini } 3299674ae819SStefano Zampini 3300dc359a40SStefano Zampini /* parameter apply transpose determines if the interface preconditioner should be applied transposed or not */ 3301674ae819SStefano Zampini #undef __FUNCT__ 3302674ae819SStefano Zampini #define __FUNCT__ "PCBDDCApplyInterfacePreconditioner" 3303dc359a40SStefano Zampini PetscErrorCode PCBDDCApplyInterfacePreconditioner(PC pc, PetscBool applytranspose) 3304674ae819SStefano Zampini { 3305674ae819SStefano Zampini PetscErrorCode ierr; 3306674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 3307674ae819SStefano Zampini PC_IS* pcis = (PC_IS*) (pc->data); 3308674ae819SStefano Zampini const PetscScalar zero = 0.0; 3309674ae819SStefano Zampini 3310674ae819SStefano Zampini PetscFunctionBegin; 3311dc359a40SStefano Zampini /* Application of PSI^T or PHI^T (depending on applytranspose, see comment above) */ 33124fee134fSStefano Zampini if (!pcbddc->benign_apply_coarse_only) { 3313dc359a40SStefano Zampini if (applytranspose) { 3314674ae819SStefano Zampini ierr = MatMultTranspose(pcbddc->coarse_phi_B,pcis->vec1_B,pcbddc->vec1_P);CHKERRQ(ierr); 33158eeda7d8SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultTransposeAdd(pcbddc->coarse_phi_D,pcis->vec1_D,pcbddc->vec1_P,pcbddc->vec1_P);CHKERRQ(ierr); } 3316dc359a40SStefano Zampini } else { 3317674ae819SStefano Zampini ierr = MatMultTranspose(pcbddc->coarse_psi_B,pcis->vec1_B,pcbddc->vec1_P);CHKERRQ(ierr); 3318674ae819SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultTransposeAdd(pcbddc->coarse_psi_D,pcis->vec1_D,pcbddc->vec1_P,pcbddc->vec1_P);CHKERRQ(ierr); } 331915aaf578SStefano Zampini } 33204fee134fSStefano Zampini } else { 33214fee134fSStefano Zampini ierr = VecSet(pcbddc->vec1_P,zero);CHKERRQ(ierr); 33224fee134fSStefano Zampini } 3323efc2fbd9SStefano Zampini 3324efc2fbd9SStefano Zampini /* add p0 to the last value of vec1_P holding the coarse dof relative to p0 */ 33254f1b2e48SStefano Zampini if (pcbddc->benign_n) { 3326efc2fbd9SStefano Zampini PetscScalar *array; 33274f1b2e48SStefano Zampini PetscInt j; 3328efc2fbd9SStefano Zampini 3329efc2fbd9SStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 33304f1b2e48SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) array[pcbddc->local_primal_size-pcbddc->benign_n+j] += pcbddc->benign_p0[j]; 3331efc2fbd9SStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 3332efc2fbd9SStefano Zampini } 3333efc2fbd9SStefano Zampini 333412edc857SStefano Zampini /* start communications from local primal nodes to rhs of coarse solver */ 333512edc857SStefano Zampini ierr = VecSet(pcbddc->coarse_vec,zero);CHKERRQ(ierr); 333612edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataBegin(pc,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 333712edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataEnd(pc,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 333812edc857SStefano Zampini 33399f00e9b4SStefano Zampini /* Coarse solution -> rhs and sol updated inside PCBDDCScattarCoarseDataBegin/End */ 334012edc857SStefano Zampini /* TODO remove null space when doing multilevel */ 334112edc857SStefano Zampini if (pcbddc->coarse_ksp) { 334251694757SStefano Zampini Mat coarse_mat; 3343964fefecSStefano Zampini Vec rhs,sol; 334451694757SStefano Zampini MatNullSpace nullsp; 334527b6a85dSStefano Zampini PetscBool isbddc = PETSC_FALSE; 3346964fefecSStefano Zampini 334727b6a85dSStefano Zampini if (pcbddc->benign_have_null) { 334827b6a85dSStefano Zampini PC coarse_pc; 334927b6a85dSStefano Zampini 335027b6a85dSStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&coarse_pc);CHKERRQ(ierr); 335127b6a85dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)coarse_pc,PCBDDC,&isbddc);CHKERRQ(ierr); 335227b6a85dSStefano Zampini /* we need to propagate to coarser levels the need for a possible benign correction */ 335327b6a85dSStefano Zampini if (isbddc && pcbddc->benign_apply_coarse_only && !pcbddc->benign_skip_correction) { 335427b6a85dSStefano Zampini PC_BDDC* coarsepcbddc = (PC_BDDC*)(coarse_pc->data); 335527b6a85dSStefano Zampini coarsepcbddc->benign_skip_correction = PETSC_FALSE; 33563bca92a6SStefano Zampini coarsepcbddc->benign_apply_coarse_only = PETSC_TRUE; 335727b6a85dSStefano Zampini } 335827b6a85dSStefano Zampini } 3359964fefecSStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&rhs);CHKERRQ(ierr); 3360964fefecSStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&sol);CHKERRQ(ierr); 336151694757SStefano Zampini ierr = KSPGetOperators(pcbddc->coarse_ksp,&coarse_mat,NULL);CHKERRQ(ierr); 336251694757SStefano Zampini ierr = MatGetNullSpace(coarse_mat,&nullsp);CHKERRQ(ierr); 336351694757SStefano Zampini if (nullsp) { 336451694757SStefano Zampini ierr = MatNullSpaceRemove(nullsp,rhs);CHKERRQ(ierr); 336551694757SStefano Zampini } 336612edc857SStefano Zampini if (applytranspose) { 33671f4df5f7SStefano Zampini if (pcbddc->benign_apply_coarse_only) { 33681f4df5f7SStefano Zampini SETERRQ(PetscObjectComm((PetscObject)pcbddc->coarse_ksp),PETSC_ERR_SUP,"Not yet implemented"); 33692701bc32SStefano Zampini } else { 3370964fefecSStefano Zampini ierr = KSPSolveTranspose(pcbddc->coarse_ksp,rhs,sol);CHKERRQ(ierr); 33712701bc32SStefano Zampini } 33722701bc32SStefano Zampini } else { 33731f4df5f7SStefano Zampini if (pcbddc->benign_apply_coarse_only && isbddc) { /* need just to apply the coarse preconditioner during presolve */ 33742701bc32SStefano Zampini PC coarse_pc; 33752701bc32SStefano Zampini 33762701bc32SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&coarse_pc);CHKERRQ(ierr); 33772701bc32SStefano Zampini ierr = PCPreSolve(coarse_pc,pcbddc->coarse_ksp);CHKERRQ(ierr); 33783e589ea0SStefano Zampini ierr = PCBDDCBenignRemoveInterior(coarse_pc,rhs,sol);CHKERRQ(ierr); 33792701bc32SStefano Zampini ierr = PCPostSolve(coarse_pc,pcbddc->coarse_ksp);CHKERRQ(ierr); 338012edc857SStefano Zampini } else { 3381964fefecSStefano Zampini ierr = KSPSolve(pcbddc->coarse_ksp,rhs,sol);CHKERRQ(ierr); 338212edc857SStefano Zampini } 33832701bc32SStefano Zampini } 338427b6a85dSStefano Zampini /* we don't the benign correction at coarser levels anymore */ 338527b6a85dSStefano Zampini if (pcbddc->benign_have_null && isbddc) { 338627b6a85dSStefano Zampini PC coarse_pc; 338727b6a85dSStefano Zampini PC_BDDC* coarsepcbddc; 338827b6a85dSStefano Zampini 338927b6a85dSStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&coarse_pc);CHKERRQ(ierr); 339027b6a85dSStefano Zampini coarsepcbddc = (PC_BDDC*)(coarse_pc->data); 339127b6a85dSStefano Zampini coarsepcbddc->benign_skip_correction = PETSC_TRUE; 33923bca92a6SStefano Zampini coarsepcbddc->benign_apply_coarse_only = PETSC_FALSE; 339327b6a85dSStefano Zampini } 339451694757SStefano Zampini if (nullsp) { 339551694757SStefano Zampini ierr = MatNullSpaceRemove(nullsp,sol);CHKERRQ(ierr); 339651694757SStefano Zampini } 339712edc857SStefano Zampini } 3398674ae819SStefano Zampini 3399674ae819SStefano Zampini /* Local solution on R nodes */ 34004fee134fSStefano Zampini if (pcis->n && !pcbddc->benign_apply_coarse_only) { 340180677318SStefano Zampini ierr = PCBDDCSolveSubstructureCorrection(pc,pcis->vec1_B,pcis->vec1_D,applytranspose);CHKERRQ(ierr); 34029f00e9b4SStefano Zampini } 34039f00e9b4SStefano Zampini /* communications from coarse sol to local primal nodes */ 34049f00e9b4SStefano Zampini ierr = PCBDDCScatterCoarseDataBegin(pc,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 340512edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataEnd(pc,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 3406674ae819SStefano Zampini 34074fee134fSStefano Zampini /* Sum contributions from the two levels */ 34084fee134fSStefano Zampini if (!pcbddc->benign_apply_coarse_only) { 3409dc359a40SStefano Zampini if (applytranspose) { 3410dc359a40SStefano Zampini ierr = MatMultAdd(pcbddc->coarse_psi_B,pcbddc->vec1_P,pcis->vec1_B,pcis->vec1_B);CHKERRQ(ierr); 3411dc359a40SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultAdd(pcbddc->coarse_psi_D,pcbddc->vec1_P,pcis->vec1_D,pcis->vec1_D);CHKERRQ(ierr); } 3412dc359a40SStefano Zampini } else { 3413674ae819SStefano Zampini ierr = MatMultAdd(pcbddc->coarse_phi_B,pcbddc->vec1_P,pcis->vec1_B,pcis->vec1_B);CHKERRQ(ierr); 34148eeda7d8SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultAdd(pcbddc->coarse_phi_D,pcbddc->vec1_P,pcis->vec1_D,pcis->vec1_D);CHKERRQ(ierr); } 3415dc359a40SStefano Zampini } 3416efc2fbd9SStefano Zampini /* store p0 */ 34174f1b2e48SStefano Zampini if (pcbddc->benign_n) { 3418efc2fbd9SStefano Zampini PetscScalar *array; 34194f1b2e48SStefano Zampini PetscInt j; 3420efc2fbd9SStefano Zampini 3421efc2fbd9SStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 34224f1b2e48SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) pcbddc->benign_p0[j] = array[pcbddc->local_primal_size-pcbddc->benign_n+j]; 3423efc2fbd9SStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 3424efc2fbd9SStefano Zampini } 34254fee134fSStefano Zampini } else { /* expand the coarse solution */ 34264fee134fSStefano Zampini if (applytranspose) { 34274fee134fSStefano Zampini ierr = MatMult(pcbddc->coarse_psi_B,pcbddc->vec1_P,pcis->vec1_B);CHKERRQ(ierr); 34284fee134fSStefano Zampini } else { 34294fee134fSStefano Zampini ierr = MatMult(pcbddc->coarse_phi_B,pcbddc->vec1_P,pcis->vec1_B);CHKERRQ(ierr); 34304fee134fSStefano Zampini } 34314fee134fSStefano Zampini } 3432674ae819SStefano Zampini PetscFunctionReturn(0); 3433674ae819SStefano Zampini } 3434674ae819SStefano Zampini 3435674ae819SStefano Zampini #undef __FUNCT__ 3436674ae819SStefano Zampini #define __FUNCT__ "PCBDDCScatterCoarseDataBegin" 343712edc857SStefano Zampini PetscErrorCode PCBDDCScatterCoarseDataBegin(PC pc,InsertMode imode, ScatterMode smode) 3438674ae819SStefano Zampini { 3439674ae819SStefano Zampini PetscErrorCode ierr; 3440674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 344158da7f69SStefano Zampini PetscScalar *array; 344212edc857SStefano Zampini Vec from,to; 3443674ae819SStefano Zampini 3444674ae819SStefano Zampini PetscFunctionBegin; 344512edc857SStefano Zampini if (smode == SCATTER_REVERSE) { /* from global to local -> get data from coarse solution */ 344612edc857SStefano Zampini from = pcbddc->coarse_vec; 344712edc857SStefano Zampini to = pcbddc->vec1_P; 344812edc857SStefano Zampini if (pcbddc->coarse_ksp) { /* get array from coarse processes */ 344912edc857SStefano Zampini Vec tvec; 345058da7f69SStefano Zampini 345158da7f69SStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 345258da7f69SStefano Zampini ierr = VecResetArray(tvec);CHKERRQ(ierr); 345312edc857SStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 345458da7f69SStefano Zampini ierr = VecGetArray(tvec,&array);CHKERRQ(ierr); 345558da7f69SStefano Zampini ierr = VecPlaceArray(from,array);CHKERRQ(ierr); 345658da7f69SStefano Zampini ierr = VecRestoreArray(tvec,&array);CHKERRQ(ierr); 345712edc857SStefano Zampini } 345812edc857SStefano Zampini } else { /* from local to global -> put data in coarse right hand side */ 345912edc857SStefano Zampini from = pcbddc->vec1_P; 346012edc857SStefano Zampini to = pcbddc->coarse_vec; 346112edc857SStefano Zampini } 346212edc857SStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,from,to,imode,smode);CHKERRQ(ierr); 3463674ae819SStefano Zampini PetscFunctionReturn(0); 3464674ae819SStefano Zampini } 3465674ae819SStefano Zampini 3466674ae819SStefano Zampini #undef __FUNCT__ 3467674ae819SStefano Zampini #define __FUNCT__ "PCBDDCScatterCoarseDataEnd" 346812edc857SStefano Zampini PetscErrorCode PCBDDCScatterCoarseDataEnd(PC pc, InsertMode imode, ScatterMode smode) 3469674ae819SStefano Zampini { 3470674ae819SStefano Zampini PetscErrorCode ierr; 3471674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 347258da7f69SStefano Zampini PetscScalar *array; 347312edc857SStefano Zampini Vec from,to; 3474674ae819SStefano Zampini 3475674ae819SStefano Zampini PetscFunctionBegin; 347612edc857SStefano Zampini if (smode == SCATTER_REVERSE) { /* from global to local -> get data from coarse solution */ 347712edc857SStefano Zampini from = pcbddc->coarse_vec; 347812edc857SStefano Zampini to = pcbddc->vec1_P; 347912edc857SStefano Zampini } else { /* from local to global -> put data in coarse right hand side */ 348012edc857SStefano Zampini from = pcbddc->vec1_P; 348112edc857SStefano Zampini to = pcbddc->coarse_vec; 348212edc857SStefano Zampini } 348312edc857SStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,from,to,imode,smode);CHKERRQ(ierr); 348412edc857SStefano Zampini if (smode == SCATTER_FORWARD) { 348512edc857SStefano Zampini if (pcbddc->coarse_ksp) { /* get array from coarse processes */ 348612edc857SStefano Zampini Vec tvec; 348758da7f69SStefano Zampini 348812edc857SStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 348958da7f69SStefano Zampini ierr = VecGetArray(to,&array);CHKERRQ(ierr); 349058da7f69SStefano Zampini ierr = VecPlaceArray(tvec,array);CHKERRQ(ierr); 349158da7f69SStefano Zampini ierr = VecRestoreArray(to,&array);CHKERRQ(ierr); 349258da7f69SStefano Zampini } 349358da7f69SStefano Zampini } else { 349458da7f69SStefano Zampini if (pcbddc->coarse_ksp) { /* restore array of pcbddc->coarse_vec */ 349558da7f69SStefano Zampini ierr = VecResetArray(from);CHKERRQ(ierr); 349612edc857SStefano Zampini } 349712edc857SStefano Zampini } 3498674ae819SStefano Zampini PetscFunctionReturn(0); 3499674ae819SStefano Zampini } 3500674ae819SStefano Zampini 3501984c4197SStefano Zampini /* uncomment for testing purposes */ 3502984c4197SStefano Zampini /* #define PETSC_MISSING_LAPACK_GESVD 1 */ 3503674ae819SStefano Zampini #undef __FUNCT__ 3504674ae819SStefano Zampini #define __FUNCT__ "PCBDDCConstraintsSetUp" 3505674ae819SStefano Zampini PetscErrorCode PCBDDCConstraintsSetUp(PC pc) 3506674ae819SStefano Zampini { 3507674ae819SStefano Zampini PetscErrorCode ierr; 3508674ae819SStefano Zampini PC_IS* pcis = (PC_IS*)(pc->data); 3509674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 3510674ae819SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 3511984c4197SStefano Zampini /* one and zero */ 3512984c4197SStefano Zampini PetscScalar one=1.0,zero=0.0; 3513984c4197SStefano Zampini /* space to store constraints and their local indices */ 35149162d606SStefano Zampini PetscScalar *constraints_data; 35159162d606SStefano Zampini PetscInt *constraints_idxs,*constraints_idxs_B; 35169162d606SStefano Zampini PetscInt *constraints_idxs_ptr,*constraints_data_ptr; 35179162d606SStefano Zampini PetscInt *constraints_n; 3518984c4197SStefano Zampini /* iterators */ 3519b3d85658SStefano Zampini PetscInt i,j,k,total_counts,total_counts_cc,cum; 3520984c4197SStefano Zampini /* BLAS integers */ 3521e310c8b4SStefano Zampini PetscBLASInt lwork,lierr; 3522e310c8b4SStefano Zampini PetscBLASInt Blas_N,Blas_M,Blas_K,Blas_one=1; 3523c4303822SStefano Zampini PetscBLASInt Blas_LDA,Blas_LDB,Blas_LDC; 3524727cdba6SStefano Zampini /* reuse */ 35250e6343abSStefano Zampini PetscInt olocal_primal_size,olocal_primal_size_cc; 35260e6343abSStefano Zampini PetscInt *olocal_primal_ref_node,*olocal_primal_ref_mult; 3527984c4197SStefano Zampini /* change of basis */ 3528b3d85658SStefano Zampini PetscBool qr_needed; 35299162d606SStefano Zampini PetscBT change_basis,qr_needed_idx; 3530984c4197SStefano Zampini /* auxiliary stuff */ 353164efe560SStefano Zampini PetscInt *nnz,*is_indices; 35328a0068c3SStefano Zampini PetscInt ncc; 3533984c4197SStefano Zampini /* some quantities */ 353445a1bb75SStefano Zampini PetscInt n_vertices,total_primal_vertices,valid_constraints; 3535a58a30b4SStefano Zampini PetscInt size_of_constraint,max_size_of_constraint=0,max_constraints,temp_constraints; 3536984c4197SStefano Zampini 3537674ae819SStefano Zampini PetscFunctionBegin; 35388e61c736SStefano Zampini /* Destroy Mat objects computed previously */ 35398e61c736SStefano Zampini ierr = MatDestroy(&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 35408e61c736SStefano Zampini ierr = MatDestroy(&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 354116909a7fSStefano Zampini ierr = MatDestroy(&pcbddc->switch_static_change);CHKERRQ(ierr); 3542088faed8SStefano Zampini /* save info on constraints from previous setup (if any) */ 3543088faed8SStefano Zampini olocal_primal_size = pcbddc->local_primal_size; 35440e6343abSStefano Zampini olocal_primal_size_cc = pcbddc->local_primal_size_cc; 35450e6343abSStefano Zampini ierr = PetscMalloc2(olocal_primal_size_cc,&olocal_primal_ref_node,olocal_primal_size_cc,&olocal_primal_ref_mult);CHKERRQ(ierr); 35460e6343abSStefano Zampini ierr = PetscMemcpy(olocal_primal_ref_node,pcbddc->local_primal_ref_node,olocal_primal_size_cc*sizeof(PetscInt));CHKERRQ(ierr); 35470e6343abSStefano Zampini ierr = PetscMemcpy(olocal_primal_ref_mult,pcbddc->local_primal_ref_mult,olocal_primal_size_cc*sizeof(PetscInt));CHKERRQ(ierr); 35480e6343abSStefano Zampini ierr = PetscFree2(pcbddc->local_primal_ref_node,pcbddc->local_primal_ref_mult);CHKERRQ(ierr); 3549088faed8SStefano Zampini ierr = PetscFree(pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 3550cf5a6209SStefano Zampini 3551cf5a6209SStefano Zampini if (!pcbddc->adaptive_selection) { 35529162d606SStefano Zampini IS ISForVertices,*ISForFaces,*ISForEdges; 3553cf5a6209SStefano Zampini MatNullSpace nearnullsp; 3554cf5a6209SStefano Zampini const Vec *nearnullvecs; 3555cf5a6209SStefano Zampini Vec *localnearnullsp; 3556cf5a6209SStefano Zampini PetscScalar *array; 3557cf5a6209SStefano Zampini PetscInt n_ISForFaces,n_ISForEdges,nnsp_size; 3558cf5a6209SStefano Zampini PetscBool nnsp_has_cnst; 3559674ae819SStefano Zampini /* LAPACK working arrays for SVD or POD */ 3560b3d85658SStefano Zampini PetscBool skip_lapack,boolforchange; 3561674ae819SStefano Zampini PetscScalar *work; 3562674ae819SStefano Zampini PetscReal *singular_vals; 3563674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX) 3564674ae819SStefano Zampini PetscReal *rwork; 3565674ae819SStefano Zampini #endif 3566674ae819SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 3567674ae819SStefano Zampini PetscScalar *temp_basis,*correlation_mat; 3568674ae819SStefano Zampini #else 3569964fefecSStefano Zampini PetscBLASInt dummy_int=1; 3570964fefecSStefano Zampini PetscScalar dummy_scalar=1.; 3571674ae819SStefano Zampini #endif 3572674ae819SStefano Zampini 3573674ae819SStefano Zampini /* Get index sets for faces, edges and vertices from graph */ 3574d06fc5fdSStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,&n_ISForFaces,&ISForFaces,&n_ISForEdges,&ISForEdges,&ISForVertices);CHKERRQ(ierr); 3575e4d548c7SStefano Zampini /* print some info */ 35761f4df5f7SStefano Zampini if (pcbddc->dbg_flag && !pcbddc->sub_schurs) { 3577e4d548c7SStefano Zampini PetscInt nv; 3578e4d548c7SStefano Zampini 3579e4d548c7SStefano Zampini ierr = PCBDDCGraphASCIIView(pcbddc->mat_graph,pcbddc->dbg_flag,pcbddc->dbg_viewer);CHKERRQ(ierr); 3580e4d548c7SStefano Zampini ierr = ISGetSize(ISForVertices,&nv);CHKERRQ(ierr); 3581e4d548c7SStefano Zampini ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 3582e4d548c7SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"--------------------------------------------------------------\n");CHKERRQ(ierr); 3583e4d548c7SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate vertices (%d)\n",PetscGlobalRank,nv,pcbddc->use_vertices);CHKERRQ(ierr); 3584e4d548c7SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate edges (%d)\n",PetscGlobalRank,n_ISForEdges,pcbddc->use_edges);CHKERRQ(ierr); 3585e4d548c7SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate faces (%d)\n",PetscGlobalRank,n_ISForFaces,pcbddc->use_faces);CHKERRQ(ierr); 3586e4d548c7SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 3587e4d548c7SStefano Zampini ierr = PetscViewerASCIIPopSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 3588e4d548c7SStefano Zampini } 3589e4d548c7SStefano Zampini 3590d06fc5fdSStefano Zampini /* free unneeded index sets */ 3591d06fc5fdSStefano Zampini if (!pcbddc->use_vertices) { 3592d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForVertices);CHKERRQ(ierr); 3593674ae819SStefano Zampini } 3594d06fc5fdSStefano Zampini if (!pcbddc->use_edges) { 3595d06fc5fdSStefano Zampini for (i=0;i<n_ISForEdges;i++) { 3596d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForEdges[i]);CHKERRQ(ierr); 3597d06fc5fdSStefano Zampini } 3598d06fc5fdSStefano Zampini ierr = PetscFree(ISForEdges);CHKERRQ(ierr); 3599d06fc5fdSStefano Zampini n_ISForEdges = 0; 3600d06fc5fdSStefano Zampini } 3601d06fc5fdSStefano Zampini if (!pcbddc->use_faces) { 3602d06fc5fdSStefano Zampini for (i=0;i<n_ISForFaces;i++) { 3603d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForFaces[i]);CHKERRQ(ierr); 3604d06fc5fdSStefano Zampini } 3605d06fc5fdSStefano Zampini ierr = PetscFree(ISForFaces);CHKERRQ(ierr); 3606d06fc5fdSStefano Zampini n_ISForFaces = 0; 3607d06fc5fdSStefano Zampini } 360870022509SStefano Zampini 3609674ae819SStefano Zampini /* check if near null space is attached to global mat */ 3610674ae819SStefano Zampini ierr = MatGetNearNullSpace(pc->pmat,&nearnullsp);CHKERRQ(ierr); 3611674ae819SStefano Zampini if (nearnullsp) { 3612674ae819SStefano Zampini ierr = MatNullSpaceGetVecs(nearnullsp,&nnsp_has_cnst,&nnsp_size,&nearnullvecs);CHKERRQ(ierr); 3613f4ddd8eeSStefano Zampini /* remove any stored info */ 3614f4ddd8eeSStefano Zampini ierr = MatNullSpaceDestroy(&pcbddc->onearnullspace);CHKERRQ(ierr); 3615f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->onearnullvecs_state);CHKERRQ(ierr); 3616f4ddd8eeSStefano Zampini /* store information for BDDC solver reuse */ 3617f4ddd8eeSStefano Zampini ierr = PetscObjectReference((PetscObject)nearnullsp);CHKERRQ(ierr); 3618f4ddd8eeSStefano Zampini pcbddc->onearnullspace = nearnullsp; 3619473ba861SJed Brown ierr = PetscMalloc1(nnsp_size,&pcbddc->onearnullvecs_state);CHKERRQ(ierr); 3620f4ddd8eeSStefano Zampini for (i=0;i<nnsp_size;i++) { 3621f4ddd8eeSStefano Zampini ierr = PetscObjectStateGet((PetscObject)nearnullvecs[i],&pcbddc->onearnullvecs_state[i]);CHKERRQ(ierr); 3622f4ddd8eeSStefano Zampini } 3623984c4197SStefano Zampini } else { /* if near null space is not provided BDDC uses constants by default */ 3624984c4197SStefano Zampini nnsp_size = 0; 3625674ae819SStefano Zampini nnsp_has_cnst = PETSC_TRUE; 3626674ae819SStefano Zampini } 3627984c4197SStefano Zampini /* get max number of constraints on a single cc */ 3628984c4197SStefano Zampini max_constraints = nnsp_size; 3629984c4197SStefano Zampini if (nnsp_has_cnst) max_constraints++; 3630984c4197SStefano Zampini 3631674ae819SStefano Zampini /* 3632674ae819SStefano Zampini Evaluate maximum storage size needed by the procedure 36339162d606SStefano Zampini - Indices for connected component i stored at "constraints_idxs + constraints_idxs_ptr[i]" 36349162d606SStefano Zampini - Values for constraints on connected component i stored at "constraints_data + constraints_data_ptr[i]" 36359162d606SStefano Zampini There can be multiple constraints per connected component 3636674ae819SStefano Zampini */ 3637674ae819SStefano Zampini n_vertices = 0; 3638674ae819SStefano Zampini if (ISForVertices) { 3639674ae819SStefano Zampini ierr = ISGetSize(ISForVertices,&n_vertices);CHKERRQ(ierr); 3640674ae819SStefano Zampini } 36419162d606SStefano Zampini ncc = n_vertices+n_ISForFaces+n_ISForEdges; 36429162d606SStefano Zampini ierr = PetscMalloc3(ncc+1,&constraints_idxs_ptr,ncc+1,&constraints_data_ptr,ncc,&constraints_n);CHKERRQ(ierr); 36439162d606SStefano Zampini 36449162d606SStefano Zampini total_counts = n_ISForFaces+n_ISForEdges; 36459162d606SStefano Zampini total_counts *= max_constraints; 3646674ae819SStefano Zampini total_counts += n_vertices; 36474641a718SStefano Zampini ierr = PetscBTCreate(total_counts,&change_basis);CHKERRQ(ierr); 36489162d606SStefano Zampini 3649674ae819SStefano Zampini total_counts = 0; 3650674ae819SStefano Zampini max_size_of_constraint = 0; 3651674ae819SStefano Zampini for (i=0;i<n_ISForEdges+n_ISForFaces;i++) { 36529162d606SStefano Zampini IS used_is; 3653674ae819SStefano Zampini if (i<n_ISForEdges) { 36549162d606SStefano Zampini used_is = ISForEdges[i]; 3655674ae819SStefano Zampini } else { 36569162d606SStefano Zampini used_is = ISForFaces[i-n_ISForEdges]; 3657674ae819SStefano Zampini } 36589162d606SStefano Zampini ierr = ISGetSize(used_is,&j);CHKERRQ(ierr); 3659674ae819SStefano Zampini total_counts += j; 3660674ae819SStefano Zampini max_size_of_constraint = PetscMax(j,max_size_of_constraint); 3661674ae819SStefano Zampini } 36629162d606SStefano 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); 36639162d606SStefano Zampini 3664984c4197SStefano Zampini /* get local part of global near null space vectors */ 3665785e854fSJed Brown ierr = PetscMalloc1(nnsp_size,&localnearnullsp);CHKERRQ(ierr); 3666984c4197SStefano Zampini for (k=0;k<nnsp_size;k++) { 3667984c4197SStefano Zampini ierr = VecDuplicate(pcis->vec1_N,&localnearnullsp[k]);CHKERRQ(ierr); 3668e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,nearnullvecs[k],localnearnullsp[k],INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 3669e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,nearnullvecs[k],localnearnullsp[k],INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 3670984c4197SStefano Zampini } 3671674ae819SStefano Zampini 3672242a89d7SStefano Zampini /* whether or not to skip lapack calls */ 3673242a89d7SStefano Zampini skip_lapack = PETSC_TRUE; 3674a773dcb8SStefano Zampini if (n_ISForFaces+n_ISForEdges && max_constraints > 1 && !pcbddc->use_nnsp_true) skip_lapack = PETSC_FALSE; 3675242a89d7SStefano Zampini 3676984c4197SStefano Zampini /* First we issue queries to allocate optimal workspace for LAPACKgesvd (or LAPACKsyev if SVD is missing) */ 3677a773dcb8SStefano Zampini if (!skip_lapack) { 3678674ae819SStefano Zampini PetscScalar temp_work; 3679911cabfeSStefano Zampini 3680674ae819SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 3681984c4197SStefano Zampini /* Proper Orthogonal Decomposition (POD) using the snapshot method */ 3682785e854fSJed Brown ierr = PetscMalloc1(max_constraints*max_constraints,&correlation_mat);CHKERRQ(ierr); 3683785e854fSJed Brown ierr = PetscMalloc1(max_constraints,&singular_vals);CHKERRQ(ierr); 3684785e854fSJed Brown ierr = PetscMalloc1(max_size_of_constraint*max_constraints,&temp_basis);CHKERRQ(ierr); 3685674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX) 3686785e854fSJed Brown ierr = PetscMalloc1(3*max_constraints,&rwork);CHKERRQ(ierr); 3687674ae819SStefano Zampini #endif 3688674ae819SStefano Zampini /* now we evaluate the optimal workspace using query with lwork=-1 */ 3689c8244a33SStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_N);CHKERRQ(ierr); 3690c8244a33SStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_LDA);CHKERRQ(ierr); 3691674ae819SStefano Zampini lwork = -1; 3692674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 3693674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 3694c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,&temp_work,&lwork,&lierr)); 3695674ae819SStefano Zampini #else 3696c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,&temp_work,&lwork,rwork,&lierr)); 3697674ae819SStefano Zampini #endif 3698674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 3699984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to SYEV Lapack routine %d",(int)lierr); 3700674ae819SStefano Zampini #else /* on missing GESVD */ 3701674ae819SStefano Zampini /* SVD */ 3702674ae819SStefano Zampini PetscInt max_n,min_n; 3703674ae819SStefano Zampini max_n = max_size_of_constraint; 3704984c4197SStefano Zampini min_n = max_constraints; 3705984c4197SStefano Zampini if (max_size_of_constraint < max_constraints) { 3706674ae819SStefano Zampini min_n = max_size_of_constraint; 3707984c4197SStefano Zampini max_n = max_constraints; 3708674ae819SStefano Zampini } 3709785e854fSJed Brown ierr = PetscMalloc1(min_n,&singular_vals);CHKERRQ(ierr); 3710674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX) 3711785e854fSJed Brown ierr = PetscMalloc1(5*min_n,&rwork);CHKERRQ(ierr); 3712674ae819SStefano Zampini #endif 3713674ae819SStefano Zampini /* now we evaluate the optimal workspace using query with lwork=-1 */ 3714674ae819SStefano Zampini lwork = -1; 3715e310c8b4SStefano Zampini ierr = PetscBLASIntCast(max_n,&Blas_M);CHKERRQ(ierr); 3716e310c8b4SStefano Zampini ierr = PetscBLASIntCast(min_n,&Blas_N);CHKERRQ(ierr); 3717b7d8b9f8SStefano Zampini ierr = PetscBLASIntCast(max_n,&Blas_LDA);CHKERRQ(ierr); 3718674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 3719674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 37209162d606SStefano 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)); 3721674ae819SStefano Zampini #else 37229162d606SStefano 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)); 3723674ae819SStefano Zampini #endif 3724674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 3725984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to GESVD Lapack routine %d",(int)lierr); 3726984c4197SStefano Zampini #endif /* on missing GESVD */ 3727674ae819SStefano Zampini /* Allocate optimal workspace */ 3728674ae819SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(temp_work),&lwork);CHKERRQ(ierr); 3729854ce69bSBarry Smith ierr = PetscMalloc1(lwork,&work);CHKERRQ(ierr); 3730674ae819SStefano Zampini } 3731674ae819SStefano Zampini /* Now we can loop on constraining sets */ 3732674ae819SStefano Zampini total_counts = 0; 37339162d606SStefano Zampini constraints_idxs_ptr[0] = 0; 37349162d606SStefano Zampini constraints_data_ptr[0] = 0; 3735674ae819SStefano Zampini /* vertices */ 37369162d606SStefano Zampini if (n_vertices) { 