1ab5c6b0cSJed Brown #include <../src/ksp/pc/impls/bddc/bddc.h> 2ab5c6b0cSJed Brown #include <../src/ksp/pc/impls/bddc/bddcprivate.h> 3674ae819SStefano Zampini #include <petscblaslapack.h> 4674ae819SStefano Zampini 5906d46d4SStefano Zampini static PetscErrorCode PCBDDCMatMultTranspose_Private(Mat A, Vec x, Vec y); 6906d46d4SStefano Zampini static PetscErrorCode PCBDDCMatMult_Private(Mat A, Vec x, Vec y); 7906d46d4SStefano Zampini 8674ae819SStefano Zampini #undef __FUNCT__ 95408967cSStefano Zampini #define __FUNCT__ "PCBDDCBenignCheck" 105408967cSStefano Zampini PetscErrorCode PCBDDCBenignCheck(PC pc, IS zerodiag) 115408967cSStefano Zampini { 125408967cSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 135408967cSStefano Zampini PC_IS* pcis = (PC_IS*)(pc->data); 14dee84bffSStefano Zampini IS dirIS = NULL; 155408967cSStefano Zampini PetscErrorCode ierr; 165408967cSStefano Zampini 175408967cSStefano Zampini PetscFunctionBegin; 18dee84bffSStefano Zampini ierr = PCBDDCGraphGetDirichletDofs(pcbddc->mat_graph,&dirIS);CHKERRQ(ierr); 195408967cSStefano Zampini if (zerodiag) { 205408967cSStefano Zampini Mat A; 215408967cSStefano Zampini Vec vec3_N; 225408967cSStefano Zampini PetscScalar *vals; 235408967cSStefano Zampini const PetscInt *idxs; 245408967cSStefano Zampini PetscInt i,nz; 255408967cSStefano Zampini 265408967cSStefano Zampini /* p0 */ 275408967cSStefano Zampini ierr = VecSet(pcis->vec1_N,0.);CHKERRQ(ierr); 285408967cSStefano Zampini ierr = PetscMalloc1(pcis->n,&vals);CHKERRQ(ierr); 295408967cSStefano Zampini ierr = ISGetLocalSize(zerodiag,&nz);CHKERRQ(ierr); 305408967cSStefano Zampini ierr = ISGetIndices(zerodiag,&idxs);CHKERRQ(ierr); 315408967cSStefano Zampini for (i=0;i<nz;i++) vals[i] = 1.; /* TODO add quadrature */ 325408967cSStefano Zampini ierr = VecSetValues(pcis->vec1_N,nz,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 335408967cSStefano Zampini ierr = VecAssemblyBegin(pcis->vec1_N);CHKERRQ(ierr); 345408967cSStefano Zampini ierr = VecAssemblyEnd(pcis->vec1_N);CHKERRQ(ierr); 355408967cSStefano Zampini /* v_I */ 365408967cSStefano Zampini ierr = VecSetRandom(pcis->vec2_N,NULL);CHKERRQ(ierr); 375408967cSStefano Zampini for (i=0;i<nz;i++) vals[i] = 0.; 385408967cSStefano Zampini ierr = VecSetValues(pcis->vec2_N,nz,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 395408967cSStefano Zampini ierr = ISRestoreIndices(zerodiag,&idxs);CHKERRQ(ierr); 405408967cSStefano Zampini ierr = ISGetIndices(pcis->is_B_local,&idxs);CHKERRQ(ierr); 415408967cSStefano Zampini for (i=0;i<pcis->n_B;i++) vals[i] = 0.; 425408967cSStefano Zampini ierr = VecSetValues(pcis->vec2_N,pcis->n_B,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 435408967cSStefano Zampini ierr = ISRestoreIndices(pcis->is_B_local,&idxs);CHKERRQ(ierr); 445408967cSStefano Zampini if (dirIS) { 455408967cSStefano Zampini PetscInt n; 465408967cSStefano Zampini 475408967cSStefano Zampini ierr = ISGetLocalSize(dirIS,&n);CHKERRQ(ierr); 485408967cSStefano Zampini ierr = ISGetIndices(dirIS,&idxs);CHKERRQ(ierr); 495408967cSStefano Zampini for (i=0;i<n;i++) vals[i] = 0.; 505408967cSStefano Zampini ierr = VecSetValues(pcis->vec2_N,n,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 515408967cSStefano Zampini ierr = ISRestoreIndices(dirIS,&idxs);CHKERRQ(ierr); 525408967cSStefano Zampini } 535408967cSStefano Zampini ierr = VecAssemblyBegin(pcis->vec2_N);CHKERRQ(ierr); 545408967cSStefano Zampini ierr = VecAssemblyEnd(pcis->vec2_N);CHKERRQ(ierr); 555408967cSStefano Zampini ierr = VecDuplicate(pcis->vec1_N,&vec3_N);CHKERRQ(ierr); 565408967cSStefano Zampini ierr = VecSet(vec3_N,0.);CHKERRQ(ierr); 575408967cSStefano Zampini ierr = MatISGetLocalMat(pc->mat,&A);CHKERRQ(ierr); 585408967cSStefano Zampini ierr = MatMult(A,pcis->vec1_N,vec3_N);CHKERRQ(ierr); 595408967cSStefano Zampini ierr = VecDot(vec3_N,pcis->vec2_N,&vals[0]);CHKERRQ(ierr); 605408967cSStefano Zampini if (PetscAbsScalar(vals[0]) > PETSC_SMALL) { /* TODO: should I add the A-norm in test? */ 615408967cSStefano Zampini SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_SUP,"Benign trick can not be applied! b(v_I,p_0) = %f (should be numerically 0.)",PetscAbsScalar(vals[0])); 625408967cSStefano Zampini } 635408967cSStefano Zampini ierr = PetscFree(vals);CHKERRQ(ierr); 645408967cSStefano Zampini ierr = VecDestroy(&vec3_N);CHKERRQ(ierr); 655408967cSStefano Zampini } 66dee84bffSStefano Zampini ierr = ISDestroy(&dirIS);CHKERRQ(ierr); 675408967cSStefano Zampini 685408967cSStefano Zampini /* check PCBDDCBenignGetOrSetP0 */ 695408967cSStefano Zampini ierr = VecSetRandom(pcis->vec1_global,NULL);CHKERRQ(ierr); 705408967cSStefano Zampini pcbddc->benign_p0 = -PetscGlobalRank; 715408967cSStefano Zampini ierr = PCBDDCBenignGetOrSetP0(pc,pcis->vec1_global,PETSC_FALSE);CHKERRQ(ierr); 725408967cSStefano Zampini pcbddc->benign_p0 = 1; 735408967cSStefano Zampini ierr = PCBDDCBenignGetOrSetP0(pc,pcis->vec1_global,PETSC_TRUE);CHKERRQ(ierr); 745408967cSStefano Zampini if (pcbddc->benign_p0_gidx >=0 && pcbddc->benign_p0 != -PetscGlobalRank) { 755408967cSStefano Zampini SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Error testing PCBDDCBenignGetOrSetP0! Found %1.4e instead of %1.4e\n",pcbddc->benign_p0,-PetscGlobalRank);CHKERRQ(ierr); 765408967cSStefano Zampini } 775408967cSStefano Zampini PetscFunctionReturn(0); 785408967cSStefano Zampini } 795408967cSStefano Zampini 805408967cSStefano Zampini #undef __FUNCT__ 81339f8db1SStefano Zampini #define __FUNCT__ "PCBDDCBenignDetectSaddlePoint" 82339f8db1SStefano Zampini PetscErrorCode PCBDDCBenignDetectSaddlePoint(PC pc, IS *zerodiaglocal) 83339f8db1SStefano Zampini { 84339f8db1SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 85*40fa8d13SStefano Zampini IS pressures,zerodiag; 86339f8db1SStefano Zampini PetscInt nz; 87339f8db1SStefano Zampini PetscBool sorted; 88339f8db1SStefano Zampini PetscErrorCode ierr; 89339f8db1SStefano Zampini 90339f8db1SStefano Zampini PetscFunctionBegin; 91339f8db1SStefano Zampini ierr = MatDestroy(&pcbddc->benign_original_mat);CHKERRQ(ierr); 92339f8db1SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->local_mat);CHKERRQ(ierr); 93339f8db1SStefano Zampini pcbddc->benign_original_mat = pcbddc->local_mat; 94*40fa8d13SStefano Zampini /* if a local info on dofs is present, assumes the last field is represented by "pressures", otherwise, uses only zerodiagonal dofs (ok if the pressure block is all zero) */ 95*40fa8d13SStefano Zampini if (pcbddc->n_ISForDofsLocal) { 96*40fa8d13SStefano Zampini PetscInt p = pcbddc->n_ISForDofsLocal-1; 97*40fa8d13SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->ISForDofsLocal[p]);CHKERRQ(ierr); 98*40fa8d13SStefano Zampini pressures = pcbddc->ISForDofsLocal[p]; 99*40fa8d13SStefano Zampini ierr = ISSorted(pressures,&sorted);CHKERRQ(ierr); 100*40fa8d13SStefano Zampini if (!sorted) { 101*40fa8d13SStefano Zampini ierr = ISSort(pressures);CHKERRQ(ierr); 102*40fa8d13SStefano Zampini } 103*40fa8d13SStefano Zampini } else { 104*40fa8d13SStefano Zampini pressures = NULL; 105*40fa8d13SStefano Zampini } 106339f8db1SStefano Zampini ierr = MatFindZeroDiagonals(pcbddc->benign_original_mat,&zerodiag);CHKERRQ(ierr); 107339f8db1SStefano Zampini ierr = ISSorted(zerodiag,&sorted);CHKERRQ(ierr); 108339f8db1SStefano Zampini if (!sorted) { 109339f8db1SStefano Zampini ierr = ISSort(zerodiag);CHKERRQ(ierr); 110339f8db1SStefano Zampini } 111*40fa8d13SStefano Zampini /* 112*40fa8d13SStefano Zampini Check if all the pressure dofs have a zero diagonal 113*40fa8d13SStefano Zampini If not, a change of basis on pressures is not needed 114*40fa8d13SStefano Zampini since the local Schur complements are SPD 115*40fa8d13SStefano Zampini */ 116339f8db1SStefano Zampini ierr = ISGetLocalSize(zerodiag,&nz);CHKERRQ(ierr); 117*40fa8d13SStefano Zampini if (pressures) { 118*40fa8d13SStefano Zampini PetscInt np; 119*40fa8d13SStefano Zampini PetscBool sameis; 120*40fa8d13SStefano Zampini 121*40fa8d13SStefano Zampini sameis = PETSC_FALSE; 122*40fa8d13SStefano Zampini ierr = ISGetLocalSize(pressures,&np);CHKERRQ(ierr); 123*40fa8d13SStefano Zampini if (nz == np) { 124*40fa8d13SStefano Zampini const PetscInt *idxs,*pidxs; 125*40fa8d13SStefano Zampini 126*40fa8d13SStefano Zampini ierr = ISGetIndices(pressures,&pidxs);CHKERRQ(ierr); 127*40fa8d13SStefano Zampini ierr = ISGetIndices(zerodiag,&idxs);CHKERRQ(ierr); 128*40fa8d13SStefano Zampini ierr = PetscMemcmp(pidxs,idxs,np*sizeof(PetscInt),&sameis);CHKERRQ(ierr); 129*40fa8d13SStefano Zampini ierr = ISRestoreIndices(zerodiag,&idxs);CHKERRQ(ierr); 130*40fa8d13SStefano Zampini ierr = ISRestoreIndices(pressures,&pidxs);CHKERRQ(ierr); 131*40fa8d13SStefano Zampini } 132*40fa8d13SStefano Zampini if (!sameis) { /* destroy index sets and set nz to 0 to avoid next code branch */ 133*40fa8d13SStefano Zampini ierr = ISDestroy(&zerodiag);CHKERRQ(ierr); 134*40fa8d13SStefano Zampini nz = 0; 135*40fa8d13SStefano Zampini } 136*40fa8d13SStefano Zampini ierr = ISDestroy(&pressures);CHKERRQ(ierr); 137*40fa8d13SStefano Zampini } 138339f8db1SStefano Zampini if (nz) { 139339f8db1SStefano Zampini IS zerodiagc; 140339f8db1SStefano Zampini PetscScalar *array; 141339f8db1SStefano Zampini const PetscInt *idxs,*idxsc; 142339f8db1SStefano Zampini PetscInt i,n,*nnz; 143339f8db1SStefano Zampini 144339f8db1SStefano Zampini /* TODO: add check for shared dofs and raise error */ 145339f8db1SStefano Zampini ierr = MatGetLocalSize(pcbddc->benign_original_mat,&n,NULL);CHKERRQ(ierr); 146339f8db1SStefano Zampini ierr = ISComplement(zerodiag,0,n,&zerodiagc);CHKERRQ(ierr); 147339f8db1SStefano Zampini ierr = ISGetIndices(zerodiag,&idxs);CHKERRQ(ierr); 148339f8db1SStefano Zampini ierr = ISGetIndices(zerodiagc,&idxsc);CHKERRQ(ierr); 149339f8db1SStefano Zampini /* local change of basis for pressures */ 150339f8db1SStefano Zampini ierr = MatDestroy(&pcbddc->benign_change);CHKERRQ(ierr); 151339f8db1SStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)pcbddc->benign_original_mat),&pcbddc->benign_change);CHKERRQ(ierr); 152339f8db1SStefano Zampini ierr = MatSetType(pcbddc->benign_change,MATAIJ);CHKERRQ(ierr); 153339f8db1SStefano Zampini ierr = MatSetSizes(pcbddc->benign_change,n,n,PETSC_DECIDE,PETSC_DECIDE);CHKERRQ(ierr); 154339f8db1SStefano Zampini ierr = PetscMalloc1(n,&nnz);CHKERRQ(ierr); 155339f8db1SStefano Zampini for (i=0;i<n-nz;i++) nnz[idxsc[i]] = 1; /* identity on velocities */ 156339f8db1SStefano Zampini for (i=0;i<nz-1;i++) nnz[idxs[i]] = 2; /* change on pressures */ 157339f8db1SStefano Zampini nnz[idxs[nz-1]] = nz; /* last local pressure dof: _0 set */ 158339f8db1SStefano Zampini ierr = MatSeqAIJSetPreallocation(pcbddc->benign_change,0,nnz);CHKERRQ(ierr); 159339f8db1SStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 160339f8db1SStefano Zampini /* set identity on velocities */ 161339f8db1SStefano Zampini for (i=0;i<n-nz;i++) { 162339f8db1SStefano Zampini ierr = MatSetValue(pcbddc->benign_change,idxsc[i],idxsc[i],1.,INSERT_VALUES);CHKERRQ(ierr); 163339f8db1SStefano Zampini } 164339f8db1SStefano Zampini /* set change on pressures */ 165339f8db1SStefano Zampini for (i=0;i<nz-1;i++) { 166339f8db1SStefano Zampini PetscScalar vals[2]; 167339f8db1SStefano Zampini PetscInt cols[2]; 168339f8db1SStefano Zampini 169339f8db1SStefano Zampini /* TODO: add quadrature */ 170339f8db1SStefano Zampini cols[0] = idxs[i]; 171339f8db1SStefano Zampini cols[1] = idxs[nz-1]; 172339f8db1SStefano Zampini vals[0] = 1.; 173339f8db1SStefano Zampini vals[1] = 1./nz; 174339f8db1SStefano Zampini ierr = MatSetValues(pcbddc->benign_change,1,idxs+i,2,cols,vals,INSERT_VALUES);CHKERRQ(ierr); 175339f8db1SStefano Zampini } 176339f8db1SStefano Zampini ierr = PetscMalloc1(nz,&array);CHKERRQ(ierr); 177339f8db1SStefano Zampini for (i=0;i<nz-1;i++) array[i] = -1.; 178339f8db1SStefano Zampini array[nz-1] = 1./nz; 179339f8db1SStefano Zampini ierr = MatSetValues(pcbddc->benign_change,1,idxs+nz-1,nz,idxs,array,INSERT_VALUES);CHKERRQ(ierr); 180339f8db1SStefano Zampini ierr = PetscFree(array);CHKERRQ(ierr); 181339f8db1SStefano Zampini ierr = MatAssemblyBegin(pcbddc->benign_change,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 182339f8db1SStefano Zampini ierr = MatAssemblyEnd(pcbddc->benign_change,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 183339f8db1SStefano Zampini /* TODO: need optimization? */ 184339f8db1SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 185339f8db1SStefano Zampini ierr = MatPtAP(pcbddc->benign_original_mat,pcbddc->benign_change,MAT_INITIAL_MATRIX,2.0,&pcbddc->local_mat);CHKERRQ(ierr); 186339f8db1SStefano Zampini /* store local and global idxs for p0 */ 187339f8db1SStefano Zampini pcbddc->benign_p0_lidx = idxs[nz-1]; 188339f8db1SStefano Zampini ierr = ISLocalToGlobalMappingApply(pc->pmat->rmap->mapping,1,&idxs[nz-1],&pcbddc->benign_p0_gidx);CHKERRQ(ierr); 189339f8db1SStefano Zampini ierr = ISRestoreIndices(zerodiag,&idxs);CHKERRQ(ierr); 190339f8db1SStefano Zampini ierr = ISRestoreIndices(zerodiagc,&idxsc);CHKERRQ(ierr); 191339f8db1SStefano Zampini ierr = ISDestroy(&zerodiagc);CHKERRQ(ierr); 192339f8db1SStefano Zampini } 193339f8db1SStefano Zampini *zerodiaglocal = zerodiag; 194339f8db1SStefano Zampini PetscFunctionReturn(0); 195339f8db1SStefano Zampini } 196339f8db1SStefano Zampini 197339f8db1SStefano Zampini #undef __FUNCT__ 198015636ebSStefano Zampini #define __FUNCT__ "PCBDDCBenignGetOrSetP0" 199015636ebSStefano Zampini PetscErrorCode PCBDDCBenignGetOrSetP0(PC pc, Vec v, PetscBool get) 200efc2fbd9SStefano Zampini { 201efc2fbd9SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 202efc2fbd9SStefano Zampini PetscErrorCode ierr; 203efc2fbd9SStefano Zampini 204efc2fbd9SStefano Zampini PetscFunctionBegin; 205efc2fbd9SStefano Zampini if (!pcbddc->benign_sf) { 206efc2fbd9SStefano Zampini ierr = PetscSFCreate(PetscObjectComm((PetscObject)pc),&pcbddc->benign_sf);CHKERRQ(ierr); 207c9ed8603SStefano Zampini if (pcbddc->benign_p0_gidx >= 0) { 208efc2fbd9SStefano Zampini ierr = PetscSFSetGraphLayout(pcbddc->benign_sf,pc->pmat->rmap,1,NULL,PETSC_OWN_POINTER,&pcbddc->benign_p0_gidx);CHKERRQ(ierr); 209c9ed8603SStefano Zampini } else { 210c9ed8603SStefano Zampini ierr = PetscSFSetGraphLayout(pcbddc->benign_sf,pc->pmat->rmap,0,NULL,PETSC_OWN_POINTER,&pcbddc->benign_p0_gidx);CHKERRQ(ierr); 211c9ed8603SStefano Zampini } 212efc2fbd9SStefano Zampini } 213015636ebSStefano Zampini if (get) { /* use SF to get values */ 214efc2fbd9SStefano Zampini PetscScalar *array; 215efc2fbd9SStefano Zampini 216efc2fbd9SStefano Zampini ierr = VecGetArrayRead(v,(const PetscScalar**)&array);CHKERRQ(ierr); 217efc2fbd9SStefano Zampini ierr = PetscSFBcastBegin(pcbddc->benign_sf,MPIU_SCALAR,array,&pcbddc->benign_p0);CHKERRQ(ierr); 218efc2fbd9SStefano Zampini ierr = PetscSFBcastEnd(pcbddc->benign_sf,MPIU_SCALAR,array,&pcbddc->benign_p0);CHKERRQ(ierr); 219efc2fbd9SStefano Zampini ierr = VecRestoreArrayRead(v,(const PetscScalar**)&array);CHKERRQ(ierr); 220efc2fbd9SStefano Zampini } else { /* use VecSetValue */ 221c9ed8603SStefano Zampini if (pcbddc->benign_p0_gidx >= 0) { 222efc2fbd9SStefano Zampini ierr = VecSetValue(v,pcbddc->benign_p0_gidx,pcbddc->benign_p0,INSERT_VALUES);CHKERRQ(ierr); 223c9ed8603SStefano Zampini } 224efc2fbd9SStefano Zampini ierr = VecAssemblyBegin(v);CHKERRQ(ierr); 225efc2fbd9SStefano Zampini ierr = VecAssemblyEnd(v);CHKERRQ(ierr); 226efc2fbd9SStefano Zampini } 227efc2fbd9SStefano Zampini PetscFunctionReturn(0); 228efc2fbd9SStefano Zampini } 229efc2fbd9SStefano Zampini 230efc2fbd9SStefano Zampini #undef __FUNCT__ 231c263805aSStefano Zampini #define __FUNCT__ "PCBDDCBenignPopOrPushB0" 232c263805aSStefano Zampini PetscErrorCode PCBDDCBenignPopOrPushB0(PC pc, PetscBool pop) 233c263805aSStefano Zampini { 234c263805aSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 235c263805aSStefano Zampini PetscErrorCode ierr; 236c263805aSStefano Zampini 237c263805aSStefano Zampini PetscFunctionBegin; 238c263805aSStefano Zampini /* TODO: add error checking 239c263805aSStefano Zampini - avoid nested pop (or push) calls. 240c263805aSStefano Zampini - cannot push before pop. 2411c604dc7SStefano Zampini - cannot call this if pcbddc->local_mat is NULL 242c263805aSStefano Zampini */ 243c9ed8603SStefano Zampini if (pcbddc->benign_p0_lidx < 0) { 244efc2fbd9SStefano Zampini PetscFunctionReturn(0); 245efc2fbd9SStefano Zampini } 246c263805aSStefano Zampini if (pop) { 247c263805aSStefano Zampini const PetscInt *cB0_cols; 248c263805aSStefano Zampini PetscInt cB0_ncol; 249c263805aSStefano Zampini const PetscScalar *cB0_vals; 250c263805aSStefano Zampini 251c263805aSStefano Zampini /* extract B_0 */ 252efc2fbd9SStefano Zampini ierr = MatGetRow(pcbddc->local_mat,pcbddc->benign_p0_lidx,&cB0_ncol,&cB0_cols,&cB0_vals);CHKERRQ(ierr); 253c263805aSStefano Zampini pcbddc->B0_ncol = cB0_ncol; 254c263805aSStefano Zampini ierr = PetscFree2(pcbddc->B0_cols,pcbddc->B0_vals);CHKERRQ(ierr); 255c263805aSStefano Zampini ierr = PetscMalloc2(cB0_ncol,&pcbddc->B0_cols,cB0_ncol,&pcbddc->B0_vals);CHKERRQ(ierr); 256c263805aSStefano Zampini ierr = PetscMemcpy(pcbddc->B0_cols,cB0_cols,cB0_ncol*sizeof(PetscInt));CHKERRQ(ierr); 257c263805aSStefano Zampini ierr = PetscMemcpy(pcbddc->B0_vals,cB0_vals,cB0_ncol*sizeof(PetscScalar));CHKERRQ(ierr); 258efc2fbd9SStefano Zampini ierr = MatRestoreRow(pcbddc->local_mat,pcbddc->benign_p0_lidx,&cB0_ncol,&cB0_cols,&cB0_vals);CHKERRQ(ierr); 259c263805aSStefano Zampini /* remove rows and cols from local problem */ 260c263805aSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_KEEP_NONZERO_PATTERN,PETSC_TRUE);CHKERRQ(ierr); 261efc2fbd9SStefano Zampini ierr = MatZeroRowsColumns(pcbddc->local_mat,1,&pcbddc->benign_p0_lidx,1.,NULL,NULL);CHKERRQ(ierr); 262c263805aSStefano Zampini } else { /* push */ 263efc2fbd9SStefano Zampini ierr = MatSetValues(pcbddc->local_mat,1,&pcbddc->benign_p0_lidx,pcbddc->B0_ncol,pcbddc->B0_cols,pcbddc->B0_vals,INSERT_VALUES);CHKERRQ(ierr); 264efc2fbd9SStefano Zampini ierr = MatSetValues(pcbddc->local_mat,pcbddc->B0_ncol,pcbddc->B0_cols,1,&pcbddc->benign_p0_lidx,pcbddc->B0_vals,INSERT_VALUES);CHKERRQ(ierr); 265efc2fbd9SStefano Zampini ierr = MatSetValue(pcbddc->local_mat,pcbddc->benign_p0_lidx,pcbddc->benign_p0_lidx,0.0,INSERT_VALUES);CHKERRQ(ierr); 266c263805aSStefano Zampini ierr = MatAssemblyBegin(pcbddc->local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 267c263805aSStefano Zampini ierr = MatAssemblyEnd(pcbddc->local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 268c263805aSStefano Zampini } 269c263805aSStefano Zampini PetscFunctionReturn(0); 270c263805aSStefano Zampini } 271c263805aSStefano Zampini 272c263805aSStefano Zampini #undef __FUNCT__ 273b1b3d7a2SStefano Zampini #define __FUNCT__ "PCBDDCAdaptiveSelection" 27408122e43SStefano Zampini PetscErrorCode PCBDDCAdaptiveSelection(PC pc) 275b1b3d7a2SStefano Zampini { 276b1b3d7a2SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 27708122e43SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 27808122e43SStefano Zampini PetscBLASInt B_dummyint,B_neigs,B_ierr,B_lwork; 27908122e43SStefano Zampini PetscBLASInt *B_iwork,*B_ifail; 28008122e43SStefano Zampini PetscScalar *work,lwork; 28108122e43SStefano Zampini PetscScalar *St,*S,*eigv; 28208122e43SStefano Zampini PetscScalar *Sarray,*Starray; 28308122e43SStefano Zampini PetscReal *eigs,thresh; 2841b968477SStefano Zampini PetscInt i,nmax,nmin,nv,cum,mss,cum2,cumarray,maxneigs; 285f6f667cfSStefano Zampini PetscBool allocated_S_St; 28608122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 28708122e43SStefano Zampini PetscReal *rwork; 28808122e43SStefano Zampini #endif 289b1b3d7a2SStefano Zampini PetscErrorCode ierr; 290b1b3d7a2SStefano Zampini 291b1b3d7a2SStefano Zampini PetscFunctionBegin; 29208122e43SStefano Zampini if (!sub_schurs->use_mumps) { 29308122e43SStefano Zampini SETERRQ(PetscObjectComm((PetscObject)pc),PETSC_ERR_SUP,"Adaptive selection of constraints requires MUMPS"); 29408122e43SStefano Zampini } 29508122e43SStefano Zampini 29606a4e24aSStefano Zampini if (sub_schurs->n_subs && (!sub_schurs->is_hermitian || !sub_schurs->is_posdef)) { 29706a4e24aSStefano Zampini SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_SUP,"Adaptive selection not yet implemented for general matrix pencils (herm %d, posdef %d)\n",sub_schurs->is_hermitian,sub_schurs->is_posdef); 29806a4e24aSStefano Zampini } 29906a4e24aSStefano Zampini 300fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 301fd14bc51SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 302fd14bc51SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 303fd14bc51SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Check adaptive selection of constraints\n");CHKERRQ(ierr); 304fd14bc51SStefano Zampini ierr = PetscViewerASCIISynchronizedAllow(pcbddc->dbg_viewer,PETSC_TRUE);CHKERRQ(ierr); 305fd14bc51SStefano Zampini } 306fd14bc51SStefano Zampini 307e496cd5dSStefano Zampini if (pcbddc->dbg_flag) { 308e496cd5dSStefano 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); 309e496cd5dSStefano Zampini } 310e496cd5dSStefano Zampini 31108122e43SStefano Zampini /* max size of subsets */ 31208122e43SStefano Zampini mss = 0; 31308122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 31408122e43SStefano Zampini PetscInt subset_size; 315862806e4SStefano Zampini 31608122e43SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 31708122e43SStefano Zampini mss = PetscMax(mss,subset_size); 31808122e43SStefano Zampini } 31908122e43SStefano Zampini 32008122e43SStefano Zampini /* min/max and threshold */ 32108122e43SStefano Zampini nmax = pcbddc->adaptive_nmax > 0 ? pcbddc->adaptive_nmax : mss; 322f6f667cfSStefano Zampini nmin = pcbddc->adaptive_nmin > 0 ? pcbddc->adaptive_nmin : 0; 32308122e43SStefano Zampini nmax = PetscMax(nmin,nmax); 324f6f667cfSStefano Zampini allocated_S_St = PETSC_FALSE; 325f6f667cfSStefano Zampini if (nmin) { 326f6f667cfSStefano Zampini allocated_S_St = PETSC_TRUE; 327f6f667cfSStefano Zampini } 32808122e43SStefano Zampini 32908122e43SStefano Zampini /* allocate lapack workspace */ 33008122e43SStefano Zampini cum = cum2 = 0; 33108122e43SStefano Zampini maxneigs = 0; 33208122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 33308122e43SStefano Zampini PetscInt n,subset_size; 334f6f667cfSStefano Zampini 33508122e43SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 33608122e43SStefano Zampini n = PetscMin(subset_size,nmax); 3379162d606SStefano Zampini cum += subset_size; 3389162d606SStefano Zampini cum2 += subset_size*n; 33908122e43SStefano Zampini maxneigs = PetscMax(maxneigs,n); 34008122e43SStefano Zampini } 34108122e43SStefano Zampini if (mss) { 3429ab7bb16SStefano Zampini if (sub_schurs->is_hermitian && sub_schurs->is_posdef) { 34308122e43SStefano Zampini PetscBLASInt B_itype = 1; 34408122e43SStefano Zampini PetscBLASInt B_N = mss; 3454c6709b3SStefano Zampini PetscReal zero = 0.0; 3464c6709b3SStefano Zampini PetscReal eps = 0.0; /* dlamch? */ 34708122e43SStefano Zampini 34808122e43SStefano Zampini B_lwork = -1; 34908122e43SStefano Zampini S = NULL; 35008122e43SStefano Zampini St = NULL; 351a58a30b4SStefano Zampini eigs = NULL; 352a58a30b4SStefano Zampini eigv = NULL; 353a58a30b4SStefano Zampini B_iwork = NULL; 354a58a30b4SStefano Zampini B_ifail = NULL; 355d1710679SStefano Zampini #if defined(PETSC_USE_COMPLEX) 356d1710679SStefano Zampini rwork = NULL; 357d1710679SStefano Zampini #endif 3588bec7fa6SStefano Zampini thresh = 1.0; 35908122e43SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 36008122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 36108122e43SStefano 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)); 36208122e43SStefano Zampini #else 36308122e43SStefano 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)); 36408122e43SStefano Zampini #endif 36508122e43SStefano Zampini if (B_ierr != 0) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to SYGVX Lapack routine %d",(int)B_ierr); 36608122e43SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 36708122e43SStefano Zampini } else { 36808122e43SStefano Zampini /* TODO */ 36908122e43SStefano Zampini } 37008122e43SStefano Zampini } else { 37108122e43SStefano Zampini lwork = 0; 37208122e43SStefano Zampini } 37308122e43SStefano Zampini 37408122e43SStefano Zampini nv = 0; 375d62866d3SStefano 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) */ 376d62866d3SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_vertices,&nv);CHKERRQ(ierr); 37708122e43SStefano Zampini } 3784c6709b3SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lwork),&B_lwork);CHKERRQ(ierr); 379f6f667cfSStefano Zampini if (allocated_S_St) { 380f6f667cfSStefano Zampini ierr = PetscMalloc2(mss*mss,&S,mss*mss,&St);CHKERRQ(ierr); 381f6f667cfSStefano Zampini } 382f6f667cfSStefano Zampini ierr = PetscMalloc5(mss*mss,&eigv,mss,&eigs,B_lwork,&work,5*mss,&B_iwork,mss,&B_ifail);CHKERRQ(ierr); 38308122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 38408122e43SStefano Zampini ierr = PetscMalloc1(7*mss,&rwork);CHKERRQ(ierr); 38508122e43SStefano Zampini #endif 3869162d606SStefano Zampini ierr = PetscMalloc5(nv+sub_schurs->n_subs,&pcbddc->adaptive_constraints_n, 3879162d606SStefano Zampini nv+sub_schurs->n_subs+1,&pcbddc->adaptive_constraints_idxs_ptr, 3889162d606SStefano Zampini nv+sub_schurs->n_subs+1,&pcbddc->adaptive_constraints_data_ptr, 38908122e43SStefano Zampini nv+cum,&pcbddc->adaptive_constraints_idxs, 3909162d606SStefano Zampini nv+cum2,&pcbddc->adaptive_constraints_data);CHKERRQ(ierr); 39108122e43SStefano Zampini ierr = PetscMemzero(pcbddc->adaptive_constraints_n,(nv+sub_schurs->n_subs)*sizeof(PetscInt));CHKERRQ(ierr); 39208122e43SStefano Zampini 39308122e43SStefano Zampini maxneigs = 0; 39408122e43SStefano Zampini cum = cum2 = cumarray = 0; 3959162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr[0] = 0; 3969162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr[0] = 0; 397d62866d3SStefano Zampini if (sub_schurs->is_vertices && pcbddc->use_vertices) { 39808122e43SStefano Zampini const PetscInt *idxs; 39908122e43SStefano Zampini 400d62866d3SStefano Zampini ierr = ISGetIndices(sub_schurs->is_vertices,&idxs);CHKERRQ(ierr); 40108122e43SStefano Zampini for (cum=0;cum<nv;cum++) { 40208122e43SStefano Zampini pcbddc->adaptive_constraints_n[cum] = 1; 40308122e43SStefano Zampini pcbddc->adaptive_constraints_idxs[cum] = idxs[cum]; 40408122e43SStefano Zampini pcbddc->adaptive_constraints_data[cum] = 1.0; 4059162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr[cum+1] = pcbddc->adaptive_constraints_idxs_ptr[cum]+1; 4069162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr[cum+1] = pcbddc->adaptive_constraints_data_ptr[cum]+1; 40708122e43SStefano Zampini } 40808122e43SStefano Zampini cum2 = cum; 409d62866d3SStefano Zampini ierr = ISRestoreIndices(sub_schurs->is_vertices,&idxs);CHKERRQ(ierr); 41008122e43SStefano Zampini } 41108122e43SStefano Zampini 41208122e43SStefano Zampini if (mss) { /* multilevel */ 41308122e43SStefano Zampini ierr = MatSeqAIJGetArray(sub_schurs->sum_S_Ej_inv_all,&Sarray);CHKERRQ(ierr); 41408122e43SStefano Zampini ierr = MatSeqAIJGetArray(sub_schurs->sum_S_Ej_tilda_all,&Starray);CHKERRQ(ierr); 41508122e43SStefano Zampini } 41608122e43SStefano Zampini 41708122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 41808122e43SStefano Zampini 41908122e43SStefano Zampini const PetscInt *idxs; 420f6f667cfSStefano Zampini PetscReal infty = PETSC_MAX_REAL; 421862806e4SStefano Zampini PetscInt j,subset_size,eigs_start = 0; 42208122e43SStefano Zampini PetscBLASInt B_N; 423aff50787SStefano Zampini PetscBool same_data = PETSC_FALSE; 42408122e43SStefano Zampini 425862806e4SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 426f6f667cfSStefano Zampini ierr = PetscBLASIntCast(subset_size,&B_N);CHKERRQ(ierr); 427f6f667cfSStefano Zampini if (allocated_S_St) { /* S and S_t should be copied since we could need them later */ 4289ab7bb16SStefano Zampini if (sub_schurs->is_hermitian) { 429aff50787SStefano Zampini PetscInt j,k; 430aff50787SStefano Zampini if (sub_schurs->n_subs == 1) { /* zeroing memory to use PetscMemcmp later */ 431aff50787SStefano Zampini ierr = PetscMemzero(S,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 432aff50787SStefano Zampini ierr = PetscMemzero(St,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 43308122e43SStefano Zampini } 43408122e43SStefano Zampini for (j=0;j<subset_size;j++) { 435aff50787SStefano Zampini for (k=j;k<subset_size;k++) { 436aff50787SStefano Zampini S [j*subset_size+k] = Sarray [cumarray+j*subset_size+k]; 437aff50787SStefano Zampini St[j*subset_size+k] = Starray[cumarray+j*subset_size+k]; 438aff50787SStefano Zampini } 43908122e43SStefano Zampini } 44008122e43SStefano Zampini } else { 44108122e43SStefano Zampini ierr = PetscMemcpy(S,Sarray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 44208122e43SStefano Zampini ierr = PetscMemcpy(St,Starray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 44308122e43SStefano Zampini } 4448bec7fa6SStefano Zampini } else { 445f6f667cfSStefano Zampini S = Sarray + cumarray; 446f6f667cfSStefano Zampini St = Starray + cumarray; 4478bec7fa6SStefano Zampini } 44808122e43SStefano Zampini 449f6f667cfSStefano Zampini ierr = ISGetIndices(sub_schurs->is_subs[i],&idxs);CHKERRQ(ierr); 450aff50787SStefano Zampini /* see if we can save some work */ 451aff50787SStefano Zampini if (sub_schurs->n_subs == 1) { 452aff50787SStefano Zampini ierr = PetscMemcmp(S,St,subset_size*subset_size*sizeof(PetscScalar),&same_data);CHKERRQ(ierr); 453aff50787SStefano Zampini } 454aff50787SStefano Zampini 455aff50787SStefano Zampini if (same_data) { /* there's no need of constraints here, deluxe scaling is enough */ 456aff50787SStefano Zampini B_neigs = 0; 457aff50787SStefano Zampini } else { 458aff50787SStefano Zampini /* Threshold: this is an heuristic for edges */ 459f6f667cfSStefano Zampini thresh = pcbddc->mat_graph->count[idxs[0]]*pcbddc->adaptive_threshold; 460f6f667cfSStefano Zampini 4619ab7bb16SStefano Zampini if (sub_schurs->is_hermitian && sub_schurs->is_posdef) { 46208122e43SStefano Zampini PetscBLASInt B_itype = 1; 463f6f667cfSStefano Zampini PetscBLASInt B_IL, B_IU; 4644c6709b3SStefano Zampini PetscReal eps = -1.0; /* dlamch? */ 4659552c7c7SStefano Zampini PetscInt nmin_s; 46608122e43SStefano Zampini 467fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 4688bec7fa6SStefano 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]]); 469fd14bc51SStefano Zampini } 470d16cbb6bSStefano Zampini 47108122e43SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 472d16cbb6bSStefano Zampini if (thresh > 1.+PETSC_SMALL) { 473d16cbb6bSStefano Zampini 474d16cbb6bSStefano Zampini /* ask for eigenvalues larger than thresh */ 47508122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 476f6f667cfSStefano Zampini PetscStackCallBLAS("LAPACKsygvx",LAPACKsygvx_(&B_itype,"V","V","L",&B_N,St,&B_N,S,&B_N,&thresh,&infty,&B_IL,&B_IU,&eps,&B_neigs,eigs,eigv,&B_N,work,&B_lwork,rwork,B_iwork,B_ifail,&B_ierr)); 47708122e43SStefano Zampini #else 478f6f667cfSStefano Zampini PetscStackCallBLAS("LAPACKsygvx",LAPACKsygvx_(&B_itype,"V","V","L",&B_N,St,&B_N,S,&B_N,&thresh,&infty,&B_IL,&B_IU,&eps,&B_neigs,eigs,eigv,&B_N,work,&B_lwork,B_iwork,B_ifail,&B_ierr)); 47908122e43SStefano Zampini #endif 480d16cbb6bSStefano Zampini } else { 481d16cbb6bSStefano Zampini B_IU = PetscMax(1,PetscMin(B_N,nmax)); 482d16cbb6bSStefano Zampini B_IL = 1; 483d16cbb6bSStefano Zampini #if defined(PETSC_USE_COMPLEX) 484d16cbb6bSStefano Zampini PetscStackCallBLAS("LAPACKsygvx",LAPACKsygvx_(&B_itype,"V","I","L",&B_N,St,&B_N,S,&B_N,&thresh,&infty,&B_IL,&B_IU,&eps,&B_neigs,eigs,eigv,&B_N,work,&B_lwork,rwork,B_iwork,B_ifail,&B_ierr)); 485d16cbb6bSStefano Zampini #else 486d16cbb6bSStefano Zampini PetscStackCallBLAS("LAPACKsygvx",LAPACKsygvx_(&B_itype,"V","I","L",&B_N,St,&B_N,S,&B_N,&thresh,&infty,&B_IL,&B_IU,&eps,&B_neigs,eigs,eigv,&B_N,work,&B_lwork,B_iwork,B_ifail,&B_ierr)); 487d16cbb6bSStefano Zampini #endif 488d16cbb6bSStefano Zampini } 48908122e43SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 49008122e43SStefano Zampini if (B_ierr) { 49108122e43SStefano Zampini if (B_ierr < 0 ) { 49208122e43SStefano Zampini SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYGVX Lapack routine: illegal value for argument %d",-(int)B_ierr); 49308122e43SStefano Zampini } else if (B_ierr <= B_N) { 49408122e43SStefano Zampini SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYGVX Lapack routine: %d eigenvalues failed to converge",(int)B_ierr); 49508122e43SStefano Zampini } else { 4969552c7c7SStefano Zampini SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYGVX Lapack routine: leading minor of order %d is not positive definite",(int)B_ierr-B_N-1); 49708122e43SStefano Zampini } 49808122e43SStefano Zampini } 49908122e43SStefano Zampini 50008122e43SStefano Zampini if (B_neigs > nmax) { 501fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 502fd14bc51SStefano Zampini PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," found %d eigs, more than maximum required %d.