3737674ae819SStefano Zampini ierr = ISGetIndices(ISForVertices,(const PetscInt**)&is_indices);CHKERRQ(ierr); 37389162d606SStefano Zampini ierr = PetscMemcpy(constraints_idxs,is_indices,n_vertices*sizeof(PetscInt));CHKERRQ(ierr); 3739674ae819SStefano Zampini for (i=0;i<n_vertices;i++) { 37409162d606SStefano Zampini constraints_n[total_counts] = 1; 37419162d606SStefano Zampini constraints_data[total_counts] = 1.0; 37429162d606SStefano Zampini constraints_idxs_ptr[total_counts+1] = constraints_idxs_ptr[total_counts]+1; 37439162d606SStefano Zampini constraints_data_ptr[total_counts+1] = constraints_data_ptr[total_counts]+1; 3744674ae819SStefano Zampini total_counts++; 3745674ae819SStefano Zampini } 3746674ae819SStefano Zampini ierr = ISRestoreIndices(ISForVertices,(const PetscInt**)&is_indices);CHKERRQ(ierr); 3747674ae819SStefano Zampini n_vertices = total_counts; 3748674ae819SStefano Zampini } 3749984c4197SStefano Zampini 3750674ae819SStefano Zampini /* edges and faces */ 37519162d606SStefano Zampini total_counts_cc = total_counts; 3752911cabfeSStefano Zampini for (ncc=0;ncc<n_ISForEdges+n_ISForFaces;ncc++) { 37539162d606SStefano Zampini IS used_is; 37549162d606SStefano Zampini PetscBool idxs_copied = PETSC_FALSE; 37559162d606SStefano Zampini 3756911cabfeSStefano Zampini if (ncc<n_ISForEdges) { 37579162d606SStefano Zampini used_is = ISForEdges[ncc]; 3758984c4197SStefano Zampini boolforchange = pcbddc->use_change_of_basis; /* change or not the basis on the edge */ 3759674ae819SStefano Zampini } else { 37609162d606SStefano Zampini used_is = ISForFaces[ncc-n_ISForEdges]; 3761984c4197SStefano Zampini boolforchange = (PetscBool)(pcbddc->use_change_of_basis && pcbddc->use_change_on_faces); /* change or not the basis on the face */ 3762674ae819SStefano Zampini } 3763674ae819SStefano Zampini temp_constraints = 0; /* zero the number of constraints I have on this conn comp */ 37649162d606SStefano Zampini 37659162d606SStefano Zampini ierr = ISGetSize(used_is,&size_of_constraint);CHKERRQ(ierr); 37669162d606SStefano Zampini ierr = ISGetIndices(used_is,(const PetscInt**)&is_indices);CHKERRQ(ierr); 3767984c4197SStefano Zampini /* change of basis should not be performed on local periodic nodes */ 3768984c4197SStefano Zampini if (pcbddc->mat_graph->mirrors && pcbddc->mat_graph->mirrors[is_indices[0]]) boolforchange = PETSC_FALSE; 3769674ae819SStefano Zampini if (nnsp_has_cnst) { 37705b08dc53SStefano Zampini PetscScalar quad_value; 37719162d606SStefano Zampini 37729162d606SStefano Zampini ierr = PetscMemcpy(constraints_idxs + constraints_idxs_ptr[total_counts_cc],is_indices,size_of_constraint*sizeof(PetscInt));CHKERRQ(ierr); 37739162d606SStefano Zampini idxs_copied = PETSC_TRUE; 37749162d606SStefano Zampini 3775a773dcb8SStefano Zampini if (!pcbddc->use_nnsp_true) { 3776674ae819SStefano Zampini quad_value = (PetscScalar)(1.0/PetscSqrtReal((PetscReal)size_of_constraint)); 3777a773dcb8SStefano Zampini } else { 3778a773dcb8SStefano Zampini quad_value = 1.0; 3779a773dcb8SStefano Zampini } 3780674ae819SStefano Zampini for (j=0;j<size_of_constraint;j++) { 37819162d606SStefano Zampini constraints_data[constraints_data_ptr[total_counts_cc]+j] = quad_value; 3782674ae819SStefano Zampini } 37839162d606SStefano Zampini temp_constraints++; 3784674ae819SStefano Zampini total_counts++; 3785674ae819SStefano Zampini } 3786674ae819SStefano Zampini for (k=0;k<nnsp_size;k++) { 3787984c4197SStefano Zampini PetscReal real_value; 37889162d606SStefano Zampini PetscScalar *ptr_to_data; 37899162d606SStefano Zampini 3790984c4197SStefano Zampini ierr = VecGetArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 37919162d606SStefano Zampini ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]+temp_constraints*size_of_constraint]; 3792674ae819SStefano Zampini for (j=0;j<size_of_constraint;j++) { 37939162d606SStefano Zampini ptr_to_data[j] = array[is_indices[j]]; 3794674ae819SStefano Zampini } 3795984c4197SStefano Zampini ierr = VecRestoreArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 3796984c4197SStefano Zampini /* check if array is null on the connected component */ 3797e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 37989162d606SStefano Zampini PetscStackCallBLAS("BLASasum",real_value = BLASasum_(&Blas_N,ptr_to_data,&Blas_one)); 37995b08dc53SStefano Zampini if (real_value > 0.0) { /* keep indices and values */ 3800674ae819SStefano Zampini temp_constraints++; 3801674ae819SStefano Zampini total_counts++; 38029162d606SStefano Zampini if (!idxs_copied) { 38039162d606SStefano Zampini ierr = PetscMemcpy(constraints_idxs + constraints_idxs_ptr[total_counts_cc],is_indices,size_of_constraint*sizeof(PetscInt));CHKERRQ(ierr); 38049162d606SStefano Zampini idxs_copied = PETSC_TRUE; 3805674ae819SStefano Zampini } 3806674ae819SStefano Zampini } 38079162d606SStefano Zampini } 38089162d606SStefano Zampini ierr = ISRestoreIndices(used_is,(const PetscInt**)&is_indices);CHKERRQ(ierr); 380945a1bb75SStefano Zampini valid_constraints = temp_constraints; 3810eb97c9d2SStefano Zampini if (!pcbddc->use_nnsp_true && temp_constraints) { 3811a773dcb8SStefano Zampini if (temp_constraints == 1) { /* just normalize the constraint */ 38129162d606SStefano Zampini PetscScalar norm,*ptr_to_data; 38139162d606SStefano Zampini 38149162d606SStefano Zampini ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]]; 3815a773dcb8SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 38169162d606SStefano Zampini PetscStackCallBLAS("BLASdot",norm = BLASdot_(&Blas_N,ptr_to_data,&Blas_one,ptr_to_data,&Blas_one)); 3817a773dcb8SStefano Zampini norm = 1.0/PetscSqrtReal(PetscRealPart(norm)); 38189162d606SStefano Zampini PetscStackCallBLAS("BLASscal",BLASscal_(&Blas_N,&norm,ptr_to_data,&Blas_one)); 3819a773dcb8SStefano Zampini } else { /* perform SVD */ 3820984c4197SStefano Zampini PetscReal tol = 1.0e-8; /* tolerance for retaining eigenmodes */ 38219162d606SStefano Zampini PetscScalar *ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]]; 3822674ae819SStefano Zampini 3823674ae819SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 3824984c4197SStefano Zampini /* SVD: Y = U*S*V^H -> U (eigenvectors of Y*Y^H) = Y*V*(S)^\dag 3825984c4197SStefano Zampini POD: Y^H*Y = V*D*V^H, D = S^H*S -> U = Y*V*D^(-1/2) 3826984c4197SStefano Zampini -> When PETSC_USE_COMPLEX and PETSC_MISSING_LAPACK_GESVD are defined 3827984c4197SStefano Zampini the constraints basis will differ (by a complex factor with absolute value equal to 1) 3828984c4197SStefano Zampini from that computed using LAPACKgesvd 3829984c4197SStefano Zampini -> This is due to a different computation of eigenvectors in LAPACKheev 3830984c4197SStefano Zampini -> The quality of the POD-computed basis will be the same */ 3831984c4197SStefano Zampini ierr = PetscMemzero(correlation_mat,temp_constraints*temp_constraints*sizeof(PetscScalar));CHKERRQ(ierr); 3832674ae819SStefano Zampini /* Store upper triangular part of correlation matrix */ 3833e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 3834984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 3835674ae819SStefano Zampini for (j=0;j<temp_constraints;j++) { 3836674ae819SStefano Zampini for (k=0;k<j+1;k++) { 38379162d606SStefano 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)); 3838674ae819SStefano Zampini } 3839674ae819SStefano Zampini } 3840e310c8b4SStefano Zampini /* compute eigenvalues and eigenvectors of correlation matrix */ 3841e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 3842e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_LDA);CHKERRQ(ierr); 3843674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 3844c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,work,&lwork,&lierr)); 3845674ae819SStefano Zampini #else 3846c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,work,&lwork,rwork,&lierr)); 3847674ae819SStefano Zampini #endif 3848674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 3849984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYEV Lapack routine %d",(int)lierr); 3850984c4197SStefano Zampini /* retain eigenvalues greater than tol: note that LAPACKsyev gives eigs in ascending order */ 3851674ae819SStefano Zampini j = 0; 3852984c4197SStefano Zampini while (j < temp_constraints && singular_vals[j] < tol) j++; 3853674ae819SStefano Zampini total_counts = total_counts-j; 385445a1bb75SStefano Zampini valid_constraints = temp_constraints-j; 3855e310c8b4SStefano Zampini /* scale and copy POD basis into used quadrature memory */ 3856c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 3857c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 3858c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_K);CHKERRQ(ierr); 3859c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 3860c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_LDB);CHKERRQ(ierr); 3861c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDC);CHKERRQ(ierr); 3862674ae819SStefano Zampini if (j<temp_constraints) { 3863984c4197SStefano Zampini PetscInt ii; 3864984c4197SStefano Zampini for (k=j;k<temp_constraints;k++) singular_vals[k] = 1.0/PetscSqrtReal(singular_vals[k]); 3865674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 38669162d606SStefano 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)); 3867674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 3868984c4197SStefano Zampini for (k=0;k<temp_constraints-j;k++) { 3869674ae819SStefano Zampini for (ii=0;ii<size_of_constraint;ii++) { 38709162d606SStefano 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]; 3871674ae819SStefano Zampini } 3872674ae819SStefano Zampini } 3873674ae819SStefano Zampini } 3874674ae819SStefano Zampini #else /* on missing GESVD */ 3875e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 3876e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 3877b7d8b9f8SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 3878674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 3879674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 38809162d606SStefano 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)); 3881674ae819SStefano Zampini #else 38829162d606SStefano 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)); 3883674ae819SStefano Zampini #endif 3884984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GESVD Lapack routine %d",(int)lierr); 3885674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 3886984c4197SStefano Zampini /* retain eigenvalues greater than tol: note that LAPACKgesvd gives eigs in descending order */ 3887e310c8b4SStefano Zampini k = temp_constraints; 3888e310c8b4SStefano Zampini if (k > size_of_constraint) k = size_of_constraint; 3889674ae819SStefano Zampini j = 0; 3890e310c8b4SStefano Zampini while (j < k && singular_vals[k-j-1] < tol) j++; 389145a1bb75SStefano Zampini valid_constraints = k-j; 3892911cabfeSStefano Zampini total_counts = total_counts-temp_constraints+valid_constraints; 3893984c4197SStefano Zampini #endif /* on missing GESVD */ 3894674ae819SStefano Zampini } 3895a773dcb8SStefano Zampini } 38969162d606SStefano Zampini /* update pointers information */ 38979162d606SStefano Zampini if (valid_constraints) { 38989162d606SStefano Zampini constraints_n[total_counts_cc] = valid_constraints; 38999162d606SStefano Zampini constraints_idxs_ptr[total_counts_cc+1] = constraints_idxs_ptr[total_counts_cc]+size_of_constraint; 39009162d606SStefano Zampini constraints_data_ptr[total_counts_cc+1] = constraints_data_ptr[total_counts_cc]+size_of_constraint*valid_constraints; 39019162d606SStefano Zampini /* set change_of_basis flag */ 390245a1bb75SStefano Zampini if (boolforchange) { 3903b3d85658SStefano Zampini PetscBTSet(change_basis,total_counts_cc); 39049162d606SStefano Zampini } 3905b3d85658SStefano Zampini total_counts_cc++; 390645a1bb75SStefano Zampini } 390745a1bb75SStefano Zampini } 3908984c4197SStefano Zampini /* free workspace */ 39098f1c130eSStefano Zampini if (!skip_lapack) { 3910984c4197SStefano Zampini ierr = PetscFree(work);CHKERRQ(ierr); 3911984c4197SStefano Zampini #if defined(PETSC_USE_COMPLEX) 3912984c4197SStefano Zampini ierr = PetscFree(rwork);CHKERRQ(ierr); 3913984c4197SStefano Zampini #endif 3914984c4197SStefano Zampini ierr = PetscFree(singular_vals);CHKERRQ(ierr); 3915984c4197SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 3916984c4197SStefano Zampini ierr = PetscFree(correlation_mat);CHKERRQ(ierr); 3917984c4197SStefano Zampini ierr = PetscFree(temp_basis);CHKERRQ(ierr); 3918984c4197SStefano Zampini #endif 3919984c4197SStefano Zampini } 3920984c4197SStefano Zampini for (k=0;k<nnsp_size;k++) { 3921984c4197SStefano Zampini ierr = VecDestroy(&localnearnullsp[k]);CHKERRQ(ierr); 3922984c4197SStefano Zampini } 3923984c4197SStefano Zampini ierr = PetscFree(localnearnullsp);CHKERRQ(ierr); 3924cf5a6209SStefano Zampini /* free index sets of faces, edges and vertices */ 3925cf5a6209SStefano Zampini for (i=0;i<n_ISForFaces;i++) { 3926cf5a6209SStefano Zampini ierr = ISDestroy(&ISForFaces[i]);CHKERRQ(ierr); 3927cf5a6209SStefano Zampini } 3928cf5a6209SStefano Zampini if (n_ISForFaces) { 3929cf5a6209SStefano Zampini ierr = PetscFree(ISForFaces);CHKERRQ(ierr); 3930cf5a6209SStefano Zampini } 3931cf5a6209SStefano Zampini for (i=0;i<n_ISForEdges;i++) { 3932cf5a6209SStefano Zampini ierr = ISDestroy(&ISForEdges[i]);CHKERRQ(ierr); 3933cf5a6209SStefano Zampini } 3934cf5a6209SStefano Zampini if (n_ISForEdges) { 3935cf5a6209SStefano Zampini ierr = PetscFree(ISForEdges);CHKERRQ(ierr); 3936cf5a6209SStefano Zampini } 3937cf5a6209SStefano Zampini ierr = ISDestroy(&ISForVertices);CHKERRQ(ierr); 393808122e43SStefano Zampini } else { 393908122e43SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 3940984c4197SStefano Zampini 394108122e43SStefano Zampini total_counts = 0; 394208122e43SStefano Zampini n_vertices = 0; 3943d62866d3SStefano Zampini if (sub_schurs->is_vertices && pcbddc->use_vertices) { 3944d62866d3SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_vertices,&n_vertices);CHKERRQ(ierr); 394508122e43SStefano Zampini } 394608122e43SStefano Zampini max_constraints = 0; 39479162d606SStefano Zampini total_counts_cc = 0; 394808122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs+n_vertices;i++) { 394908122e43SStefano Zampini total_counts += pcbddc->adaptive_constraints_n[i]; 39509162d606SStefano Zampini if (pcbddc->adaptive_constraints_n[i]) total_counts_cc++; 395108122e43SStefano Zampini max_constraints = PetscMax(max_constraints,pcbddc->adaptive_constraints_n[i]); 395208122e43SStefano Zampini } 39539162d606SStefano Zampini constraints_idxs_ptr = pcbddc->adaptive_constraints_idxs_ptr; 39549162d606SStefano Zampini constraints_data_ptr = pcbddc->adaptive_constraints_data_ptr; 39559162d606SStefano Zampini constraints_idxs = pcbddc->adaptive_constraints_idxs; 39569162d606SStefano Zampini constraints_data = pcbddc->adaptive_constraints_data; 395774d5cdf7SStefano Zampini /* constraints_n differs from pcbddc->adaptive_constraints_n */ 39589162d606SStefano Zampini ierr = PetscMalloc1(total_counts_cc,&constraints_n);CHKERRQ(ierr); 39599162d606SStefano Zampini total_counts_cc = 0; 39609162d606SStefano Zampini for (i=0;i<sub_schurs->n_subs+n_vertices;i++) { 39619162d606SStefano Zampini if (pcbddc->adaptive_constraints_n[i]) { 39629162d606SStefano Zampini constraints_n[total_counts_cc++] = pcbddc->adaptive_constraints_n[i]; 396308122e43SStefano Zampini } 396408122e43SStefano Zampini } 39659162d606SStefano Zampini #if 0 39669162d606SStefano Zampini printf("Found %d totals (%d)\n",total_counts_cc,total_counts); 39679162d606SStefano Zampini for (i=0;i<total_counts_cc;i++) { 39689162d606SStefano Zampini printf("const %d, start %d",i,constraints_idxs_ptr[i]); 39699162d606SStefano Zampini printf(" end %d:\n",constraints_idxs_ptr[i+1]); 39709162d606SStefano Zampini for (j=constraints_idxs_ptr[i];j<constraints_idxs_ptr[i+1];j++) { 39719162d606SStefano Zampini printf(" %d",constraints_idxs[j]); 39729162d606SStefano Zampini } 39739162d606SStefano Zampini printf("\n"); 39749162d606SStefano Zampini printf("number of cc: %d\n",constraints_n[i]); 39759162d606SStefano Zampini } 39761b968477SStefano Zampini for (i=0;i<n_vertices;i++) { 39778bec7fa6SStefano Zampini PetscPrintf(PETSC_COMM_SELF,"[%d] vertex %d, n %d\n",PetscGlobalRank,i,pcbddc->adaptive_constraints_n[i]); 39781b968477SStefano Zampini } 39791b968477SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 39808bec7fa6SStefano 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]); 39811b968477SStefano Zampini } 398208122e43SStefano Zampini #endif 398308122e43SStefano Zampini 39848bec7fa6SStefano Zampini max_size_of_constraint = 0; 39859162d606SStefano 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]); 39869162d606SStefano Zampini ierr = PetscMalloc1(constraints_idxs_ptr[total_counts_cc],&constraints_idxs_B);CHKERRQ(ierr); 398708122e43SStefano Zampini /* Change of basis */ 3988b3d85658SStefano Zampini ierr = PetscBTCreate(total_counts_cc,&change_basis);CHKERRQ(ierr); 398908122e43SStefano Zampini if (pcbddc->use_change_of_basis) { 399008122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 399108122e43SStefano Zampini if (PetscBTLookup(sub_schurs->is_edge,i) || pcbddc->use_change_on_faces) { 3992b3d85658SStefano Zampini ierr = PetscBTSet(change_basis,i+n_vertices);CHKERRQ(ierr); 399308122e43SStefano Zampini } 399408122e43SStefano Zampini } 399508122e43SStefano Zampini } 399608122e43SStefano Zampini } 3997984c4197SStefano Zampini pcbddc->local_primal_size = total_counts; 39984f1b2e48SStefano Zampini ierr = PetscMalloc1(pcbddc->local_primal_size+pcbddc->benign_n,&pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 399908122e43SStefano Zampini 40009162d606SStefano Zampini /* map constraints_idxs in boundary numbering */ 40019162d606SStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,constraints_idxs_ptr[total_counts_cc],constraints_idxs,&i,constraints_idxs_B);CHKERRQ(ierr); 40026c4ed002SBarry Smith if (i != constraints_idxs_ptr[total_counts_cc]) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_SUP,"Error in boundary numbering for constraints indices %D != %D\n",constraints_idxs_ptr[total_counts_cc],i); 4003674ae819SStefano Zampini 4004674ae819SStefano Zampini /* Create constraint matrix */ 4005674ae819SStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 400616f15bc4SStefano Zampini ierr = MatSetType(pcbddc->ConstraintMatrix,MATAIJ);CHKERRQ(ierr); 4007984c4197SStefano Zampini ierr = MatSetSizes(pcbddc->ConstraintMatrix,pcbddc->local_primal_size,pcis->n,pcbddc->local_primal_size,pcis->n);CHKERRQ(ierr); 4008984c4197SStefano Zampini 4009984c4197SStefano Zampini /* find primal_dofs: subdomain corners plus dofs selected as primal after change of basis */ 4010a717540cSStefano Zampini /* determine if a QR strategy is needed for change of basis */ 4011a717540cSStefano Zampini qr_needed = PETSC_FALSE; 401274d5cdf7SStefano Zampini ierr = PetscBTCreate(total_counts_cc,&qr_needed_idx);CHKERRQ(ierr); 4013984c4197SStefano Zampini total_primal_vertices=0; 4014b3d85658SStefano Zampini pcbddc->local_primal_size_cc = 0; 40159162d606SStefano Zampini for (i=0;i<total_counts_cc;i++) { 40169162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 401772b8c272SStefano Zampini if (size_of_constraint == 1 && pcbddc->mat_graph->custom_minimal_size) { 40189162d606SStefano Zampini pcbddc->primal_indices_local_idxs[total_primal_vertices++] = constraints_idxs[constraints_idxs_ptr[i]]; 4019b3d85658SStefano Zampini pcbddc->local_primal_size_cc += 1; 402064efe560SStefano Zampini } else if (PetscBTLookup(change_basis,i)) { 40219162d606SStefano Zampini for (k=0;k<constraints_n[i];k++) { 40229162d606SStefano Zampini pcbddc->primal_indices_local_idxs[total_primal_vertices++] = constraints_idxs[constraints_idxs_ptr[i]+k]; 4023a717540cSStefano Zampini } 4024b3d85658SStefano Zampini pcbddc->local_primal_size_cc += constraints_n[i]; 402591af6908SStefano Zampini if (constraints_n[i] > 1 || pcbddc->use_qr_single) { 4026a717540cSStefano Zampini PetscBTSet(qr_needed_idx,i); 4027a717540cSStefano Zampini qr_needed = PETSC_TRUE; 4028a717540cSStefano Zampini } 4029fa434743SStefano Zampini } else { 4030b3d85658SStefano Zampini pcbddc->local_primal_size_cc += 1; 4031fa434743SStefano Zampini } 4032a717540cSStefano Zampini } 4033b371cd4fSStefano Zampini /* note that the local variable n_vertices used below stores the number of pointwise constraints */ 4034b371cd4fSStefano Zampini pcbddc->n_vertices = total_primal_vertices; 4035674ae819SStefano Zampini /* permute indices in order to have a sorted set of vertices */ 403670022509SStefano Zampini ierr = PetscSortInt(total_primal_vertices,pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 4037b3d85658SStefano Zampini 40384f1b2e48SStefano 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); 40390e6343abSStefano Zampini ierr = PetscMemcpy(pcbddc->local_primal_ref_node,pcbddc->primal_indices_local_idxs,total_primal_vertices*sizeof(PetscInt));CHKERRQ(ierr); 40400e6343abSStefano Zampini for (i=0;i<total_primal_vertices;i++) pcbddc->local_primal_ref_mult[i] = 1; 4041984c4197SStefano Zampini 4042984c4197SStefano Zampini /* nonzero structure of constraint matrix */ 404374d5cdf7SStefano Zampini /* and get reference dof for local constraints */ 4044785e854fSJed Brown ierr = PetscMalloc1(pcbddc->local_primal_size,&nnz);CHKERRQ(ierr); 4045984c4197SStefano Zampini for (i=0;i<total_primal_vertices;i++) nnz[i] = 1; 404674d5cdf7SStefano Zampini 4047984c4197SStefano Zampini j = total_primal_vertices; 404874d5cdf7SStefano Zampini total_counts = total_primal_vertices; 4049b3d85658SStefano Zampini cum = total_primal_vertices; 40509162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 40514641a718SStefano Zampini if (!PetscBTLookup(change_basis,i)) { 4052b3d85658SStefano Zampini pcbddc->local_primal_ref_node[cum] = constraints_idxs[constraints_idxs_ptr[i]]; 4053b3d85658SStefano Zampini pcbddc->local_primal_ref_mult[cum] = constraints_n[i]; 4054b3d85658SStefano Zampini cum++; 40559162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 405674d5cdf7SStefano Zampini for (k=0;k<constraints_n[i];k++) { 405774d5cdf7SStefano Zampini pcbddc->primal_indices_local_idxs[total_counts++] = constraints_idxs[constraints_idxs_ptr[i]+k]; 405874d5cdf7SStefano Zampini nnz[j+k] = size_of_constraint; 405974d5cdf7SStefano Zampini } 40609162d606SStefano Zampini j += constraints_n[i]; 4061674ae819SStefano Zampini } 4062674ae819SStefano Zampini } 4063674ae819SStefano Zampini ierr = MatSeqAIJSetPreallocation(pcbddc->ConstraintMatrix,0,nnz);CHKERRQ(ierr); 4064674ae819SStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 4065088faed8SStefano Zampini 4066674ae819SStefano Zampini /* set values in constraint matrix */ 4067984c4197SStefano Zampini for (i=0;i<total_primal_vertices;i++) { 40680e6343abSStefano Zampini ierr = MatSetValue(pcbddc->ConstraintMatrix,i,pcbddc->local_primal_ref_node[i],1.0,INSERT_VALUES);CHKERRQ(ierr); 4069674ae819SStefano Zampini } 4070984c4197SStefano Zampini total_counts = total_primal_vertices; 40719162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 40724641a718SStefano Zampini if (!PetscBTLookup(change_basis,i)) { 40739162d606SStefano Zampini PetscInt *cols; 40749162d606SStefano Zampini 40759162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 40769162d606SStefano Zampini cols = constraints_idxs+constraints_idxs_ptr[i]; 40779162d606SStefano Zampini for (k=0;k<constraints_n[i];k++) { 40789162d606SStefano Zampini PetscInt row = total_counts+k; 40799162d606SStefano Zampini PetscScalar *vals; 40809162d606SStefano Zampini 40819162d606SStefano Zampini vals = constraints_data+constraints_data_ptr[i]+k*size_of_constraint; 40829162d606SStefano Zampini ierr = MatSetValues(pcbddc->ConstraintMatrix,1,&row,size_of_constraint,cols,vals,INSERT_VALUES);CHKERRQ(ierr); 40839162d606SStefano Zampini } 40849162d606SStefano Zampini total_counts += constraints_n[i]; 4085674ae819SStefano Zampini } 4086674ae819SStefano Zampini } 4087674ae819SStefano Zampini /* assembling */ 4088674ae819SStefano Zampini ierr = MatAssemblyBegin(pcbddc->ConstraintMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4089674ae819SStefano Zampini ierr = MatAssemblyEnd(pcbddc->ConstraintMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4090088faed8SStefano Zampini 4091984c4197SStefano Zampini /* 40926a9046bcSBarry Smith ierr = PetscViewerPushFormat(PETSC_VIEWER_STDOUT_SELF,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 4093984c4197SStefano Zampini ierr = MatView(pcbddc->ConstraintMatrix,(PetscViewer)0);CHKERRQ(ierr); 4094f159cad9SBarry Smith ierr = PetscViewerPopFormat(PETSC_VIEWER_STDOUT_SELF);CHKERRQ(ierr); 4095984c4197SStefano Zampini */ 4096674ae819SStefano Zampini /* Create matrix for change of basis. We don't need it in case pcbddc->use_change_of_basis is FALSE */ 4097674ae819SStefano Zampini if (pcbddc->use_change_of_basis) { 4098026de310SStefano Zampini /* dual and primal dofs on a single cc */ 4099984c4197SStefano Zampini PetscInt dual_dofs,primal_dofs; 4100984c4197SStefano Zampini /* working stuff for GEQRF */ 410181d9aea3SBarry Smith PetscScalar *qr_basis,*qr_tau = NULL,*qr_work,lqr_work_t; 4102984c4197SStefano Zampini PetscBLASInt lqr_work; 4103984c4197SStefano Zampini /* working stuff for UNGQR */ 4104984c4197SStefano Zampini PetscScalar *gqr_work,lgqr_work_t; 4105984c4197SStefano Zampini PetscBLASInt lgqr_work; 4106984c4197SStefano Zampini /* working stuff for TRTRS */ 4107984c4197SStefano Zampini PetscScalar *trs_rhs; 41083f08241aSStefano Zampini PetscBLASInt Blas_NRHS; 4109984c4197SStefano Zampini /* pointers for values insertion into change of basis matrix */ 4110984c4197SStefano Zampini PetscInt *start_rows,*start_cols; 4111984c4197SStefano Zampini PetscScalar *start_vals; 4112984c4197SStefano Zampini /* working stuff for values insertion */ 41134641a718SStefano Zampini PetscBT is_primal; 411464efe560SStefano Zampini PetscInt *aux_primal_numbering_B; 4115906d46d4SStefano Zampini /* matrix sizes */ 4116906d46d4SStefano Zampini PetscInt global_size,local_size; 4117906d46d4SStefano Zampini /* temporary change of basis */ 4118906d46d4SStefano Zampini Mat localChangeOfBasisMatrix; 4119cf5a6209SStefano Zampini /* extra space for debugging */ 4120cf5a6209SStefano Zampini PetscScalar *dbg_work; 4121984c4197SStefano Zampini 4122906d46d4SStefano Zampini /* local temporary change of basis acts on local interfaces -> dimension is n_B x n_B */ 4123906d46d4SStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&localChangeOfBasisMatrix);CHKERRQ(ierr); 412416f15bc4SStefano Zampini ierr = MatSetType(localChangeOfBasisMatrix,MATAIJ);CHKERRQ(ierr); 4125bbb9e6c6SStefano Zampini ierr = MatSetSizes(localChangeOfBasisMatrix,pcis->n,pcis->n,pcis->n,pcis->n);CHKERRQ(ierr); 4126906d46d4SStefano Zampini /* nonzeros for local mat */ 4127bbb9e6c6SStefano Zampini ierr = PetscMalloc1(pcis->n,&nnz);CHKERRQ(ierr); 41281dd7afcfSStefano Zampini if (!pcbddc->benign_change || pcbddc->fake_change) { 4129bbb9e6c6SStefano Zampini for (i=0;i<pcis->n;i++) nnz[i]=1; 41301dd7afcfSStefano Zampini } else { 41311dd7afcfSStefano Zampini const PetscInt *ii; 41321dd7afcfSStefano Zampini PetscInt n; 41331dd7afcfSStefano Zampini PetscBool flg_row; 41341dd7afcfSStefano Zampini ierr = MatGetRowIJ(pcbddc->benign_change,0,PETSC_FALSE,PETSC_FALSE,&n,&ii,NULL,&flg_row);CHKERRQ(ierr); 41351dd7afcfSStefano Zampini for (i=0;i<n;i++) nnz[i] = ii[i+1]-ii[i]; 41361dd7afcfSStefano Zampini ierr = MatRestoreRowIJ(pcbddc->benign_change,0,PETSC_FALSE,PETSC_FALSE,&n,&ii,NULL,&flg_row);CHKERRQ(ierr); 41371dd7afcfSStefano Zampini } 41389162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 4139a717540cSStefano Zampini if (PetscBTLookup(change_basis,i)) { 41409162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 4141a717540cSStefano Zampini if (PetscBTLookup(qr_needed_idx,i)) { 41429162d606SStefano Zampini for (j=0;j<size_of_constraint;j++) nnz[constraints_idxs[constraints_idxs_ptr[i]+j]] = size_of_constraint; 4143a717540cSStefano Zampini } else { 41449162d606SStefano Zampini nnz[constraints_idxs[constraints_idxs_ptr[i]]] = size_of_constraint; 41459162d606SStefano Zampini for (j=1;j<size_of_constraint;j++) nnz[constraints_idxs[constraints_idxs_ptr[i]+j]] = 2; 4146a717540cSStefano Zampini } 4147a717540cSStefano Zampini } 4148a717540cSStefano Zampini } 4149906d46d4SStefano Zampini ierr = MatSeqAIJSetPreallocation(localChangeOfBasisMatrix,0,nnz);CHKERRQ(ierr); 4150bbb9e6c6SStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 41511dd7afcfSStefano Zampini /* Set interior change in the matrix */ 41521dd7afcfSStefano Zampini if (!pcbddc->benign_change || pcbddc->fake_change) { 4153bbb9e6c6SStefano Zampini for (i=0;i<pcis->n;i++) { 4154906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,i,i,1.0,INSERT_VALUES);CHKERRQ(ierr); 4155a717540cSStefano Zampini } 41561dd7afcfSStefano Zampini } else { 41571dd7afcfSStefano Zampini const PetscInt *ii,*jj; 41581dd7afcfSStefano Zampini PetscScalar *aa; 41591dd7afcfSStefano Zampini PetscInt n; 41601dd7afcfSStefano Zampini PetscBool flg_row; 41611dd7afcfSStefano Zampini ierr = MatGetRowIJ(pcbddc->benign_change,0,PETSC_FALSE,PETSC_FALSE,&n,&ii,&jj,&flg_row);CHKERRQ(ierr); 41621dd7afcfSStefano Zampini ierr = MatSeqAIJGetArray(pcbddc->benign_change,&aa);CHKERRQ(ierr); 41631dd7afcfSStefano Zampini for (i=0;i<n;i++) { 41641dd7afcfSStefano Zampini ierr = MatSetValues(localChangeOfBasisMatrix,1,&i,ii[i+1]-ii[i],jj+ii[i],aa+ii[i],INSERT_VALUES);CHKERRQ(ierr); 41651dd7afcfSStefano Zampini } 41661dd7afcfSStefano Zampini ierr = MatSeqAIJRestoreArray(pcbddc->benign_change,&aa);CHKERRQ(ierr); 41671dd7afcfSStefano Zampini ierr = MatRestoreRowIJ(pcbddc->benign_change,0,PETSC_FALSE,PETSC_FALSE,&n,&ii,&jj,&flg_row);CHKERRQ(ierr); 41681dd7afcfSStefano Zampini } 4169a717540cSStefano Zampini 4170a717540cSStefano Zampini if (pcbddc->dbg_flag) { 4171a717540cSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"--------------------------------------------------------------\n");CHKERRQ(ierr); 4172a717540cSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Checking change of basis computation for subdomain %04d\n",PetscGlobalRank);CHKERRQ(ierr); 4173a717540cSStefano Zampini } 4174a717540cSStefano Zampini 4175a717540cSStefano Zampini 4176a717540cSStefano Zampini /* Now we loop on the constraints which need a change of basis */ 4177a717540cSStefano Zampini /* 4178a717540cSStefano Zampini Change of basis matrix is evaluated similarly to the FIRST APPROACH in 4179a717540cSStefano Zampini Klawonn and Widlund, Dual-primal FETI-DP methods for linear elasticity, (see Sect 6.2.1) 4180a717540cSStefano Zampini 4181a6b551f4SStefano Zampini Basic blocks of change of basis matrix T computed by 4182a717540cSStefano Zampini 4183a6b551f4SStefano 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) 4184a6b551f4SStefano Zampini 4185a6b551f4SStefano Zampini | 1 0 ... 0 s_1/S | 4186a6b551f4SStefano Zampini | 0 1 ... 0 s_2/S | 4187a717540cSStefano Zampini | ... | 4188a6b551f4SStefano Zampini | 0 ... 1 s_{n-1}/S | 4189a6b551f4SStefano Zampini | -s_1/s_n ... -s_{n-1}/s_n s_n/S | 4190a717540cSStefano Zampini 4191a6b551f4SStefano Zampini with S = \sum_{i=1}^n s_i^2 4192a6b551f4SStefano Zampini NOTE: in the above example, the primal dof is the last one of the edge in LOCAL ordering 4193a6b551f4SStefano Zampini in the current implementation, the primal dof is the first one of the edge in GLOBAL ordering 4194a6b551f4SStefano Zampini 4195a6b551f4SStefano Zampini - QR decomposition of constraints otherwise 4196a717540cSStefano Zampini */ 4197a717540cSStefano Zampini if (qr_needed) { 4198984c4197SStefano Zampini /* space to store Q */ 4199854ce69bSBarry Smith ierr = PetscMalloc1(max_size_of_constraint*max_size_of_constraint,&qr_basis);CHKERRQ(ierr); 4200984c4197SStefano Zampini /* first we issue queries for optimal work */ 42013f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_M);CHKERRQ(ierr); 42023f08241aSStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_N);CHKERRQ(ierr); 42033f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 4204984c4197SStefano Zampini lqr_work = -1; 42053f08241aSStefano Zampini PetscStackCallBLAS("LAPACKgeqrf",LAPACKgeqrf_(&Blas_M,&Blas_N,qr_basis,&Blas_LDA,qr_tau,&lqr_work_t,&lqr_work,&lierr)); 4206984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to GEQRF Lapack routine %d",(int)lierr); 4207984c4197SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lqr_work_t),&lqr_work);CHKERRQ(ierr); 4208785e854fSJed Brown ierr = PetscMalloc1((PetscInt)PetscRealPart(lqr_work_t),&qr_work);CHKERRQ(ierr); 4209984c4197SStefano Zampini lgqr_work = -1; 42103f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_M);CHKERRQ(ierr); 42113f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_N);CHKERRQ(ierr); 42123f08241aSStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_K);CHKERRQ(ierr); 42133f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 42143f08241aSStefano Zampini if (Blas_K>Blas_M) Blas_K=Blas_M; /* adjust just for computing optimal work */ 42153f08241aSStefano Zampini PetscStackCallBLAS("LAPACKungqr",LAPACKungqr_(&Blas_M,&Blas_N,&Blas_K,qr_basis,&Blas_LDA,qr_tau,&lgqr_work_t,&lgqr_work,&lierr)); 4216984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to UNGQR Lapack routine %d",(int)lierr); 4217984c4197SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lgqr_work_t),&lgqr_work);CHKERRQ(ierr); 4218785e854fSJed Brown ierr = PetscMalloc1((PetscInt)PetscRealPart(lgqr_work_t),&gqr_work);CHKERRQ(ierr); 4219984c4197SStefano Zampini /* array to store scaling factors for reflectors */ 4220785e854fSJed Brown ierr = PetscMalloc1(max_constraints,&qr_tau);CHKERRQ(ierr); 4221984c4197SStefano Zampini /* array to store rhs and solution of triangular solver */ 4222785e854fSJed Brown ierr = PetscMalloc1(max_constraints*max_constraints,&trs_rhs);CHKERRQ(ierr); 4223a717540cSStefano Zampini /* allocating workspace for check */ 4224a717540cSStefano Zampini if (pcbddc->dbg_flag) { 4225cf5a6209SStefano Zampini ierr = PetscMalloc1(max_size_of_constraint*(max_constraints+max_size_of_constraint),&dbg_work);CHKERRQ(ierr); 4226a717540cSStefano Zampini } 4227a717540cSStefano Zampini } 4228984c4197SStefano Zampini /* array to store whether a node is primal or not */ 42294641a718SStefano Zampini ierr = PetscBTCreate(pcis->n_B,&is_primal);CHKERRQ(ierr); 4230473ba861SJed Brown ierr = PetscMalloc1(total_primal_vertices,&aux_primal_numbering_B);CHKERRQ(ierr); 42310e6343abSStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,total_primal_vertices,pcbddc->local_primal_ref_node,&i,aux_primal_numbering_B);CHKERRQ(ierr); 42326c4ed002SBarry Smith if (i != total_primal_vertices) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_SUP,"Error in boundary numbering for BDDC vertices! %D != %D\n",total_primal_vertices,i); 423339e2fb2aSStefano Zampini for (i=0;i<total_primal_vertices;i++) { 423439e2fb2aSStefano Zampini ierr = PetscBTSet(is_primal,aux_primal_numbering_B[i]);CHKERRQ(ierr); 423539e2fb2aSStefano Zampini } 423639e2fb2aSStefano Zampini ierr = PetscFree(aux_primal_numbering_B);CHKERRQ(ierr); 4237984c4197SStefano Zampini 4238a717540cSStefano Zampini /* loop on constraints and see whether or not they need a change of basis and compute it */ 42399162d606SStefano Zampini for (total_counts=n_vertices;total_counts<total_counts_cc;total_counts++) { 42409162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[total_counts+1]-constraints_idxs_ptr[total_counts]; 42414641a718SStefano Zampini if (PetscBTLookup(change_basis,total_counts)) { 4242984c4197SStefano Zampini /* get constraint info */ 42439162d606SStefano Zampini primal_dofs = constraints_n[total_counts]; 4244984c4197SStefano Zampini dual_dofs = size_of_constraint-primal_dofs; 4245984c4197SStefano Zampini 4246984c4197SStefano Zampini if (pcbddc->dbg_flag) { 42479162d606SStefano 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); 4248674ae819SStefano Zampini } 4249984c4197SStefano Zampini 4250fa434743SStefano Zampini if (PetscBTLookup(qr_needed_idx,total_counts)) { /* QR */ 4251a717540cSStefano Zampini 4252a717540cSStefano Zampini /* copy quadrature constraints for change of basis check */ 4253a717540cSStefano Zampini if (pcbddc->dbg_flag) { 42549162d606SStefano Zampini ierr = PetscMemcpy(dbg_work,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 4255a717540cSStefano Zampini } 4256984c4197SStefano Zampini /* copy temporary constraints into larger work vector (in order to store all columns of Q) */ 42579162d606SStefano Zampini ierr = PetscMemcpy(qr_basis,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 4258984c4197SStefano Zampini 4259984c4197SStefano Zampini /* compute QR decomposition of constraints */ 42603f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 42613f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 42623f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 4263674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 42643f08241aSStefano Zampini PetscStackCallBLAS("LAPACKgeqrf",LAPACKgeqrf_(&Blas_M,&Blas_N,qr_basis,&Blas_LDA,qr_tau,qr_work,&lqr_work,&lierr)); 4265984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GEQRF Lapack routine %d",(int)lierr); 4266674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 4267984c4197SStefano Zampini 4268984c4197SStefano Zampini /* explictly compute R^-T */ 4269984c4197SStefano Zampini ierr = PetscMemzero(trs_rhs,primal_dofs*primal_dofs*sizeof(*trs_rhs));CHKERRQ(ierr); 4270984c4197SStefano Zampini for (j=0;j<primal_dofs;j++) trs_rhs[j*(primal_dofs+1)] = 1.0; 42713f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 42723f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_NRHS);CHKERRQ(ierr); 42733f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 42743f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDB);CHKERRQ(ierr); 4275984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 42763f08241aSStefano Zampini PetscStackCallBLAS("LAPACKtrtrs",LAPACKtrtrs_("U","T","N",&Blas_N,&Blas_NRHS,qr_basis,&Blas_LDA,trs_rhs,&Blas_LDB,&lierr)); 4277984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in TRTRS Lapack routine %d",(int)lierr); 4278984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 4279984c4197SStefano Zampini 4280a717540cSStefano Zampini /* explicitly compute all columns of Q (Q = [Q1 | Q2] ) overwriting QR factorization in qr_basis */ 42813f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 42823f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 42833f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_K);CHKERRQ(ierr); 42843f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 4285984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 42863f08241aSStefano Zampini PetscStackCallBLAS("LAPACKungqr",LAPACKungqr_(&Blas_M,&Blas_N,&Blas_K,qr_basis,&Blas_LDA,qr_tau,gqr_work,&lgqr_work,&lierr)); 4287984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in UNGQR Lapack routine %d",(int)lierr); 4288984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 4289984c4197SStefano Zampini 4290984c4197SStefano Zampini /* first primal_dofs columns of Q need to be re-scaled in order to be unitary w.r.t constraints 4291984c4197SStefano Zampini i.e. C_{pxn}*Q_{nxn} should be equal to [I_pxp | 0_pxd] (see check below) 4292984c4197SStefano Zampini where n=size_of_constraint, p=primal_dofs, d=dual_dofs (n=p+d), I and 0 identity and null matrix resp. */ 42933f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 42943f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 42953f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_K);CHKERRQ(ierr); 42963f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 42973f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDB);CHKERRQ(ierr); 42983f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDC);CHKERRQ(ierr); 4299984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 43009162d606SStefano 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)); 4301984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 43029162d606SStefano Zampini ierr = PetscMemcpy(qr_basis,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 4303984c4197SStefano Zampini 4304984c4197SStefano Zampini /* insert values in change of basis matrix respecting global ordering of new primal dofs */ 43059162d606SStefano Zampini start_rows = &constraints_idxs[constraints_idxs_ptr[total_counts]]; 4306984c4197SStefano Zampini /* insert cols for primal dofs */ 4307984c4197SStefano Zampini for (j=0;j<primal_dofs;j++) { 4308984c4197SStefano Zampini start_vals = &qr_basis[j*size_of_constraint]; 43099162d606SStefano Zampini start_cols = &constraints_idxs[constraints_idxs_ptr[total_counts]+j]; 4310906d46d4SStefano Zampini ierr = MatSetValues(localChangeOfBasisMatrix,size_of_constraint,start_rows,1,start_cols,start_vals,INSERT_VALUES);CHKERRQ(ierr); 4311984c4197SStefano Zampini } 4312984c4197SStefano Zampini /* insert cols for dual dofs */ 4313984c4197SStefano Zampini for (j=0,k=0;j<dual_dofs;k++) { 43149162d606SStefano Zampini if (!PetscBTLookup(is_primal,constraints_idxs_B[constraints_idxs_ptr[total_counts]+k])) { 4315984c4197SStefano Zampini start_vals = &qr_basis[(primal_dofs+j)*size_of_constraint]; 43169162d606SStefano Zampini start_cols = &constraints_idxs[constraints_idxs_ptr[total_counts]+k]; 4317906d46d4SStefano Zampini ierr = MatSetValues(localChangeOfBasisMatrix,size_of_constraint,start_rows,1,start_cols,start_vals,INSERT_VALUES);CHKERRQ(ierr); 4318984c4197SStefano Zampini j++; 4319674ae819SStefano Zampini } 4320674ae819SStefano Zampini } 4321984c4197SStefano Zampini 4322984c4197SStefano Zampini /* check change of basis */ 4323984c4197SStefano Zampini if (pcbddc->dbg_flag) { 4324984c4197SStefano Zampini PetscInt ii,jj; 4325984c4197SStefano Zampini PetscBool valid_qr=PETSC_TRUE; 4326c4303822SStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_M);CHKERRQ(ierr); 4327c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 4328c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_K);CHKERRQ(ierr); 4329c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 4330c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDB);CHKERRQ(ierr); 4331c4303822SStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDC);CHKERRQ(ierr); 4332984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 4333cf5a6209SStefano 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)); 4334984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 4335984c4197SStefano Zampini for (jj=0;jj<size_of_constraint;jj++) { 4336984c4197SStefano Zampini for (ii=0;ii<primal_dofs;ii++) { 4337cf5a6209SStefano Zampini if (ii != jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]) > 1.e-12) valid_qr = PETSC_FALSE; 4338cf5a6209SStefano 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; 4339674ae819SStefano Zampini } 4340674ae819SStefano Zampini } 4341984c4197SStefano Zampini if (!valid_qr) { 434222d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> wrong change of basis!\n");CHKERRQ(ierr); 4343984c4197SStefano Zampini for (jj=0;jj<size_of_constraint;jj++) { 4344984c4197SStefano Zampini for (ii=0;ii<primal_dofs;ii++) { 4345cf5a6209SStefano Zampini if (ii != jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]) > 1.e-12) { 4346cf5a6209SStefano 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])); 4347674ae819SStefano Zampini } 4348cf5a6209SStefano Zampini if (ii == jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]-1.0) > 1.e-12) { 4349cf5a6209SStefano 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])); 4350984c4197SStefano Zampini } 4351984c4197SStefano Zampini } 4352984c4197SStefano Zampini } 4353674ae819SStefano Zampini } else { 435422d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> right change of basis!\n");CHKERRQ(ierr); 4355674ae819SStefano Zampini } 4356674ae819SStefano Zampini } 4357a717540cSStefano Zampini } else { /* simple transformation block */ 4358a717540cSStefano Zampini PetscInt row,col; 4359a6b551f4SStefano Zampini PetscScalar val,norm; 4360a6b551f4SStefano Zampini 4361a6b551f4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 43629162d606SStefano 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)); 4363a717540cSStefano Zampini for (j=0;j<size_of_constraint;j++) { 43649162d606SStefano Zampini PetscInt row_B = constraints_idxs_B[constraints_idxs_ptr[total_counts]+j]; 43659162d606SStefano Zampini row = constraints_idxs[constraints_idxs_ptr[total_counts]+j]; 4366bbb9e6c6SStefano Zampini if (!PetscBTLookup(is_primal,row_B)) { 43679162d606SStefano Zampini col = constraints_idxs[constraints_idxs_ptr[total_counts]]; 4368906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,row,1.0,INSERT_VALUES);CHKERRQ(ierr); 43699162d606SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,col,constraints_data[constraints_data_ptr[total_counts]+j]/norm,INSERT_VALUES);CHKERRQ(ierr); 4370a717540cSStefano Zampini } else { 4371a717540cSStefano Zampini for (k=0;k<size_of_constraint;k++) { 43729162d606SStefano Zampini col = constraints_idxs[constraints_idxs_ptr[total_counts]+k]; 4373a717540cSStefano Zampini if (row != col) { 43749162d606SStefano Zampini val = -constraints_data[constraints_data_ptr[total_counts]+k]/constraints_data[constraints_data_ptr[total_counts]]; 4375a717540cSStefano Zampini } else { 43769162d606SStefano Zampini val = constraints_data[constraints_data_ptr[total_counts]]/norm; 4377a717540cSStefano Zampini } 4378906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,col,val,INSERT_VALUES);CHKERRQ(ierr); 4379a717540cSStefano Zampini } 4380a717540cSStefano Zampini } 4381a717540cSStefano Zampini } 438298a51de6SStefano Zampini if (pcbddc->dbg_flag) { 438322d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> using standard change of basis\n");CHKERRQ(ierr); 4384a717540cSStefano Zampini } 4385674ae819SStefano Zampini } 4386984c4197SStefano Zampini } else { 4387984c4197SStefano Zampini if (pcbddc->dbg_flag) { 43889162d606SStefano 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); 4389674ae819SStefano Zampini } 4390674ae819SStefano Zampini } 4391674ae819SStefano Zampini } 4392a717540cSStefano Zampini 4393a717540cSStefano Zampini /* free workspace */ 4394a717540cSStefano Zampini if (qr_needed) { 4395984c4197SStefano Zampini if (pcbddc->dbg_flag) { 4396cf5a6209SStefano Zampini ierr = PetscFree(dbg_work);CHKERRQ(ierr); 4397984c4197SStefano Zampini } 4398984c4197SStefano Zampini ierr = PetscFree(trs_rhs);CHKERRQ(ierr); 4399984c4197SStefano Zampini ierr = PetscFree(qr_tau);CHKERRQ(ierr); 4400984c4197SStefano Zampini ierr = PetscFree(qr_work);CHKERRQ(ierr); 4401984c4197SStefano Zampini ierr = PetscFree(gqr_work);CHKERRQ(ierr); 4402984c4197SStefano Zampini ierr = PetscFree(qr_basis);CHKERRQ(ierr); 4403674ae819SStefano Zampini } 4404a717540cSStefano Zampini ierr = PetscBTDestroy(&is_primal);CHKERRQ(ierr); 4405906d46d4SStefano Zampini ierr = MatAssemblyBegin(localChangeOfBasisMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4406906d46d4SStefano Zampini ierr = MatAssemblyEnd(localChangeOfBasisMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4407906d46d4SStefano Zampini 4408906d46d4SStefano Zampini /* assembling of global change of variable */ 440988c03ad3SStefano Zampini if (!pcbddc->fake_change) { 4410bbb9e6c6SStefano Zampini Mat tmat; 441116f15bc4SStefano Zampini PetscInt bs; 441216f15bc4SStefano Zampini 4413906d46d4SStefano Zampini ierr = VecGetSize(pcis->vec1_global,&global_size);CHKERRQ(ierr); 4414906d46d4SStefano Zampini ierr = VecGetLocalSize(pcis->vec1_global,&local_size);CHKERRQ(ierr); 4415bbb9e6c6SStefano Zampini ierr = MatDuplicate(pc->pmat,MAT_DO_NOT_COPY_VALUES,&tmat);CHKERRQ(ierr); 4416bbb9e6c6SStefano Zampini ierr = MatISSetLocalMat(tmat,localChangeOfBasisMatrix);CHKERRQ(ierr); 4417bbb9e6c6SStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)pc),&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 4418bbb9e6c6SStefano Zampini ierr = MatSetType(pcbddc->ChangeOfBasisMatrix,MATAIJ);CHKERRQ(ierr); 441916f15bc4SStefano Zampini ierr = MatGetBlockSize(pc->pmat,&bs);CHKERRQ(ierr); 442016f15bc4SStefano Zampini ierr = MatSetBlockSize(pcbddc->ChangeOfBasisMatrix,bs);CHKERRQ(ierr); 4421906d46d4SStefano Zampini ierr = MatSetSizes(pcbddc->ChangeOfBasisMatrix,local_size,local_size,global_size,global_size);CHKERRQ(ierr); 4422bbb9e6c6SStefano Zampini ierr = MatISSetMPIXAIJPreallocation_Private(tmat,pcbddc->ChangeOfBasisMatrix,PETSC_TRUE);CHKERRQ(ierr); 4423bbb9e6c6SStefano Zampini ierr = MatISGetMPIXAIJ(tmat,MAT_REUSE_MATRIX,&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 4424bbb9e6c6SStefano Zampini ierr = MatDestroy(&tmat);CHKERRQ(ierr); 4425bbb9e6c6SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 4426bbb9e6c6SStefano Zampini ierr = VecSet(pcis->vec1_N,1.0);CHKERRQ(ierr); 4427e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4428e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4429bbb9e6c6SStefano Zampini ierr = VecReciprocal(pcis->vec1_global);CHKERRQ(ierr); 4430bbb9e6c6SStefano Zampini ierr = MatDiagonalScale(pcbddc->ChangeOfBasisMatrix,pcis->vec1_global,NULL);CHKERRQ(ierr); 443188c03ad3SStefano Zampini 4432906d46d4SStefano Zampini /* check */ 4433906d46d4SStefano Zampini if (pcbddc->dbg_flag) { 4434906d46d4SStefano Zampini PetscReal error; 4435906d46d4SStefano Zampini Vec x,x_change; 4436906d46d4SStefano Zampini 4437906d46d4SStefano Zampini ierr = VecDuplicate(pcis->vec1_global,&x);CHKERRQ(ierr); 4438906d46d4SStefano Zampini ierr = VecDuplicate(pcis->vec1_global,&x_change);CHKERRQ(ierr); 4439906d46d4SStefano Zampini ierr = VecSetRandom(x,NULL);CHKERRQ(ierr); 4440906d46d4SStefano Zampini ierr = VecCopy(x,pcis->vec1_global);CHKERRQ(ierr); 4441e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,x,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4442e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,x,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4443bbb9e6c6SStefano Zampini ierr = MatMult(localChangeOfBasisMatrix,pcis->vec1_N,pcis->vec2_N);CHKERRQ(ierr); 4444e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec2_N,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4445e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec2_N,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4446906d46d4SStefano Zampini ierr = MatMult(pcbddc->ChangeOfBasisMatrix,pcis->vec1_global,x_change);CHKERRQ(ierr); 4447906d46d4SStefano Zampini ierr = VecAXPY(x,-1.0,x_change);CHKERRQ(ierr); 4448906d46d4SStefano Zampini ierr = VecNorm(x,NORM_INFINITY,&error);CHKERRQ(ierr); 4449906d46d4SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4450bbb9e6c6SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Error global vs local change: %1.6e\n",error);CHKERRQ(ierr); 4451906d46d4SStefano Zampini ierr = VecDestroy(&x);CHKERRQ(ierr); 4452906d46d4SStefano Zampini ierr = VecDestroy(&x_change);CHKERRQ(ierr); 4453906d46d4SStefano Zampini } 4454b96c3477SStefano Zampini /* adapt sub_schurs computed (if any) */ 4455b96c3477SStefano Zampini if (pcbddc->use_deluxe_scaling) { 4456b96c3477SStefano Zampini PCBDDCSubSchurs sub_schurs=pcbddc->sub_schurs; 4457bf3a8328SStefano Zampini 4458bf3a8328SStefano Zampini if (pcbddc->use_change_of_basis && pcbddc->adaptive_userdefined) { 4459bf3a8328SStefano Zampini SETERRQ(PetscObjectComm((PetscObject)pc),PETSC_ERR_SUP,"Cannot mix automatic change of basis, adaptive selection and user-defined constraints");CHKERRQ(ierr); 4460bf3a8328SStefano Zampini } 4461b96c3477SStefano Zampini if (sub_schurs->S_Ej_all) { 4462ac632422SStefano Zampini Mat S_new,tmat; 4463bf3a8328SStefano Zampini IS is_all_N,is_V_Sall = NULL; 4464bbb9e6c6SStefano Zampini 4465bbb9e6c6SStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(pcis->BtoNmap,sub_schurs->is_Ej_all,&is_all_N);CHKERRQ(ierr); 44666816873aSStefano Zampini ierr = MatGetSubMatrix(localChangeOfBasisMatrix,is_all_N,is_all_N,MAT_INITIAL_MATRIX,&tmat);CHKERRQ(ierr); 4467bf3a8328SStefano Zampini if (pcbddc->deluxe_zerorows) { 4468bf3a8328SStefano Zampini ISLocalToGlobalMapping NtoSall; 4469bf3a8328SStefano Zampini IS is_V; 4470b087196eSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,pcbddc->n_vertices,pcbddc->local_primal_ref_node,PETSC_COPY_VALUES,&is_V);CHKERRQ(ierr); 4471b087196eSStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(is_all_N,&NtoSall);CHKERRQ(ierr); 4472b087196eSStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(NtoSall,IS_GTOLM_DROP,is_V,&is_V_Sall);CHKERRQ(ierr); 4473b087196eSStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&NtoSall);CHKERRQ(ierr); 4474b087196eSStefano Zampini ierr = ISDestroy(&is_V);CHKERRQ(ierr); 4475bf3a8328SStefano Zampini } 4476bf3a8328SStefano Zampini ierr = ISDestroy(&is_all_N);CHKERRQ(ierr); 4477ac632422SStefano Zampini ierr = MatPtAP(sub_schurs->S_Ej_all,tmat,MAT_INITIAL_MATRIX,1.0,&S_new);CHKERRQ(ierr); 4478b96c3477SStefano Zampini ierr = MatDestroy(&sub_schurs->S_Ej_all);CHKERRQ(ierr); 4479ac632422SStefano Zampini ierr = PetscObjectReference((PetscObject)S_new);CHKERRQ(ierr); 4480bf3a8328SStefano Zampini if (pcbddc->deluxe_zerorows) { 4481bf3a8328SStefano Zampini const PetscScalar *array; 4482bf3a8328SStefano Zampini const PetscInt *idxs_V,*idxs_all; 4483bf3a8328SStefano Zampini PetscInt i,n_V; 4484bf3a8328SStefano Zampini 4485b087196eSStefano Zampini ierr = MatZeroRowsColumnsIS(S_new,is_V_Sall,1.,NULL,NULL);CHKERRQ(ierr); 4486b087196eSStefano Zampini ierr = ISGetLocalSize(is_V_Sall,&n_V);CHKERRQ(ierr); 4487b087196eSStefano Zampini ierr = ISGetIndices(is_V_Sall,&idxs_V);CHKERRQ(ierr); 4488b087196eSStefano Zampini ierr = ISGetIndices(sub_schurs->is_Ej_all,&idxs_all);CHKERRQ(ierr); 4489b087196eSStefano Zampini ierr = VecGetArrayRead(pcis->D,&array);CHKERRQ(ierr); 4490b087196eSStefano Zampini for (i=0;i<n_V;i++) { 4491b087196eSStefano Zampini PetscScalar val; 4492b087196eSStefano Zampini PetscInt idx; 4493b087196eSStefano Zampini 4494b087196eSStefano Zampini idx = idxs_V[i]; 4495b087196eSStefano Zampini val = array[idxs_all[idxs_V[i]]]; 4496b087196eSStefano Zampini ierr = MatSetValue(S_new,idx,idx,val,INSERT_VALUES);CHKERRQ(ierr); 4497b087196eSStefano Zampini } 4498b087196eSStefano Zampini ierr = MatAssemblyBegin(S_new,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4499b087196eSStefano Zampini ierr = MatAssemblyEnd(S_new,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4500bf3a8328SStefano Zampini ierr = VecRestoreArrayRead(pcis->D,&array);CHKERRQ(ierr); 4501bf3a8328SStefano Zampini ierr = ISRestoreIndices(sub_schurs->is_Ej_all,&idxs_all);CHKERRQ(ierr); 4502bf3a8328SStefano Zampini ierr = ISRestoreIndices(is_V_Sall,&idxs_V);CHKERRQ(ierr); 4503bf3a8328SStefano Zampini } 4504ac632422SStefano Zampini sub_schurs->S_Ej_all = S_new; 4505ac632422SStefano Zampini ierr = MatDestroy(&S_new);CHKERRQ(ierr); 4506ac632422SStefano Zampini if (sub_schurs->sum_S_Ej_all) { 4507ac632422SStefano Zampini ierr = MatPtAP(sub_schurs->sum_S_Ej_all,tmat,MAT_INITIAL_MATRIX,1.0,&S_new);CHKERRQ(ierr); 4508b96c3477SStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_all);CHKERRQ(ierr); 4509ac632422SStefano Zampini ierr = PetscObjectReference((PetscObject)S_new);CHKERRQ(ierr); 4510bf3a8328SStefano Zampini if (pcbddc->deluxe_zerorows) { 4511b087196eSStefano Zampini ierr = MatZeroRowsColumnsIS(S_new,is_V_Sall,1.,NULL,NULL);CHKERRQ(ierr); 4512bf3a8328SStefano Zampini } 4513ac632422SStefano Zampini sub_schurs->sum_S_Ej_all = S_new; 4514ac632422SStefano Zampini ierr = MatDestroy(&S_new);CHKERRQ(ierr); 4515ac632422SStefano Zampini } 4516b087196eSStefano Zampini ierr = ISDestroy(&is_V_Sall);CHKERRQ(ierr); 451788c03ad3SStefano Zampini ierr = MatDestroy(&tmat);CHKERRQ(ierr); 4518b96c3477SStefano Zampini } 4519c9db6a07SStefano Zampini /* destroy any change of basis context in sub_schurs */ 4520c9db6a07SStefano Zampini if (sub_schurs->change) { 4521c9db6a07SStefano Zampini PetscInt i; 4522c9db6a07SStefano Zampini 4523c9db6a07SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 4524c9db6a07SStefano Zampini ierr = KSPDestroy(&sub_schurs->change[i]);CHKERRQ(ierr); 4525c9db6a07SStefano Zampini } 4526c9db6a07SStefano Zampini ierr = PetscFree(sub_schurs->change);CHKERRQ(ierr); 4527c9db6a07SStefano Zampini } 4528b96c3477SStefano Zampini } 452916909a7fSStefano Zampini if (pcbddc->switch_static) { /* need to save the local change */ 453016909a7fSStefano Zampini pcbddc->switch_static_change = localChangeOfBasisMatrix; 453116909a7fSStefano Zampini } else { 4532906d46d4SStefano Zampini ierr = MatDestroy(&localChangeOfBasisMatrix);CHKERRQ(ierr); 453316909a7fSStefano Zampini } 45341dd7afcfSStefano Zampini /* determine if any process has changed the pressures locally */ 453527b6a85dSStefano Zampini pcbddc->change_interior = pcbddc->benign_have_null; 453672b8c272SStefano Zampini } else { /* fake change (get back change of basis into ConstraintMatrix and info on qr) */ 453772b8c272SStefano Zampini ierr = MatDestroy(&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 453872b8c272SStefano Zampini pcbddc->ConstraintMatrix = localChangeOfBasisMatrix; 453972b8c272SStefano Zampini pcbddc->use_qr_single = qr_needed; 454072b8c272SStefano Zampini } 45411dd7afcfSStefano Zampini } else if (pcbddc->user_ChangeOfBasisMatrix || pcbddc->benign_saddle_point) { 454227b6a85dSStefano Zampini if (!pcbddc->benign_have_null && pcbddc->user_ChangeOfBasisMatrix) { 4543b9b85e73SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->user_ChangeOfBasisMatrix);CHKERRQ(ierr); 4544b9b85e73SStefano Zampini pcbddc->ChangeOfBasisMatrix = pcbddc->user_ChangeOfBasisMatrix; 4545906d46d4SStefano Zampini } else { 45461dd7afcfSStefano Zampini Mat benign_global = NULL; 454727b6a85dSStefano Zampini if (pcbddc->benign_have_null) { 45481dd7afcfSStefano Zampini Mat tmat; 45491dd7afcfSStefano Zampini 45501dd7afcfSStefano Zampini pcbddc->change_interior = PETSC_TRUE; 45511dd7afcfSStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 45521dd7afcfSStefano Zampini ierr = VecSet(pcis->vec1_N,1.0);CHKERRQ(ierr); 45531dd7afcfSStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 45541dd7afcfSStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 45551dd7afcfSStefano Zampini ierr = VecReciprocal(pcis->vec1_global);CHKERRQ(ierr); 45561dd7afcfSStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_global,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 45571dd7afcfSStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_global,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 45581dd7afcfSStefano Zampini ierr = MatDuplicate(pc->pmat,MAT_DO_NOT_COPY_VALUES,&tmat);CHKERRQ(ierr); 45591dd7afcfSStefano Zampini if (pcbddc->benign_change) { 45601dd7afcfSStefano Zampini Mat M; 45611dd7afcfSStefano Zampini 45621dd7afcfSStefano Zampini ierr = MatDuplicate(pcbddc->benign_change,MAT_COPY_VALUES,&M);CHKERRQ(ierr); 45631dd7afcfSStefano Zampini ierr = MatDiagonalScale(M,pcis->vec1_N,NULL);CHKERRQ(ierr); 45641dd7afcfSStefano Zampini ierr = MatISSetLocalMat(tmat,M);CHKERRQ(ierr); 45651dd7afcfSStefano Zampini ierr = MatDestroy(&M);CHKERRQ(ierr); 4566906d46d4SStefano Zampini } else { 45671dd7afcfSStefano Zampini Mat eye; 45681dd7afcfSStefano Zampini PetscScalar *array; 45691dd7afcfSStefano Zampini 45701dd7afcfSStefano Zampini ierr = VecGetArray(pcis->vec1_N,&array);CHKERRQ(ierr); 