\n",B_neigs,nmax); 503fd14bc51SStefano Zampini } 504f6f667cfSStefano Zampini eigs_start = B_neigs -nmax; 50508122e43SStefano Zampini B_neigs = nmax; 50608122e43SStefano Zampini } 50708122e43SStefano Zampini 5089552c7c7SStefano Zampini nmin_s = PetscMin(nmin,B_N); 5099552c7c7SStefano Zampini if (B_neigs < nmin_s) { 51008122e43SStefano Zampini PetscBLASInt B_neigs2; 51108122e43SStefano Zampini 512f6f667cfSStefano Zampini B_IU = B_N - B_neigs; 513f6f667cfSStefano Zampini B_IL = B_N - nmin_s + 1; 514fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 515fd14bc51SStefano 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); 516fd14bc51SStefano Zampini } 5179ab7bb16SStefano Zampini if (sub_schurs->is_hermitian) { 51808122e43SStefano Zampini PetscInt j; 51908122e43SStefano Zampini for (j=0;j<subset_size;j++) { 52008122e43SStefano Zampini ierr = PetscMemcpy(S+j*(subset_size+1),Sarray+cumarray+j*(subset_size+1),(subset_size-j)*sizeof(PetscScalar));CHKERRQ(ierr); 52108122e43SStefano Zampini } 52208122e43SStefano Zampini for (j=0;j<subset_size;j++) { 52308122e43SStefano Zampini ierr = PetscMemcpy(St+j*(subset_size+1),Starray+cumarray+j*(subset_size+1),(subset_size-j)*sizeof(PetscScalar));CHKERRQ(ierr); 52408122e43SStefano Zampini } 52508122e43SStefano Zampini } else { 52608122e43SStefano Zampini ierr = PetscMemcpy(S,Sarray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 52708122e43SStefano Zampini ierr = PetscMemcpy(St,Starray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 52808122e43SStefano Zampini } 52908122e43SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 53008122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 531f6f667cfSStefano Zampini PetscStackCallBLAS("LAPACKsygvx",LAPACKsygvx_(&B_itype,"V","I","L",&B_N,St,&B_N,S,&B_N,&thresh,&infty,&B_IL,&B_IU,&eps,&B_neigs2,eigs+B_neigs,eigv+B_neigs*subset_size,&B_N,work,&B_lwork,rwork,B_iwork,B_ifail,&B_ierr)); 53208122e43SStefano Zampini #else 533f6f667cfSStefano Zampini PetscStackCallBLAS("LAPACKsygvx",LAPACKsygvx_(&B_itype,"V","I","L",&B_N,St,&B_N,S,&B_N,&thresh,&infty,&B_IL,&B_IU,&eps,&B_neigs2,eigs+B_neigs,eigv+B_neigs*subset_size,&B_N,work,&B_lwork,B_iwork,B_ifail,&B_ierr)); 53408122e43SStefano Zampini #endif 53508122e43SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 53608122e43SStefano Zampini B_neigs += B_neigs2; 53708122e43SStefano Zampini } 53808122e43SStefano Zampini if (B_ierr) { 53908122e43SStefano Zampini if (B_ierr < 0 ) { 54008122e43SStefano Zampini SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYGVX Lapack routine: illegal value for argument %d",-(int)B_ierr); 54108122e43SStefano Zampini } else if (B_ierr <= B_N) { 54208122e43SStefano Zampini SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYGVX Lapack routine: %d eigenvalues failed to converge",(int)B_ierr); 54308122e43SStefano Zampini } else { 5449552c7c7SStefano Zampini SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYGVX Lapack routine: leading minor of order %d is not positive definite",(int)B_ierr-B_N-1); 54508122e43SStefano Zampini } 54608122e43SStefano Zampini } 547fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 548ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," -> Got %d eigs\n",B_neigs);CHKERRQ(ierr); 54908122e43SStefano Zampini for (j=0;j<B_neigs;j++) { 55008122e43SStefano Zampini if (eigs[j] == 0.0) { 551ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," Inf\n");CHKERRQ(ierr); 55208122e43SStefano Zampini } else { 553ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," %1.6e\n",eigs[j+eigs_start]);CHKERRQ(ierr); 554fd14bc51SStefano Zampini } 55508122e43SStefano Zampini } 55608122e43SStefano Zampini } 55708122e43SStefano Zampini } else { 55808122e43SStefano Zampini /* TODO */ 55908122e43SStefano Zampini } 560aff50787SStefano Zampini } 5618bec7fa6SStefano Zampini maxneigs = PetscMax(B_neigs,maxneigs); 5628bec7fa6SStefano Zampini pcbddc->adaptive_constraints_n[i+nv] = B_neigs; 5639162d606SStefano Zampini if (B_neigs) { 5649162d606SStefano 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); 565fd14bc51SStefano Zampini 566fd14bc51SStefano Zampini if (pcbddc->dbg_flag > 1) { 5679552c7c7SStefano Zampini PetscInt ii; 5689552c7c7SStefano Zampini for (ii=0;ii<B_neigs;ii++) { 569ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," -> Eigenvector %d/%d (%d)\n",ii,B_neigs,B_N);CHKERRQ(ierr); 5709552c7c7SStefano Zampini for (j=0;j<B_N;j++) { 571ac47001eSStefano Zampini #if defined(PETSC_USE_COMPLEX) 572ac47001eSStefano Zampini PetscReal r = PetscRealPart(pcbddc->adaptive_constraints_data[ii*subset_size+j+pcbddc->adaptive_constraints_data_ptr[cum]]); 573ac47001eSStefano Zampini PetscReal c = PetscImaginaryPart(pcbddc->adaptive_constraints_data[ii*subset_size+j+pcbddc->adaptive_constraints_data_ptr[cum]]); 574ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," %1.4e + %1.4e i\n",r,c);CHKERRQ(ierr); 575ac47001eSStefano Zampini #else 576ac47001eSStefano 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); 577ac47001eSStefano Zampini #endif 5789552c7c7SStefano Zampini } 5799552c7c7SStefano Zampini } 580fd14bc51SStefano Zampini } 58108122e43SStefano Zampini #if 0 5829162d606SStefano Zampini for (j=0;j<B_neigs;j++) { 58308122e43SStefano Zampini PetscBLASInt Blas_N,Blas_one = 1.0; 58408122e43SStefano Zampini PetscScalar norm; 58508122e43SStefano Zampini ierr = PetscBLASIntCast(subset_size,&Blas_N);CHKERRQ(ierr); 5869162d606SStefano Zampini PetscStackCallBLAS("BLASdot",norm = BLASdot_(&Blas_N,pcbddc->adaptive_constraints_data+pcbddc->adaptive_constraints_data_ptr[cum]+j*subset_size, 5879162d606SStefano Zampini &Blas_one,pcbddc->adaptive_constraints_data+pcbddc->adaptive_constraints_data_ptr[cum]+j*subset_size,&Blas_one)); 58808122e43SStefano Zampini if (pcbddc->adaptive_constraints_data[cum2] > 0.0) { 58908122e43SStefano Zampini norm = 1.0/PetscSqrtReal(PetscRealPart(norm)); 59008122e43SStefano Zampini } else { 59108122e43SStefano Zampini norm = -1.0/PetscSqrtReal(PetscRealPart(norm)); 59208122e43SStefano Zampini } 5939162d606SStefano Zampini PetscStackCallBLAS("BLASscal",BLASscal_(&Blas_N,&norm,pcbddc->adaptive_constraints_data+pcbddc->adaptive_constraints_data_ptr[cum]+j*subset_size,&Blas_one)); 594b1b3d7a2SStefano Zampini } 595b1b3d7a2SStefano Zampini #endif 5969162d606SStefano Zampini ierr = PetscMemcpy(pcbddc->adaptive_constraints_idxs+pcbddc->adaptive_constraints_idxs_ptr[cum],idxs,subset_size*sizeof(PetscInt));CHKERRQ(ierr); 5979162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr[cum+1] = pcbddc->adaptive_constraints_idxs_ptr[cum] + subset_size; 5989162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr[cum+1] = pcbddc->adaptive_constraints_data_ptr[cum] + subset_size*B_neigs; 5999162d606SStefano Zampini cum++; 60008122e43SStefano Zampini } 60108122e43SStefano Zampini ierr = ISRestoreIndices(sub_schurs->is_subs[i],&idxs);CHKERRQ(ierr); 60208122e43SStefano Zampini /* shift for next computation */ 60308122e43SStefano Zampini cumarray += subset_size*subset_size; 60408122e43SStefano Zampini } 605fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 606fd14bc51SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 607fd14bc51SStefano Zampini } 60808122e43SStefano Zampini 60908122e43SStefano Zampini if (mss) { 61008122e43SStefano Zampini ierr = MatSeqAIJRestoreArray(sub_schurs->sum_S_Ej_inv_all,&Sarray);CHKERRQ(ierr); 61108122e43SStefano Zampini ierr = MatSeqAIJRestoreArray(sub_schurs->sum_S_Ej_tilda_all,&Starray);CHKERRQ(ierr); 612f6f667cfSStefano Zampini /* destroy matrices (junk) */ 613f6f667cfSStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_inv_all);CHKERRQ(ierr); 614f6f667cfSStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_tilda_all);CHKERRQ(ierr); 61508122e43SStefano Zampini } 616f6f667cfSStefano Zampini if (allocated_S_St) { 617f6f667cfSStefano Zampini ierr = PetscFree2(S,St);CHKERRQ(ierr); 618f6f667cfSStefano Zampini } 619f6f667cfSStefano Zampini ierr = PetscFree5(eigv,eigs,work,B_iwork,B_ifail);CHKERRQ(ierr); 62008122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 62108122e43SStefano Zampini ierr = PetscFree(rwork);CHKERRQ(ierr); 62208122e43SStefano Zampini #endif 62308122e43SStefano Zampini if (pcbddc->dbg_flag) { 6241b968477SStefano Zampini PetscInt maxneigs_r; 62508122e43SStefano Zampini ierr = MPI_Allreduce(&maxneigs,&maxneigs_r,1,MPIU_INT,MPI_MAX,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 6269b28b941SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Maximum number of constraints per cc %d\n",maxneigs_r);CHKERRQ(ierr); 62708122e43SStefano Zampini } 62808122e43SStefano Zampini PetscFunctionReturn(0); 62908122e43SStefano Zampini } 630b1b3d7a2SStefano Zampini 631674ae819SStefano Zampini #undef __FUNCT__ 632c8587f34SStefano Zampini #define __FUNCT__ "PCBDDCSetUpSolvers" 633c8587f34SStefano Zampini PetscErrorCode PCBDDCSetUpSolvers(PC pc) 634c8587f34SStefano Zampini { 635c8587f34SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 6368629588bSStefano Zampini PetscScalar *coarse_submat_vals; 637c8587f34SStefano Zampini PetscErrorCode ierr; 638c8587f34SStefano Zampini 639c8587f34SStefano Zampini PetscFunctionBegin; 640f4ddd8eeSStefano Zampini /* Setup local scatters R_to_B and (optionally) R_to_D */ 6415e8657edSStefano Zampini /* PCBDDCSetUpLocalWorkVectors should be called first! */ 642c8587f34SStefano Zampini ierr = PCBDDCSetUpLocalScatters(pc);CHKERRQ(ierr); 643c8587f34SStefano Zampini 644684f6988SStefano Zampini /* Setup local neumann solver ksp_R */ 6450fccc4e9SStefano Zampini /* PCBDDCSetUpLocalScatters should be called first! */ 646684f6988SStefano Zampini ierr = PCBDDCSetUpLocalSolvers(pc,PETSC_FALSE,PETSC_TRUE);CHKERRQ(ierr); 647c8587f34SStefano Zampini 648c8587f34SStefano Zampini /* Change global null space passed in by the user if change of basis has been requested */ 649b9b85e73SStefano Zampini if (pcbddc->NullSpace && pcbddc->ChangeOfBasisMatrix) { 650c8587f34SStefano Zampini ierr = PCBDDCNullSpaceAdaptGlobal(pc);CHKERRQ(ierr); 651c8587f34SStefano Zampini } 652c8587f34SStefano Zampini 6538629588bSStefano Zampini /* 6548629588bSStefano Zampini Setup local correction and local part of coarse basis. 6558629588bSStefano Zampini Gives back the dense local part of the coarse matrix in column major ordering 6568629588bSStefano Zampini */ 65747f4ddc3SStefano Zampini ierr = PCBDDCSetUpCorrection(pc,&coarse_submat_vals);CHKERRQ(ierr); 6588629588bSStefano Zampini 6598629588bSStefano Zampini /* Compute total number of coarse nodes and setup coarse solver */ 6608629588bSStefano Zampini ierr = PCBDDCSetUpCoarseSolver(pc,coarse_submat_vals);CHKERRQ(ierr); 6618629588bSStefano Zampini 6628629588bSStefano Zampini /* free */ 6638629588bSStefano Zampini ierr = PetscFree(coarse_submat_vals);CHKERRQ(ierr); 664c8587f34SStefano Zampini PetscFunctionReturn(0); 665c8587f34SStefano Zampini } 666c8587f34SStefano Zampini 667c8587f34SStefano Zampini #undef __FUNCT__ 668674ae819SStefano Zampini #define __FUNCT__ "PCBDDCResetCustomization" 669674ae819SStefano Zampini PetscErrorCode PCBDDCResetCustomization(PC pc) 670674ae819SStefano Zampini { 671674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 672674ae819SStefano Zampini PetscErrorCode ierr; 673674ae819SStefano Zampini 674674ae819SStefano Zampini PetscFunctionBegin; 675674ae819SStefano Zampini ierr = PCBDDCGraphResetCSR(pcbddc->mat_graph);CHKERRQ(ierr); 676674ae819SStefano Zampini ierr = ISDestroy(&pcbddc->user_primal_vertices);CHKERRQ(ierr); 677674ae819SStefano Zampini ierr = MatNullSpaceDestroy(&pcbddc->NullSpace);CHKERRQ(ierr); 678674ae819SStefano Zampini ierr = ISDestroy(&pcbddc->NeumannBoundaries);CHKERRQ(ierr); 679785d1243SStefano Zampini ierr = ISDestroy(&pcbddc->NeumannBoundariesLocal);CHKERRQ(ierr); 680674ae819SStefano Zampini ierr = ISDestroy(&pcbddc->DirichletBoundaries);CHKERRQ(ierr); 681f4ddd8eeSStefano Zampini ierr = MatNullSpaceDestroy(&pcbddc->onearnullspace);CHKERRQ(ierr); 682f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->onearnullvecs_state);CHKERRQ(ierr); 683785d1243SStefano Zampini ierr = ISDestroy(&pcbddc->DirichletBoundariesLocal);CHKERRQ(ierr); 68463602bcaSStefano Zampini ierr = PCBDDCSetDofsSplitting(pc,0,NULL);CHKERRQ(ierr); 68563602bcaSStefano Zampini ierr = PCBDDCSetDofsSplittingLocal(pc,0,NULL);CHKERRQ(ierr); 686674ae819SStefano Zampini PetscFunctionReturn(0); 687674ae819SStefano Zampini } 688674ae819SStefano Zampini 689674ae819SStefano Zampini #undef __FUNCT__ 690674ae819SStefano Zampini #define __FUNCT__ "PCBDDCResetTopography" 691674ae819SStefano Zampini PetscErrorCode PCBDDCResetTopography(PC pc) 692674ae819SStefano Zampini { 693674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 694674ae819SStefano Zampini PetscErrorCode ierr; 695674ae819SStefano Zampini 696674ae819SStefano Zampini PetscFunctionBegin; 697b9b85e73SStefano Zampini ierr = MatDestroy(&pcbddc->user_ChangeOfBasisMatrix);CHKERRQ(ierr); 698674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 699674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 700674ae819SStefano Zampini ierr = PCBDDCGraphReset(pcbddc->mat_graph);CHKERRQ(ierr); 701b96c3477SStefano Zampini ierr = PCBDDCSubSchursReset(pcbddc->sub_schurs);CHKERRQ(ierr); 702674ae819SStefano Zampini PetscFunctionReturn(0); 703674ae819SStefano Zampini } 704674ae819SStefano Zampini 705674ae819SStefano Zampini #undef __FUNCT__ 706674ae819SStefano Zampini #define __FUNCT__ "PCBDDCResetSolvers" 707674ae819SStefano Zampini PetscErrorCode PCBDDCResetSolvers(PC pc) 708674ae819SStefano Zampini { 709674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 71006656605SStefano Zampini PetscScalar *array; 711674ae819SStefano Zampini PetscErrorCode ierr; 712674ae819SStefano Zampini 713674ae819SStefano Zampini PetscFunctionBegin; 714674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->coarse_vec);CHKERRQ(ierr); 71558da7f69SStefano Zampini if (pcbddc->coarse_phi_B) { 71606656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&array);CHKERRQ(ierr); 71706656605SStefano Zampini ierr = PetscFree(array);CHKERRQ(ierr); 71858da7f69SStefano Zampini } 719674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_B);CHKERRQ(ierr); 720674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_D);CHKERRQ(ierr); 72115aaf578SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_B);CHKERRQ(ierr); 72215aaf578SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_D);CHKERRQ(ierr); 723674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec1_P);CHKERRQ(ierr); 724674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec1_C);CHKERRQ(ierr); 725674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat2);CHKERRQ(ierr); 72606656605SStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat1);CHKERRQ(ierr); 727674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec1_R);CHKERRQ(ierr); 728674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec2_R);CHKERRQ(ierr); 7298ce42a96SStefano Zampini ierr = ISDestroy(&pcbddc->is_R_local);CHKERRQ(ierr); 730674ae819SStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_B);CHKERRQ(ierr); 731674ae819SStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_D);CHKERRQ(ierr); 732674ae819SStefano Zampini ierr = VecScatterDestroy(&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 733f4ddd8eeSStefano Zampini ierr = KSPDestroy(&pcbddc->ksp_D);CHKERRQ(ierr); 734f4ddd8eeSStefano Zampini ierr = KSPDestroy(&pcbddc->ksp_R);CHKERRQ(ierr); 735f4ddd8eeSStefano Zampini ierr = KSPDestroy(&pcbddc->coarse_ksp);CHKERRQ(ierr); 736f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 737727cdba6SStefano Zampini ierr = PetscFree(pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 7380e6343abSStefano Zampini ierr = PetscFree2(pcbddc->local_primal_ref_node,pcbddc->local_primal_ref_mult);CHKERRQ(ierr); 739f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->global_primal_indices);CHKERRQ(ierr); 74070cf5478SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling);CHKERRQ(ierr); 7416e683305SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 74281d14e9dSStefano Zampini ierr = MatDestroy(&pcbddc->benign_change);CHKERRQ(ierr); 7430369aaf7SStefano Zampini ierr = VecDestroy(&pcbddc->benign_vec);CHKERRQ(ierr); 7448b9f24d4SStefano Zampini ierr = MatDestroy(&pcbddc->benign_original_mat);CHKERRQ(ierr); 7458b9f24d4SStefano Zampini ierr = PetscSFDestroy(&pcbddc->benign_sf);CHKERRQ(ierr); 74681d14e9dSStefano Zampini ierr = PetscFree2(pcbddc->B0_cols,pcbddc->B0_vals);CHKERRQ(ierr); 747674ae819SStefano Zampini PetscFunctionReturn(0); 748674ae819SStefano Zampini } 749674ae819SStefano Zampini 750674ae819SStefano Zampini #undef __FUNCT__ 751f4ddd8eeSStefano Zampini #define __FUNCT__ "PCBDDCSetUpLocalWorkVectors" 752f4ddd8eeSStefano Zampini PetscErrorCode PCBDDCSetUpLocalWorkVectors(PC pc) 7536bfb1811SStefano Zampini { 7546bfb1811SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 7556bfb1811SStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 7566bfb1811SStefano Zampini VecType impVecType; 757019a44ceSStefano Zampini PetscInt n_constraints,n_R,old_size,n_benign; 7586bfb1811SStefano Zampini PetscErrorCode ierr; 7596bfb1811SStefano Zampini 7606bfb1811SStefano Zampini PetscFunctionBegin; 761f4ddd8eeSStefano Zampini if (!pcbddc->ConstraintMatrix) { 762019a44ceSStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"BDDC Constraint matrix has not been created"); 763f4ddd8eeSStefano Zampini } 764e7b262bdSStefano Zampini /* get sizes */ 765c9ed8603SStefano Zampini n_benign = 0; 766c9ed8603SStefano Zampini if (pcbddc->benign_p0_lidx >= 0) n_benign = 1; 767019a44ceSStefano Zampini n_constraints = pcbddc->local_primal_size - n_benign - pcbddc->n_vertices; 768b371cd4fSStefano Zampini n_R = pcis->n - pcbddc->n_vertices; 7696bfb1811SStefano Zampini ierr = VecGetType(pcis->vec1_N,&impVecType);CHKERRQ(ierr); 770e7b262bdSStefano Zampini /* local work vectors (try to avoid unneeded work)*/ 771e7b262bdSStefano Zampini /* R nodes */ 772e7b262bdSStefano Zampini old_size = -1; 773e7b262bdSStefano Zampini if (pcbddc->vec1_R) { 774e7b262bdSStefano Zampini ierr = VecGetSize(pcbddc->vec1_R,&old_size);CHKERRQ(ierr); 775e7b262bdSStefano Zampini } 776e7b262bdSStefano Zampini if (n_R != old_size) { 777e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec1_R);CHKERRQ(ierr); 778e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec2_R);CHKERRQ(ierr); 7796bfb1811SStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_R);CHKERRQ(ierr); 7806bfb1811SStefano Zampini ierr = VecSetSizes(pcbddc->vec1_R,PETSC_DECIDE,n_R);CHKERRQ(ierr); 7816bfb1811SStefano Zampini ierr = VecSetType(pcbddc->vec1_R,impVecType);CHKERRQ(ierr); 7826bfb1811SStefano Zampini ierr = VecDuplicate(pcbddc->vec1_R,&pcbddc->vec2_R);CHKERRQ(ierr); 783e7b262bdSStefano Zampini } 784e7b262bdSStefano Zampini /* local primal dofs */ 785e7b262bdSStefano Zampini old_size = -1; 786e7b262bdSStefano Zampini if (pcbddc->vec1_P) { 787e7b262bdSStefano Zampini ierr = VecGetSize(pcbddc->vec1_P,&old_size);CHKERRQ(ierr); 788e7b262bdSStefano Zampini } 789e9189074SStefano Zampini if (pcbddc->local_primal_size != old_size) { 790e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec1_P);CHKERRQ(ierr); 79183b7ccabSStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_P);CHKERRQ(ierr); 792e9189074SStefano Zampini ierr = VecSetSizes(pcbddc->vec1_P,PETSC_DECIDE,pcbddc->local_primal_size);CHKERRQ(ierr); 7936bfb1811SStefano Zampini ierr = VecSetType(pcbddc->vec1_P,impVecType);CHKERRQ(ierr); 794e7b262bdSStefano Zampini } 795e7b262bdSStefano Zampini /* local explicit constraints */ 796e7b262bdSStefano Zampini old_size = -1; 797e7b262bdSStefano Zampini if (pcbddc->vec1_C) { 798e7b262bdSStefano Zampini ierr = VecGetSize(pcbddc->vec1_C,&old_size);CHKERRQ(ierr); 799e7b262bdSStefano Zampini } 800e7b262bdSStefano Zampini if (n_constraints && n_constraints != old_size) { 801e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec1_C);CHKERRQ(ierr); 80283b7ccabSStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_C);CHKERRQ(ierr); 80383b7ccabSStefano Zampini ierr = VecSetSizes(pcbddc->vec1_C,PETSC_DECIDE,n_constraints);CHKERRQ(ierr); 80483b7ccabSStefano Zampini ierr = VecSetType(pcbddc->vec1_C,impVecType);CHKERRQ(ierr); 80583b7ccabSStefano Zampini } 8066bfb1811SStefano Zampini PetscFunctionReturn(0); 8076bfb1811SStefano Zampini } 8086bfb1811SStefano Zampini 8096bfb1811SStefano Zampini #undef __FUNCT__ 81047f4ddc3SStefano Zampini #define __FUNCT__ "PCBDDCSetUpCorrection" 81147f4ddc3SStefano Zampini PetscErrorCode PCBDDCSetUpCorrection(PC pc, PetscScalar **coarse_submat_vals_n) 81288ebb749SStefano Zampini { 81325084f0cSStefano Zampini PetscErrorCode ierr; 81425084f0cSStefano Zampini /* pointers to pcis and pcbddc */ 81588ebb749SStefano Zampini PC_IS* pcis = (PC_IS*)pc->data; 81688ebb749SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 817d62866d3SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 81825084f0cSStefano Zampini /* submatrices of local problem */ 81980677318SStefano Zampini Mat A_RV,A_VR,A_VV,local_auxmat2_R; 820019a44ceSStefano Zampini /* submatrices of benign trick */ 821d16cbb6bSStefano Zampini Mat B0_V = NULL; 82206656605SStefano Zampini /* submatrices of local coarse problem */ 82306656605SStefano Zampini Mat S_VV,S_CV,S_VC,S_CC; 82425084f0cSStefano Zampini /* working matrices */ 82506656605SStefano Zampini Mat C_CR; 82625084f0cSStefano Zampini /* additional working stuff */ 82706656605SStefano Zampini PC pc_R; 828d12edf2fSStefano Zampini Mat F,B0 = NULL; 82906656605SStefano Zampini PetscBool isLU,isCHOL,isILU; 83006656605SStefano Zampini 83125084f0cSStefano Zampini PetscScalar *coarse_submat_vals; /* TODO: use a PETSc matrix */ 83206656605SStefano Zampini PetscScalar *work; 83306656605SStefano Zampini PetscInt *idx_V_B; 834d12edf2fSStefano Zampini PetscInt n,n_vertices,n_constraints,n_benign,p0_lidx_I = 0; 83506656605SStefano Zampini PetscInt i,n_R,n_D,n_B; 836b9d89cd5SStefano Zampini PetscBool unsymmetric_check; 83745a1bb75SStefano Zampini /* matrix type (vector type propagated downstream from vec1_C and local matrix type) */ 83888ebb749SStefano Zampini MatType impMatType; 83925084f0cSStefano Zampini /* some shortcuts to scalars */ 84006656605SStefano Zampini PetscScalar one=1.0,m_one=-1.0; 84188ebb749SStefano Zampini 84288ebb749SStefano Zampini PetscFunctionBegin; 843c9ed8603SStefano Zampini n_benign = 0; 844c9ed8603SStefano Zampini if (pcbddc->benign_p0_lidx >= 0) n_benign = 1; 845b371cd4fSStefano Zampini n_vertices = pcbddc->n_vertices; 846019a44ceSStefano Zampini n_constraints = pcbddc->local_primal_size - n_benign - n_vertices; 84788ebb749SStefano Zampini /* Set Non-overlapping dimensions */ 848b371cd4fSStefano Zampini n_B = pcis->n_B; 849b371cd4fSStefano Zampini n_D = pcis->n - n_B; 85088ebb749SStefano Zampini n_R = pcis->n - n_vertices; 85188ebb749SStefano Zampini 85288ebb749SStefano Zampini /* Set types for local objects needed by BDDC precondtioner */ 85388ebb749SStefano Zampini impMatType = MATSEQDENSE; 85488ebb749SStefano Zampini 85588ebb749SStefano Zampini /* vertices in boundary numbering */ 856785e854fSJed Brown ierr = PetscMalloc1(n_vertices,&idx_V_B);CHKERRQ(ierr); 8570e6343abSStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,n_vertices,pcbddc->local_primal_ref_node,&i,idx_V_B);CHKERRQ(ierr); 85888ebb749SStefano Zampini if (i != n_vertices) { 85922d5777bSStefano Zampini SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Error in boundary numbering for BDDC vertices! %d != %d\n",n_vertices,i); 86088ebb749SStefano Zampini } 86188ebb749SStefano Zampini 86206656605SStefano Zampini /* Subdomain contribution (Non-overlapping) to coarse matrix */ 863019a44ceSStefano Zampini ierr = PetscCalloc1(pcbddc->local_primal_size*pcbddc->local_primal_size,&coarse_submat_vals);CHKERRQ(ierr); 86406656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_vertices,n_vertices,coarse_submat_vals,&S_VV);CHKERRQ(ierr); 86506656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_VV,pcbddc->local_primal_size);CHKERRQ(ierr); 86606656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_constraints,n_vertices,coarse_submat_vals+n_vertices,&S_CV);CHKERRQ(ierr); 86706656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_CV,pcbddc->local_primal_size);CHKERRQ(ierr); 86806656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_vertices,n_constraints,coarse_submat_vals+pcbddc->local_primal_size*n_vertices,&S_VC);CHKERRQ(ierr); 86906656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_VC,pcbddc->local_primal_size);CHKERRQ(ierr); 87006656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_constraints,n_constraints,coarse_submat_vals+(pcbddc->local_primal_size+1)*n_vertices,&S_CC);CHKERRQ(ierr); 87106656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_CC,pcbddc->local_primal_size);CHKERRQ(ierr); 87206656605SStefano Zampini 87306656605SStefano Zampini unsymmetric_check = PETSC_FALSE; 87406656605SStefano Zampini /* allocate workspace */ 87506656605SStefano Zampini n = 0; 87606656605SStefano Zampini if (n_constraints) { 87706656605SStefano Zampini n += n_R*n_constraints; 87806656605SStefano Zampini } 87906656605SStefano Zampini if (n_vertices) { 88006656605SStefano Zampini n = PetscMax(2*n_R*n_vertices,n); 88180677318SStefano Zampini n = PetscMax((n_R+n_B)*n_vertices,n); 88206656605SStefano Zampini } 8833301b35fSStefano Zampini if (!pcbddc->symmetric_primal) { 88406656605SStefano Zampini n = PetscMax(2*n_R*pcbddc->local_primal_size,n); 88506656605SStefano Zampini unsymmetric_check = PETSC_TRUE; 88606656605SStefano Zampini } 88706656605SStefano Zampini ierr = PetscMalloc1(n,&work);CHKERRQ(ierr); 88806656605SStefano Zampini 88906656605SStefano Zampini /* determine if can use MatSolve routines instead of calling KSPSolve on ksp_R */ 89006656605SStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_R);CHKERRQ(ierr); 89106656605SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_R,PCLU,&isLU);CHKERRQ(ierr); 89206656605SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_R,PCILU,&isILU);CHKERRQ(ierr); 89306656605SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_R,PCCHOLESKY,&isCHOL);CHKERRQ(ierr); 89406656605SStefano Zampini if (isLU || isILU || isCHOL) { 89506656605SStefano Zampini ierr = PCFactorGetMatrix(pc_R,&F);CHKERRQ(ierr); 896d62866d3SStefano Zampini } else if (sub_schurs->reuse_mumps) { 897d62866d3SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 898d62866d3SStefano Zampini MatFactorType type; 899d62866d3SStefano Zampini 9006816873aSStefano Zampini F = reuse_mumps->F; 9016816873aSStefano Zampini ierr = MatGetFactorType(F,&type);CHKERRQ(ierr); 902d62866d3SStefano Zampini if (type == MAT_FACTOR_CHOLESKY) isCHOL = PETSC_TRUE; 90306656605SStefano Zampini } else { 90406656605SStefano Zampini F = NULL; 90506656605SStefano Zampini } 90606656605SStefano Zampini 90788ebb749SStefano Zampini /* Precompute stuffs needed for preprocessing and application of BDDC*/ 90888ebb749SStefano Zampini if (n_constraints) { 90906656605SStefano Zampini Mat M1,M2,M3; 91080677318SStefano Zampini Mat auxmat; 91106656605SStefano Zampini IS is_aux; 91280677318SStefano Zampini PetscScalar *array,*array2; 91306656605SStefano Zampini 914f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat1);CHKERRQ(ierr); 91580677318SStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat2);CHKERRQ(ierr); 91688ebb749SStefano Zampini 91725084f0cSStefano Zampini /* Extract constraints on R nodes: C_{CR} */ 91825084f0cSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,n_constraints,n_vertices,1,&is_aux);CHKERRQ(ierr); 9198ce42a96SStefano Zampini ierr = MatGetSubMatrix(pcbddc->ConstraintMatrix,is_aux,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&C_CR);CHKERRQ(ierr); 92080677318SStefano Zampini ierr = MatGetSubMatrix(pcbddc->ConstraintMatrix,is_aux,pcis->is_B_local,MAT_INITIAL_MATRIX,&auxmat);CHKERRQ(ierr); 92188ebb749SStefano Zampini 92280677318SStefano Zampini /* Assemble local_auxmat2_R = (- A_{RR}^{-1} C^T_{CR}) needed by BDDC setup */ 92380677318SStefano Zampini /* Assemble pcbddc->local_auxmat2 = R_to_B (- A_{RR}^{-1} C^T_{CR}) needed by BDDC application */ 92406656605SStefano Zampini ierr = PetscMemzero(work,n_R*n_constraints*sizeof(PetscScalar));CHKERRQ(ierr); 92588ebb749SStefano Zampini for (i=0;i<n_constraints;i++) { 92606656605SStefano Zampini const PetscScalar *row_cmat_values; 92706656605SStefano Zampini const PetscInt *row_cmat_indices; 92806656605SStefano Zampini PetscInt size_of_constraint,j; 92988ebb749SStefano Zampini 93006656605SStefano Zampini ierr = MatGetRow(C_CR,i,&size_of_constraint,&row_cmat_indices,&row_cmat_values);CHKERRQ(ierr); 93106656605SStefano Zampini for (j=0;j<size_of_constraint;j++) { 93206656605SStefano Zampini work[row_cmat_indices[j]+i*n_R] = -row_cmat_values[j]; 93306656605SStefano Zampini } 93406656605SStefano Zampini ierr = MatRestoreRow(C_CR,i,&size_of_constraint,&row_cmat_indices,&row_cmat_values);CHKERRQ(ierr); 93506656605SStefano Zampini } 93680677318SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_constraints,NULL,&local_auxmat2_R);CHKERRQ(ierr); 93706656605SStefano Zampini if (F) { 93806656605SStefano Zampini Mat B; 93906656605SStefano Zampini 94006656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_constraints,work,&B);CHKERRQ(ierr); 94180677318SStefano Zampini ierr = MatMatSolve(F,B,local_auxmat2_R);CHKERRQ(ierr); 94206656605SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 94306656605SStefano Zampini } else { 94480677318SStefano Zampini PetscScalar *marr; 94580677318SStefano Zampini 94680677318SStefano Zampini ierr = MatDenseGetArray(local_auxmat2_R,&marr);CHKERRQ(ierr); 94706656605SStefano Zampini for (i=0;i<n_constraints;i++) { 94806656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*n_R);CHKERRQ(ierr); 94980677318SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,marr+i*n_R);CHKERRQ(ierr); 95006656605SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 95106656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 95206656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 95306656605SStefano Zampini } 95480677318SStefano Zampini ierr = MatDenseRestoreArray(local_auxmat2_R,&marr);CHKERRQ(ierr); 95506656605SStefano Zampini } 95680677318SStefano Zampini if (!pcbddc->switch_static) { 95780677318SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,n_constraints,NULL,&pcbddc->local_auxmat2);CHKERRQ(ierr); 95880677318SStefano Zampini ierr = MatDenseGetArray(pcbddc->local_auxmat2,&array);CHKERRQ(ierr); 95980677318SStefano Zampini ierr = MatDenseGetArray(local_auxmat2_R,&array2);CHKERRQ(ierr); 96080677318SStefano Zampini for (i=0;i<n_constraints;i++) { 