45711dd7afcfSStefano Zampini ierr = MatCreateSeqAIJ(PETSC_COMM_SELF,pcis->n,pcis->n,1,NULL,&eye);CHKERRQ(ierr); 45721dd7afcfSStefano Zampini for (i=0;i<pcis->n;i++) { 45731dd7afcfSStefano Zampini ierr = MatSetValue(eye,i,i,array[i],INSERT_VALUES);CHKERRQ(ierr); 4574906d46d4SStefano Zampini } 45751dd7afcfSStefano Zampini ierr = VecRestoreArray(pcis->vec1_N,&array);CHKERRQ(ierr); 45761dd7afcfSStefano Zampini ierr = MatAssemblyBegin(eye,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 45771dd7afcfSStefano Zampini ierr = MatAssemblyEnd(eye,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 45781dd7afcfSStefano Zampini ierr = MatISSetLocalMat(tmat,eye);CHKERRQ(ierr); 45791dd7afcfSStefano Zampini ierr = MatDestroy(&eye);CHKERRQ(ierr); 45801dd7afcfSStefano Zampini } 45811dd7afcfSStefano Zampini ierr = MatISGetMPIXAIJ(tmat,MAT_INITIAL_MATRIX,&benign_global);CHKERRQ(ierr); 45821dd7afcfSStefano Zampini ierr = MatDestroy(&tmat);CHKERRQ(ierr); 45831dd7afcfSStefano Zampini } 45841dd7afcfSStefano Zampini if (pcbddc->user_ChangeOfBasisMatrix) { 45851dd7afcfSStefano Zampini ierr = MatMatMult(pcbddc->user_ChangeOfBasisMatrix,benign_global,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 45861dd7afcfSStefano Zampini ierr = MatDestroy(&benign_global);CHKERRQ(ierr); 458727b6a85dSStefano Zampini } else if (pcbddc->benign_have_null) { 45881dd7afcfSStefano Zampini pcbddc->ChangeOfBasisMatrix = benign_global; 45891dd7afcfSStefano Zampini } 45901dd7afcfSStefano Zampini } 459116909a7fSStefano Zampini if (pcbddc->switch_static && pcbddc->ChangeOfBasisMatrix) { /* need to save the local change */ 459216909a7fSStefano Zampini IS is_global; 459316909a7fSStefano Zampini const PetscInt *gidxs; 459416909a7fSStefano Zampini 459516909a7fSStefano Zampini ierr = ISLocalToGlobalMappingGetIndices(pc->pmat->rmap->mapping,&gidxs);CHKERRQ(ierr); 459616909a7fSStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),pcis->n,gidxs,PETSC_COPY_VALUES,&is_global);CHKERRQ(ierr); 459716909a7fSStefano Zampini ierr = ISLocalToGlobalMappingRestoreIndices(pc->pmat->rmap->mapping,&gidxs);CHKERRQ(ierr); 459816909a7fSStefano Zampini ierr = MatGetSubMatrixUnsorted(pcbddc->ChangeOfBasisMatrix,is_global,is_global,&pcbddc->switch_static_change);CHKERRQ(ierr); 459916909a7fSStefano Zampini ierr = ISDestroy(&is_global);CHKERRQ(ierr); 460016909a7fSStefano Zampini } 46011dd7afcfSStefano Zampini } 46021dd7afcfSStefano Zampini if (!pcbddc->fake_change && pcbddc->ChangeOfBasisMatrix && !pcbddc->work_change) { 46031dd7afcfSStefano Zampini ierr = VecDuplicate(pcis->vec1_global,&pcbddc->work_change);CHKERRQ(ierr); 4604b9b85e73SStefano Zampini } 4605a717540cSStefano Zampini 460672b8c272SStefano Zampini if (!pcbddc->fake_change) { 46074f1b2e48SStefano Zampini /* add pressure dofs to set of primal nodes for numbering purposes */ 46084f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 46094f1b2e48SStefano Zampini pcbddc->local_primal_ref_node[pcbddc->local_primal_size_cc] = pcbddc->benign_p0_lidx[i]; 46104f1b2e48SStefano Zampini pcbddc->primal_indices_local_idxs[pcbddc->local_primal_size] = pcbddc->benign_p0_lidx[i]; 4611019a44ceSStefano Zampini pcbddc->local_primal_ref_mult[pcbddc->local_primal_size_cc] = 1; 4612019a44ceSStefano Zampini pcbddc->local_primal_size_cc++; 4613019a44ceSStefano Zampini pcbddc->local_primal_size++; 4614019a44ceSStefano Zampini } 4615019a44ceSStefano Zampini 4616019a44ceSStefano Zampini /* check if a new primal space has been introduced (also take into account benign trick) */ 4617727cdba6SStefano Zampini pcbddc->new_primal_space_local = PETSC_TRUE; 4618727cdba6SStefano Zampini if (olocal_primal_size == pcbddc->local_primal_size) { 46199f47a83aSStefano 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); 4620c1c8e736SStefano Zampini pcbddc->new_primal_space_local = (PetscBool)(!pcbddc->new_primal_space_local); 46210e6343abSStefano Zampini if (!pcbddc->new_primal_space_local) { 46229f47a83aSStefano 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); 4623727cdba6SStefano Zampini pcbddc->new_primal_space_local = (PetscBool)(!pcbddc->new_primal_space_local); 4624727cdba6SStefano Zampini } 46250e6343abSStefano Zampini } 4626727cdba6SStefano Zampini /* new_primal_space will be used for numbering of coarse dofs, so it should be the same across all subdomains */ 4627b2566f29SBarry Smith ierr = MPIU_Allreduce(&pcbddc->new_primal_space_local,&pcbddc->new_primal_space,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 462872b8c272SStefano Zampini } 462972b8c272SStefano Zampini ierr = PetscFree2(olocal_primal_ref_node,olocal_primal_ref_mult);CHKERRQ(ierr); 4630727cdba6SStefano Zampini 4631a717540cSStefano Zampini /* flush dbg viewer */ 4632b8ffe317SStefano Zampini if (pcbddc->dbg_flag) { 4633b8ffe317SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4634b8ffe317SStefano Zampini } 4635a717540cSStefano Zampini 4636e310c8b4SStefano Zampini /* free workspace */ 4637a717540cSStefano Zampini ierr = PetscBTDestroy(&qr_needed_idx);CHKERRQ(ierr); 46384641a718SStefano Zampini ierr = PetscBTDestroy(&change_basis);CHKERRQ(ierr); 463908122e43SStefano Zampini if (!pcbddc->adaptive_selection) { 46409162d606SStefano Zampini ierr = PetscFree3(constraints_idxs_ptr,constraints_data_ptr,constraints_n);CHKERRQ(ierr); 46419162d606SStefano Zampini ierr = PetscFree3(constraints_data,constraints_idxs,constraints_idxs_B);CHKERRQ(ierr); 464208122e43SStefano Zampini } else { 46439162d606SStefano Zampini ierr = PetscFree5(pcbddc->adaptive_constraints_n, 46449162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr, 46459162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr, 464608122e43SStefano Zampini pcbddc->adaptive_constraints_idxs, 464708122e43SStefano Zampini pcbddc->adaptive_constraints_data);CHKERRQ(ierr); 46489162d606SStefano Zampini ierr = PetscFree(constraints_n);CHKERRQ(ierr); 46499162d606SStefano Zampini ierr = PetscFree(constraints_idxs_B);CHKERRQ(ierr); 465008122e43SStefano Zampini } 4651674ae819SStefano Zampini PetscFunctionReturn(0); 4652674ae819SStefano Zampini } 4653674ae819SStefano Zampini 4654674ae819SStefano Zampini #undef __FUNCT__ 4655674ae819SStefano Zampini #define __FUNCT__ "PCBDDCAnalyzeInterface" 4656674ae819SStefano Zampini PetscErrorCode PCBDDCAnalyzeInterface(PC pc) 4657674ae819SStefano Zampini { 4658674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 4659674ae819SStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 4660674ae819SStefano Zampini Mat_IS *matis = (Mat_IS*)pc->pmat->data; 466114f95afaSStefano Zampini PetscInt ierr,i,N; 4662674ae819SStefano Zampini 4663674ae819SStefano Zampini PetscFunctionBegin; 46648e61c736SStefano Zampini /* Reset previously computed graph */ 46658e61c736SStefano Zampini ierr = PCBDDCGraphReset(pcbddc->mat_graph);CHKERRQ(ierr); 4666674ae819SStefano Zampini /* Init local Graph struct */ 46677fb0e2dbSStefano Zampini ierr = MatGetSize(pc->pmat,&N,NULL);CHKERRQ(ierr); 46683bbff08aSStefano Zampini ierr = PCBDDCGraphInit(pcbddc->mat_graph,pcis->mapping,N);CHKERRQ(ierr); 4669674ae819SStefano Zampini 4670575ad6abSStefano Zampini /* Check validity of the csr graph passed in by the user */ 46715099eff2SStefano Zampini if (pcbddc->mat_graph->nvtxs_csr && pcbddc->mat_graph->nvtxs_csr != pcbddc->mat_graph->nvtxs) { 46725099eff2SStefano 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); 4673575ad6abSStefano Zampini } 46749577ea80SStefano Zampini 4675674ae819SStefano Zampini /* Set default CSR adjacency of local dofs if not provided by the user with PCBDDCSetLocalAdjacencyGraph */ 4676d4d8cf7bSStefano Zampini if ( (!pcbddc->mat_graph->xadj || !pcbddc->mat_graph->adjncy) && pcbddc->use_local_adj) { 46774d379d7bSStefano Zampini PetscInt *xadj,*adjncy; 46784d379d7bSStefano Zampini PetscInt nvtxs; 4679e496cd5dSStefano Zampini PetscBool flg_row=PETSC_FALSE; 4680674ae819SStefano Zampini 46812fffb893SStefano Zampini ierr = MatGetRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 46822fffb893SStefano Zampini if (flg_row) { 46834d379d7bSStefano Zampini ierr = PCBDDCSetLocalAdjacencyGraph(pc,nvtxs,xadj,adjncy,PETSC_COPY_VALUES);CHKERRQ(ierr); 4684b96c3477SStefano Zampini pcbddc->computed_rowadj = PETSC_TRUE; 46852fffb893SStefano Zampini } 46862fffb893SStefano Zampini ierr = MatRestoreRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 4687674ae819SStefano Zampini } 46889b28b941SStefano Zampini if (pcbddc->dbg_flag) { 46899b28b941SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4690674ae819SStefano Zampini } 4691674ae819SStefano Zampini 4692674ae819SStefano Zampini /* Setup of Graph */ 469314f95afaSStefano Zampini ierr = PCBDDCGraphSetUp(pcbddc->mat_graph,pcbddc->vertex_size,pcbddc->NeumannBoundariesLocal,pcbddc->DirichletBoundariesLocal,pcbddc->n_ISForDofsLocal,pcbddc->ISForDofsLocal,pcbddc->user_primal_vertices_local);CHKERRQ(ierr); 4694674ae819SStefano Zampini 46954f1b2e48SStefano Zampini /* attach info on disconnected subdomains if present */ 46964f1b2e48SStefano Zampini if (pcbddc->n_local_subs) { 46974f1b2e48SStefano Zampini PetscInt *local_subs; 46984f1b2e48SStefano Zampini 46994f1b2e48SStefano Zampini ierr = PetscMalloc1(N,&local_subs);CHKERRQ(ierr); 47004f1b2e48SStefano Zampini for (i=0;i<pcbddc->n_local_subs;i++) { 47014f1b2e48SStefano Zampini const PetscInt *idxs; 47024f1b2e48SStefano Zampini PetscInt nl,j; 47034f1b2e48SStefano Zampini 47044f1b2e48SStefano Zampini ierr = ISGetLocalSize(pcbddc->local_subs[i],&nl);CHKERRQ(ierr); 47054f1b2e48SStefano Zampini ierr = ISGetIndices(pcbddc->local_subs[i],&idxs);CHKERRQ(ierr); 47064f1b2e48SStefano Zampini for (j=0;j<nl;j++) { 47074f1b2e48SStefano Zampini local_subs[idxs[j]] = i; 47084f1b2e48SStefano Zampini } 47094f1b2e48SStefano Zampini ierr = ISRestoreIndices(pcbddc->local_subs[i],&idxs);CHKERRQ(ierr); 47104f1b2e48SStefano Zampini } 47114f1b2e48SStefano Zampini pcbddc->mat_graph->n_local_subs = pcbddc->n_local_subs; 47124f1b2e48SStefano Zampini pcbddc->mat_graph->local_subs = local_subs; 47134f1b2e48SStefano Zampini } 47144f1b2e48SStefano Zampini 4715674ae819SStefano Zampini /* Graph's connected components analysis */ 4716674ae819SStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(pcbddc->mat_graph);CHKERRQ(ierr); 4717674ae819SStefano Zampini PetscFunctionReturn(0); 4718674ae819SStefano Zampini } 4719674ae819SStefano Zampini 4720dc456d91SStefano Zampini /* given an index sets possibly with holes, renumbers the indexes removing the holes */ 4721674ae819SStefano Zampini #undef __FUNCT__ 4722674ae819SStefano Zampini #define __FUNCT__ "PCBDDCSubsetNumbering" 4723dc456d91SStefano Zampini PetscErrorCode PCBDDCSubsetNumbering(IS subset, IS subset_mult, PetscInt *N_n, IS *subset_n) 4724674ae819SStefano Zampini { 4725dc456d91SStefano Zampini PetscSF sf; 4726dc456d91SStefano Zampini PetscLayout map; 4727dc456d91SStefano Zampini const PetscInt *idxs; 4728dc456d91SStefano Zampini PetscInt *leaf_data,*root_data,*gidxs; 4729dc456d91SStefano Zampini PetscInt N,n,i,lbounds[2],gbounds[2],Nl; 4730dc456d91SStefano Zampini PetscInt n_n,nlocals,start,first_index; 4731dc456d91SStefano Zampini PetscMPIInt commsize; 4732674ae819SStefano Zampini PetscBool first_found; 4733674ae819SStefano Zampini PetscErrorCode ierr; 4734674ae819SStefano Zampini 4735674ae819SStefano Zampini PetscFunctionBegin; 4736dc456d91SStefano Zampini ierr = ISGetLocalSize(subset,&n);CHKERRQ(ierr); 4737dc456d91SStefano Zampini if (subset_mult) { 4738dc456d91SStefano Zampini PetscCheckSameComm(subset,1,subset_mult,2); 4739dc456d91SStefano Zampini ierr = ISGetLocalSize(subset,&i);CHKERRQ(ierr); 4740dc456d91SStefano Zampini if (i != n) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Local subset and multiplicity sizes don't match! %d != %d",n,i); 4741674ae819SStefano Zampini } 4742dc456d91SStefano Zampini /* create workspace layout for computing global indices of subset */ 4743dc456d91SStefano Zampini ierr = ISGetIndices(subset,&idxs);CHKERRQ(ierr); 4744dc456d91SStefano Zampini lbounds[0] = lbounds[1] = 0; 4745dc456d91SStefano Zampini for (i=0;i<n;i++) { 4746dc456d91SStefano Zampini if (idxs[i] < lbounds[0]) lbounds[0] = idxs[i]; 4747dc456d91SStefano Zampini else if (idxs[i] > lbounds[1]) lbounds[1] = idxs[i]; 4748674ae819SStefano Zampini } 4749dc456d91SStefano Zampini lbounds[0] = -lbounds[0]; 4750b2566f29SBarry Smith ierr = MPIU_Allreduce(lbounds,gbounds,2,MPIU_INT,MPI_MAX,PetscObjectComm((PetscObject)subset));CHKERRQ(ierr); 4751dc456d91SStefano Zampini gbounds[0] = -gbounds[0]; 4752dc456d91SStefano Zampini N = gbounds[1] - gbounds[0] + 1; 4753dc456d91SStefano Zampini ierr = PetscLayoutCreate(PetscObjectComm((PetscObject)subset),&map);CHKERRQ(ierr); 4754dc456d91SStefano Zampini ierr = PetscLayoutSetBlockSize(map,1);CHKERRQ(ierr); 4755dc456d91SStefano Zampini ierr = PetscLayoutSetSize(map,N);CHKERRQ(ierr); 4756dc456d91SStefano Zampini ierr = PetscLayoutSetUp(map);CHKERRQ(ierr); 4757dc456d91SStefano Zampini ierr = PetscLayoutGetLocalSize(map,&Nl);CHKERRQ(ierr); 4758dc456d91SStefano Zampini 4759dc456d91SStefano Zampini /* create sf : leaf_data == multiplicity of indexes, root data == global index in layout */ 4760dc456d91SStefano Zampini ierr = PetscMalloc2(n,&leaf_data,Nl,&root_data);CHKERRQ(ierr); 4761dc456d91SStefano Zampini if (subset_mult) { 4762dc456d91SStefano Zampini const PetscInt* idxs_mult; 4763dc456d91SStefano Zampini 4764dc456d91SStefano Zampini ierr = ISGetIndices(subset_mult,&idxs_mult);CHKERRQ(ierr); 4765dc456d91SStefano Zampini ierr = PetscMemcpy(leaf_data,idxs_mult,n*sizeof(PetscInt));CHKERRQ(ierr); 4766dc456d91SStefano Zampini ierr = ISRestoreIndices(subset_mult,&idxs_mult);CHKERRQ(ierr); 4767674ae819SStefano Zampini } else { 4768dc456d91SStefano Zampini for (i=0;i<n;i++) leaf_data[i] = 1; 4769674ae819SStefano Zampini } 4770dc456d91SStefano Zampini /* local size of new subset */ 4771dc456d91SStefano Zampini n_n = 0; 4772dc456d91SStefano Zampini for (i=0;i<n;i++) n_n += leaf_data[i]; 4773dc456d91SStefano Zampini 4774dc456d91SStefano Zampini /* global indexes in layout */ 4775dc456d91SStefano Zampini ierr = PetscMalloc1(n_n,&gidxs);CHKERRQ(ierr); /* allocating possibly extra space in gidxs which will be used later */ 4776dc456d91SStefano Zampini for (i=0;i<n;i++) gidxs[i] = idxs[i] - gbounds[0]; 4777dc456d91SStefano Zampini ierr = ISRestoreIndices(subset,&idxs);CHKERRQ(ierr); 4778dc456d91SStefano Zampini ierr = PetscSFCreate(PetscObjectComm((PetscObject)subset),&sf);CHKERRQ(ierr); 4779dc456d91SStefano Zampini ierr = PetscSFSetGraphLayout(sf,map,n,NULL,PETSC_COPY_VALUES,gidxs);CHKERRQ(ierr); 4780dc456d91SStefano Zampini ierr = PetscLayoutDestroy(&map);CHKERRQ(ierr); 4781dc456d91SStefano Zampini 4782dc456d91SStefano Zampini /* reduce from leaves to roots */ 4783dc456d91SStefano Zampini ierr = PetscMemzero(root_data,Nl*sizeof(PetscInt));CHKERRQ(ierr); 478464a8e5bcSStefano Zampini ierr = PetscSFReduceBegin(sf,MPIU_INT,leaf_data,root_data,MPI_MAX);CHKERRQ(ierr); 478564a8e5bcSStefano Zampini ierr = PetscSFReduceEnd(sf,MPIU_INT,leaf_data,root_data,MPI_MAX);CHKERRQ(ierr); 4786dc456d91SStefano Zampini 4787dc456d91SStefano Zampini /* count indexes in local part of layout */ 4788674ae819SStefano Zampini nlocals = 0; 4789674ae819SStefano Zampini first_index = -1; 4790674ae819SStefano Zampini first_found = PETSC_FALSE; 4791dc456d91SStefano Zampini for (i=0;i<Nl;i++) { 4792dc456d91SStefano Zampini if (!first_found && root_data[i]) { 4793674ae819SStefano Zampini first_found = PETSC_TRUE; 4794674ae819SStefano Zampini first_index = i; 4795674ae819SStefano Zampini } 4796dc456d91SStefano Zampini nlocals += root_data[i]; 4797674ae819SStefano Zampini } 4798dc456d91SStefano Zampini 4799dc456d91SStefano Zampini /* cumulative of number of indexes and size of subset without holes */ 48005fa240b1SStefano Zampini #if defined(PETSC_HAVE_MPI_EXSCAN) 4801dc456d91SStefano Zampini start = 0; 480264a8e5bcSStefano Zampini ierr = MPI_Exscan(&nlocals,&start,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)subset));CHKERRQ(ierr); 48035fa240b1SStefano Zampini #else 480464a8e5bcSStefano Zampini ierr = MPI_Scan(&nlocals,&start,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)subset));CHKERRQ(ierr); 48055fa240b1SStefano Zampini start = start-nlocals; 48065fa240b1SStefano Zampini #endif 48075fa240b1SStefano Zampini 4808dc456d91SStefano Zampini if (N_n) { /* compute total size of new subset if requested */ 4809dc456d91SStefano Zampini *N_n = start + nlocals; 4810dc456d91SStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)subset),&commsize);CHKERRQ(ierr); 4811dc456d91SStefano Zampini ierr = MPI_Bcast(N_n,1,MPIU_INT,commsize-1,PetscObjectComm((PetscObject)subset));CHKERRQ(ierr); 4812674ae819SStefano Zampini } 48135fa240b1SStefano Zampini 48145fa240b1SStefano Zampini /* adapt root data with cumulative */ 4815674ae819SStefano Zampini if (first_found) { 4816dc456d91SStefano Zampini PetscInt old_index; 4817dc456d91SStefano Zampini 4818dc456d91SStefano Zampini root_data[first_index] += start; 4819674ae819SStefano Zampini old_index = first_index; 4820dc456d91SStefano Zampini for (i=first_index+1;i<Nl;i++) { 4821dc456d91SStefano Zampini if (root_data[i]) { 4822dc456d91SStefano Zampini root_data[i] += root_data[old_index]; 4823674ae819SStefano Zampini old_index = i; 4824674ae819SStefano Zampini } 4825674ae819SStefano Zampini } 4826674ae819SStefano Zampini } 4827dc456d91SStefano Zampini 4828dc456d91SStefano Zampini /* from roots to leaves */ 4829dc456d91SStefano Zampini ierr = PetscSFBcastBegin(sf,MPIU_INT,root_data,leaf_data);CHKERRQ(ierr); 4830dc456d91SStefano Zampini ierr = PetscSFBcastEnd(sf,MPIU_INT,root_data,leaf_data);CHKERRQ(ierr); 4831dc456d91SStefano Zampini ierr = PetscSFDestroy(&sf);CHKERRQ(ierr); 4832dc456d91SStefano Zampini 4833dc456d91SStefano Zampini /* create new IS with global indexes without holes */ 4834dc456d91SStefano Zampini if (subset_mult) { 4835dc456d91SStefano Zampini const PetscInt* idxs_mult; 4836dc456d91SStefano Zampini PetscInt cum; 4837dc456d91SStefano Zampini 4838dc456d91SStefano Zampini cum = 0; 4839dc456d91SStefano Zampini ierr = ISGetIndices(subset_mult,&idxs_mult);CHKERRQ(ierr); 4840dc456d91SStefano Zampini for (i=0;i<n;i++) { 4841dc456d91SStefano Zampini PetscInt j; 4842dc456d91SStefano Zampini for (j=0;j<idxs_mult[i];j++) gidxs[cum++] = leaf_data[i] - idxs_mult[i] + j; 4843674ae819SStefano Zampini } 4844dc456d91SStefano Zampini ierr = ISRestoreIndices(subset_mult,&idxs_mult);CHKERRQ(ierr); 4845674ae819SStefano Zampini } else { 4846dc456d91SStefano Zampini for (i=0;i<n;i++) { 4847dc456d91SStefano Zampini gidxs[i] = leaf_data[i]-1; 4848674ae819SStefano Zampini } 4849674ae819SStefano Zampini } 4850dc456d91SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)subset),n_n,gidxs,PETSC_OWN_POINTER,subset_n);CHKERRQ(ierr); 4851dc456d91SStefano Zampini ierr = PetscFree2(leaf_data,root_data);CHKERRQ(ierr); 4852674ae819SStefano Zampini PetscFunctionReturn(0); 4853674ae819SStefano Zampini } 48549a7d3425SStefano Zampini 48559a7d3425SStefano Zampini #undef __FUNCT__ 48569a7d3425SStefano Zampini #define __FUNCT__ "PCBDDCOrthonormalizeVecs" 48579a7d3425SStefano Zampini PetscErrorCode PCBDDCOrthonormalizeVecs(PetscInt n, Vec vecs[]) 48589a7d3425SStefano Zampini { 48599a7d3425SStefano Zampini PetscInt i,j; 48609a7d3425SStefano Zampini PetscScalar *alphas; 48619a7d3425SStefano Zampini PetscErrorCode ierr; 48629a7d3425SStefano Zampini 48639a7d3425SStefano Zampini PetscFunctionBegin; 48649a7d3425SStefano Zampini /* this implements stabilized Gram-Schmidt */ 4865785e854fSJed Brown ierr = PetscMalloc1(n,&alphas);CHKERRQ(ierr); 48669a7d3425SStefano Zampini for (i=0;i<n;i++) { 48679a7d3425SStefano Zampini ierr = VecNormalize(vecs[i],NULL);CHKERRQ(ierr); 48689a7d3425SStefano Zampini if (i<n) { ierr = VecMDot(vecs[i],n-i-1,&vecs[i+1],&alphas[i+1]);CHKERRQ(ierr); } 48699a7d3425SStefano Zampini for (j=i+1;j<n;j++) { ierr = VecAXPY(vecs[j],PetscConj(-alphas[j]),vecs[i]);CHKERRQ(ierr); } 48709a7d3425SStefano Zampini } 48719a7d3425SStefano Zampini ierr = PetscFree(alphas);CHKERRQ(ierr); 48729a7d3425SStefano Zampini PetscFunctionReturn(0); 48739a7d3425SStefano Zampini } 48749a7d3425SStefano Zampini 4875e7931f94SStefano Zampini #undef __FUNCT__ 487670cf5478SStefano Zampini #define __FUNCT__ "MatISGetSubassemblingPattern" 487757de7509SStefano Zampini PetscErrorCode MatISGetSubassemblingPattern(Mat mat, PetscInt *n_subdomains, PetscInt redprocs, IS* is_sends, PetscBool *have_void) 4878e7931f94SStefano Zampini { 487957de7509SStefano Zampini Mat A; 4880e7931f94SStefano Zampini PetscInt n_neighs,*neighs,*n_shared,**shared; 4881e7931f94SStefano Zampini PetscMPIInt size,rank,color; 488252e5ac9dSStefano Zampini PetscInt *xadj,*adjncy; 488352e5ac9dSStefano Zampini PetscInt *adjncy_wgt,*v_wgt,*ranks_send_to_idx; 488427b6a85dSStefano Zampini PetscInt im_active,active_procs,n,i,j,local_size,threshold = 2; 488557de7509SStefano Zampini PetscInt void_procs,*procs_candidates = NULL; 488627b6a85dSStefano Zampini PetscInt xadj_count, *count; 488727b6a85dSStefano Zampini PetscBool ismatis,use_vwgt=PETSC_FALSE; 488827b6a85dSStefano Zampini PetscSubcomm psubcomm; 488927b6a85dSStefano Zampini MPI_Comm subcomm; 489052e5ac9dSStefano Zampini PetscErrorCode ierr; 4891a57a6d2fSStefano Zampini 4892e7931f94SStefano Zampini PetscFunctionBegin; 489357de7509SStefano Zampini PetscValidHeaderSpecific(mat,MAT_CLASSID,1); 489457de7509SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)mat,MATIS,&ismatis);CHKERRQ(ierr); 489557de7509SStefano Zampini if (!ismatis) SETERRQ1(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot use %s on a matrix object which is not of type MATIS",__FUNCT__); 489657de7509SStefano Zampini PetscValidLogicalCollectiveInt(mat,*n_subdomains,2); 489757de7509SStefano Zampini PetscValidLogicalCollectiveInt(mat,redprocs,3); 489857de7509SStefano Zampini if (*n_subdomains <=0) SETERRQ1(PetscObjectComm((PetscObject)mat),PETSC_ERR_ARG_WRONG,"Invalid number of subdomains requested %d\n",*n_subdomains); 489957de7509SStefano Zampini 490057de7509SStefano Zampini if (have_void) *have_void = PETSC_FALSE; 490157de7509SStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)mat),&size);CHKERRQ(ierr); 490257de7509SStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)mat),&rank);CHKERRQ(ierr); 490357de7509SStefano Zampini ierr = MatISGetLocalMat(mat,&A);CHKERRQ(ierr); 490457de7509SStefano Zampini ierr = MatGetLocalSize(A,&n,NULL);CHKERRQ(ierr); 490557de7509SStefano Zampini im_active = !!(n); 490657de7509SStefano Zampini ierr = MPIU_Allreduce(&im_active,&active_procs,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 490757de7509SStefano Zampini void_procs = size - active_procs; 490857de7509SStefano Zampini /* get ranks of of non-active processes in mat communicator */ 490957de7509SStefano Zampini if (void_procs) { 491057de7509SStefano Zampini PetscInt ncand; 491157de7509SStefano Zampini 491257de7509SStefano Zampini if (have_void) *have_void = PETSC_TRUE; 491357de7509SStefano Zampini ierr = PetscMalloc1(size,&procs_candidates);CHKERRQ(ierr); 491457de7509SStefano Zampini ierr = MPI_Allgather(&im_active,1,MPIU_INT,procs_candidates,1,MPIU_INT,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 491557de7509SStefano Zampini for (i=0,ncand=0;i<size;i++) { 491657de7509SStefano Zampini if (!procs_candidates[i]) { 491757de7509SStefano Zampini procs_candidates[ncand++] = i; 491857de7509SStefano Zampini } 491957de7509SStefano Zampini } 492057de7509SStefano Zampini /* force n_subdomains to be not greater that the number of non-active processes */ 492157de7509SStefano Zampini *n_subdomains = PetscMin(void_procs,*n_subdomains); 492257de7509SStefano Zampini } 492357de7509SStefano Zampini 492457de7509SStefano Zampini /* number of subdomains requested greater than active processes -> just shift the matrix */ 492557de7509SStefano Zampini if (active_procs < *n_subdomains) { 492657de7509SStefano Zampini PetscInt issize,isidx; 492757de7509SStefano Zampini if (im_active) { 492857de7509SStefano Zampini issize = 1; 492957de7509SStefano Zampini if (procs_candidates) { /* shift the pattern on non-active candidates (if any) */ 493057de7509SStefano Zampini isidx = procs_candidates[rank]; 493157de7509SStefano Zampini } else { 493257de7509SStefano Zampini isidx = rank; 493357de7509SStefano Zampini } 493457de7509SStefano Zampini } else { 493557de7509SStefano Zampini issize = 0; 493657de7509SStefano Zampini isidx = -1; 493757de7509SStefano Zampini } 493857de7509SStefano Zampini *n_subdomains = active_procs; 493957de7509SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)mat),issize,&isidx,PETSC_COPY_VALUES,is_sends);CHKERRQ(ierr); 4940daf8a457SStefano Zampini ierr = PetscFree(procs_candidates);CHKERRQ(ierr); 494157de7509SStefano Zampini PetscFunctionReturn(0); 494257de7509SStefano Zampini } 4943c5929fdfSBarry Smith ierr = PetscOptionsGetBool(NULL,NULL,"-matis_partitioning_use_vwgt",&use_vwgt,NULL);CHKERRQ(ierr); 4944c5929fdfSBarry Smith ierr = PetscOptionsGetInt(NULL,NULL,"-matis_partitioning_threshold",&threshold,NULL);CHKERRQ(ierr); 494527b6a85dSStefano Zampini threshold = PetscMax(threshold,2); 4946e7931f94SStefano Zampini 4947e7931f94SStefano Zampini /* Get info on mapping */ 49483bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetSize(mat->rmap->mapping,&local_size);CHKERRQ(ierr); 49493bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetInfo(mat->rmap->mapping,&n_neighs,&neighs,&n_shared,&shared);CHKERRQ(ierr); 4950e7931f94SStefano Zampini 4951e7931f94SStefano Zampini /* build local CSR graph of subdomains' connectivity */ 4952785e854fSJed Brown ierr = PetscMalloc1(2,&xadj);CHKERRQ(ierr); 4953e7931f94SStefano Zampini xadj[0] = 0; 4954e7931f94SStefano Zampini xadj[1] = PetscMax(n_neighs-1,0); 4955785e854fSJed Brown ierr = PetscMalloc1(xadj[1],&adjncy);CHKERRQ(ierr); 4956785e854fSJed Brown ierr = PetscMalloc1(xadj[1],&adjncy_wgt);CHKERRQ(ierr); 495727b6a85dSStefano Zampini ierr = PetscCalloc1(local_size,&count);CHKERRQ(ierr); 495827b6a85dSStefano Zampini for (i=1;i<n_neighs;i++) 495927b6a85dSStefano Zampini for (j=0;j<n_shared[i];j++) 496027b6a85dSStefano Zampini count[shared[i][j]] += 1; 4961e7931f94SStefano Zampini 496227b6a85dSStefano Zampini xadj_count = 0; 49632b510759SStefano Zampini for (i=1;i<n_neighs;i++) { 496427b6a85dSStefano Zampini for (j=0;j<n_shared[i];j++) { 496527b6a85dSStefano Zampini if (count[shared[i][j]] < threshold) { 4966d023bfaeSStefano Zampini adjncy[xadj_count] = neighs[i]; 4967d023bfaeSStefano Zampini adjncy_wgt[xadj_count] = n_shared[i]; 4968d023bfaeSStefano Zampini xadj_count++; 496927b6a85dSStefano Zampini break; 497027b6a85dSStefano Zampini } 4971e7931f94SStefano Zampini } 4972e7931f94SStefano Zampini } 4973d023bfaeSStefano Zampini xadj[1] = xadj_count; 497427b6a85dSStefano Zampini ierr = PetscFree(count);CHKERRQ(ierr); 49753bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingRestoreInfo(mat->rmap->mapping,&n_neighs,&neighs,&n_shared,&shared);CHKERRQ(ierr); 4976e7931f94SStefano Zampini ierr = PetscSortIntWithArray(xadj[1],adjncy,adjncy_wgt);CHKERRQ(ierr); 4977e7931f94SStefano Zampini 49783837a79fSStefano Zampini ierr = PetscMalloc1(1,&ranks_send_to_idx);CHKERRQ(ierr); 4979e7931f94SStefano Zampini 498027b6a85dSStefano Zampini /* Restrict work on active processes only */ 498127b6a85dSStefano Zampini ierr = PetscMPIIntCast(im_active,&color);CHKERRQ(ierr); 498227b6a85dSStefano Zampini if (void_procs) { 498327b6a85dSStefano Zampini ierr = PetscSubcommCreate(PetscObjectComm((PetscObject)mat),&psubcomm);CHKERRQ(ierr); 498427b6a85dSStefano Zampini ierr = PetscSubcommSetNumber(psubcomm,2);CHKERRQ(ierr); /* 2 groups, active process and not active processes */ 498527b6a85dSStefano Zampini ierr = PetscSubcommSetTypeGeneral(psubcomm,color,rank);CHKERRQ(ierr); 498627b6a85dSStefano Zampini subcomm = PetscSubcommChild(psubcomm); 498727b6a85dSStefano Zampini } else { 498827b6a85dSStefano Zampini psubcomm = NULL; 498927b6a85dSStefano Zampini subcomm = PetscObjectComm((PetscObject)mat); 499027b6a85dSStefano Zampini } 499127b6a85dSStefano Zampini 499227b6a85dSStefano Zampini v_wgt = NULL; 499327b6a85dSStefano Zampini if (!color) { 4994e7931f94SStefano Zampini ierr = PetscFree(xadj);CHKERRQ(ierr); 4995e7931f94SStefano Zampini ierr = PetscFree(adjncy);CHKERRQ(ierr); 4996e7931f94SStefano Zampini ierr = PetscFree(adjncy_wgt);CHKERRQ(ierr); 4997c8587f34SStefano Zampini } else { 499852e5ac9dSStefano Zampini Mat subdomain_adj; 499952e5ac9dSStefano Zampini IS new_ranks,new_ranks_contig; 500052e5ac9dSStefano Zampini MatPartitioning partitioner; 500127b6a85dSStefano Zampini PetscInt rstart=0,rend=0; 500252e5ac9dSStefano Zampini PetscInt *is_indices,*oldranks; 500357de7509SStefano Zampini PetscMPIInt size; 5004b0c7d250SStefano Zampini PetscBool aggregate; 5005b0c7d250SStefano Zampini 500627b6a85dSStefano Zampini ierr = MPI_Comm_size(subcomm,&size);CHKERRQ(ierr); 500727b6a85dSStefano Zampini if (void_procs) { 500827b6a85dSStefano Zampini PetscInt prank = rank; 5009785e854fSJed Brown ierr = PetscMalloc1(size,&oldranks);CHKERRQ(ierr); 501027b6a85dSStefano Zampini ierr = MPI_Allgather(&prank,1,MPIU_INT,oldranks,1,MPIU_INT,subcomm);CHKERRQ(ierr); 5011e7931f94SStefano Zampini for (i=0;i<xadj[1];i++) { 5012e7931f94SStefano Zampini ierr = PetscFindInt(adjncy[i],size,oldranks,&adjncy[i]);CHKERRQ(ierr); 5013c8587f34SStefano Zampini } 5014e7931f94SStefano Zampini ierr = PetscSortIntWithArray(xadj[1],adjncy,adjncy_wgt);CHKERRQ(ierr); 501527b6a85dSStefano Zampini } else { 501627b6a85dSStefano Zampini oldranks = NULL; 501727b6a85dSStefano Zampini } 5018b0c7d250SStefano