96180677318SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,array2+i*n_R);CHKERRQ(ierr); 96280677318SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,array+i*n_B);CHKERRQ(ierr); 96380677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 96480677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 96580677318SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 96680677318SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 96780677318SStefano Zampini } 96880677318SStefano Zampini ierr = MatDenseRestoreArray(local_auxmat2_R,&array2);CHKERRQ(ierr); 96980677318SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->local_auxmat2,&array);CHKERRQ(ierr); 97080677318SStefano Zampini ierr = MatMatMult(auxmat,pcbddc->local_auxmat2,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&M3);CHKERRQ(ierr); 97180677318SStefano Zampini } else { 97280677318SStefano Zampini ierr = PetscObjectReference((PetscObject)local_auxmat2_R);CHKERRQ(ierr); 97380677318SStefano Zampini pcbddc->local_auxmat2 = local_auxmat2_R; 97425084f0cSStefano Zampini ierr = MatMatMult(C_CR,pcbddc->local_auxmat2,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&M3);CHKERRQ(ierr); 97580677318SStefano Zampini } 97680677318SStefano Zampini ierr = ISDestroy(&is_aux);CHKERRQ(ierr); 97780677318SStefano Zampini /* Assemble explicitly S_CC = ( C_{CR} A_{RR}^{-1} C^T_{CR} )^{-1} */ 97880677318SStefano Zampini ierr = MatScale(M3,m_one);CHKERRQ(ierr); 97906656605SStefano Zampini ierr = MatDuplicate(M3,MAT_DO_NOT_COPY_VALUES,&M1);CHKERRQ(ierr); 98006656605SStefano Zampini ierr = MatDuplicate(M3,MAT_DO_NOT_COPY_VALUES,&M2);CHKERRQ(ierr); 98180677318SStefano Zampini if (isCHOL) { 98280677318SStefano Zampini ierr = MatCholeskyFactor(M3,NULL,NULL);CHKERRQ(ierr); 98380677318SStefano Zampini } else { 98425084f0cSStefano Zampini ierr = MatLUFactor(M3,NULL,NULL,NULL);CHKERRQ(ierr); 98580677318SStefano Zampini } 98680677318SStefano Zampini ierr = VecSet(pcbddc->vec1_C,one);CHKERRQ(ierr); 98706656605SStefano Zampini ierr = MatDiagonalSet(M2,pcbddc->vec1_C,INSERT_VALUES);CHKERRQ(ierr); 98825084f0cSStefano Zampini ierr = MatMatSolve(M3,M2,M1);CHKERRQ(ierr); 98925084f0cSStefano Zampini ierr = MatDestroy(&M2);CHKERRQ(ierr); 99025084f0cSStefano Zampini ierr = MatDestroy(&M3);CHKERRQ(ierr); 99180677318SStefano Zampini /* Assemble local_auxmat1 = S_CC*C_{CB} needed by BDDC application in KSP and in preproc */ 99280677318SStefano Zampini ierr = MatMatMult(M1,auxmat,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&pcbddc->local_auxmat1);CHKERRQ(ierr); 99380677318SStefano Zampini ierr = MatDestroy(&auxmat);CHKERRQ(ierr); 99406656605SStefano Zampini ierr = MatCopy(M1,S_CC,SAME_NONZERO_PATTERN);CHKERRQ(ierr); /* S_CC can have a different LDA, MatMatSolve doesn't support it */ 99506656605SStefano Zampini ierr = MatDestroy(&M1);CHKERRQ(ierr); 996f4ddd8eeSStefano Zampini } 99788ebb749SStefano Zampini /* Get submatrices from subdomain matrix */ 998d16cbb6bSStefano Zampini if (n_benign) { 999d16cbb6bSStefano Zampini IS dummy; 1000d16cbb6bSStefano Zampini Mat B0_R; 1001d16cbb6bSStefano Zampini PetscReal norm; 1002d16cbb6bSStefano Zampini PetscInt ii[2]; 1003d16cbb6bSStefano Zampini 1004d16cbb6bSStefano Zampini ii[0] = 0; 1005d16cbb6bSStefano Zampini ii[1] = pcbddc->B0_ncol; 1006d16cbb6bSStefano Zampini ierr = MatCreateSeqAIJWithArrays(PETSC_COMM_SELF,1,pcis->n,ii,pcbddc->B0_cols,pcbddc->B0_vals,&B0);CHKERRQ(ierr); 1007d16cbb6bSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,1,0,1,&dummy);CHKERRQ(ierr); 1008d16cbb6bSStefano Zampini ierr = MatGetSubMatrix(B0,dummy,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&B0_R);CHKERRQ(ierr); 1009d16cbb6bSStefano Zampini ierr = MatNorm(B0_R,NORM_INFINITY,&norm);CHKERRQ(ierr); 1010d16cbb6bSStefano Zampini if (norm > PETSC_SMALL) { 1011d16cbb6bSStefano Zampini SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_SUP,"Benign trick can not be applied! ||B0_R|| = %f (should be numerically 0.)",norm); 1012d16cbb6bSStefano Zampini } 1013d16cbb6bSStefano Zampini ierr = MatDestroy(&B0_R);CHKERRQ(ierr); 1014d16cbb6bSStefano Zampini ierr = ISDestroy(&dummy);CHKERRQ(ierr); 1015d16cbb6bSStefano Zampini } 1016d16cbb6bSStefano Zampini 101788ebb749SStefano Zampini if (n_vertices) { 101806656605SStefano Zampini IS is_aux; 10193a50541eSStefano Zampini 10206816873aSStefano Zampini if (sub_schurs->reuse_mumps) { /* is_R_local is not sorted, ISComplement doesn't like it */ 10216816873aSStefano Zampini IS tis; 10226816873aSStefano Zampini 10236816873aSStefano Zampini ierr = ISDuplicate(pcbddc->is_R_local,&tis);CHKERRQ(ierr); 10246816873aSStefano Zampini ierr = ISSort(tis);CHKERRQ(ierr); 10256816873aSStefano Zampini ierr = ISComplement(tis,0,pcis->n,&is_aux);CHKERRQ(ierr); 10266816873aSStefano Zampini ierr = ISDestroy(&tis);CHKERRQ(ierr); 10276816873aSStefano Zampini } else { 10283a50541eSStefano Zampini ierr = ISComplement(pcbddc->is_R_local,0,pcis->n,&is_aux);CHKERRQ(ierr); 10296816873aSStefano Zampini } 10309577ea80SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,pcbddc->is_R_local,is_aux,MAT_INITIAL_MATRIX,&A_RV);CHKERRQ(ierr); 10319577ea80SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,is_aux,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&A_VR);CHKERRQ(ierr); 103204708bb6SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,is_aux,is_aux,MAT_INITIAL_MATRIX,&A_VV);CHKERRQ(ierr); 1033019a44ceSStefano Zampini if (n_benign) { 1034019a44ceSStefano Zampini IS dummy; 1035019a44ceSStefano Zampini 1036019a44ceSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,1,0,1,&dummy);CHKERRQ(ierr); 1037019a44ceSStefano Zampini ierr = MatGetSubMatrix(B0,dummy,is_aux,MAT_INITIAL_MATRIX,&B0_V);CHKERRQ(ierr); 1038019a44ceSStefano Zampini ierr = ISDestroy(&dummy);CHKERRQ(ierr); 1039019a44ceSStefano Zampini } 104025084f0cSStefano Zampini ierr = ISDestroy(&is_aux);CHKERRQ(ierr); 104188ebb749SStefano Zampini } 104288ebb749SStefano Zampini 104388ebb749SStefano Zampini /* Matrix of coarse basis functions (local) */ 1044f4ddd8eeSStefano Zampini if (pcbddc->coarse_phi_B) { 104506656605SStefano Zampini PetscInt on_B,on_primal,on_D=n_D; 104606656605SStefano Zampini if (pcbddc->coarse_phi_D) { 104706656605SStefano Zampini ierr = MatGetSize(pcbddc->coarse_phi_D,&on_D,NULL);CHKERRQ(ierr); 104806656605SStefano Zampini } 1049f4ddd8eeSStefano Zampini ierr = MatGetSize(pcbddc->coarse_phi_B,&on_B,&on_primal);CHKERRQ(ierr); 105006656605SStefano Zampini if (on_B != n_B || on_primal != pcbddc->local_primal_size || on_D != n_D) { 105106656605SStefano Zampini PetscScalar *marray; 105206656605SStefano Zampini 105306656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&marray);CHKERRQ(ierr); 105406656605SStefano Zampini ierr = PetscFree(marray);CHKERRQ(ierr); 1055f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_B);CHKERRQ(ierr); 1056f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_B);CHKERRQ(ierr); 1057f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_D);CHKERRQ(ierr); 1058f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_D);CHKERRQ(ierr); 1059f4ddd8eeSStefano Zampini } 1060f4ddd8eeSStefano Zampini } 106106656605SStefano Zampini 1062f4ddd8eeSStefano Zampini if (!pcbddc->coarse_phi_B) { 106306656605SStefano Zampini PetscScalar *marray; 106488ebb749SStefano Zampini 106506656605SStefano Zampini n = n_B*pcbddc->local_primal_size; 10668eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 106706656605SStefano Zampini n += n_D*pcbddc->local_primal_size; 106888ebb749SStefano Zampini } 10693301b35fSStefano Zampini if (!pcbddc->symmetric_primal) { 107006656605SStefano Zampini n *= 2; 107188ebb749SStefano Zampini } 107206656605SStefano Zampini ierr = PetscCalloc1(n,&marray);CHKERRQ(ierr); 107306656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,pcbddc->local_primal_size,marray,&pcbddc->coarse_phi_B);CHKERRQ(ierr); 107406656605SStefano Zampini n = n_B*pcbddc->local_primal_size; 10758eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 107606656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_D,pcbddc->local_primal_size,marray+n,&pcbddc->coarse_phi_D);CHKERRQ(ierr); 107706656605SStefano Zampini n += n_D*pcbddc->local_primal_size; 107888ebb749SStefano Zampini } 10793301b35fSStefano Zampini if (!pcbddc->symmetric_primal) { 108006656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,pcbddc->local_primal_size,marray+n,&pcbddc->coarse_psi_B);CHKERRQ(ierr); 10818eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 108206656605SStefano Zampini n = n_B*pcbddc->local_primal_size; 108306656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_D,pcbddc->local_primal_size,marray+n,&pcbddc->coarse_psi_D);CHKERRQ(ierr); 108488ebb749SStefano Zampini } 108588ebb749SStefano Zampini } else { 1086c0553b1fSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->coarse_phi_B);CHKERRQ(ierr); 1087c0553b1fSStefano Zampini pcbddc->coarse_psi_B = pcbddc->coarse_phi_B; 10881b968477SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 1089c0553b1fSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->coarse_phi_D);CHKERRQ(ierr); 1090c0553b1fSStefano Zampini pcbddc->coarse_psi_D = pcbddc->coarse_phi_D; 1091c0553b1fSStefano Zampini } 109288ebb749SStefano Zampini } 109306656605SStefano Zampini } 1094019a44ceSStefano Zampini 109506656605SStefano Zampini /* We are now ready to evaluate coarse basis functions and subdomain contribution to coarse problem */ 1096d12edf2fSStefano Zampini if (n_benign && (pcbddc->switch_static || pcbddc->dbg_flag)) { 1097d12edf2fSStefano Zampini const PetscInt *idxs; 1098d12edf2fSStefano Zampini 1099d12edf2fSStefano Zampini ierr = ISGetIndices(pcis->is_I_local,&idxs);CHKERRQ(ierr); 1100d12edf2fSStefano Zampini ierr = PetscFindInt(pcbddc->benign_p0_lidx,pcis->n-pcis->n_B,idxs,&p0_lidx_I);CHKERRQ(ierr); 1101d12edf2fSStefano Zampini ierr = ISRestoreIndices(pcis->is_I_local,&idxs);CHKERRQ(ierr); 1102d12edf2fSStefano Zampini } 1103d16cbb6bSStefano Zampini 110406656605SStefano Zampini /* vertices */ 110506656605SStefano Zampini if (n_vertices) { 110616f15bc4SStefano Zampini 110704708bb6SStefano Zampini ierr = MatConvert(A_VV,impMatType,MAT_REUSE_MATRIX,&A_VV);CHKERRQ(ierr); 110804708bb6SStefano Zampini 110916f15bc4SStefano Zampini if (n_R) { 111006656605SStefano Zampini Mat A_RRmA_RV,S_VVt; /* S_VVt with LDA=N */ 111106656605SStefano Zampini PetscBLASInt B_N,B_one = 1; 111216f15bc4SStefano Zampini PetscScalar *x,*y; 111304708bb6SStefano Zampini PetscBool isseqaij; 111406656605SStefano Zampini 111521eccb56SStefano Zampini ierr = MatScale(A_RV,m_one);CHKERRQ(ierr); 111606656605SStefano Zampini ierr = MatConvert(A_RV,impMatType,MAT_REUSE_MATRIX,&A_RV);CHKERRQ(ierr); 