Zampini aggregate = ((redprocs > 0 && redprocs < size) ? PETSC_TRUE : PETSC_FALSE); 501927b6a85dSStefano Zampini if (aggregate) { /* TODO: all this part could be made more efficient */ 5020b0c7d250SStefano Zampini PetscInt lrows,row,ncols,*cols; 5021b0c7d250SStefano Zampini PetscMPIInt nrank; 5022b0c7d250SStefano Zampini PetscScalar *vals; 5023b0c7d250SStefano Zampini 502427b6a85dSStefano Zampini ierr = MPI_Comm_rank(subcomm,&nrank);CHKERRQ(ierr); 5025b0c7d250SStefano Zampini lrows = 0; 5026b0c7d250SStefano Zampini if (nrank<redprocs) { 5027b0c7d250SStefano Zampini lrows = size/redprocs; 5028b0c7d250SStefano Zampini if (nrank<size%redprocs) lrows++; 5029b0c7d250SStefano Zampini } 503027b6a85dSStefano Zampini ierr = MatCreateAIJ(subcomm,lrows,lrows,size,size,50,NULL,50,NULL,&subdomain_adj);CHKERRQ(ierr); 5031b0c7d250SStefano Zampini ierr = MatGetOwnershipRange(subdomain_adj,&rstart,&rend);CHKERRQ(ierr); 5032b0c7d250SStefano Zampini ierr = MatSetOption(subdomain_adj,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_FALSE);CHKERRQ(ierr); 5033b0c7d250SStefano Zampini ierr = MatSetOption(subdomain_adj,MAT_NEW_NONZERO_ALLOCATION_ERR,PETSC_FALSE);CHKERRQ(ierr); 5034b0c7d250SStefano Zampini row = nrank; 5035b0c7d250SStefano Zampini ncols = xadj[1]-xadj[0]; 5036b0c7d250SStefano Zampini cols = adjncy; 5037b0c7d250SStefano Zampini ierr = PetscMalloc1(ncols,&vals);CHKERRQ(ierr); 5038b0c7d250SStefano Zampini for (i=0;i<ncols;i++) vals[i] = adjncy_wgt[i]; 5039b0c7d250SStefano Zampini ierr = MatSetValues(subdomain_adj,1,&row,ncols,cols,vals,INSERT_VALUES);CHKERRQ(ierr); 5040b0c7d250SStefano Zampini ierr = MatAssemblyBegin(subdomain_adj,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5041b0c7d250SStefano Zampini ierr = MatAssemblyEnd(subdomain_adj,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 504252e5ac9dSStefano Zampini ierr = PetscFree(xadj);CHKERRQ(ierr); 504352e5ac9dSStefano Zampini ierr = PetscFree(adjncy);CHKERRQ(ierr); 504452e5ac9dSStefano Zampini ierr = PetscFree(adjncy_wgt);CHKERRQ(ierr); 5045b0c7d250SStefano Zampini ierr = PetscFree(vals);CHKERRQ(ierr); 504627b6a85dSStefano Zampini if (use_vwgt) { 504727b6a85dSStefano Zampini Vec v; 504827b6a85dSStefano Zampini const PetscScalar *array; 504927b6a85dSStefano Zampini PetscInt nl; 505027b6a85dSStefano Zampini 505127b6a85dSStefano Zampini ierr = MatCreateVecs(subdomain_adj,&v,NULL);CHKERRQ(ierr); 505227b6a85dSStefano Zampini ierr = VecSetValue(v,row,(PetscScalar)local_size,INSERT_VALUES);CHKERRQ(ierr); 505327b6a85dSStefano Zampini ierr = VecAssemblyBegin(v);CHKERRQ(ierr); 505427b6a85dSStefano Zampini ierr = VecAssemblyEnd(v);CHKERRQ(ierr); 505527b6a85dSStefano Zampini ierr = VecGetLocalSize(v,&nl);CHKERRQ(ierr); 505627b6a85dSStefano Zampini ierr = VecGetArrayRead(v,&array);CHKERRQ(ierr); 505727b6a85dSStefano Zampini ierr = PetscMalloc1(nl,&v_wgt);CHKERRQ(ierr); 505822db5ddcSStefano Zampini for (i=0;i<nl;i++) v_wgt[i] = (PetscInt)PetscRealPart(array[i]); 505927b6a85dSStefano Zampini ierr = VecRestoreArrayRead(v,&array);CHKERRQ(ierr); 506027b6a85dSStefano Zampini ierr = VecDestroy(&v);CHKERRQ(ierr); 506127b6a85dSStefano Zampini } 5062b0c7d250SStefano Zampini } else { 506327b6a85dSStefano Zampini ierr = MatCreateMPIAdj(subcomm,1,(PetscInt)size,xadj,adjncy,adjncy_wgt,&subdomain_adj);CHKERRQ(ierr); 506427b6a85dSStefano Zampini if (use_vwgt) { 506527b6a85dSStefano Zampini ierr = PetscMalloc1(1,&v_wgt);CHKERRQ(ierr); 506627b6a85dSStefano Zampini v_wgt[0] = local_size; 506727b6a85dSStefano Zampini } 5068b0c7d250SStefano Zampini } 506922b6e8a2SStefano Zampini /* ierr = MatView(subdomain_adj,0);CHKERRQ(ierr); */ 5070e7931f94SStefano Zampini 5071e7931f94SStefano Zampini /* Partition */ 507227b6a85dSStefano Zampini ierr = MatPartitioningCreate(subcomm,&partitioner);CHKERRQ(ierr); 5073e7931f94SStefano Zampini ierr = MatPartitioningSetAdjacency(partitioner,subdomain_adj);CHKERRQ(ierr); 507427b6a85dSStefano Zampini if (v_wgt) { 5075e7931f94SStefano Zampini ierr = MatPartitioningSetVertexWeights(partitioner,v_wgt);CHKERRQ(ierr); 5076c8587f34SStefano Zampini } 507757de7509SStefano Zampini *n_subdomains = PetscMin((PetscInt)size,*n_subdomains); 507857de7509SStefano Zampini ierr = MatPartitioningSetNParts(partitioner,*n_subdomains);CHKERRQ(ierr); 5079e7931f94SStefano Zampini ierr = MatPartitioningSetFromOptions(partitioner);CHKERRQ(ierr); 5080e7931f94SStefano Zampini ierr = MatPartitioningApply(partitioner,&new_ranks);CHKERRQ(ierr); 508122b6e8a2SStefano Zampini /* ierr = MatPartitioningView(partitioner,0);CHKERRQ(ierr); */ 5082e7931f94SStefano Zampini 508352e5ac9dSStefano Zampini /* renumber new_ranks to avoid "holes" in new set of processors */ 508452e5ac9dSStefano Zampini ierr = PCBDDCSubsetNumbering(new_ranks,NULL,NULL,&new_ranks_contig);CHKERRQ(ierr); 508552e5ac9dSStefano Zampini ierr = ISDestroy(&new_ranks);CHKERRQ(ierr); 508652e5ac9dSStefano Zampini ierr = ISGetIndices(new_ranks_contig,(const PetscInt**)&is_indices);CHKERRQ(ierr); 508757de7509SStefano Zampini if (!aggregate) { 508857de7509SStefano Zampini if (procs_candidates) { /* shift the pattern on non-active candidates (if any) */ 508927b6a85dSStefano Zampini #if defined(PETSC_USE_DEBUG) 509027b6a85dSStefano Zampini if (!oldranks) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"This should not happen"); 509127b6a85dSStefano Zampini #endif 509257de7509SStefano Zampini ranks_send_to_idx[0] = procs_candidates[oldranks[is_indices[0]]]; 509327b6a85dSStefano Zampini } else if (oldranks) { 5094b0c7d250SStefano Zampini ranks_send_to_idx[0] = oldranks[is_indices[0]]; 509527b6a85dSStefano Zampini } else { 509627b6a85dSStefano Zampini ranks_send_to_idx[0] = is_indices[0]; 509757de7509SStefano Zampini } 509828143c3dSStefano Zampini } else { 5099b0c7d250SStefano Zampini PetscInt idxs[1]; 5100b0c7d250SStefano Zampini PetscMPIInt tag; 5101b0c7d250SStefano Zampini MPI_Request *reqs; 5102b0c7d250SStefano Zampini 5103b0c7d250SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)subdomain_adj,&tag);CHKERRQ(ierr); 5104b0c7d250SStefano Zampini ierr = PetscMalloc1(rend-rstart,&reqs);CHKERRQ(ierr); 5105b0c7d250SStefano Zampini for (i=rstart;i<rend;i++) { 510627b6a85dSStefano Zampini ierr = MPI_Isend(is_indices+i-rstart,1,MPIU_INT,i,tag,subcomm,&reqs[i-rstart]);CHKERRQ(ierr); 510728143c3dSStefano Zampini } 510827b6a85dSStefano Zampini ierr = MPI_Recv(idxs,1,MPIU_INT,MPI_ANY_SOURCE,tag,subcomm,MPI_STATUS_IGNORE);CHKERRQ(ierr); 5109b0c7d250SStefano Zampini ierr = MPI_Waitall(rend-rstart,reqs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 5110b0c7d250SStefano Zampini ierr = PetscFree(reqs);CHKERRQ(ierr); 511157de7509SStefano Zampini if (procs_candidates) { /* shift the pattern on non-active candidates (if any) */ 511227b6a85dSStefano Zampini #if defined(PETSC_USE_DEBUG) 511327b6a85dSStefano Zampini if (!oldranks) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"This should not happen"); 511427b6a85dSStefano Zampini #endif 511557de7509SStefano Zampini ranks_send_to_idx[0] = procs_candidates[oldranks[idxs[0]]]; 511627b6a85dSStefano Zampini } else if (oldranks) { 5117b0c7d250SStefano Zampini ranks_send_to_idx[0] = oldranks[idxs[0]]; 511827b6a85dSStefano Zampini } else { 511927b6a85dSStefano Zampini ranks_send_to_idx[0] = idxs[0]; 5120e7931f94SStefano Zampini } 512157de7509SStefano Zampini } 512252e5ac9dSStefano Zampini ierr = ISRestoreIndices(new_ranks_contig,(const PetscInt**)&is_indices);CHKERRQ(ierr); 5123e7931f94SStefano Zampini /* clean up */ 5124e7931f94SStefano Zampini ierr = PetscFree(oldranks);CHKERRQ(ierr); 512552e5ac9dSStefano Zampini ierr = ISDestroy(&new_ranks_contig);CHKERRQ(ierr); 5126e7931f94SStefano Zampini ierr = MatDestroy(&subdomain_adj);CHKERRQ(ierr); 5127e7931f94SStefano Zampini ierr = MatPartitioningDestroy(&partitioner);CHKERRQ(ierr); 5128e7931f94SStefano Zampini } 512927b6a85dSStefano Zampini ierr = PetscSubcommDestroy(&psubcomm);CHKERRQ(ierr); 513057de7509SStefano Zampini ierr = PetscFree(procs_candidates);CHKERRQ(ierr); 5131e7931f94SStefano Zampini 5132e7931f94SStefano Zampini /* assemble parallel IS for sends */ 5133e7931f94SStefano Zampini i = 1; 513427b6a85dSStefano Zampini if (!color) i=0; 513557de7509SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)mat),i,ranks_send_to_idx,PETSC_OWN_POINTER,is_sends);CHKERRQ(ierr); 5136e7931f94SStefano Zampini PetscFunctionReturn(0); 5137e7931f94SStefano Zampini } 5138e7931f94SStefano Zampini 5139e7931f94SStefano Zampini typedef enum {MATDENSE_PRIVATE=0,MATAIJ_PRIVATE,MATBAIJ_PRIVATE,MATSBAIJ_PRIVATE}MatTypePrivate; 5140e7931f94SStefano Zampini 5141e7931f94SStefano Zampini #undef __FUNCT__ 5142e7931f94SStefano Zampini #define __FUNCT__ "MatISSubassemble" 514357de7509SStefano Zampini PetscErrorCode MatISSubassemble(Mat mat, IS is_sends, PetscInt n_subdomains, PetscBool restrict_comm, PetscBool restrict_full, PetscBool reuse, Mat *mat_n, PetscInt nis, IS isarray[]) 5144e7931f94SStefano Zampini { 514570cf5478SStefano Zampini Mat local_mat; 5146e7931f94SStefano Zampini IS is_sends_internal; 51479d30be91SStefano Zampini PetscInt rows,cols,new_local_rows; 514828143c3dSStefano Zampini PetscInt i,bs,buf_size_idxs,buf_size_idxs_is,buf_size_vals; 51499d30be91SStefano Zampini PetscBool ismatis,isdense,newisdense,destroy_mat; 5150e7931f94SStefano Zampini ISLocalToGlobalMapping l2gmap; 5151e7931f94SStefano Zampini PetscInt* l2gmap_indices; 5152e7931f94SStefano Zampini const PetscInt* is_indices; 5153e7931f94SStefano Zampini MatType new_local_type; 5154e7931f94SStefano Zampini /* buffers */ 5155e7931f94SStefano Zampini PetscInt *ptr_idxs,*send_buffer_idxs,*recv_buffer_idxs; 515628143c3dSStefano Zampini PetscInt *ptr_idxs_is,*send_buffer_idxs_is,*recv_buffer_idxs_is; 51579d30be91SStefano Zampini PetscInt *recv_buffer_idxs_local; 5158e7931f94SStefano Zampini PetscScalar *ptr_vals,*send_buffer_vals,*recv_buffer_vals; 5159e7931f94SStefano Zampini /* MPI */ 516028143c3dSStefano Zampini MPI_Comm comm,comm_n; 516128143c3dSStefano Zampini PetscSubcomm subcomm; 5162e7931f94SStefano Zampini PetscMPIInt n_sends,n_recvs,commsize; 516328143c3dSStefano Zampini PetscMPIInt *iflags,*ilengths_idxs,*ilengths_vals,*ilengths_idxs_is; 516428143c3dSStefano Zampini PetscMPIInt *onodes,*onodes_is,*olengths_idxs,*olengths_idxs_is,*olengths_vals; 516528143c3dSStefano Zampini PetscMPIInt len,tag_idxs,tag_idxs_is,tag_vals,source_dest; 516628143c3dSStefano Zampini MPI_Request *send_req_idxs,*send_req_idxs_is,*send_req_vals; 516728143c3dSStefano Zampini MPI_Request *recv_req_idxs,*recv_req_idxs_is,*recv_req_vals; 5168e7931f94SStefano Zampini PetscErrorCode ierr; 5169e7931f94SStefano Zampini 5170e7931f94SStefano Zampini PetscFunctionBegin; 517157de7509SStefano Zampini PetscValidHeaderSpecific(mat,MAT_CLASSID,1); 5172e7931f94SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)mat,MATIS,&ismatis);CHKERRQ(ierr); 517328143c3dSStefano Zampini if (!ismatis) SETERRQ1(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot use %s on a matrix object which is not of type MATIS",__FUNCT__); 517457de7509SStefano Zampini PetscValidLogicalCollectiveInt(mat,n_subdomains,3); 517557de7509SStefano Zampini PetscValidLogicalCollectiveBool(mat,restrict_comm,4); 517657de7509SStefano Zampini PetscValidLogicalCollectiveBool(mat,restrict_full,5); 517757de7509SStefano Zampini PetscValidLogicalCollectiveBool(mat,reuse,6); 517857de7509SStefano Zampini PetscValidLogicalCollectiveInt(mat,nis,8); 517957de7509SStefano Zampini 518057de7509SStefano Zampini /* further checks */ 5181e7931f94SStefano Zampini ierr = MatISGetLocalMat(mat,&local_mat);CHKERRQ(ierr); 5182e7931f94SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)local_mat,MATSEQDENSE,&isdense);CHKERRQ(ierr); 5183e7931f94SStefano Zampini if (!isdense) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Currently cannot subassemble MATIS when local matrix type is not of type SEQDENSE"); 5184e7931f94SStefano Zampini ierr = MatGetSize(local_mat,&rows,&cols);CHKERRQ(ierr); 5185e7931f94SStefano Zampini if (rows != cols) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Local MATIS matrices should be square"); 518657de7509SStefano Zampini if (reuse && *mat_n) { 518770cf5478SStefano Zampini PetscInt mrows,mcols,mnrows,mncols; 518857de7509SStefano Zampini PetscValidHeaderSpecific(*mat_n,MAT_CLASSID,7); 518970cf5478SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)*mat_n,MATIS,&ismatis);CHKERRQ(ierr); 519028143c3dSStefano Zampini if (!ismatis) SETERRQ(PetscObjectComm((PetscObject)*mat_n),PETSC_ERR_SUP,"Cannot reuse a matrix which is not of type MATIS"); 519170cf5478SStefano Zampini ierr = MatGetSize(mat,&mrows,&mcols);CHKERRQ(ierr); 519270cf5478SStefano Zampini ierr = MatGetSize(*mat_n,&mnrows,&mncols);CHKERRQ(ierr); 519370cf5478SStefano Zampini if (mrows != mnrows) SETERRQ2(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot reuse matrix! Wrong number of rows %D != %D",mrows,mnrows); 519470cf5478SStefano Zampini if (mcols != mncols) SETERRQ2(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot reuse matrix! Wrong number of cols %D != %D",mcols,mncols); 519570cf5478SStefano Zampini } 5196e7931f94SStefano Zampini ierr = MatGetBlockSize(local_mat,&bs);CHKERRQ(ierr); 5197e7931f94SStefano Zampini PetscValidLogicalCollectiveInt(mat,bs,0); 519857de7509SStefano Zampini 5199e7931f94SStefano Zampini /* prepare IS for sending if not provided */ 5200e7931f94SStefano Zampini if (!is_sends) { 520128143c3dSStefano Zampini if (!n_subdomains) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"You should specify either an IS or a target number of subdomains"); 520257de7509SStefano Zampini ierr = MatISGetSubassemblingPattern(mat,&n_subdomains,0,&is_sends_internal,NULL);CHKERRQ(ierr); 5203c8587f34SStefano Zampini } else { 5204e7931f94SStefano Zampini ierr = PetscObjectReference((PetscObject)is_sends);CHKERRQ(ierr); 5205e7931f94SStefano Zampini is_sends_internal = is_sends; 5206c8587f34SStefano Zampini } 5207e7931f94SStefano Zampini 5208e7931f94SStefano Zampini /* get comm */ 5209a316fed8SStefano Zampini ierr = PetscObjectGetComm((PetscObject)mat,&comm);CHKERRQ(ierr); 5210e7931f94SStefano Zampini 5211e7931f94SStefano Zampini /* compute number of sends */ 5212e7931f94SStefano Zampini ierr = ISGetLocalSize(is_sends_internal,&i);CHKERRQ(ierr); 5213e7931f94SStefano Zampini ierr = PetscMPIIntCast(i,&n_sends);CHKERRQ(ierr); 5214e7931f94SStefano Zampini 5215e7931f94SStefano Zampini /* compute number of receives */ 5216e7931f94SStefano Zampini ierr = MPI_Comm_size(comm,&commsize);CHKERRQ(ierr); 5217785e854fSJed Brown ierr = PetscMalloc1(commsize,&iflags);CHKERRQ(ierr); 5218e7931f94SStefano Zampini ierr = PetscMemzero(iflags,commsize*sizeof(*iflags));CHKERRQ(ierr); 5219e7931f94SStefano Zampini ierr = ISGetIndices(is_sends_internal,&is_indices);CHKERRQ(ierr); 5220e7931f94SStefano Zampini for (i=0;i<n_sends;i++) iflags[is_indices[i]] = 1; 5221e7931f94SStefano Zampini ierr = PetscGatherNumberOfMessages(comm,iflags,NULL,&n_recvs);CHKERRQ(ierr); 5222e7931f94SStefano Zampini ierr = PetscFree(iflags);CHKERRQ(ierr); 5223e7931f94SStefano Zampini 522428143c3dSStefano Zampini /* restrict comm if requested */ 522528143c3dSStefano Zampini subcomm = 0; 522628143c3dSStefano Zampini destroy_mat = PETSC_FALSE; 522728143c3dSStefano Zampini if (restrict_comm) { 5228779c1cceSStefano Zampini PetscMPIInt color,subcommsize; 5229779c1cceSStefano Zampini 523028143c3dSStefano Zampini color = 0; 523153a05cb3SStefano Zampini if (restrict_full) { 523253a05cb3SStefano Zampini if (!n_recvs) color = 1; /* processes not receiving anything will not partecipate in new comm (full restriction) */ 523353a05cb3SStefano Zampini } else { 523453a05cb3SStefano Zampini if (!n_recvs && n_sends) color = 1; /* just those processes that are sending but not receiving anything will not partecipate in new comm */ 523553a05cb3SStefano Zampini } 5236b2566f29SBarry Smith ierr = MPIU_Allreduce(&color,&subcommsize,1,MPI_INT,MPI_SUM,comm);CHKERRQ(ierr); 523728143c3dSStefano Zampini subcommsize = commsize - subcommsize; 523828143c3dSStefano Zampini /* check if reuse has been requested */ 523957de7509SStefano Zampini if (reuse) { 524028143c3dSStefano Zampini if (*mat_n) { 524128143c3dSStefano Zampini PetscMPIInt subcommsize2; 524228143c3dSStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)*mat_n),&subcommsize2);CHKERRQ(ierr); 524328143c3dSStefano Zampini if (subcommsize != subcommsize2) SETERRQ2(PetscObjectComm((PetscObject)*mat_n),PETSC_ERR_PLIB,"Cannot reuse matrix! wrong subcomm size %d != %d",subcommsize,subcommsize2); 524428143c3dSStefano Zampini comm_n = PetscObjectComm((PetscObject)*mat_n); 524528143c3dSStefano Zampini } else { 524628143c3dSStefano Zampini comm_n = PETSC_COMM_SELF; 524728143c3dSStefano Zampini } 524828143c3dSStefano Zampini } else { /* MAT_INITIAL_MATRIX */ 5249779c1cceSStefano Zampini PetscMPIInt rank; 5250779c1cceSStefano Zampini 5251779c1cceSStefano Zampini ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 525228143c3dSStefano Zampini ierr = PetscSubcommCreate(comm,&subcomm);CHKERRQ(ierr); 525328143c3dSStefano Zampini ierr = PetscSubcommSetNumber(subcomm,2);CHKERRQ(ierr); 525428143c3dSStefano Zampini ierr = PetscSubcommSetTypeGeneral(subcomm,color,rank);CHKERRQ(ierr); 5255306c2d5bSBarry Smith comm_n = PetscSubcommChild(subcomm); 525628143c3dSStefano Zampini } 525728143c3dSStefano Zampini /* flag to destroy *mat_n if not significative */ 525828143c3dSStefano Zampini if (color) destroy_mat = PETSC_TRUE; 525928143c3dSStefano Zampini } else { 526028143c3dSStefano Zampini comm_n = comm; 526128143c3dSStefano Zampini } 526228143c3dSStefano Zampini 5263e7931f94SStefano Zampini /* prepare send/receive buffers */ 5264785e854fSJed Brown ierr = PetscMalloc1(commsize,&ilengths_idxs);CHKERRQ(ierr); 5265e7931f94SStefano Zampini ierr = PetscMemzero(ilengths_idxs,commsize*sizeof(*ilengths_idxs));CHKERRQ(ierr); 5266785e854fSJed Brown ierr = PetscMalloc1(commsize,&ilengths_vals);CHKERRQ(ierr); 5267e7931f94SStefano Zampini ierr = PetscMemzero(ilengths_vals,commsize*sizeof(*ilengths_vals));CHKERRQ(ierr); 526828143c3dSStefano Zampini if (nis) { 5269854ce69bSBarry Smith ierr = PetscCalloc1(commsize,&ilengths_idxs_is);CHKERRQ(ierr); 527028143c3dSStefano Zampini } 5271e7931f94SStefano Zampini 527228143c3dSStefano Zampini /* Get data from local matrices */ 52736c4ed002SBarry Smith if (!isdense) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Subassembling of AIJ local matrices not yet implemented"); 5274e7931f94SStefano Zampini /* TODO: See below some guidelines on how to prepare the local buffers */ 5275e7931f94SStefano Zampini /* 5276e7931f94SStefano Zampini send_buffer_vals should contain the raw values of the local matrix 5277e7931f94SStefano Zampini send_buffer_idxs should contain: 5278e7931f94SStefano Zampini - MatType_PRIVATE type 5279e7931f94SStefano Zampini - PetscInt size_of_l2gmap 5280e7931f94SStefano Zampini - PetscInt global_row_indices[size_of_l2gmap] 5281e7931f94SStefano Zampini - PetscInt all_other_info_which_is_needed_to_compute_preallocation_and_set_values 5282e7931f94SStefano Zampini */ 52836c4ed002SBarry Smith else { 5284e7931f94SStefano Zampini ierr = MatDenseGetArray(local_mat,&send_buffer_vals);CHKERRQ(ierr); 52853bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetSize(mat->rmap->mapping,&i);CHKERRQ(ierr); 5286854ce69bSBarry Smith ierr = PetscMalloc1(i+2,&send_buffer_idxs);CHKERRQ(ierr); 5287e7931f94SStefano Zampini send_buffer_idxs[0] = (PetscInt)MATDENSE_PRIVATE; 5288e7931f94SStefano Zampini send_buffer_idxs[1] = i; 52893bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetIndices(mat->rmap->mapping,(const PetscInt**)&ptr_idxs);CHKERRQ(ierr); 5290e7931f94SStefano Zampini ierr = PetscMemcpy(&send_buffer_idxs[2],ptr_idxs,i*sizeof(PetscInt));CHKERRQ(ierr); 52913bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingRestoreIndices(mat->rmap->mapping,(const PetscInt**)&ptr_idxs);CHKERRQ(ierr); 5292e7931f94SStefano Zampini ierr = PetscMPIIntCast(i,&len);CHKERRQ(ierr); 5293e7931f94SStefano Zampini for (i=0;i<n_sends;i++) { 5294e7931f94SStefano Zampini ilengths_vals[is_indices[i]] = len*len; 5295e7931f94SStefano Zampini ilengths_idxs[is_indices[i]] = len+2; 5296c8587f34SStefano Zampini } 5297c8587f34SStefano Zampini } 5298e7931f94SStefano Zampini ierr = PetscGatherMessageLengths2(comm,n_sends,n_recvs,ilengths_idxs,ilengths_vals,&onodes,&olengths_idxs,&olengths_vals);CHKERRQ(ierr); 529928143c3dSStefano Zampini /* additional is (if any) */ 530028143c3dSStefano Zampini if (nis) { 530128143c3dSStefano Zampini PetscMPIInt psum; 530228143c3dSStefano Zampini PetscInt j; 530328143c3dSStefano Zampini for (j=0,psum=0;j<nis;j++) { 530428143c3dSStefano Zampini PetscInt plen; 530528143c3dSStefano Zampini ierr = ISGetLocalSize(isarray[j],&plen);CHKERRQ(ierr); 530628143c3dSStefano Zampini ierr = PetscMPIIntCast(plen,&len);CHKERRQ(ierr); 530728143c3dSStefano Zampini psum += len+1; /* indices + lenght */ 530828143c3dSStefano Zampini } 5309854ce69bSBarry Smith ierr = PetscMalloc1(psum,&send_buffer_idxs_is);CHKERRQ(ierr); 531028143c3dSStefano Zampini for (j=0,psum=0;j<nis;j++) { 531128143c3dSStefano Zampini PetscInt plen; 531228143c3dSStefano Zampini const PetscInt *is_array_idxs; 531328143c3dSStefano Zampini ierr = ISGetLocalSize(isarray[j],&plen);CHKERRQ(ierr); 531428143c3dSStefano Zampini send_buffer_idxs_is[psum] = plen; 531528143c3dSStefano Zampini ierr = ISGetIndices(isarray[j],&is_array_idxs);CHKERRQ(ierr); 531628143c3dSStefano Zampini ierr = PetscMemcpy(&send_buffer_idxs_is[psum+1],is_array_idxs,plen*sizeof(PetscInt));CHKERRQ(ierr); 531728143c3dSStefano Zampini ierr = ISRestoreIndices(isarray[j],&is_array_idxs);CHKERRQ(ierr); 531828143c3dSStefano Zampini psum += plen+1; /* indices + lenght */ 531928143c3dSStefano Zampini } 532028143c3dSStefano Zampini for (i=0;i<n_sends;i++) { 532128143c3dSStefano Zampini ilengths_idxs_is[is_indices[i]] = psum; 532228143c3dSStefano Zampini } 532328143c3dSStefano Zampini ierr = PetscGatherMessageLengths(comm,n_sends,n_recvs,ilengths_idxs_is,&onodes_is,&olengths_idxs_is);CHKERRQ(ierr); 532428143c3dSStefano Zampini } 532528143c3dSStefano Zampini 5326e7931f94SStefano Zampini buf_size_idxs = 0; 5327e7931f94SStefano Zampini buf_size_vals = 0; 532828143c3dSStefano Zampini buf_size_idxs_is = 0; 5329e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 5330e7931f94SStefano Zampini buf_size_idxs += (PetscInt)olengths_idxs[i]; 5331e7931f94SStefano Zampini buf_size_vals += (PetscInt)olengths_vals[i]; 533228143c3dSStefano Zampini if (nis) buf_size_idxs_is += (PetscInt)olengths_idxs_is[i]; 5333e7931f94SStefano Zampini } 5334785e854fSJed Brown ierr = PetscMalloc1(buf_size_idxs,&recv_buffer_idxs);CHKERRQ(ierr); 5335785e854fSJed Brown ierr = PetscMalloc1(buf_size_vals,&recv_buffer_vals);CHKERRQ(ierr); 533695ecbf38SStefano Zampini ierr = PetscMalloc1(buf_size_idxs_is,&recv_buffer_idxs_is);CHKERRQ(ierr); 5337e7931f94SStefano Zampini 5338e7931f94SStefano Zampini /* get new tags for clean communications */ 5339e7931f94SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_idxs);CHKERRQ(ierr); 5340e7931f94SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_vals);CHKERRQ(ierr); 534128143c3dSStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_idxs_is);CHKERRQ(ierr); 5342e7931f94SStefano Zampini 5343e7931f94SStefano Zampini /* allocate for requests */ 5344785e854fSJed Brown ierr = PetscMalloc1(n_sends,&send_req_idxs);CHKERRQ(ierr); 5345785e854fSJed Brown ierr = PetscMalloc1(n_sends,&send_req_vals);CHKERRQ(ierr); 534695ecbf38SStefano Zampini ierr = PetscMalloc1(n_sends,&send_req_idxs_is);CHKERRQ(ierr); 5347785e854fSJed Brown ierr = PetscMalloc1(n_recvs,&recv_req_idxs);CHKERRQ(ierr); 5348785e854fSJed Brown ierr = PetscMalloc1(n_recvs,&recv_req_vals);CHKERRQ(ierr); 534995ecbf38SStefano Zampini ierr = PetscMalloc1(n_recvs,&recv_req_idxs_is);CHKERRQ(ierr); 5350e7931f94SStefano Zampini 5351e7931f94SStefano Zampini /* communications */ 5352e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 5353e7931f94SStefano Zampini ptr_vals = recv_buffer_vals; 535428143c3dSStefano Zampini ptr_idxs_is = recv_buffer_idxs_is; 5355e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 5356e7931f94SStefano Zampini source_dest = onodes[i]; 5357e7931f94SStefano Zampini ierr = MPI_Irecv(ptr_idxs,olengths_idxs[i],MPIU_INT,source_dest,tag_idxs,comm,&recv_req_idxs[i]);CHKERRQ(ierr); 5358e7931f94SStefano Zampini ierr = MPI_Irecv(ptr_vals,olengths_vals[i],MPIU_SCALAR,source_dest,tag_vals,comm,&recv_req_vals[i]);CHKERRQ(ierr); 5359e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 5360e7931f94SStefano Zampini ptr_vals += olengths_vals[i]; 536128143c3dSStefano Zampini if (nis) { 536257de7509SStefano Zampini source_dest = onodes_is[i]; 536328143c3dSStefano 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); 536428143c3dSStefano Zampini ptr_idxs_is += olengths_idxs_is[i]; 536528143c3dSStefano Zampini } 5366e7931f94SStefano Zampini } 5367e7931f94SStefano Zampini for (i=0;i<n_sends;i++) { 5368e7931f94SStefano Zampini ierr = PetscMPIIntCast(is_indices[i],&source_dest);CHKERRQ(ierr); 5369e7931f94SStefano Zampini ierr = MPI_Isend(send_buffer_idxs,ilengths_idxs[source_dest],MPIU_INT,source_dest,tag_idxs,comm,&send_req_idxs[i]);CHKERRQ(ierr); 5370e7931f94SStefano Zampini ierr = MPI_Isend(send_buffer_vals,ilengths_vals[source_dest],MPIU_SCALAR,source_dest,tag_vals,comm,&send_req_vals[i]);CHKERRQ(ierr); 537128143c3dSStefano Zampini if (nis) { 537228143c3dSStefano 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); 537328143c3dSStefano Zampini } 5374e7931f94SStefano Zampini } 5375e7931f94SStefano Zampini ierr = ISRestoreIndices(is_sends_internal,&is_indices);CHKERRQ(ierr); 5376e7931f94SStefano Zampini ierr = ISDestroy(&is_sends_internal);CHKERRQ(ierr); 5377e7931f94SStefano Zampini 5378e7931f94SStefano Zampini /* assemble new l2g map */ 5379e7931f94SStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_idxs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 5380e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 53819d30be91SStefano Zampini new_local_rows = 0; 5382e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 53839d30be91SStefano Zampini new_local_rows += *(ptr_idxs+1); /* second element is the local size of the l2gmap */ 5384e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 5385e7931f94SStefano Zampini } 53869d30be91SStefano Zampini ierr = PetscMalloc1(new_local_rows,&l2gmap_indices);CHKERRQ(ierr); 5387e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 53889d30be91SStefano Zampini new_local_rows = 0; 5389e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 53909d30be91SStefano Zampini ierr = PetscMemcpy(&l2gmap_indices[new_local_rows],ptr_idxs+2,(*(ptr_idxs+1))*sizeof(PetscInt));CHKERRQ(ierr); 53919d30be91SStefano Zampini new_local_rows += *(ptr_idxs+1); /* second element is the local size of the l2gmap */ 5392e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 5393e7931f94SStefano Zampini } 53949d30be91SStefano Zampini ierr = PetscSortRemoveDupsInt(&new_local_rows,l2gmap_indices);CHKERRQ(ierr); 53959d30be91SStefano Zampini ierr = ISLocalToGlobalMappingCreate(comm_n,1,new_local_rows,l2gmap_indices,PETSC_COPY_VALUES,&l2gmap);CHKERRQ(ierr); 5396e7931f94SStefano Zampini ierr = PetscFree(l2gmap_indices);CHKERRQ(ierr); 5397e7931f94SStefano Zampini 5398e7931f94SStefano Zampini /* infer new local matrix type from received local matrices type */ 5399e7931f94SStefano Zampini /* currently if all local matrices are of type X, then the resulting matrix will be of type X, except for the dense case */ 5400e7931f94SStefano 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) */ 5401e7931f94SStefano Zampini if (n_recvs) { 540228143c3dSStefano Zampini MatTypePrivate new_local_type_private = (MatTypePrivate)send_buffer_idxs[0]; 5403e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 5404e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 5405e7931f94SStefano Zampini if ((PetscInt)new_local_type_private != *ptr_idxs) { 5406e7931f94SStefano Zampini new_local_type_private = MATAIJ_PRIVATE; 5407e7931f94SStefano Zampini break; 5408e7931f94SStefano Zampini } 5409e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 5410e7931f94SStefano Zampini } 5411e7931f94SStefano Zampini switch (new_local_type_private) { 541228143c3dSStefano Zampini case MATDENSE_PRIVATE: 541328143c3dSStefano Zampini if (n_recvs>1) { /* subassembling of dense matrices does not give a dense matrix! */ 5414e7931f94SStefano Zampini new_local_type = MATSEQAIJ; 5415e7931f94SStefano Zampini bs = 1; 541628143c3dSStefano Zampini } else { /* if I receive only 1 dense matrix */ 541728143c3dSStefano Zampini new_local_type = MATSEQDENSE; 541828143c3dSStefano Zampini bs = 1; 541928143c3dSStefano Zampini } 5420e7931f94SStefano Zampini break; 5421e7931f94SStefano Zampini case MATAIJ_PRIVATE: 5422e7931f94SStefano Zampini new_local_type = MATSEQAIJ; 5423e7931f94SStefano Zampini bs = 1; 5424e7931f94SStefano Zampini break; 5425e7931f94SStefano Zampini case MATBAIJ_PRIVATE: 5426e7931f94SStefano Zampini new_local_type = MATSEQBAIJ; 5427e7931f94SStefano Zampini break; 5428e7931f94SStefano Zampini case MATSBAIJ_PRIVATE: 5429e7931f94SStefano Zampini new_local_type = MATSEQSBAIJ; 5430e7931f94SStefano Zampini break; 5431e7931f94SStefano Zampini default: 54329d30be91SStefano Zampini SETERRQ2(comm,PETSC_ERR_SUP,"Unsupported private type %d in %s",new_local_type_private,__FUNCT__); 5433e7931f94SStefano Zampini break; 5434e7931f94SStefano Zampini } 543528143c3dSStefano Zampini } else { /* by default, new_local_type is seqdense */ 543628143c3dSStefano Zampini new_local_type = MATSEQDENSE; 543728143c3dSStefano Zampini bs = 1; 5438e7931f94SStefano Zampini } 5439e7931f94SStefano Zampini 544070cf5478SStefano Zampini /* create MATIS object if needed */ 544157de7509SStefano Zampini if (!reuse) { 5442e7931f94SStefano Zampini ierr = MatGetSize(mat,&rows,&cols);CHKERRQ(ierr); 5443e176bc59SStefano Zampini ierr = MatCreateIS(comm_n,bs,PETSC_DECIDE,PETSC_DECIDE,rows,cols,l2gmap,NULL,mat_n);CHKERRQ(ierr); 544470cf5478SStefano Zampini } else { 544570cf5478SStefano Zampini /* it also destroys the local matrices */ 544657de7509SStefano Zampini if (*mat_n) { 544770cf5478SStefano Zampini ierr = MatSetLocalToGlobalMapping(*mat_n,l2gmap,l2gmap);CHKERRQ(ierr); 544857de7509SStefano Zampini } else { /* this is a fake object */ 544957de7509SStefano Zampini ierr = MatCreateIS(comm_n,bs,PETSC_DECIDE,PETSC_DECIDE,rows,cols,l2gmap,NULL,mat_n);CHKERRQ(ierr); 545057de7509SStefano Zampini } 545170cf5478SStefano Zampini } 545270cf5478SStefano Zampini ierr = MatISGetLocalMat(*mat_n,&local_mat);CHKERRQ(ierr); 5453e7931f94SStefano Zampini ierr = MatSetType(local_mat,new_local_type);CHKERRQ(ierr); 54549d30be91SStefano Zampini 54559d30be91SStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_vals,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 54569d30be91SStefano Zampini 54579d30be91SStefano Zampini /* Global to local map of received indices */ 54589d30be91SStefano Zampini ierr = PetscMalloc1(buf_size_idxs,&recv_buffer_idxs_local);CHKERRQ(ierr); /* needed for values insertion */ 54599d30be91SStefano Zampini ierr = ISGlobalToLocalMappingApply(l2gmap,IS_GTOLM_MASK,buf_size_idxs,recv_buffer_idxs,&i,recv_buffer_idxs_local);CHKERRQ(ierr); 54609d30be91SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&l2gmap);CHKERRQ(ierr); 54619d30be91SStefano Zampini 54629d30be91SStefano Zampini /* restore attributes -> type of incoming data and its size */ 54639d30be91SStefano Zampini buf_size_idxs = 0; 54649d30be91SStefano Zampini for (i=0;i<n_recvs;i++) { 54659d30be91SStefano Zampini recv_buffer_idxs_local[buf_size_idxs] = recv_buffer_idxs[buf_size_idxs]; 54669d30be91SStefano Zampini recv_buffer_idxs_local[buf_size_idxs+1] = recv_buffer_idxs[buf_size_idxs+1]; 54679d30be91SStefano Zampini buf_size_idxs += (PetscInt)olengths_idxs[i]; 54689d30be91SStefano Zampini } 54699d30be91SStefano Zampini ierr = PetscFree(recv_buffer_idxs);CHKERRQ(ierr); 54709d30be91SStefano Zampini 54719d30be91SStefano Zampini /* set preallocation */ 54729d30be91SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)local_mat,MATSEQDENSE,&newisdense);CHKERRQ(ierr); 54739d30be91SStefano Zampini if (!newisdense) { 54749d30be91SStefano Zampini PetscInt *new_local_nnz=0; 54759d30be91SStefano Zampini 54769d30be91SStefano Zampini ptr_idxs = recv_buffer_idxs_local; 54779d30be91SStefano Zampini if (n_recvs) { 54789d30be91SStefano Zampini ierr = PetscCalloc1(new_local_rows,&new_local_nnz);CHKERRQ(ierr); 54799d30be91SStefano Zampini } 54809d30be91SStefano Zampini for (i=0;i<n_recvs;i++) { 54819d30be91SStefano Zampini PetscInt j; 54829d30be91SStefano Zampini if (*ptr_idxs == (PetscInt)MATDENSE_PRIVATE) { /* preallocation provided for dense case only */ 54839d30be91SStefano Zampini for (j=0;j<*(ptr_idxs+1);j++) { 54849d30be91SStefano Zampini new_local_nnz[*(ptr_idxs+2+j)] += *(ptr_idxs+1); 54859d30be91SStefano Zampini } 54869d30be91SStefano Zampini } else { 54879d30be91SStefano Zampini /* TODO */ 54889d30be91SStefano Zampini } 54899d30be91SStefano Zampini ptr_idxs += olengths_idxs[i]; 54909d30be91SStefano Zampini } 54919d30be91SStefano Zampini if (new_local_nnz) { 54929d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] = PetscMin(new_local_nnz[i],new_local_rows); 54939d30be91SStefano Zampini ierr = MatSeqAIJSetPreallocation(local_mat,0,new_local_nnz);CHKERRQ(ierr); 54949d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] /= bs; 54959d30be91SStefano Zampini ierr = MatSeqBAIJSetPreallocation(local_mat,bs,0,new_local_nnz);CHKERRQ(ierr); 54969d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] = PetscMax(new_local_nnz[i]-i,0); 54979d30be91SStefano Zampini ierr = MatSeqSBAIJSetPreallocation(local_mat,bs,0,new_local_nnz);CHKERRQ(ierr); 54989d30be91SStefano Zampini } else { 54999d30be91SStefano Zampini ierr = MatSetUp(local_mat);CHKERRQ(ierr); 55009d30be91SStefano Zampini } 55019d30be91SStefano Zampini ierr = PetscFree(new_local_nnz);CHKERRQ(ierr); 55029d30be91SStefano Zampini } else { 55039d30be91SStefano Zampini ierr = MatSetUp(local_mat);CHKERRQ(ierr); 55049d30be91SStefano Zampini } 5505e7931f94SStefano Zampini 5506e7931f94SStefano Zampini /* set values */ 5507e7931f94SStefano Zampini ptr_vals = recv_buffer_vals; 55089d30be91SStefano Zampini ptr_idxs = recv_buffer_idxs_local; 5509e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 5510e7931f94SStefano Zampini if (*ptr_idxs == (PetscInt)MATDENSE_PRIVATE) { /* values insertion provided for dense case only */ 5511e7931f94SStefano Zampini ierr = MatSetOption(local_mat,MAT_ROW_ORIENTED,PETSC_FALSE);CHKERRQ(ierr); 55129d30be91SStefano Zampini ierr = MatSetValues(local_mat,*(ptr_idxs+1),ptr_idxs+2,*(ptr_idxs+1),ptr_idxs+2,ptr_vals,ADD_VALUES);CHKERRQ(ierr); 5513e7931f94SStefano Zampini ierr = MatAssemblyBegin(local_mat,MAT_FLUSH_ASSEMBLY);CHKERRQ(ierr); 5514e7931f94SStefano Zampini ierr = MatAssemblyEnd(local_mat,MAT_FLUSH_ASSEMBLY);CHKERRQ(ierr); 5515e7931f94SStefano Zampini ierr = MatSetOption(local_mat,MAT_ROW_ORIENTED,PETSC_TRUE);CHKERRQ(ierr); 551628143c3dSStefano Zampini } else { 551728143c3dSStefano Zampini /* TODO */ 5518e7931f94SStefano Zampini } 5519e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 5520e7931f94SStefano Zampini ptr_vals += olengths_vals[i]; 5521e7931f94SStefano Zampini } 5522e7931f94SStefano Zampini ierr = MatAssemblyBegin(local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5523e7931f94SStefano Zampini ierr = MatAssemblyEnd(local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 552470cf5478SStefano Zampini ierr = MatAssemblyBegin(*mat_n,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 552570cf5478SStefano Zampini ierr = MatAssemblyEnd(*mat_n,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 55269d30be91SStefano Zampini ierr = PetscFree(recv_buffer_vals);CHKERRQ(ierr); 55279d30be91SStefano Zampini ierr = PetscFree(recv_buffer_idxs_local);CHKERRQ(ierr); 5528e7931f94SStefano Zampini 5529dfd14d43SStefano Zampini #if 0 553028143c3dSStefano Zampini if (!restrict_comm) { /* check */ 5531e7931f94SStefano Zampini Vec lvec,rvec; 5532e7931f94SStefano Zampini PetscReal infty_error; 5533e7931f94SStefano Zampini 55342a7a6963SBarry Smith ierr = MatCreateVecs(mat,&rvec,&lvec);CHKERRQ(ierr); 5535e7931f94SStefano Zampini ierr = VecSetRandom(rvec,NULL);CHKERRQ(ierr); 5536e7931f94SStefano Zampini ierr = MatMult(mat,rvec,lvec);CHKERRQ(ierr); 5537e7931f94SStefano Zampini ierr = VecScale(lvec,-1.0);CHKERRQ(ierr); 553870cf5478SStefano Zampini ierr = MatMultAdd(*mat_n,rvec,lvec,lvec);CHKERRQ(ierr); 5539e7931f94SStefano Zampini ierr = VecNorm(lvec,NORM_INFINITY,&infty_error);CHKERRQ(ierr); 5540e7931f94SStefano Zampini ierr = PetscPrintf(PetscObjectComm((PetscObject)mat),"Infinity error subassembling %1.6e\n",infty_error); 5541e7931f94SStefano Zampini ierr = VecDestroy(&rvec);CHKERRQ(ierr); 5542e7931f94SStefano Zampini ierr = VecDestroy(&lvec);CHKERRQ(ierr); 5543e7931f94SStefano Zampini } 554428143c3dSStefano Zampini #endif 5545e7931f94SStefano Zampini 554628143c3dSStefano Zampini /* assemble new additional is (if any) */ 554728143c3dSStefano Zampini if (nis) { 554828143c3dSStefano Zampini PetscInt **temp_idxs,*count_is,j,psum; 554928143c3dSStefano Zampini 555028143c3dSStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_idxs_is,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 5551854ce69bSBarry Smith ierr = PetscCalloc1(nis,&count_is);CHKERRQ(ierr); 555228143c3dSStefano Zampini ptr_idxs = recv_buffer_idxs_is; 555328143c3dSStefano Zampini psum = 0; 555428143c3dSStefano Zampini for (i=0;i<n_recvs;i++) { 555528143c3dSStefano Zampini for (j=0;j<nis;j++) { 555628143c3dSStefano Zampini PetscInt plen = *(ptr_idxs); /* first element is the local size of IS's indices */ 555728143c3dSStefano Zampini count_is[j] += plen; /* increment counting of buffer for j-th IS */ 555828143c3dSStefano Zampini psum += plen; 555928143c3dSStefano Zampini ptr_idxs += plen+1; /* shift pointer to received data */ 556028143c3dSStefano Zampini } 556128143c3dSStefano Zampini } 5562854ce69bSBarry Smith ierr = PetscMalloc1(nis,&temp_idxs);CHKERRQ(ierr); 5563854ce69bSBarry Smith ierr = PetscMalloc1(psum,&temp_idxs[0]);CHKERRQ(ierr); 556428143c3dSStefano Zampini for (i=1;i<nis;i++) { 556528143c3dSStefano Zampini temp_idxs[i] = temp_idxs[i-1]+count_is[i-1]; 556628143c3dSStefano Zampini } 556728143c3dSStefano Zampini ierr = PetscMemzero(count_is,nis*sizeof(PetscInt));CHKERRQ(ierr); 556828143c3dSStefano Zampini ptr_idxs = recv_buffer_idxs_is; 556928143c3dSStefano Zampini for (i=0;i<n_recvs;i++) { 557028143c3dSStefano Zampini for (j=0;j<nis;j++) { 557128143c3dSStefano Zampini PetscInt plen = *(ptr_idxs); /* first element is the local size of IS's indices */ 557228143c3dSStefano Zampini ierr = PetscMemcpy(&temp_idxs[j][count_is[j]],ptr_idxs+1,plen*sizeof(PetscInt));CHKERRQ(ierr); 557328143c3dSStefano Zampini count_is[j] += plen; /* increment starting point of buffer for j-th IS */ 557428143c3dSStefano Zampini ptr_idxs += plen+1; /* shift pointer to received data */ 557528143c3dSStefano Zampini } 557628143c3dSStefano Zampini } 557728143c3dSStefano Zampini for (i=0;i<nis;i++) { 557828143c3dSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 557928143c3dSStefano Zampini ierr = PetscSortRemoveDupsInt(&count_is[i],temp_idxs[i]);CHKERRQ(ierr);CHKERRQ(ierr); 558028143c3dSStefano Zampini ierr = ISCreateGeneral(comm_n,count_is[i],temp_idxs[i],PETSC_COPY_VALUES,&isarray[i]);CHKERRQ(ierr); 558128143c3dSStefano Zampini } 558228143c3dSStefano Zampini ierr = PetscFree(count_is);CHKERRQ(ierr); 558328143c3dSStefano Zampini ierr = PetscFree(temp_idxs[0]);CHKERRQ(ierr); 558428143c3dSStefano Zampini ierr = PetscFree(temp_idxs);CHKERRQ(ierr); 558528143c3dSStefano Zampini } 5586e7931f94SStefano Zampini /* free workspace */ 558728143c3dSStefano Zampini ierr = PetscFree(recv_buffer_idxs_is);CHKERRQ(ierr); 5588e7931f94SStefano Zampini ierr = MPI_Waitall(n_sends,send_req_idxs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 5589e7931f94SStefano Zampini ierr = PetscFree(send_buffer_idxs);CHKERRQ(ierr); 5590e7931f94SStefano Zampini ierr = MPI_Waitall(n_sends,send_req_vals,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 5591e7931f94SStefano Zampini if (isdense) { 5592e7931f94SStefano Zampini ierr = MatISGetLocalMat(mat,&local_mat);CHKERRQ(ierr); 5593e7931f94SStefano Zampini ierr = MatDenseRestoreArray(local_mat,&send_buffer_vals);CHKERRQ(ierr); 5594e7931f94SStefano Zampini } else { 5595e7931f94SStefano Zampini /* ierr = PetscFree(send_buffer_vals);CHKERRQ(ierr); */ 5596e7931f94SStefano Zampini } 559728143c3dSStefano Zampini if (nis) { 559828143c3dSStefano Zampini ierr = MPI_Waitall(n_sends,send_req_idxs_is,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 559928143c3dSStefano Zampini ierr = PetscFree(send_buffer_idxs_is);CHKERRQ(ierr); 560028143c3dSStefano Zampini } 5601e7931f94SStefano Zampini ierr = PetscFree(recv_req_idxs);CHKERRQ(ierr); 5602e7931f94SStefano Zampini ierr = PetscFree(recv_req_vals);CHKERRQ(ierr); 560328143c3dSStefano Zampini ierr = PetscFree(recv_req_idxs_is);CHKERRQ(ierr); 5604e7931f94SStefano Zampini ierr = PetscFree(send_req_idxs);CHKERRQ(ierr); 5605e7931f94SStefano Zampini ierr = PetscFree(send_req_vals);CHKERRQ(ierr); 560628143c3dSStefano Zampini ierr = PetscFree(send_req_idxs_is);CHKERRQ(ierr); 5607e7931f94SStefano Zampini ierr = PetscFree(ilengths_vals);CHKERRQ(ierr); 5608e7931f94SStefano Zampini ierr = PetscFree(ilengths_idxs);CHKERRQ(ierr); 5609e7931f94SStefano Zampini ierr = PetscFree(olengths_vals);CHKERRQ(ierr); 5610e7931f94SStefano Zampini ierr = PetscFree(olengths_idxs);CHKERRQ(ierr); 5611e7931f94SStefano Zampini ierr = PetscFree(onodes);CHKERRQ(ierr); 561228143c3dSStefano Zampini if (nis) { 561328143c3dSStefano Zampini ierr = PetscFree(ilengths_idxs_is);CHKERRQ(ierr); 561428143c3dSStefano Zampini ierr = PetscFree(olengths_idxs_is);CHKERRQ(ierr); 561528143c3dSStefano Zampini ierr = PetscFree(onodes_is);CHKERRQ(ierr); 561628143c3dSStefano Zampini } 561728143c3dSStefano Zampini ierr = PetscSubcommDestroy(&subcomm);CHKERRQ(ierr); 561828143c3dSStefano Zampini if (destroy_mat) { /* destroy mat is true only if restrict comm is true and process will not partecipate */ 561928143c3dSStefano Zampini ierr = MatDestroy(mat_n);CHKERRQ(ierr); 562028143c3dSStefano Zampini for (i=0;i<nis;i++) { 562128143c3dSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 562228143c3dSStefano Zampini } 562353a05cb3SStefano Zampini *mat_n = NULL; 562428143c3dSStefano Zampini } 5625e7931f94SStefano Zampini PetscFunctionReturn(0); 5626e7931f94SStefano Zampini } 5627a57a6d2fSStefano Zampini 562812edc857SStefano Zampini /* temporary hack into ksp private data structure */ 5629af0996ceSBarry Smith #include <petsc/private/kspimpl.h> 563012edc857SStefano Zampini 5631c8587f34SStefano Zampini #undef __FUNCT__ 5632c8587f34SStefano Zampini #define __FUNCT__ "PCBDDCSetUpCoarseSolver" 5633c8587f34SStefano Zampini PetscErrorCode PCBDDCSetUpCoarseSolver(PC pc,PetscScalar* coarse_submat_vals) 5634c8587f34SStefano Zampini { 5635c8587f34SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 5636c8587f34SStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 563720a2ab83SStefano Zampini Mat coarse_mat,coarse_mat_is,coarse_submat_dense; 56389881197aSStefano Zampini MatNullSpace CoarseNullSpace = NULL; 563920a2ab83SStefano Zampini ISLocalToGlobalMapping coarse_islg; 56406e683305SStefano Zampini IS coarse_is,*isarray; 56416e683305SStefano Zampini PetscInt i,im_active=-1,active_procs=-1; 564230368db7SStefano Zampini PetscInt nis,nisdofs,nisneu,nisvert; 5643f9eb5b7dSStefano Zampini PC pc_temp; 5644c8587f34SStefano Zampini PCType coarse_pc_type; 5645c8587f34SStefano Zampini KSPType coarse_ksp_type; 5646f9eb5b7dSStefano Zampini PetscBool multilevel_requested,multilevel_allowed; 56474f3a063dSStefano Zampini PetscBool isredundant,isbddc,isnn,coarse_reuse; 56486e683305SStefano Zampini Mat t_coarse_mat_is; 564957de7509SStefano Zampini PetscInt ncoarse; 565068457ee5SStefano Zampini PetscBool compute_vecs = PETSC_FALSE; 565122bc73bbSStefano Zampini PetscScalar *array; 565257de7509SStefano Zampini MatReuse coarse_mat_reuse; 565357de7509SStefano Zampini PetscBool restr, full_restr, have_void; 56549881197aSStefano Zampini PetscErrorCode ierr; 5655fdc09c96SStefano Zampini 5656c8587f34SStefano Zampini PetscFunctionBegin; 5657c8587f34SStefano Zampini /* Assign global numbering to coarse dofs */ 565868457ee5SStefano 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 */ 5659fa7f1dd8SStefano Zampini PetscInt ocoarse_size; 56605a75c04eSSatish Balay compute_vecs = PETSC_TRUE; 5661fa7f1dd8SStefano Zampini ocoarse_size = pcbddc->coarse_size; 5662f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->global_primal_indices);CHKERRQ(ierr); 5663f4ddd8eeSStefano Zampini ierr = PCBDDCComputePrimalNumbering(pc,&pcbddc->coarse_size,&pcbddc->global_primal_indices);CHKERRQ(ierr); 5664f4ddd8eeSStefano Zampini /* see if we can avoid some work */ 5665fa7f1dd8SStefano Zampini if (pcbddc->coarse_ksp) { /* coarse ksp has already been created */ 566651bea450SStefano Zampini /* if the coarse size is different or we are using adaptive selection, better to not reuse the coarse matrix */ 566751bea450SStefano Zampini if (ocoarse_size != pcbddc->coarse_size || pcbddc->adaptive_selection) { 5668dc4bcba2SStefano Zampini PC pc; 5669dc4bcba2SStefano Zampini PetscBool isbddc; 5670dc4bcba2SStefano Zampini 5671dc4bcba2SStefano Zampini /* temporary workaround since PCBDDC does not have a reset method so far */ 5672dc4bcba2SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&pc);CHKERRQ(ierr); 5673dc4bcba2SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc,PCBDDC,&isbddc);CHKERRQ(ierr); 5674dc4bcba2SStefano Zampini if (isbddc) { 567563c961adSStefano Zampini ierr = KSPDestroy(&pcbddc->coarse_ksp);CHKERRQ(ierr); 567663c961adSStefano Zampini } else { 5677727cdba6SStefano Zampini ierr = KSPReset(pcbddc->coarse_ksp);CHKERRQ(ierr); 567863c961adSStefano Zampini } 5679fa7f1dd8SStefano Zampini coarse_reuse = PETSC_FALSE; 5680fa7f1dd8SStefano Zampini } else { /* we can safely reuse already computed coarse matrix */ 5681fa7f1dd8SStefano Zampini coarse_reuse = PETSC_TRUE; 5682f4ddd8eeSStefano Zampini } 5683fa7f1dd8SStefano Zampini } else { /* there's no coarse ksp, so we need to create the coarse matrix too */ 5684fa7f1dd8SStefano Zampini coarse_reuse = PETSC_FALSE; 5685f4ddd8eeSStefano Zampini } 568670cf5478SStefano Zampini /* reset any subassembling information */ 568757de7509SStefano Zampini if (!coarse_reuse || pcbddc->recompute_topography) { 568870cf5478SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling);CHKERRQ(ierr); 568957de7509SStefano Zampini } 56906e683305SStefano Zampini } else { /* primal space is unchanged, so we can reuse coarse matrix */ 5691fa7f1dd8SStefano Zampini coarse_reuse = PETSC_TRUE; 5692f4ddd8eeSStefano Zampini } 569357de7509SStefano Zampini /* assemble coarse matrix */ 569457de7509SStefano Zampini if (coarse_reuse && pcbddc->coarse_ksp) { 569557de7509SStefano Zampini ierr = KSPGetOperators(pcbddc->coarse_ksp,&coarse_mat,NULL);CHKERRQ(ierr); 569657de7509SStefano Zampini ierr = PetscObjectReference((PetscObject)coarse_mat);CHKERRQ(ierr); 569757de7509SStefano Zampini coarse_mat_reuse = MAT_REUSE_MATRIX; 569818a45a71SStefano Zampini } else { 569957de7509SStefano Zampini coarse_mat = NULL; 570057de7509SStefano Zampini coarse_mat_reuse = MAT_INITIAL_MATRIX; 57016e683305SStefano Zampini } 5702e7931f94SStefano Zampini 5703abbbba34SStefano Zampini /* creates temporary l2gmap and IS for coarse indexes */ 5704abbbba34SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),pcbddc->local_primal_size,pcbddc->global_primal_indices,PETSC_COPY_VALUES,&coarse_is);CHKERRQ(ierr); 5705abbbba34SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(coarse_is,&coarse_islg);CHKERRQ(ierr); 5706abbbba34SStefano Zampini 5707abbbba34SStefano Zampini /* creates temporary MATIS object for coarse matrix */ 570822bc73bbSStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,pcbddc->local_primal_size,pcbddc->local_primal_size,NULL,&coarse_submat_dense);CHKERRQ(ierr); 570922bc73bbSStefano Zampini ierr = MatDenseGetArray(coarse_submat_dense,&array);CHKERRQ(ierr); 571022bc73bbSStefano Zampini ierr = PetscMemcpy(array,coarse_submat_vals,sizeof(*coarse_submat_vals)*pcbddc->local_primal_size*pcbddc->local_primal_size);CHKERRQ(ierr); 571122bc73bbSStefano Zampini ierr = MatDenseRestoreArray(coarse_submat_dense,&array);CHKERRQ(ierr); 5712e176bc59SStefano 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); 57136e683305SStefano Zampini ierr = MatISSetLocalMat(t_coarse_mat_is,coarse_submat_dense);CHKERRQ(ierr); 57146e683305SStefano Zampini ierr = MatAssemblyBegin(t_coarse_mat_is,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 57156e683305SStefano Zampini ierr = MatAssemblyEnd(t_coarse_mat_is,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5716abbbba34SStefano Zampini ierr = MatDestroy(&coarse_submat_dense);CHKERRQ(ierr); 5717abbbba34SStefano Zampini 571857de7509SStefano Zampini /* count "active" (i.e. with positive local size) and "void" processes */ 571957de7509SStefano Zampini im_active = !!(pcis->n); 572057de7509SStefano Zampini ierr = MPIU_Allreduce(&im_active,&active_procs,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 572157de7509SStefano Zampini 572257de7509SStefano Zampini /* determine number of process partecipating to coarse solver and compute subassembling pattern */ 572357de7509SStefano Zampini /* restr : whether if we want to exclude senders (which are not receivers) from the subassembling pattern */ 572457de7509SStefano Zampini /* full_restr : just use the receivers from the subassembling pattern */ 572557de7509SStefano Zampini coarse_mat_is = NULL; 572657de7509SStefano Zampini multilevel_allowed = PETSC_FALSE; 572757de7509SStefano Zampini multilevel_requested = PETSC_FALSE; 572857de7509SStefano Zampini full_restr = PETSC_TRUE; 572957de7509SStefano Zampini pcbddc->coarse_eqs_per_proc = PetscMin(pcbddc->coarse_size,pcbddc->coarse_eqs_per_proc); 573057de7509SStefano Zampini if (pcbddc->current_level < pcbddc->max_levels) multilevel_requested = PETSC_TRUE; 573157de7509SStefano Zampini if (multilevel_requested) { 573257de7509SStefano Zampini ncoarse = active_procs/pcbddc->coarsening_ratio; 573357de7509SStefano Zampini restr = PETSC_FALSE; 573457de7509SStefano Zampini full_restr = PETSC_FALSE; 573557de7509SStefano Zampini } else { 573657de7509SStefano Zampini ncoarse = pcbddc->coarse_size/pcbddc->coarse_eqs_per_proc; 573757de7509SStefano Zampini restr = PETSC_TRUE; 573857de7509SStefano Zampini full_restr = PETSC_TRUE; 573957de7509SStefano Zampini } 574057de7509SStefano Zampini ncoarse = PetscMax(1,ncoarse); 574157de7509SStefano Zampini if (!pcbddc->coarse_subassembling) { 574257de7509SStefano Zampini ierr = MatISGetSubassemblingPattern(t_coarse_mat_is,&ncoarse,pcbddc->coarse_adj_red,&pcbddc->coarse_subassembling,&have_void);CHKERRQ(ierr); 574357de7509SStefano Zampini } else { /* if a subassembling pattern exists, then we can reuse the coarse ksp and compute the number of process involved */ 574457de7509SStefano Zampini PetscInt psum; 574557de7509SStefano Zampini PetscMPIInt size; 574657de7509SStefano Zampini if (pcbddc->coarse_ksp) psum = 1; 574757de7509SStefano Zampini else psum = 0; 574857de7509SStefano Zampini ierr = MPIU_Allreduce(&psum,&ncoarse,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 574957de7509SStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)pc),&size);CHKERRQ(ierr); 575057de7509SStefano Zampini if (ncoarse < size) have_void = PETSC_TRUE; 575157de7509SStefano Zampini } 575257de7509SStefano Zampini /* determine if we can go multilevel */ 575357de7509SStefano Zampini if (multilevel_requested) { 575457de7509SStefano Zampini if (ncoarse > 1) multilevel_allowed = PETSC_TRUE; /* found enough processes */ 575557de7509SStefano Zampini else restr = full_restr = PETSC_TRUE; /* 1 subdomain, use a direct solver */ 575657de7509SStefano Zampini } 575757de7509SStefano Zampini if (multilevel_allowed && have_void) restr = PETSC_TRUE; 575857de7509SStefano Zampini 5759e4d548c7SStefano Zampini /* dump subassembling pattern */ 5760e4d548c7SStefano Zampini if (pcbddc->dbg_flag && multilevel_allowed) { 5761e4d548c7SStefano Zampini ierr = ISView(pcbddc->coarse_subassembling,pcbddc->dbg_viewer);CHKERRQ(ierr); 5762e4d548c7SStefano Zampini } 5763e4d548c7SStefano Zampini 57646e683305SStefano Zampini /* compute dofs splitting and neumann boundaries for coarse dofs */ 576527b6a85dSStefano Zampini if (multilevel_allowed && (pcbddc->n_ISForDofsLocal || pcbddc->NeumannBoundariesLocal)) { /* protects from unneded computations */ 57666e683305SStefano Zampini PetscInt *tidxs,*tidxs2,nout,tsize,i; 57676e683305SStefano Zampini const PetscInt *idxs; 57686e683305SStefano Zampini ISLocalToGlobalMapping tmap; 57696e683305SStefano Zampini 57706e683305SStefano Zampini /* create map between primal indices (in local representative ordering) and local primal numbering */ 57710be93d54SStefano Zampini ierr = ISLocalToGlobalMappingCreate(PETSC_COMM_SELF,1,pcbddc->local_primal_size,pcbddc->primal_indices_local_idxs,PETSC_COPY_VALUES,&tmap);CHKERRQ(ierr); 57726e683305SStefano Zampini /* allocate space for temporary storage */ 5773854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->local_primal_size,&tidxs);CHKERRQ(ierr); 5774854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->local_primal_size,&tidxs2);CHKERRQ(ierr); 57756e683305SStefano Zampini /* allocate for IS array */ 57766e683305SStefano Zampini nisdofs = pcbddc->n_ISForDofsLocal; 57776e683305SStefano Zampini nisneu = !!pcbddc->NeumannBoundariesLocal; 577827b6a85dSStefano Zampini nisvert = 0; /* nisvert is not used */ 577930368db7SStefano Zampini nis = nisdofs + nisneu + nisvert; 5780854ce69bSBarry Smith ierr = PetscMalloc1(nis,&isarray);CHKERRQ(ierr); 57816e683305SStefano Zampini /* dofs splitting */ 57826e683305SStefano Zampini for (i=0;i<nisdofs;i++) { 57836e683305SStefano Zampini /* ierr = ISView(pcbddc->ISForDofsLocal[i],0);CHKERRQ(ierr); */ 57846e683305SStefano Zampini ierr = ISGetLocalSize(pcbddc->ISForDofsLocal[i],&tsize);CHKERRQ(ierr); 57856e683305SStefano Zampini ierr = ISGetIndices(pcbddc->ISForDofsLocal[i],&idxs);CHKERRQ(ierr); 57866e683305SStefano Zampini ierr = ISGlobalToLocalMappingApply(tmap,IS_GTOLM_DROP,tsize,idxs,&nout,tidxs);CHKERRQ(ierr); 57876e683305SStefano Zampini ierr = ISRestoreIndices(pcbddc->ISForDofsLocal[i],&idxs);CHKERRQ(ierr); 57886e683305SStefano Zampini ierr = ISLocalToGlobalMappingApply(coarse_islg,nout,tidxs,tidxs2);CHKERRQ(ierr); 578930368db7SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),nout,tidxs2,PETSC_COPY_VALUES,&isarray[i]);CHKERRQ(ierr); 57906e683305SStefano Zampini /* ierr = ISView(isarray[i],0);CHKERRQ(ierr); */ 57916e683305SStefano Zampini } 57926e683305SStefano Zampini /* neumann boundaries */ 57936e683305SStefano Zampini if (pcbddc->NeumannBoundariesLocal) { 57946e683305SStefano Zampini /* ierr = ISView(pcbddc->NeumannBoundariesLocal,0);CHKERRQ(ierr); */ 57956e683305SStefano Zampini ierr = ISGetLocalSize(pcbddc->NeumannBoundariesLocal,&tsize);CHKERRQ(ierr); 57966e683305SStefano Zampini ierr = ISGetIndices(pcbddc->NeumannBoundariesLocal,&idxs);CHKERRQ(ierr); 57976e683305SStefano Zampini ierr = ISGlobalToLocalMappingApply(tmap,IS_GTOLM_DROP,tsize,idxs,&nout,tidxs);CHKERRQ(ierr); 57986e683305SStefano Zampini ierr = ISRestoreIndices(pcbddc->NeumannBoundariesLocal,&idxs);CHKERRQ(ierr); 57996e683305SStefano Zampini ierr = ISLocalToGlobalMappingApply(coarse_islg,nout,tidxs,tidxs2);CHKERRQ(ierr); 580030368db7SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),nout,tidxs2,PETSC_COPY_VALUES,&isarray[nisdofs]);CHKERRQ(ierr); 58016e683305SStefano Zampini /* ierr = ISView(isarray[nisdofs],0);CHKERRQ(ierr); */ 58026e683305SStefano Zampini } 58036e683305SStefano Zampini /* free memory */ 58046e683305SStefano Zampini ierr = PetscFree(tidxs);CHKERRQ(ierr); 58056e683305SStefano Zampini ierr = PetscFree(tidxs2);CHKERRQ(ierr); 58066e683305SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&tmap);CHKERRQ(ierr); 58076e683305SStefano Zampini } else { 58086e683305SStefano Zampini nis = 0; 58096e683305SStefano Zampini nisdofs = 0; 58106e683305SStefano Zampini nisneu = 0; 581130368db7SStefano Zampini nisvert = 0; 58126e683305SStefano Zampini isarray = NULL; 58136e683305SStefano Zampini } 58146e683305SStefano Zampini /* destroy no longer needed map */ 58156e683305SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&coarse_islg);CHKERRQ(ierr); 58166e683305SStefano Zampini 581757de7509SStefano Zampini /* subassemble */ 581857de7509SStefano Zampini if (multilevel_allowed) { 581957de7509SStefano Zampini PetscBool reuse,reuser; 582057de7509SStefano Zampini if (coarse_mat) reuse = PETSC_TRUE; 582157de7509SStefano Zampini else reuse = PETSC_FALSE; 582257de7509SStefano Zampini ierr = MPIU_Allreduce(&reuse,&reuser,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 582357de7509SStefano Zampini if (reuser) { 582457de7509SStefano Zampini ierr = MatISSubassemble(t_coarse_mat_is,pcbddc->coarse_subassembling,0,restr,full_restr,PETSC_TRUE,&coarse_mat,nis,isarray);CHKERRQ(ierr); 582574e2c79eSStefano Zampini } else { 582657de7509SStefano Zampini ierr = MatISSubassemble(t_coarse_mat_is,pcbddc->coarse_subassembling,0,restr,full_restr,PETSC_FALSE,&coarse_mat_is,nis,isarray);CHKERRQ(ierr); 582774e2c79eSStefano Zampini } 58281f4df5f7SStefano Zampini /* TODO: if (pcbddc->benign_have_null) -> give a hint to the coarser levels if they have to locally apply the benign trick or not */ 582974e2c79eSStefano Zampini } else { 583057de7509SStefano Zampini ierr = MatISSubassemble(t_coarse_mat_is,pcbddc->coarse_subassembling,0,restr,full_restr,PETSC_FALSE,&coarse_mat_is,nis,isarray);CHKERRQ(ierr); 58316e683305SStefano Zampini } 583257de7509SStefano Zampini if (coarse_mat_is || coarse_mat) { 583357de7509SStefano Zampini PetscMPIInt size; 583457de7509SStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)coarse_mat_is),&size); 583557de7509SStefano Zampini if (!multilevel_allowed) { 583657de7509SStefano Zampini ierr = MatISGetMPIXAIJ(coarse_mat_is,coarse_mat_reuse,&coarse_mat);CHKERRQ(ierr); 58376e683305SStefano Zampini } else { 583857de7509SStefano Zampini Mat A; 5839779c1cceSStefano Zampini 584057de7509SStefano Zampini /* if this matrix is present, it means we are not reusing the coarse matrix */ 584157de7509SStefano Zampini if (coarse_mat_is) { 584257de7509SStefano Zampini if (coarse_mat) SETERRQ(PetscObjectComm((PetscObject)coarse_mat_is),PETSC_ERR_PLIB,"This should not happen"); 584357de7509SStefano Zampini ierr = PetscObjectReference((PetscObject)coarse_mat_is);CHKERRQ(ierr); 584457de7509SStefano Zampini coarse_mat = coarse_mat_is; 584557de7509SStefano Zampini } 584657de7509SStefano Zampini /* be sure we don't have MatSeqDENSE as local mat */ 584757de7509SStefano Zampini ierr = MatISGetLocalMat(coarse_mat,&A);CHKERRQ(ierr); 584857de7509SStefano Zampini ierr = MatConvert(A,MATSEQAIJ,MAT_INPLACE_MATRIX,&A);CHKERRQ(ierr); 5849779c1cceSStefano Zampini } 5850779c1cceSStefano Zampini } 585157de7509SStefano Zampini ierr = MatDestroy(&t_coarse_mat_is);CHKERRQ(ierr); 585257de7509SStefano Zampini ierr = MatDestroy(&coarse_mat_is);CHKERRQ(ierr); 58536e683305SStefano Zampini 58546e683305SStefano Zampini /* create local to global scatters for coarse problem */ 585568457ee5SStefano Zampini if (compute_vecs) { 58566e683305SStefano Zampini PetscInt lrows; 58576e683305SStefano Zampini ierr = VecDestroy(&pcbddc->coarse_vec);CHKERRQ(ierr); 585857de7509SStefano Zampini if (coarse_mat) { 585957de7509SStefano Zampini ierr = MatGetLocalSize(coarse_mat,&lrows,NULL);CHKERRQ(ierr); 58606e683305SStefano Zampini } else { 58616e683305SStefano Zampini lrows = 0; 58626e683305SStefano Zampini } 58636e683305SStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pc),&pcbddc->coarse_vec);CHKERRQ(ierr); 58646e683305SStefano Zampini ierr = VecSetSizes(pcbddc->coarse_vec,lrows,PETSC_DECIDE);CHKERRQ(ierr); 58656e683305SStefano Zampini ierr = VecSetType(pcbddc->coarse_vec,VECSTANDARD);CHKERRQ(ierr); 58666e683305SStefano Zampini ierr = VecScatterDestroy(&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 58676e683305SStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_P,NULL,pcbddc->coarse_vec,coarse_is,&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 58686e683305SStefano Zampini } 58696e683305SStefano Zampini ierr = ISDestroy(&coarse_is);CHKERRQ(ierr); 5870c8587f34SStefano Zampini 5871f9eb5b7dSStefano Zampini /* set defaults for coarse KSP and PC */ 5872f9eb5b7dSStefano Zampini if (multilevel_allowed) { 5873f9eb5b7dSStefano Zampini coarse_ksp_type = KSPRICHARDSON; 5874f9eb5b7dSStefano Zampini coarse_pc_type = PCBDDC; 5875f9eb5b7dSStefano Zampini } else { 5876f9eb5b7dSStefano Zampini coarse_ksp_type = KSPPREONLY; 5877f9eb5b7dSStefano Zampini coarse_pc_type = PCREDUNDANT; 5878c8587f34SStefano Zampini } 5879c8587f34SStefano Zampini 58806e683305SStefano Zampini /* print some info if requested */ 58816e683305SStefano Zampini if (pcbddc->dbg_flag) { 58826e683305SStefano Zampini if (!multilevel_allowed) { 58836e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 58846e683305SStefano Zampini if (multilevel_requested) { 58856e683305SStefano 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); 58866e683305SStefano Zampini } else if (pcbddc->max_levels) { 58876e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Maximum number of requested levels reached (%d)\n",pcbddc->max_levels);CHKERRQ(ierr); 58886e683305SStefano Zampini } 58896e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 58906e683305SStefano Zampini } 58916e683305SStefano Zampini } 58926e683305SStefano Zampini 5893f9eb5b7dSStefano Zampini /* create the coarse KSP object only once with defaults */ 589457de7509SStefano Zampini if (coarse_mat) { 58956a1308c2SStefano Zampini PetscViewer dbg_viewer = NULL; 58966e683305SStefano Zampini if (pcbddc->dbg_flag) { 589757de7509SStefano Zampini dbg_viewer = PETSC_VIEWER_STDOUT_(PetscObjectComm((PetscObject)coarse_mat)); 58986e683305SStefano Zampini ierr = PetscViewerASCIIAddTab(dbg_viewer,2*pcbddc->current_level);CHKERRQ(ierr); 58996e683305SStefano Zampini } 5900f9eb5b7dSStefano Zampini if (!pcbddc->coarse_ksp) { 5901312be037SStefano Zampini char prefix[256],str_level[16]; 5902e604994aSStefano Zampini size_t len; 590357de7509SStefano Zampini ierr = KSPCreate(PetscObjectComm((PetscObject)coarse_mat),&pcbddc->coarse_ksp);CHKERRQ(ierr); 5904422a814eSBarry Smith ierr = KSPSetErrorIfNotConverged(pcbddc->coarse_ksp,pc->erroriffailure);CHKERRQ(ierr); 5905c8587f34SStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->coarse_ksp,(PetscObject)pc,1);CHKERRQ(ierr); 5906f9eb5b7dSStefano Zampini ierr = KSPSetTolerances(pcbddc->coarse_ksp,PETSC_DEFAULT,PETSC_DEFAULT,PETSC_DEFAULT,1);CHKERRQ(ierr); 590757de7509SStefano Zampini ierr = KSPSetOperators(pcbddc->coarse_ksp,coarse_mat,coarse_mat);CHKERRQ(ierr); 5908c8587f34SStefano Zampini ierr = KSPSetType(pcbddc->coarse_ksp,coarse_ksp_type);CHKERRQ(ierr); 59096e683305SStefano Zampini ierr = KSPSetNormType(pcbddc->coarse_ksp,KSP_NORM_NONE);CHKERRQ(ierr); 5910c8587f34SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&pc_temp);CHKERRQ(ierr); 5911c8587f34SStefano Zampini ierr = PCSetType(pc_temp,coarse_pc_type);CHKERRQ(ierr); 5912e604994aSStefano Zampini /* prefix */ 5913e604994aSStefano Zampini ierr = PetscStrcpy(prefix,"");CHKERRQ(ierr); 5914e604994aSStefano Zampini ierr = PetscStrcpy(str_level,"");CHKERRQ(ierr); 5915e604994aSStefano Zampini if (!pcbddc->current_level) { 5916e604994aSStefano Zampini ierr = PetscStrcpy(prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 5917e604994aSStefano Zampini ierr = PetscStrcat(prefix,"pc_bddc_coarse_");CHKERRQ(ierr); 5918c8587f34SStefano Zampini } else { 5919e604994aSStefano Zampini ierr = PetscStrlen(((PetscObject)pc)->prefix,&len);CHKERRQ(ierr); 5920312be037SStefano Zampini if (pcbddc->current_level>1) len -= 3; /* remove "lX_" with X level number */ 5921312be037SStefano Zampini if (pcbddc->current_level>10) len -= 1; /* remove another char from level number */ 592234d6797cSStefano Zampini ierr = PetscStrncpy(prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 5923312be037SStefano Zampini sprintf(str_level,"l%d_",(int)(pcbddc->current_level)); 5924e604994aSStefano Zampini ierr = PetscStrcat(prefix,str_level);CHKERRQ(ierr); 5925e604994aSStefano Zampini } 5926e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->coarse_ksp,prefix);CHKERRQ(ierr); 59273e3c6dadSStefano Zampini /* propagate BDDC info to the next level (these are dummy calls if pc_temp is not of type PCBDDC) */ 59283e3c6dadSStefano Zampini ierr = PCBDDCSetLevel(pc_temp,pcbddc->current_level+1);CHKERRQ(ierr); 59293e3c6dadSStefano Zampini ierr = PCBDDCSetCoarseningRatio(pc_temp,pcbddc->coarsening_ratio);CHKERRQ(ierr); 59303e3c6dadSStefano Zampini ierr = PCBDDCSetLevels(pc_temp,pcbddc->max_levels);CHKERRQ(ierr); 5931f9eb5b7dSStefano Zampini /* allow user customization */ 5932f9eb5b7dSStefano Zampini ierr = KSPSetFromOptions(pcbddc->coarse_ksp);CHKERRQ(ierr); 59333e3c6dadSStefano Zampini } 59343e3c6dadSStefano Zampini /* propagate BDDC info to the next level (these are dummy calls if pc_temp is not of type PCBDDC) */ 593551bea450SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&pc_temp);CHKERRQ(ierr); 59363e3c6dadSStefano Zampini if (nisdofs) { 59373e3c6dadSStefano Zampini ierr = PCBDDCSetDofsSplitting(pc_temp,nisdofs,isarray);CHKERRQ(ierr); 59383e3c6dadSStefano Zampini for (i=0;i<nisdofs;i++) { 59393e3c6dadSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 59403e3c6dadSStefano Zampini } 59413e3c6dadSStefano Zampini } 59423e3c6dadSStefano Zampini if (nisneu) { 59433e3c6dadSStefano Zampini ierr = PCBDDCSetNeumannBoundaries(pc_temp,isarray[nisdofs]);CHKERRQ(ierr); 59443e3c6dadSStefano Zampini ierr = ISDestroy(&isarray[nisdofs]);CHKERRQ(ierr); 5945312be037SStefano Zampini } 594630368db7SStefano Zampini if (nisvert) { 594730368db7SStefano Zampini ierr = PCBDDCSetPrimalVerticesIS(pc_temp,isarray[nis-1]);CHKERRQ(ierr); 594830368db7SStefano Zampini ierr = ISDestroy(&isarray[nis-1]);CHKERRQ(ierr); 594930368db7SStefano Zampini } 5950f9eb5b7dSStefano Zampini 5951f9eb5b7dSStefano Zampini /* get some info after set from options */ 5952f9eb5b7dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCNN,&isnn);CHKERRQ(ierr); 5953f9eb5b7dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCBDDC,&isbddc);CHKERRQ(ierr); 59544f3a063dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCREDUNDANT,&isredundant);CHKERRQ(ierr); 59556e683305SStefano Zampini if (isbddc && !multilevel_allowed) { /* multilevel can only be requested via pc_bddc_set_levels */ 5956f9eb5b7dSStefano Zampini ierr = PCSetType(pc_temp,coarse_pc_type);CHKERRQ(ierr); 5957f9eb5b7dSStefano Zampini isbddc = PETSC_FALSE; 5958f9eb5b7dSStefano Zampini } 595939f0f02cSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 59604f3a063dSStefano Zampini if (isredundant) { 59614f3a063dSStefano Zampini KSP inner_ksp; 59624f3a063dSStefano Zampini PC inner_pc; 59634f3a063dSStefano Zampini ierr = PCRedundantGetKSP(pc_temp,&inner_ksp);CHKERRQ(ierr); 59644f3a063dSStefano Zampini ierr = KSPGetPC(inner_ksp,&inner_pc);CHKERRQ(ierr); 59654f3a063dSStefano Zampini ierr = PCFactorSetReuseFill(inner_pc,PETSC_TRUE);CHKERRQ(ierr); 59664f3a063dSStefano Zampini } 5967f9eb5b7dSStefano Zampini 596857de7509SStefano Zampini /* parameters which miss an API */ 596957de7509SStefano Zampini if (isbddc) { 5970720d30f9SStefano Zampini PC_BDDC* pcbddc_coarse = (PC_BDDC*)pc_temp->data; 5971720d30f9SStefano Zampini pcbddc_coarse->detect_disconnected = PETSC_TRUE; 597257de7509SStefano Zampini pcbddc_coarse->coarse_eqs_per_proc = pcbddc->coarse_eqs_per_proc; 597327b6a85dSStefano Zampini pcbddc_coarse->benign_saddle_point = pcbddc->benign_have_null; 597427b6a85dSStefano Zampini if (pcbddc_coarse->benign_saddle_point) { 597527b6a85dSStefano Zampini pcbddc_coarse->benign_compute_nonetflux = PETSC_TRUE; 5976720d30f9SStefano Zampini pcbddc_coarse->adaptive_userdefined = PETSC_TRUE; 597759e48ca4SStefano Zampini if (pcbddc->adaptive_threshold < 1.0) pcbddc_coarse->deluxe_zerorows = PETSC_TRUE; 5978720d30f9SStefano Zampini } 5979d4d8cf7bSStefano Zampini } 59809881197aSStefano Zampini 59813301b35fSStefano Zampini /* propagate symmetry info of coarse matrix */ 59825a16e3a0SStefano Zampini ierr = MatSetOption(coarse_mat,MAT_STRUCTURALLY_SYMMETRIC,PETSC_TRUE);CHKERRQ(ierr); 59833301b35fSStefano Zampini if (pc->pmat->symmetric_set) { 59843301b35fSStefano Zampini ierr = MatSetOption(coarse_mat,MAT_SYMMETRIC,pc->pmat->symmetric);CHKERRQ(ierr); 59853301b35fSStefano Zampini } 59863301b35fSStefano Zampini if (pc->pmat->hermitian_set) { 59873301b35fSStefano Zampini ierr = MatSetOption(coarse_mat,MAT_HERMITIAN,pc->pmat->hermitian);CHKERRQ(ierr); 59883301b35fSStefano Zampini } 59893301b35fSStefano Zampini if (pc->pmat->spd_set) { 59903301b35fSStefano Zampini ierr = MatSetOption(coarse_mat,MAT_SPD,pc->pmat->spd);CHKERRQ(ierr); 59913301b35fSStefano Zampini } 599227b6a85dSStefano Zampini if (pcbddc->benign_saddle_point && !pcbddc->benign_have_null) { 599327b6a85dSStefano Zampini ierr = MatSetOption(coarse_mat,MAT_SPD,PETSC_TRUE);CHKERRQ(ierr); 599427b6a85dSStefano Zampini } 59956e683305SStefano Zampini /* set operators */ 59965f76c7aeSStefano Zampini ierr = KSPSetOperators(pcbddc->coarse_ksp,coarse_mat,coarse_mat);CHKERRQ(ierr); 59976e683305SStefano Zampini if (pcbddc->dbg_flag) { 59986e683305SStefano Zampini ierr = PetscViewerASCIISubtractTab(dbg_viewer,2*pcbddc->current_level);CHKERRQ(ierr); 59996e683305SStefano Zampini } 60006e683305SStefano Zampini } 60016e683305SStefano Zampini ierr = PetscFree(isarray);CHKERRQ(ierr); 6002b1ecc7b1SStefano Zampini #if 0 6003b9b85e73SStefano Zampini { 6004b9b85e73SStefano Zampini PetscViewer viewer; 6005b9b85e73SStefano Zampini char filename[256]; 6006b1ecc7b1SStefano Zampini sprintf(filename,"coarse_mat_level%d.m",pcbddc->current_level); 6007b1ecc7b1SStefano Zampini ierr = PetscViewerASCIIOpen(PetscObjectComm((PetscObject)coarse_mat),filename,&viewer);CHKERRQ(ierr); 60086a9046bcSBarry Smith ierr = PetscViewerPushFormat(viewer,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 6009b9b85e73SStefano Zampini ierr = MatView(coarse_mat,viewer);CHKERRQ(ierr); 6010f159cad9SBarry Smith ierr = PetscViewerPopFormat(viewer);CHKERRQ(ierr); 6011b9b85e73SStefano Zampini ierr = PetscViewerDestroy(&viewer);CHKERRQ(ierr); 6012b9b85e73SStefano Zampini } 6013b9b85e73SStefano Zampini #endif 6014f9eb5b7dSStefano Zampini 601598a51de6SStefano Zampini if (pcbddc->coarse_ksp) { 601698a51de6SStefano Zampini Vec crhs,csol; 601704708bb6SStefano Zampini 6018f347579bSStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&csol);CHKERRQ(ierr); 6019f347579bSStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&crhs);CHKERRQ(ierr); 6020f347579bSStefano Zampini if (!csol) { 60212a7a6963SBarry Smith ierr = MatCreateVecs(coarse_mat,&((pcbddc->coarse_ksp)->vec_sol),NULL);CHKERRQ(ierr); 6022f9eb5b7dSStefano Zampini } 6023f347579bSStefano Zampini if (!crhs) { 60242a7a6963SBarry Smith ierr = MatCreateVecs(coarse_mat,NULL,&((pcbddc->coarse_ksp)->vec_rhs));CHKERRQ(ierr); 6025f347579bSStefano Zampini } 6026b0f5fe93SStefano Zampini } 6027b0f5fe93SStefano Zampini 6028b0f5fe93SStefano Zampini /* compute null space for coarse solver if the benign trick has been requested */ 6029b0f5fe93SStefano Zampini if (pcbddc->benign_null) { 6030b0f5fe93SStefano Zampini 6031b0f5fe93SStefano Zampini ierr = VecSet(pcbddc->vec1_P,0.);CHKERRQ(ierr); 60324f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 60334f1b2e48SStefano Zampini ierr = VecSetValue(pcbddc->vec1_P,pcbddc->local_primal_size-pcbddc->benign_n+i,1.0,INSERT_VALUES);CHKERRQ(ierr); 60344f1b2e48SStefano Zampini } 6035b0f5fe93SStefano Zampini ierr = VecAssemblyBegin(pcbddc->vec1_P);CHKERRQ(ierr); 6036b0f5fe93SStefano Zampini ierr = VecAssemblyEnd(pcbddc->vec1_P);CHKERRQ(ierr); 6037b0f5fe93SStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6038b0f5fe93SStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6039b0f5fe93SStefano Zampini if (coarse_mat) { 6040b0f5fe93SStefano Zampini Vec nullv; 6041b0f5fe93SStefano Zampini PetscScalar *array,*array2; 6042b0f5fe93SStefano Zampini PetscInt nl; 6043b0f5fe93SStefano Zampini 6044b0f5fe93SStefano Zampini ierr = MatCreateVecs(coarse_mat,&nullv,NULL);CHKERRQ(ierr); 6045b0f5fe93SStefano Zampini ierr = VecGetLocalSize(nullv,&nl);CHKERRQ(ierr); 6046b0f5fe93SStefano Zampini ierr = VecGetArrayRead(pcbddc->coarse_vec,(const PetscScalar**)&array);CHKERRQ(ierr); 6047b0f5fe93SStefano Zampini ierr = VecGetArray(nullv,&array2);CHKERRQ(ierr); 6048b0f5fe93SStefano Zampini ierr = PetscMemcpy(array2,array,nl*sizeof(*array));CHKERRQ(ierr); 6049b0f5fe93SStefano Zampini ierr = VecRestoreArray(nullv,&array2);CHKERRQ(ierr); 6050b0f5fe93SStefano Zampini ierr = VecRestoreArrayRead(pcbddc->coarse_vec,(const PetscScalar**)&array);CHKERRQ(ierr); 6051b0f5fe93SStefano Zampini ierr = VecNormalize(nullv,NULL);CHKERRQ(ierr); 6052b0f5fe93SStefano Zampini ierr = MatNullSpaceCreate(PetscObjectComm((PetscObject)coarse_mat),PETSC_FALSE,1,&nullv,&CoarseNullSpace);CHKERRQ(ierr); 6053b0f5fe93SStefano Zampini ierr = VecDestroy(&nullv);CHKERRQ(ierr); 6054b0f5fe93SStefano Zampini } 6055b0f5fe93SStefano Zampini } 6056b0f5fe93SStefano Zampini 6057b0f5fe93SStefano Zampini if (pcbddc->coarse_ksp) { 6058b0f5fe93SStefano Zampini PetscBool ispreonly; 6059b0f5fe93SStefano Zampini 6060b0f5fe93SStefano Zampini if (CoarseNullSpace) { 6061b0f5fe93SStefano Zampini PetscBool isnull; 6062b0f5fe93SStefano Zampini ierr = MatNullSpaceTest(CoarseNullSpace,coarse_mat,&isnull);CHKERRQ(ierr); 6063*bef83e63SStefano Zampini if (isnull) { 6064b0f5fe93SStefano Zampini ierr = MatSetNullSpace(coarse_mat,CoarseNullSpace);CHKERRQ(ierr); 6065b0f5fe93SStefano Zampini } 6066*bef83e63SStefano Zampini /* TODO: add local nullspaces (if any) */ 6067b0f5fe93SStefano Zampini } 6068b0f5fe93SStefano Zampini /* setup coarse ksp */ 6069b0f5fe93SStefano Zampini ierr = KSPSetUp(pcbddc->coarse_ksp);CHKERRQ(ierr); 6070cbcc2c2aSStefano Zampini /* Check coarse problem if in debug mode or if solving with an iterative method */ 6071cbcc2c2aSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcbddc->coarse_ksp,KSPPREONLY,&ispreonly);CHKERRQ(ierr); 60726e683305SStefano Zampini if (pcbddc->dbg_flag || (!ispreonly && pcbddc->use_coarse_estimates) ) { 6073c8587f34SStefano Zampini KSP check_ksp; 60742b510759SStefano Zampini KSPType check_ksp_type; 6075c8587f34SStefano Zampini PC check_pc; 60766e683305SStefano Zampini Vec check_vec,coarse_vec; 60776a1308c2SStefano Zampini PetscReal abs_infty_error,infty_error,lambda_min=1.0,lambda_max=1.0; 60782b510759SStefano Zampini PetscInt its; 60796e683305SStefano Zampini PetscBool compute_eigs; 60806e683305SStefano Zampini PetscReal *eigs_r,*eigs_c; 60816e683305SStefano Zampini PetscInt neigs; 60828e185a42SStefano Zampini const char *prefix; 6083c8587f34SStefano Zampini 60842b510759SStefano Zampini /* Create ksp object suitable for estimation of extreme eigenvalues */ 60856e683305SStefano Zampini ierr = KSPCreate(PetscObjectComm((PetscObject)pcbddc->coarse_ksp),&check_ksp);CHKERRQ(ierr); 6086422a814eSBarry Smith ierr = KSPSetErrorIfNotConverged(pcbddc->coarse_ksp,pc->erroriffailure);CHKERRQ(ierr); 608723ee1639SBarry Smith ierr = KSPSetOperators(check_ksp,coarse_mat,coarse_mat);CHKERRQ(ierr); 6088f4ddd8eeSStefano Zampini ierr = KSPSetTolerances(check_ksp,1.e-12,1.e-12,PETSC_DEFAULT,pcbddc->coarse_size);CHKERRQ(ierr); 6089e4d548c7SStefano Zampini /* prevent from setup unneeded object */ 6090e4d548c7SStefano Zampini ierr = KSPGetPC(check_ksp,&check_pc);CHKERRQ(ierr); 6091e4d548c7SStefano Zampini ierr = PCSetType(check_pc,PCNONE);CHKERRQ(ierr); 60922b510759SStefano Zampini if (ispreonly) { 60932b510759SStefano Zampini check_ksp_type = KSPPREONLY; 60946e683305SStefano Zampini compute_eigs = PETSC_FALSE; 60952b510759SStefano Zampini } else { 6096cbcc2c2aSStefano Zampini check_ksp_type = KSPGMRES; 60976e683305SStefano Zampini compute_eigs = PETSC_TRUE; 6098c8587f34SStefano Zampini } 6099c8587f34SStefano Zampini ierr = KSPSetType(check_ksp,check_ksp_type);CHKERRQ(ierr); 61006e683305SStefano Zampini ierr = KSPSetComputeSingularValues(check_ksp,compute_eigs);CHKERRQ(ierr); 61016e683305SStefano Zampini ierr = KSPSetComputeEigenvalues(check_ksp,compute_eigs);CHKERRQ(ierr); 61026e683305SStefano Zampini ierr = KSPGMRESSetRestart(check_ksp,pcbddc->coarse_size+1);CHKERRQ(ierr); 6103a7dc3881SStefano Zampini ierr = KSPGetOptionsPrefix(pcbddc->coarse_ksp,&prefix);CHKERRQ(ierr); 6104a7dc3881SStefano Zampini ierr = KSPSetOptionsPrefix(check_ksp,prefix);CHKERRQ(ierr); 6105a7dc3881SStefano Zampini ierr = KSPAppendOptionsPrefix(check_ksp,"check_");CHKERRQ(ierr); 6106a7dc3881SStefano Zampini ierr = KSPSetFromOptions(check_ksp);CHKERRQ(ierr); 6107c8587f34SStefano Zampini ierr = KSPSetUp(check_ksp);CHKERRQ(ierr); 6108c8587f34SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&check_pc);CHKERRQ(ierr); 6109c8587f34SStefano Zampini ierr = KSPSetPC(check_ksp,check_pc);CHKERRQ(ierr); 6110c8587f34SStefano Zampini /* create random vec */ 61112701bc32SStefano Zampini ierr = MatCreateVecs(coarse_mat,&coarse_vec,&check_vec);CHKERRQ(ierr); 6112c8587f34SStefano Zampini ierr = VecSetRandom(check_vec,NULL);CHKERRQ(ierr); 61136e683305SStefano Zampini ierr = MatMult(coarse_mat,check_vec,coarse_vec);CHKERRQ(ierr); 6114c8587f34SStefano Zampini /* solve coarse problem */ 61156e683305SStefano Zampini ierr = KSPSolve(check_ksp,coarse_vec,coarse_vec);CHKERRQ(ierr); 6116cbcc2c2aSStefano Zampini /* set eigenvalue estimation if preonly has not been requested */ 61176e683305SStefano Zampini if (compute_eigs) { 6118854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->coarse_size+1,&eigs_r);CHKERRQ(ierr); 6119854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->coarse_size+1,&eigs_c);CHKERRQ(ierr); 61206e683305SStefano Zampini ierr = KSPComputeEigenvalues(check_ksp,pcbddc->coarse_size+1,eigs_r,eigs_c,&neigs);CHKERRQ(ierr); 61216e683305SStefano Zampini lambda_max = eigs_r[neigs-1]; 61226e683305SStefano Zampini lambda_min = eigs_r[0]; 61236e683305SStefano Zampini if (pcbddc->use_coarse_estimates) { 61242701bc32SStefano Zampini if (lambda_max>=lambda_min) { /* using PETSC_SMALL since lambda_max == lambda_min is not allowed by KSPChebyshevSetEigenvalues */ 61252701bc32SStefano Zampini ierr = KSPChebyshevSetEigenvalues(pcbddc->coarse_ksp,lambda_max+PETSC_SMALL,lambda_min);CHKERRQ(ierr); 6126cbcc2c2aSStefano Zampini ierr = KSPRichardsonSetScale(pcbddc->coarse_ksp,2.0/(lambda_max+lambda_min));CHKERRQ(ierr); 6127cbcc2c2aSStefano Zampini } 6128c8587f34SStefano Zampini } 6129c8587f34SStefano Zampini } 6130cbcc2c2aSStefano Zampini 6131c8587f34SStefano Zampini /* check coarse problem residual error */ 61326e683305SStefano Zampini if (pcbddc->dbg_flag) { 61336e683305SStefano Zampini PetscViewer dbg_viewer = PETSC_VIEWER_STDOUT_(PetscObjectComm((PetscObject)pcbddc->coarse_ksp)); 61346e683305SStefano Zampini ierr = PetscViewerASCIIAddTab(dbg_viewer,2*(pcbddc->current_level+1));CHKERRQ(ierr); 61356e683305SStefano Zampini ierr = VecAXPY(check_vec,-1.0,coarse_vec);CHKERRQ(ierr); 6136c8587f34SStefano Zampini ierr = VecNorm(check_vec,NORM_INFINITY,&infty_error);CHKERRQ(ierr); 61376e683305SStefano Zampini ierr = MatMult(coarse_mat,check_vec,coarse_vec);CHKERRQ(ierr); 61386e683305SStefano Zampini ierr = VecNorm(coarse_vec,NORM_INFINITY,&abs_infty_error);CHKERRQ(ierr); 6139779c1cceSStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem details (use estimates %d)\n",pcbddc->use_coarse_estimates);CHKERRQ(ierr); 61406e683305SStefano Zampini ierr = PetscObjectPrintClassNamePrefixType((PetscObject)(pcbddc->coarse_ksp),dbg_viewer);CHKERRQ(ierr); 61416e683305SStefano Zampini ierr = PetscObjectPrintClassNamePrefixType((PetscObject)(check_pc),dbg_viewer);CHKERRQ(ierr); 61426e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem exact infty_error : %1.6e\n",infty_error);CHKERRQ(ierr); 61436e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem residual infty_error: %1.6e\n",abs_infty_error);CHKERRQ(ierr); 6144b0f5fe93SStefano Zampini if (CoarseNullSpace) { 6145b0f5fe93SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem is singular\n");CHKERRQ(ierr); 6146b0f5fe93SStefano Zampini } 61476e683305SStefano Zampini if (compute_eigs) { 61486e683305SStefano Zampini PetscReal lambda_max_s,lambda_min_s; 6149deec49d1SStefano Zampini ierr = KSPGetType(check_ksp,&check_ksp_type);CHKERRQ(ierr); 6150c8587f34SStefano Zampini ierr = KSPGetIterationNumber(check_ksp,&its);CHKERRQ(ierr); 61516e683305SStefano Zampini ierr = KSPComputeExtremeSingularValues(check_ksp,&lambda_max_s,&lambda_min_s);CHKERRQ(ierr); 61526e683305SStefano 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); 61536e683305SStefano Zampini for (i=0;i<neigs;i++) { 61546e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"%1.6e %1.6ei\n",eigs_r[i],eigs_c[i]);CHKERRQ(ierr); 6155c8587f34SStefano Zampini } 61566e683305SStefano Zampini } 61576e683305SStefano Zampini ierr = PetscViewerFlush(dbg_viewer);CHKERRQ(ierr); 61586e683305SStefano Zampini ierr = PetscViewerASCIISubtractTab(dbg_viewer,2*(pcbddc->current_level+1));CHKERRQ(ierr); 61596e683305SStefano Zampini } 6160e4d548c7SStefano Zampini ierr = VecDestroy(&check_vec);CHKERRQ(ierr); 61612701bc32SStefano Zampini ierr = VecDestroy(&coarse_vec);CHKERRQ(ierr); 6162c8587f34SStefano Zampini ierr = KSPDestroy(&check_ksp);CHKERRQ(ierr); 61636e683305SStefano Zampini if (compute_eigs) { 61646e683305SStefano Zampini ierr = PetscFree(eigs_r);CHKERRQ(ierr); 61656e683305SStefano Zampini ierr = PetscFree(eigs_c);CHKERRQ(ierr); 6166c8587f34SStefano Zampini } 61676e683305SStefano Zampini } 61686e683305SStefano Zampini } 6169*bef83e63SStefano Zampini ierr = MatNullSpaceDestroy(&CoarseNullSpace);CHKERRQ(ierr); 6170cbcc2c2aSStefano Zampini /* print additional info */ 6171cbcc2c2aSStefano Zampini if (pcbddc->dbg_flag) { 61726e683305SStefano Zampini /* waits until all processes reaches this point */ 61736e683305SStefano Zampini ierr = PetscBarrier((PetscObject)pc);CHKERRQ(ierr); 6174cbcc2c2aSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Coarse solver setup completed at level %d\n",pcbddc->current_level);CHKERRQ(ierr); 6175cbcc2c2aSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 6176cbcc2c2aSStefano Zampini } 6177cbcc2c2aSStefano Zampini 61782b510759SStefano Zampini /* free memory */ 6179fdc635d7SStefano Zampini ierr = MatDestroy(&coarse_mat);CHKERRQ(ierr); 6180c8587f34SStefano Zampini PetscFunctionReturn(0); 6181c8587f34SStefano Zampini } 6182674ae819SStefano Zampini 6183f34684f1SStefano Zampini #undef __FUNCT__ 6184f34684f1SStefano Zampini #define __FUNCT__ "PCBDDCComputePrimalNumbering" 6185f34684f1SStefano Zampini PetscErrorCode PCBDDCComputePrimalNumbering(PC pc,PetscInt* coarse_size_n,PetscInt** local_primal_indices_n) 6186f34684f1SStefano Zampini { 6187f34684f1SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 6188f34684f1SStefano Zampini PC_IS* pcis = (PC_IS*)pc->data; 6189f34684f1SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 6190dc456d91SStefano Zampini IS subset,subset_mult,subset_n; 6191dc456d91SStefano Zampini PetscInt local_size,coarse_size=0; 619273be2a3aSStefano Zampini PetscInt *local_primal_indices=NULL; 6193dc456d91SStefano Zampini const PetscInt *t_local_primal_indices; 6194f34684f1SStefano Zampini PetscErrorCode ierr; 6195f34684f1SStefano Zampini 6196f34684f1SStefano Zampini PetscFunctionBegin; 6197f34684f1SStefano Zampini /* Compute global number of coarse dofs */ 61986c4ed002SBarry Smith if (pcbddc->local_primal_size && !pcbddc->local_primal_ref_node) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"BDDC ConstraintsSetUp should be called first"); 6199dc456d91SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)(pc->pmat)),pcbddc->local_primal_size_cc,pcbddc->local_primal_ref_node,PETSC_COPY_VALUES,&subset_n);CHKERRQ(ierr); 62003bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(pcis->mapping,subset_n,&subset);CHKERRQ(ierr); 6201dc456d91SStefano Zampini ierr = ISDestroy(&subset_n);CHKERRQ(ierr); 6202dc456d91SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)(pc->pmat)),pcbddc->local_primal_size_cc,pcbddc->local_primal_ref_mult,PETSC_COPY_VALUES,&subset_mult);CHKERRQ(ierr); 6203dc456d91SStefano Zampini ierr = PCBDDCSubsetNumbering(subset,subset_mult,&coarse_size,&subset_n);CHKERRQ(ierr); 6204dc456d91SStefano Zampini ierr = ISDestroy(&subset);CHKERRQ(ierr); 6205dc456d91SStefano Zampini ierr = ISDestroy(&subset_mult);CHKERRQ(ierr); 6206dc456d91SStefano Zampini ierr = ISGetLocalSize(subset_n,&local_size);CHKERRQ(ierr); 62076c4ed002SBarry Smith if (local_size != pcbddc->local_primal_size) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Invalid number of local primal indices computed %D != %D",local_size,pcbddc->local_primal_size); 6208dc456d91SStefano Zampini ierr = PetscMalloc1(local_size,&local_primal_indices);CHKERRQ(ierr); 6209dc456d91SStefano Zampini ierr = ISGetIndices(subset_n,&t_local_primal_indices);CHKERRQ(ierr); 6210dc456d91SStefano Zampini ierr = PetscMemcpy(local_primal_indices,t_local_primal_indices,local_size*sizeof(PetscInt));CHKERRQ(ierr); 6211dc456d91SStefano Zampini ierr = ISRestoreIndices(subset_n,&t_local_primal_indices);CHKERRQ(ierr); 6212dc456d91SStefano Zampini ierr = ISDestroy(&subset_n);CHKERRQ(ierr); 6213f34684f1SStefano Zampini 6214f34684f1SStefano Zampini /* check numbering */ 6215f34684f1SStefano Zampini if (pcbddc->dbg_flag) { 6216019a44ceSStefano Zampini PetscScalar coarsesum,*array,*array2; 6217dc456d91SStefano Zampini PetscInt i; 6218b9b85e73SStefano Zampini PetscBool set_error = PETSC_FALSE,set_error_reduced = PETSC_FALSE; 6219f34684f1SStefano Zampini 6220f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 6221f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 6222f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Check coarse indices\n");CHKERRQ(ierr); 62231575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 6224019a44ceSStefano Zampini /* counter */ 6225019a44ceSStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 6226019a44ceSStefano Zampini ierr = VecSet(pcis->vec1_N,1.0);CHKERRQ(ierr); 6227019a44ceSStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6228019a44ceSStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6229019a44ceSStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_global,pcis->vec2_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6230019a44ceSStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_global,pcis->vec2_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6231f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_N,0.0);CHKERRQ(ierr); 6232f34684f1SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 6233727cdba6SStefano Zampini ierr = VecSetValue(pcis->vec1_N,pcbddc->primal_indices_local_idxs[i],1.0,INSERT_VALUES);CHKERRQ(ierr); 6234f34684f1SStefano Zampini } 6235f34684f1SStefano Zampini ierr = VecAssemblyBegin(pcis->vec1_N);CHKERRQ(ierr); 6236f34684f1SStefano Zampini ierr = VecAssemblyEnd(pcis->vec1_N);CHKERRQ(ierr); 6237f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 6238e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6239e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6240e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_global,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6241e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_global,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6242f34684f1SStefano Zampini ierr = VecGetArray(pcis->vec1_N,&array);CHKERRQ(ierr); 6243019a44ceSStefano Zampini ierr = VecGetArray(pcis->vec2_N,&array2);CHKERRQ(ierr); 6244f34684f1SStefano Zampini for (i=0;i<pcis->n;i++) { 6245019a44ceSStefano Zampini if (array[i] != 0.0 && array[i] != array2[i]) { 62462c66d082SStefano Zampini PetscInt owned = (PetscInt)PetscRealPart(array[i]),gi; 624775c01103SStefano Zampini PetscInt neigh = (PetscInt)PetscRealPart(array2[i]); 6248b9b85e73SStefano Zampini set_error = PETSC_TRUE; 62492c66d082SStefano Zampini ierr = ISLocalToGlobalMappingApply(pcis->mapping,1,&i,&gi);CHKERRQ(ierr); 62502c66d082SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d: local index %d (gid %d) owned by %d processes instead of %d!\n",PetscGlobalRank,i,gi,owned,neigh);CHKERRQ(ierr); 6251f34684f1SStefano Zampini } 6252f34684f1SStefano Zampini } 6253019a44ceSStefano Zampini ierr = VecRestoreArray(pcis->vec2_N,&array2);CHKERRQ(ierr); 6254b2566f29SBarry Smith ierr = MPIU_Allreduce(&set_error,&set_error_reduced,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 6255f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 6256f34684f1SStefano Zampini for (i=0;i<pcis->n;i++) { 6257f34684f1SStefano Zampini if (PetscRealPart(array[i]) > 0.0) array[i] = 1.0/PetscRealPart(array[i]); 6258f34684f1SStefano Zampini } 6259f34684f1SStefano Zampini ierr = VecRestoreArray(pcis->vec1_N,&array);CHKERRQ(ierr); 6260f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 6261e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6262e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6263f34684f1SStefano Zampini ierr = VecSum(pcis->vec1_global,&coarsesum);CHKERRQ(ierr); 6264f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Size of coarse problem is %d (%lf)\n",coarse_size,PetscRealPart(coarsesum));CHKERRQ(ierr); 6265b9b85e73SStefano Zampini if (pcbddc->dbg_flag > 1 || set_error_reduced) { 6266ca8b9ea9SStefano Zampini PetscInt *gidxs; 6267ca8b9ea9SStefano Zampini 6268ca8b9ea9SStefano Zampini ierr = PetscMalloc1(pcbddc->local_primal_size,&gidxs);CHKERRQ(ierr); 62693bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingApply(pcis->mapping,pcbddc->local_primal_size,pcbddc->primal_indices_local_idxs,gidxs);CHKERRQ(ierr); 6270f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Distribution of local primal indices\n");CHKERRQ(ierr); 6271f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 6272f34684f1SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d\n",PetscGlobalRank);CHKERRQ(ierr); 6273f34684f1SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 62744bc2dc4bSStefano 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); 6275f34684f1SStefano Zampini } 6276f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 6277ca8b9ea9SStefano Zampini ierr = PetscFree(gidxs);CHKERRQ(ierr); 6278f34684f1SStefano Zampini } 6279f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 62801575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 6281302440fdSBarry Smith if (set_error_reduced) SETERRQ(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"BDDC Numbering of coarse dofs failed"); 6282f34684f1SStefano Zampini } 62838bec7fa6SStefano Zampini /* ierr = PetscPrintf(PetscObjectComm((PetscObject)pc),"Size of coarse problem is %d\n",coarse_size);CHKERRQ(ierr); */ 6284f34684f1SStefano Zampini /* get back data */ 6285f34684f1SStefano Zampini *coarse_size_n = coarse_size; 6286f34684f1SStefano Zampini *local_primal_indices_n = local_primal_indices; 6287674ae819SStefano Zampini PetscFunctionReturn(0); 6288674ae819SStefano Zampini } 6289674ae819SStefano Zampini 6290e456f2a8SStefano Zampini #undef __FUNCT__ 6291e456f2a8SStefano Zampini #define __FUNCT__ "PCBDDCGlobalToLocal" 6292a7dc3881SStefano Zampini PetscErrorCode PCBDDCGlobalToLocal(VecScatter g2l_ctx,Vec gwork, Vec lwork, IS globalis, IS* localis) 6293e456f2a8SStefano Zampini { 6294e456f2a8SStefano Zampini IS localis_t; 6295a7dc3881SStefano Zampini PetscInt i,lsize,*idxs,n; 6296e456f2a8SStefano Zampini PetscScalar *vals; 6297e456f2a8SStefano Zampini PetscErrorCode ierr; 6298e456f2a8SStefano Zampini 6299e456f2a8SStefano Zampini PetscFunctionBegin; 6300a7dc3881SStefano Zampini /* get indices in local ordering exploiting local to global map */ 6301e456f2a8SStefano Zampini ierr = ISGetLocalSize(globalis,&lsize);CHKERRQ(ierr); 6302854ce69bSBarry Smith ierr = PetscMalloc1(lsize,&vals);CHKERRQ(ierr); 6303e456f2a8SStefano Zampini for (i=0;i<lsize;i++) vals[i] = 1.0; 6304e456f2a8SStefano Zampini ierr = ISGetIndices(globalis,(const PetscInt**)&idxs);CHKERRQ(ierr); 6305a7dc3881SStefano Zampini ierr = VecSet(gwork,0.0);CHKERRQ(ierr); 6306a7dc3881SStefano Zampini ierr = VecSet(lwork,0.0);CHKERRQ(ierr); 63071035eff8SStefano Zampini if (idxs) { /* multilevel guard */ 6308a7dc3881SStefano Zampini ierr = VecSetValues(gwork,lsize,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 63091035eff8SStefano Zampini } 6310a7dc3881SStefano Zampini ierr = VecAssemblyBegin(gwork);CHKERRQ(ierr); 6311e456f2a8SStefano Zampini ierr = ISRestoreIndices(globalis,(const PetscInt**)&idxs);CHKERRQ(ierr); 6312e456f2a8SStefano Zampini ierr = PetscFree(vals);CHKERRQ(ierr); 6313a7dc3881SStefano Zampini ierr = VecAssemblyEnd(gwork);CHKERRQ(ierr); 6314a7dc3881SStefano Zampini /* now compute set in local ordering */ 6315a7dc3881SStefano Zampini ierr = VecScatterBegin(g2l_ctx,gwork,lwork,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6316a7dc3881SStefano Zampini ierr = VecScatterEnd(g2l_ctx,gwork,lwork,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6317a7dc3881SStefano Zampini ierr = VecGetArrayRead(lwork,(const PetscScalar**)&vals);CHKERRQ(ierr); 6318a7dc3881SStefano Zampini ierr = VecGetSize(lwork,&n);CHKERRQ(ierr); 6319a7dc3881SStefano Zampini for (i=0,lsize=0;i<n;i++) { 6320ff92baa0SMatthew G. Knepley if (PetscRealPart(vals[i]) > 0.5) { 6321e456f2a8SStefano Zampini lsize++; 6322e456f2a8SStefano Zampini } 6323e456f2a8SStefano Zampini } 6324854ce69bSBarry Smith ierr = PetscMalloc1(lsize,&idxs);CHKERRQ(ierr); 6325a7dc3881SStefano Zampini for (i=0,lsize=0;i<n;i++) { 6326ff92baa0SMatthew G. Knepley if (PetscRealPart(vals[i]) > 0.5) { 6327e456f2a8SStefano Zampini idxs[lsize++] = i; 6328e456f2a8SStefano Zampini } 6329e456f2a8SStefano Zampini } 6330a7dc3881SStefano Zampini ierr = VecRestoreArrayRead(lwork,(const PetscScalar**)&vals);CHKERRQ(ierr); 6331a7dc3881SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)gwork),lsize,idxs,PETSC_OWN_POINTER,&localis_t);CHKERRQ(ierr); 6332e456f2a8SStefano Zampini *localis = localis_t; 6333e456f2a8SStefano Zampini PetscFunctionReturn(0); 6334e456f2a8SStefano Zampini } 6335906d46d4SStefano Zampini 6336b96c3477SStefano Zampini #undef __FUNCT__ 6337b96c3477SStefano Zampini #define __FUNCT__ "PCBDDCSetUpSubSchurs" 633808122e43SStefano Zampini PetscErrorCode PCBDDCSetUpSubSchurs(PC pc) 6339b96c3477SStefano Zampini { 6340a64f4aa4SStefano Zampini PC_IS *pcis=(PC_IS*)pc->data; 6341b96c3477SStefano Zampini PC_BDDC *pcbddc=(PC_BDDC*)pc->data; 6342b96c3477SStefano Zampini PCBDDCSubSchurs sub_schurs=pcbddc->sub_schurs; 6343a64f4aa4SStefano Zampini Mat S_j; 6344b96c3477SStefano Zampini PetscInt *used_xadj,*used_adjncy; 6345b96c3477SStefano Zampini PetscBool free_used_adj; 6346b96c3477SStefano Zampini PetscErrorCode ierr; 6347b96c3477SStefano Zampini 6348b96c3477SStefano Zampini PetscFunctionBegin; 6349b96c3477SStefano Zampini /* decide the adjacency to be used for determining internal problems for local schur on subsets */ 6350b96c3477SStefano Zampini free_used_adj = PETSC_FALSE; 635108122e43SStefano Zampini if (pcbddc->sub_schurs_layers == -1) { 6352b96c3477SStefano Zampini used_xadj = NULL; 6353b96c3477SStefano Zampini used_adjncy = NULL; 6354b96c3477SStefano Zampini } else { 635508122e43SStefano Zampini if (pcbddc->sub_schurs_use_useradj && pcbddc->mat_graph->xadj) { 635608122e43SStefano Zampini used_xadj = pcbddc->mat_graph->xadj; 635708122e43SStefano Zampini used_adjncy = pcbddc->mat_graph->adjncy; 635808122e43SStefano Zampini } else if (pcbddc->computed_rowadj) { 6359b96c3477SStefano Zampini used_xadj = pcbddc->mat_graph->xadj; 6360b96c3477SStefano Zampini used_adjncy = pcbddc->mat_graph->adjncy; 6361b96c3477SStefano Zampini } else { 63622fffb893SStefano Zampini PetscBool flg_row=PETSC_FALSE; 6363b96c3477SStefano Zampini const PetscInt *xadj,*adjncy; 6364b96c3477SStefano Zampini PetscInt nvtxs; 6365b96c3477SStefano Zampini 63662fffb893SStefano Zampini ierr = MatGetRowIJ(pcbddc->local_mat,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 63672fffb893SStefano Zampini if (flg_row) { 6368b96c3477SStefano Zampini ierr = PetscMalloc2(nvtxs+1,&used_xadj,xadj[nvtxs],&used_adjncy);CHKERRQ(ierr); 6369b96c3477SStefano Zampini ierr = PetscMemcpy(used_xadj,xadj,(nvtxs+1)*sizeof(*xadj));CHKERRQ(ierr); 6370b96c3477SStefano Zampini ierr = PetscMemcpy(used_adjncy,adjncy,(xadj[nvtxs])*sizeof(*adjncy));CHKERRQ(ierr); 6371b96c3477SStefano Zampini free_used_adj = PETSC_TRUE; 63722fffb893SStefano Zampini } else { 63732fffb893SStefano Zampini pcbddc->sub_schurs_layers = -1; 63742fffb893SStefano Zampini used_xadj = NULL; 63752fffb893SStefano Zampini used_adjncy = NULL; 63762fffb893SStefano Zampini } 63772fffb893SStefano Zampini ierr = MatRestoreRowIJ(pcbddc->local_mat,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 6378b96c3477SStefano Zampini } 6379b96c3477SStefano Zampini } 6380d5574798SStefano Zampini 6381d5574798SStefano Zampini /* setup sub_schurs data */ 6382a64f4aa4SStefano Zampini ierr = MatCreateSchurComplement(pcis->A_II,pcis->A_II,pcis->A_IB,pcis->A_BI,pcis->A_BB,&S_j);CHKERRQ(ierr); 6383df4d28bfSStefano Zampini if (!sub_schurs->schur_explicit) { 6384df4d28bfSStefano Zampini /* pcbddc->ksp_D up to date only if not using MatFactor with Schur complement support */ 6385a64f4aa4SStefano Zampini ierr = MatSchurComplementSetKSP(S_j,pcbddc->ksp_D);CHKERRQ(ierr); 638691af6908SStefano Zampini ierr = PCBDDCSubSchursSetUp(sub_schurs,NULL,S_j,PETSC_FALSE,used_xadj,used_adjncy,pcbddc->sub_schurs_layers,NULL,pcbddc->adaptive_selection,PETSC_FALSE,PETSC_FALSE,0,NULL,NULL,NULL,NULL);CHKERRQ(ierr); 6387a64f4aa4SStefano Zampini } else { 63886816873aSStefano Zampini PetscBool reuse_solvers = (PetscBool)!pcbddc->use_change_of_basis; 6389b7ab4a40SStefano Zampini PetscBool isseqaij,need_change = PETSC_FALSE;; 6390a3df083aSStefano Zampini PetscInt benign_n; 639172b8c272SStefano Zampini Mat change = NULL; 63929d54b7f4SStefano Zampini Vec scaling = NULL; 639372b8c272SStefano Zampini IS change_primal = NULL; 6394a3df083aSStefano Zampini 63955feab87aSStefano Zampini if (!pcbddc->use_vertices && reuse_solvers) { 63965feab87aSStefano Zampini PetscInt n_vertices; 63975feab87aSStefano Zampini 63985feab87aSStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_vertices,&n_vertices);CHKERRQ(ierr); 63992034aafcSStefano Zampini reuse_solvers = (PetscBool)!n_vertices; 64005feab87aSStefano Zampini } 640104708bb6SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcbddc->local_mat,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 640204708bb6SStefano Zampini if (!isseqaij) { 640304708bb6SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 640404708bb6SStefano Zampini if (matis->A == pcbddc->local_mat) { 640504708bb6SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 640604708bb6SStefano Zampini ierr = MatConvert(matis->A,MATSEQAIJ,MAT_INITIAL_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 640704708bb6SStefano Zampini } else { 6408511c6705SHong Zhang ierr = MatConvert(pcbddc->local_mat,MATSEQAIJ,MAT_INPLACE_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 640904708bb6SStefano Zampini } 641004708bb6SStefano Zampini } 6411a3df083aSStefano Zampini if (!pcbddc->benign_change_explicit) { 6412a3df083aSStefano Zampini benign_n = pcbddc->benign_n; 6413ca92afb2SStefano Zampini } else { 6414a3df083aSStefano Zampini benign_n = 0; 6415ca92afb2SStefano Zampini } 6416b7ab4a40SStefano Zampini /* sub_schurs->change is a local object; instead, PCBDDCConstraintsSetUp and the quantities used in the test below are logically collective on pc. 6417b7ab4a40SStefano Zampini We need a global reduction to avoid possible deadlocks. 6418b7ab4a40SStefano Zampini We assume that sub_schurs->change is created once, and then reused for different solves, unless the topography has been recomputed */ 641972b8c272SStefano Zampini if (pcbddc->adaptive_userdefined || (pcbddc->deluxe_zerorows && !pcbddc->use_change_of_basis)) { 642022db5ddcSStefano Zampini PetscBool have_loc_change = (PetscBool)(!!sub_schurs->change); 6421b7ab4a40SStefano Zampini ierr = MPIU_Allreduce(&have_loc_change,&need_change,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 642222db5ddcSStefano Zampini need_change = (PetscBool)(!need_change); 6423b7ab4a40SStefano Zampini } 6424b7ab4a40SStefano Zampini /* If the user defines additional constraints, we import them here. 6425b7ab4a40SStefano Zampini We need to compute the change of basis according to the quadrature weights attached to pmat via MatSetNearNullSpace, and this could not be done (at the moment) without some hacking */ 6426b7ab4a40SStefano Zampini if (need_change) { 642788c03ad3SStefano Zampini PC_IS *pcisf; 642888c03ad3SStefano Zampini PC_BDDC *pcbddcf; 642988c03ad3SStefano Zampini PC pcf; 643088c03ad3SStefano Zampini 6431e4d548c7SStefano Zampini if (pcbddc->sub_schurs_rebuild) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Cannot compute change of basis with a different graph"); 643288c03ad3SStefano Zampini ierr = PCCreate(PetscObjectComm((PetscObject)pc),&pcf);CHKERRQ(ierr); 643388c03ad3SStefano Zampini ierr = PCSetOperators(pcf,pc->mat,pc->pmat);CHKERRQ(ierr); 643488c03ad3SStefano Zampini ierr = PCSetType(pcf,PCBDDC);CHKERRQ(ierr); 643588c03ad3SStefano Zampini /* hacks */ 643688c03ad3SStefano Zampini pcisf = (PC_IS*)pcf->data; 643772b8c272SStefano Zampini pcisf->is_B_local = pcis->is_B_local; 643872b8c272SStefano Zampini pcisf->vec1_N = pcis->vec1_N; 643972b8c272SStefano Zampini pcisf->BtoNmap = pcis->BtoNmap; 644072b8c272SStefano Zampini pcisf->n = pcis->n; 644172b8c272SStefano Zampini pcisf->n_B = pcis->n_B; 644288c03ad3SStefano Zampini pcbddcf = (PC_BDDC*)pcf->data; 644388c03ad3SStefano Zampini ierr = PetscFree(pcbddcf->mat_graph);CHKERRQ(ierr); 644488c03ad3SStefano Zampini pcbddcf->mat_graph = pcbddc->mat_graph; 644588c03ad3SStefano Zampini pcbddcf->use_faces = PETSC_TRUE; 644688c03ad3SStefano Zampini pcbddcf->use_change_of_basis = PETSC_TRUE; 644788c03ad3SStefano Zampini pcbddcf->use_change_on_faces = PETSC_TRUE; 644872b8c272SStefano Zampini pcbddcf->use_qr_single = PETSC_TRUE; 644988c03ad3SStefano Zampini pcbddcf->fake_change = PETSC_TRUE; 645088c03ad3SStefano Zampini ierr = PCBDDCConstraintsSetUp(pcf);CHKERRQ(ierr); 645172b8c272SStefano Zampini /* store information on primal vertices and change of basis (in local numbering) */ 645272b8c272SStefano Zampini sub_schurs->change_with_qr = pcbddcf->use_qr_single; 645372b8c272SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,pcbddcf->n_vertices,pcbddcf->local_primal_ref_node,PETSC_COPY_VALUES,&change_primal);CHKERRQ(ierr); 645472b8c272SStefano Zampini change = pcbddcf->ConstraintMatrix; 645572b8c272SStefano Zampini pcbddcf->ConstraintMatrix = NULL; 645688c03ad3SStefano Zampini /* free unneeded memory allocated in PCBDDCConstraintsSetUp */ 645772b8c272SStefano Zampini ierr = PetscFree(pcbddcf->sub_schurs);CHKERRQ(ierr); 645888c03ad3SStefano Zampini ierr = MatNullSpaceDestroy(&pcbddcf->onearnullspace);CHKERRQ(ierr); 645988c03ad3SStefano Zampini ierr = PetscFree2(pcbddcf->local_primal_ref_node,pcbddcf->local_primal_ref_mult);CHKERRQ(ierr); 646088c03ad3SStefano Zampini ierr = PetscFree(pcbddcf->primal_indices_local_idxs);CHKERRQ(ierr); 646188c03ad3SStefano Zampini ierr = PetscFree(pcbddcf->onearnullvecs_state);CHKERRQ(ierr); 646288c03ad3SStefano Zampini ierr = PetscFree(pcf->data);CHKERRQ(ierr); 646388c03ad3SStefano Zampini pcf->ops->destroy = NULL; 646488c03ad3SStefano Zampini ierr = PCDestroy(&pcf);CHKERRQ(ierr); 646588c03ad3SStefano Zampini } 64669d54b7f4SStefano Zampini if (!pcbddc->use_deluxe_scaling) scaling = pcis->D; 646791af6908SStefano Zampini ierr = PCBDDCSubSchursSetUp(sub_schurs,pcbddc->local_mat,S_j,pcbddc->sub_schurs_exact_schur,used_xadj,used_adjncy,pcbddc->sub_schurs_layers,scaling,pcbddc->adaptive_selection,reuse_solvers,pcbddc->benign_saddle_point,benign_n,pcbddc->benign_p0_lidx,pcbddc->benign_zerodiag_subs,change,change_primal);CHKERRQ(ierr); 646872b8c272SStefano Zampini ierr = MatDestroy(&change);CHKERRQ(ierr); 646972b8c272SStefano Zampini ierr = ISDestroy(&change_primal);CHKERRQ(ierr); 6470ca92afb2SStefano Zampini } 6471d12d3064SStefano Zampini ierr = MatDestroy(&S_j);CHKERRQ(ierr); 6472b96c3477SStefano Zampini 6473b96c3477SStefano Zampini /* free adjacency */ 6474b96c3477SStefano Zampini if (free_used_adj) { 6475b96c3477SStefano Zampini ierr = PetscFree2(used_xadj,used_adjncy);CHKERRQ(ierr); 6476b96c3477SStefano Zampini } 6477b96c3477SStefano Zampini PetscFunctionReturn(0); 6478b96c3477SStefano Zampini } 6479b96c3477SStefano Zampini 6480b96c3477SStefano Zampini #undef __FUNCT__ 6481b96c3477SStefano Zampini #define __FUNCT__ "PCBDDCInitSubSchurs" 648208122e43SStefano Zampini PetscErrorCode PCBDDCInitSubSchurs(PC pc) 6483b96c3477SStefano Zampini { 6484b96c3477SStefano Zampini PC_IS *pcis=(PC_IS*)pc->data; 6485b96c3477SStefano Zampini PC_BDDC *pcbddc=(PC_BDDC*)pc->data; 6486b96c3477SStefano Zampini PCBDDCSubSchurs sub_schurs=pcbddc->sub_schurs; 6487b96c3477SStefano Zampini PCBDDCGraph graph; 6488b96c3477SStefano Zampini PetscErrorCode ierr; 6489b96c3477SStefano Zampini 6490b96c3477SStefano Zampini PetscFunctionBegin; 6491b96c3477SStefano Zampini /* attach interface graph for determining subsets */ 649208122e43SStefano Zampini if (pcbddc->sub_schurs_rebuild) { /* in case rebuild has been requested, it uses a graph generated only by the neighbouring information */ 64933301b35fSStefano Zampini IS verticesIS,verticescomm; 64943301b35fSStefano Zampini PetscInt vsize,*idxs; 6495b96c3477SStefano Zampini 6496b96c3477SStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,NULL,NULL,NULL,NULL,&verticesIS);CHKERRQ(ierr); 64973301b35fSStefano Zampini ierr = ISGetSize(verticesIS,&vsize);CHKERRQ(ierr); 64983301b35fSStefano Zampini ierr = ISGetIndices(verticesIS,(const PetscInt**)&idxs);CHKERRQ(ierr); 64993301b35fSStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),vsize,idxs,PETSC_COPY_VALUES,&verticescomm);CHKERRQ(ierr); 65003301b35fSStefano Zampini ierr = ISRestoreIndices(verticesIS,(const PetscInt**)&idxs);CHKERRQ(ierr); 65013301b35fSStefano Zampini ierr = ISDestroy(&verticesIS);CHKERRQ(ierr); 6502b96c3477SStefano Zampini ierr = PCBDDCGraphCreate(&graph);CHKERRQ(ierr); 65037fb0e2dbSStefano Zampini ierr = PCBDDCGraphInit(graph,pcbddc->mat_graph->l2gmap,pcbddc->mat_graph->nvtxs_global);CHKERRQ(ierr); 6504441e0de0SStefano Zampini ierr = PCBDDCGraphSetUp(graph,pcbddc->mat_graph->custom_minimal_size,NULL,pcbddc->DirichletBoundariesLocal,0,NULL,verticescomm);CHKERRQ(ierr); 65053301b35fSStefano Zampini ierr = ISDestroy(&verticescomm);CHKERRQ(ierr); 6506b96c3477SStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(graph);CHKERRQ(ierr); 6507b96c3477SStefano Zampini } else { 6508b96c3477SStefano Zampini graph = pcbddc->mat_graph; 6509b96c3477SStefano Zampini } 6510e4d548c7SStefano Zampini /* print some info */ 6511e4d548c7SStefano Zampini if (pcbddc->dbg_flag) { 6512e4d548c7SStefano Zampini IS vertices; 6513e4d548c7SStefano Zampini PetscInt nv,nedges,nfaces; 6514e4d548c7SStefano Zampini ierr = PCBDDCGraphASCIIView(graph,pcbddc->dbg_flag,pcbddc->dbg_viewer);CHKERRQ(ierr); 6515e4d548c7SStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(graph,&nfaces,NULL,&nedges,NULL,&vertices);CHKERRQ(ierr); 6516e4d548c7SStefano Zampini ierr = ISGetSize(vertices,&nv);CHKERRQ(ierr); 6517e4d548c7SStefano Zampini ierr = ISDestroy(&vertices);CHKERRQ(ierr); 6518e4d548c7SStefano Zampini ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 6519e4d548c7SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"--------------------------------------------------------------\n");CHKERRQ(ierr); 6520e4d548c7SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate vertices (%d)\n",PetscGlobalRank,nv,pcbddc->use_vertices);CHKERRQ(ierr); 6521e4d548c7SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate edges (%d)\n",PetscGlobalRank,nedges,pcbddc->use_edges);CHKERRQ(ierr); 6522e4d548c7SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate faces (%d)\n",PetscGlobalRank,nfaces,pcbddc->use_faces);CHKERRQ(ierr); 6523e4d548c7SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 6524e4d548c7SStefano Zampini ierr = PetscViewerASCIIPopSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 6525e4d548c7SStefano Zampini } 6526b96c3477SStefano Zampini 6527b96c3477SStefano Zampini /* sub_schurs init */ 6528a64f4aa4SStefano Zampini ierr = PCBDDCSubSchursInit(sub_schurs,pcis->is_I_local,pcis->is_B_local,graph,pcis->BtoNmap);CHKERRQ(ierr); 6529a64f4aa4SStefano Zampini 6530b96c3477SStefano Zampini /* free graph struct */ 653108122e43SStefano Zampini if (pcbddc->sub_schurs_rebuild) { 6532b96c3477SStefano Zampini ierr = PCBDDCGraphDestroy(&graph);CHKERRQ(ierr); 6533b96c3477SStefano Zampini } 6534b96c3477SStefano Zampini PetscFunctionReturn(0); 6535b96c3477SStefano Zampini } 6536fa34dd3eSStefano Zampini 6537fa34dd3eSStefano Zampini #undef __FUNCT__ 6538fa34dd3eSStefano Zampini #define __FUNCT__ "PCBDDCCheckOperator" 6539fa34dd3eSStefano Zampini PetscErrorCode PCBDDCCheckOperator(PC pc) 6540fa34dd3eSStefano Zampini { 6541fa34dd3eSStefano Zampini PC_IS *pcis=(PC_IS*)pc->data; 6542fa34dd3eSStefano Zampini PC_BDDC *pcbddc=(PC_BDDC*)pc->data; 6543fa34dd3eSStefano Zampini PetscErrorCode ierr; 6544fa34dd3eSStefano Zampini 6545fa34dd3eSStefano Zampini PetscFunctionBegin; 6546fa34dd3eSStefano Zampini if (pcbddc->n_vertices == pcbddc->local_primal_size) { 6547fa34dd3eSStefano Zampini IS zerodiag = NULL; 65484f1b2e48SStefano Zampini Mat S_j,B0_B=NULL; 6549fa34dd3eSStefano Zampini Vec dummy_vec=NULL,vec_check_B,vec_scale_P; 65504f1b2e48SStefano Zampini PetscScalar *p0_check,*array,*array2; 655175c01103SStefano Zampini PetscReal norm; 6552fa34dd3eSStefano Zampini PetscInt i; 6553fa34dd3eSStefano Zampini 6554fa34dd3eSStefano Zampini /* B0 and B0_B */ 6555fa34dd3eSStefano Zampini if (zerodiag) { 6556fa34dd3eSStefano Zampini IS dummy; 6557fa34dd3eSStefano Zampini 65584f1b2e48SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcbddc->benign_n,0,1,&dummy);CHKERRQ(ierr); 65594f1b2e48SStefano Zampini ierr = MatGetSubMatrix(pcbddc->benign_B0,dummy,pcis->is_B_local,MAT_INITIAL_MATRIX,&B0_B);CHKERRQ(ierr); 6560fa34dd3eSStefano Zampini ierr = MatCreateVecs(B0_B,NULL,&dummy_vec);CHKERRQ(ierr); 6561fa34dd3eSStefano Zampini ierr = ISDestroy(&dummy);CHKERRQ(ierr); 6562fa34dd3eSStefano Zampini } 6563fa34dd3eSStefano Zampini /* I need a primal vector to scale primal nodes since BDDC sums contibutions */ 6564fa34dd3eSStefano Zampini ierr = VecDuplicate(pcbddc->vec1_P,&vec_scale_P);CHKERRQ(ierr); 6565fa34dd3eSStefano Zampini ierr = VecSet(pcbddc->vec1_P,1.0);CHKERRQ(ierr); 6566fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6567fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6568fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,vec_scale_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6569fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,vec_scale_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6570fa34dd3eSStefano Zampini ierr = VecReciprocal(vec_scale_P);CHKERRQ(ierr); 6571fa34dd3eSStefano Zampini /* S_j */ 6572fa34dd3eSStefano Zampini ierr = MatCreateSchurComplement(pcis->A_II,pcis->A_II,pcis->A_IB,pcis->A_BI,pcis->A_BB,&S_j);CHKERRQ(ierr); 6573fa34dd3eSStefano Zampini ierr = MatSchurComplementSetKSP(S_j,pcbddc->ksp_D);CHKERRQ(ierr); 6574fa34dd3eSStefano Zampini 6575fa34dd3eSStefano Zampini /* mimic vector in \widetilde{W}_\Gamma */ 6576fa34dd3eSStefano Zampini ierr = VecSetRandom(pcis->vec1_N,NULL);CHKERRQ(ierr); 6577fa34dd3eSStefano Zampini /* continuous in primal space */ 6578fa34dd3eSStefano Zampini ierr = VecSetRandom(pcbddc->coarse_vec,NULL);CHKERRQ(ierr); 6579fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,pcbddc->vec1_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6580fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,pcbddc->vec1_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6581fa34dd3eSStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 65824f1b2e48SStefano Zampini ierr = PetscCalloc1(pcbddc->benign_n,&p0_check);CHKERRQ(ierr); 65834f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) p0_check[i] = array[pcbddc->local_primal_size-pcbddc->benign_n+i]; 6584fa34dd3eSStefano Zampini ierr = VecSetValues(pcis->vec1_N,pcbddc->local_primal_size,pcbddc->local_primal_ref_node,array,INSERT_VALUES);CHKERRQ(ierr); 6585fa34dd3eSStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 6586fa34dd3eSStefano Zampini ierr = VecAssemblyBegin(pcis->vec1_N);CHKERRQ(ierr); 6587fa34dd3eSStefano Zampini ierr = VecAssemblyEnd(pcis->vec1_N);CHKERRQ(ierr); 6588fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcis->N_to_B,pcis->vec1_N,pcis->vec2_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6589fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcis->N_to_B,pcis->vec1_N,pcis->vec2_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6590fa34dd3eSStefano Zampini ierr = VecDuplicate(pcis->vec2_B,&vec_check_B);CHKERRQ(ierr); 6591fa34dd3eSStefano Zampini ierr = VecCopy(pcis->vec2_B,vec_check_B);CHKERRQ(ierr); 6592fa34dd3eSStefano Zampini 6593fa34dd3eSStefano Zampini /* assemble rhs for coarse problem */ 6594fa34dd3eSStefano Zampini /* widetilde{S}_\Gamma w_\Gamma + \widetilde{B0}^T_B p0 */ 6595fa34dd3eSStefano Zampini /* local with Schur */ 6596fa34dd3eSStefano Zampini ierr = MatMult(S_j,pcis->vec2_B,pcis->vec1_B);CHKERRQ(ierr); 6597fa34dd3eSStefano Zampini if (zerodiag) { 6598fa34dd3eSStefano Zampini ierr = VecGetArray(dummy_vec,&array);CHKERRQ(ierr); 65994f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) array[i] = p0_check[i]; 6600fa34dd3eSStefano Zampini ierr = VecRestoreArray(dummy_vec,&array);CHKERRQ(ierr); 6601fa34dd3eSStefano Zampini ierr = MatMultTransposeAdd(B0_B,dummy_vec,pcis->vec1_B,pcis->vec1_B);CHKERRQ(ierr); 6602fa34dd3eSStefano Zampini } 6603fa34dd3eSStefano Zampini /* sum on primal nodes the local contributions */ 6604fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcis->N_to_B,pcis->vec1_B,pcis->vec1_N,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6605fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcis->N_to_B,pcis->vec1_B,pcis->vec1_N,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6606fa34dd3eSStefano Zampini ierr = VecGetArray(pcis->vec1_N,&array);CHKERRQ(ierr); 6607fa34dd3eSStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array2);CHKERRQ(ierr); 6608fa34dd3eSStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) array2[i] = array[pcbddc->local_primal_ref_node[i]]; 6609fa34dd3eSStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array2);CHKERRQ(ierr); 6610fa34dd3eSStefano Zampini ierr = VecRestoreArray(pcis->vec1_N,&array);CHKERRQ(ierr); 6611fa34dd3eSStefano Zampini ierr = VecSet(pcbddc->coarse_vec,0.);CHKERRQ(ierr); 6612fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6613fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6614fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,pcbddc->vec1_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6615fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,pcbddc->vec1_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6616fa34dd3eSStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 6617fa34dd3eSStefano Zampini /* scale primal nodes (BDDC sums contibutions) */ 6618fa34dd3eSStefano Zampini ierr = VecPointwiseMult(pcbddc->vec1_P,vec_scale_P,pcbddc->vec1_P);CHKERRQ(ierr); 6619fa34dd3eSStefano Zampini ierr = VecSetValues(pcis->vec1_N,pcbddc->local_primal_size,pcbddc->local_primal_ref_node,array,INSERT_VALUES);CHKERRQ(ierr); 6620fa34dd3eSStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 6621fa34dd3eSStefano Zampini ierr = VecAssemblyBegin(pcis->vec1_N);CHKERRQ(ierr); 6622fa34dd3eSStefano Zampini ierr = VecAssemblyEnd(pcis->vec1_N);CHKERRQ(ierr); 6623fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcis->N_to_B,pcis->vec1_N,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6624fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcis->N_to_B,pcis->vec1_N,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6625fa34dd3eSStefano Zampini /* global: \widetilde{B0}_B w_\Gamma */ 6626fa34dd3eSStefano Zampini if (zerodiag) { 6627fa34dd3eSStefano Zampini ierr = MatMult(B0_B,pcis->vec2_B,dummy_vec);CHKERRQ(ierr); 6628fa34dd3eSStefano Zampini ierr = VecGetArray(dummy_vec,&array);CHKERRQ(ierr); 66294f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) pcbddc->benign_p0[i] = array[i]; 6630fa34dd3eSStefano Zampini ierr = VecRestoreArray(dummy_vec,&array);CHKERRQ(ierr); 6631fa34dd3eSStefano Zampini } 6632fa34dd3eSStefano Zampini /* BDDC */ 6633fa34dd3eSStefano Zampini ierr = VecSet(pcis->vec1_D,0.);CHKERRQ(ierr); 6634fa34dd3eSStefano Zampini ierr = PCBDDCApplyInterfacePreconditioner(pc,PETSC_FALSE);CHKERRQ(ierr); 6635fa34dd3eSStefano Zampini 6636fa34dd3eSStefano Zampini ierr = VecCopy(pcis->vec1_B,pcis->vec2_B);CHKERRQ(ierr); 6637fa34dd3eSStefano Zampini ierr = VecAXPY(pcis->vec1_B,-1.0,vec_check_B);CHKERRQ(ierr); 6638fa34dd3eSStefano Zampini ierr = VecNorm(pcis->vec1_B,NORM_INFINITY,&norm);CHKERRQ(ierr); 6639fa34dd3eSStefano Zampini PetscPrintf(PETSC_COMM_SELF,"[%d] BDDC local error is %1.4e\n",PetscGlobalRank,norm); 66404f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 66414f1b2e48SStefano Zampini PetscPrintf(PETSC_COMM_SELF,"[%d] BDDC p0[%d] error is %1.4e\n",PetscGlobalRank,i,PetscAbsScalar(pcbddc->benign_p0[i]-p0_check[i])); 6642fa34dd3eSStefano Zampini } 66434f1b2e48SStefano Zampini ierr = PetscFree(p0_check);CHKERRQ(ierr); 6644fa34dd3eSStefano Zampini ierr = VecDestroy(&vec_scale_P);CHKERRQ(ierr); 6645fa34dd3eSStefano Zampini ierr = VecDestroy(&vec_check_B);CHKERRQ(ierr); 6646fa34dd3eSStefano Zampini ierr = VecDestroy(&dummy_vec);CHKERRQ(ierr); 6647fa34dd3eSStefano Zampini ierr = MatDestroy(&S_j);CHKERRQ(ierr); 6648fa34dd3eSStefano Zampini ierr = MatDestroy(&B0_B);CHKERRQ(ierr); 6649fa34dd3eSStefano Zampini } 6650fa34dd3eSStefano Zampini PetscFunctionReturn(0); 6651fa34dd3eSStefano Zampini } 6652