111704708bb6SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_vertices,work,&A_RRmA_RV);CHKERRQ(ierr); 11186816873aSStefano Zampini if (F) { /* TODO could be optimized for symmetric problems */ 111906656605SStefano Zampini ierr = MatMatSolve(F,A_RV,A_RRmA_RV);CHKERRQ(ierr); 112006656605SStefano Zampini } else { 112106656605SStefano Zampini ierr = MatDenseGetArray(A_RV,&y);CHKERRQ(ierr); 112206656605SStefano Zampini for (i=0;i<n_vertices;i++) { 112306656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,y+i*n_R);CHKERRQ(ierr); 112406656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,work+i*n_R);CHKERRQ(ierr); 112506656605SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 112606656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 112706656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 112806656605SStefano Zampini } 112906656605SStefano Zampini ierr = MatDenseRestoreArray(A_RV,&y);CHKERRQ(ierr); 113006656605SStefano Zampini } 113180677318SStefano Zampini ierr = MatDestroy(&A_RV);CHKERRQ(ierr); 113206656605SStefano Zampini /* S_VV and S_CV are the subdomain contribution to coarse matrix. WARNING -> column major ordering */ 113306656605SStefano Zampini if (n_constraints) { 113406656605SStefano Zampini Mat B; 113580677318SStefano Zampini 1136b3d85658SStefano Zampini ierr = PetscMemzero(work+n_R*n_vertices,n_B*n_vertices*sizeof(PetscScalar));CHKERRQ(ierr); 113780677318SStefano Zampini for (i=0;i<n_vertices;i++) { 113880677318SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*n_R);CHKERRQ(ierr); 113980677318SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,work+n_R*n_vertices+i*n_B);CHKERRQ(ierr); 114080677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 114180677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 114280677318SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 114380677318SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 114480677318SStefano Zampini } 114580677318SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,n_vertices,work+n_R*n_vertices,&B);CHKERRQ(ierr); 114680677318SStefano Zampini ierr = MatMatMult(pcbddc->local_auxmat1,B,MAT_REUSE_MATRIX,PETSC_DEFAULT,&S_CV);CHKERRQ(ierr); 114780677318SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 114806656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_vertices,work+n_R*n_vertices,&B);CHKERRQ(ierr); 114980677318SStefano Zampini ierr = MatMatMult(local_auxmat2_R,S_CV,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B);CHKERRQ(ierr); 115006656605SStefano Zampini ierr = MatScale(S_CV,m_one);CHKERRQ(ierr); 115106656605SStefano Zampini ierr = PetscBLASIntCast(n_R*n_vertices,&B_N);CHKERRQ(ierr); 115206656605SStefano Zampini PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&B_N,&one,work+n_R*n_vertices,&B_one,work,&B_one)); 115306656605SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 115406656605SStefano Zampini } 115504708bb6SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)A_VR,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 115604708bb6SStefano Zampini if (!isseqaij) { /* MatMatMult with SEQ(S)BAIJ below will raise an error */ 115704708bb6SStefano Zampini ierr = MatConvert(A_VR,MATSEQAIJ,MAT_REUSE_MATRIX,&A_VR);CHKERRQ(ierr); 115804708bb6SStefano Zampini } 115906656605SStefano Zampini ierr = MatMatMult(A_VR,A_RRmA_RV,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&S_VVt);CHKERRQ(ierr); 116080677318SStefano Zampini ierr = MatDestroy(&A_RRmA_RV);CHKERRQ(ierr); 116106656605SStefano Zampini ierr = PetscBLASIntCast(n_vertices*n_vertices,&B_N);CHKERRQ(ierr); 116206656605SStefano Zampini ierr = MatDenseGetArray(A_VV,&x);CHKERRQ(ierr); 116306656605SStefano Zampini ierr = MatDenseGetArray(S_VVt,&y);CHKERRQ(ierr); 116406656605SStefano Zampini PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&B_N,&one,x,&B_one,y,&B_one)); 116506656605SStefano Zampini ierr = MatDenseRestoreArray(A_VV,&x);CHKERRQ(ierr); 116606656605SStefano Zampini ierr = MatDenseRestoreArray(S_VVt,&y);CHKERRQ(ierr); 116706656605SStefano Zampini ierr = MatCopy(S_VVt,S_VV,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 1168d16cbb6bSStefano Zampini ierr = MatDestroy(&S_VVt);CHKERRQ(ierr); 1169019a44ceSStefano Zampini } else { 1170d16cbb6bSStefano Zampini ierr = MatCopy(A_VV,S_VV,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 1171d16cbb6bSStefano Zampini } 1172d16cbb6bSStefano Zampini if (n_benign) { 1173019a44ceSStefano Zampini const PetscScalar *vals; 1174019a44ceSStefano Zampini const PetscInt *idxs; 1175019a44ceSStefano Zampini PetscInt n,j; 1176019a44ceSStefano Zampini 1177019a44ceSStefano Zampini ierr = MatGetRow(B0_V,0,&n,&idxs,&vals);CHKERRQ(ierr); 1178d16cbb6bSStefano Zampini for (j=0;j<n;j++) { 1179d16cbb6bSStefano Zampini coarse_submat_vals[(pcbddc->local_primal_size-1)*pcbddc->local_primal_size+idxs[j]] = vals[j]; 1180d16cbb6bSStefano Zampini coarse_submat_vals[(idxs[j]+1)*pcbddc->local_primal_size-1] = vals[j]; 1181019a44ceSStefano Zampini } 1182019a44ceSStefano Zampini ierr = MatRestoreRow(B0_V,0,&n,&idxs,&vals);CHKERRQ(ierr); 118316f15bc4SStefano Zampini } 118421eccb56SStefano Zampini ierr = MatDestroy(&A_VV);CHKERRQ(ierr); 1185d16cbb6bSStefano Zampini 118606656605SStefano Zampini /* coarse basis functions */ 118706656605SStefano Zampini for (i=0;i<n_vertices;i++) { 118816f15bc4SStefano Zampini PetscScalar *y; 118916f15bc4SStefano Zampini 119006656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+n_R*i);CHKERRQ(ierr); 119106656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 119206656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*i);CHKERRQ(ierr); 119306656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 119406656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 119506656605SStefano Zampini y[n_B*i+idx_V_B[i]] = 1.0; 119606656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 119706656605SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 119806656605SStefano Zampini 119906656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 120006656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 120106656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*i);CHKERRQ(ierr); 120206656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 120306656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 120406656605SStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 1205a0aff763SStefano Zampini if (n_benign) y[n_D*i+p0_lidx_I] = 0.0; 120606656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 120706656605SStefano Zampini } 120806656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 120906656605SStefano Zampini } 121004708bb6SStefano Zampini /* if n_R == 0 the object is not destroyed */ 121104708bb6SStefano Zampini ierr = MatDestroy(&A_RV);CHKERRQ(ierr); 121206656605SStefano Zampini } 121306656605SStefano Zampini 121406656605SStefano Zampini if (n_constraints) { 121506656605SStefano Zampini Mat B; 121606656605SStefano Zampini 121706656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_constraints,work,&B);CHKERRQ(ierr); 121806656605SStefano Zampini ierr = MatScale(S_CC,m_one);CHKERRQ(ierr); 121980677318SStefano Zampini ierr = MatMatMult(local_auxmat2_R,S_CC,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B);CHKERRQ(ierr); 122006656605SStefano Zampini ierr = MatScale(S_CC,m_one);CHKERRQ(ierr); 122106656605SStefano Zampini if (n_vertices) { 122280677318SStefano Zampini if (isCHOL) { /* if we can solve the interior problem with cholesky, we should also be fine with transposing here */ 122380677318SStefano Zampini ierr = MatTranspose(S_CV,MAT_REUSE_MATRIX,&S_VC);CHKERRQ(ierr); 122480677318SStefano Zampini } else { 122580677318SStefano Zampini Mat S_VCt; 122680677318SStefano Zampini 122780677318SStefano Zampini ierr = MatMatMult(A_VR,B,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&S_VCt);CHKERRQ(ierr); 122880677318SStefano Zampini ierr = MatCopy(S_VCt,S_VC,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 122980677318SStefano Zampini ierr = MatDestroy(&S_VCt);CHKERRQ(ierr); 123080677318SStefano Zampini } 123106656605SStefano Zampini } 123206656605SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 123306656605SStefano Zampini /* coarse basis functions */ 123406656605SStefano Zampini for (i=0;i<n_constraints;i++) { 123506656605SStefano Zampini PetscScalar *y; 123606656605SStefano Zampini 123706656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+n_R*i);CHKERRQ(ierr); 123806656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 123906656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*(i+n_vertices));CHKERRQ(ierr); 124006656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 124106656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 124206656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 124306656605SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 124406656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 124506656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 124606656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*(i+n_vertices));CHKERRQ(ierr); 124706656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 124806656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 124906656605SStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 1250d16cbb6bSStefano Zampini if (n_benign) y[n_D*(i+n_vertices)+p0_lidx_I] = 0.0; 125106656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 125206656605SStefano Zampini } 125306656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 125406656605SStefano Zampini } 125506656605SStefano Zampini } 125680677318SStefano Zampini if (n_constraints) { 125780677318SStefano Zampini ierr = MatDestroy(&local_auxmat2_R);CHKERRQ(ierr); 125880677318SStefano Zampini } 125906656605SStefano Zampini 1260019a44ceSStefano Zampini ierr = MatDestroy(&B0_V);CHKERRQ(ierr); 1261019a44ceSStefano Zampini 126206656605SStefano Zampini /* compute other basis functions for non-symmetric problems */ 12633301b35fSStefano Zampini if (!pcbddc->symmetric_primal) { 126406656605SStefano Zampini 126506656605SStefano Zampini if (n_constraints) { 126616f15bc4SStefano Zampini Mat S_CCT,B_C; 126706656605SStefano Zampini 126880677318SStefano Zampini /* this is a lazy thing */ 126980677318SStefano Zampini ierr = MatConvert(C_CR,impMatType,MAT_REUSE_MATRIX,&C_CR);CHKERRQ(ierr); 127006656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_constraints,work+n_vertices*n_R,&B_C);CHKERRQ(ierr); 127106656605SStefano Zampini ierr = MatTranspose(S_CC,MAT_INITIAL_MATRIX,&S_CCT);CHKERRQ(ierr); 127206656605SStefano Zampini ierr = MatTransposeMatMult(C_CR,S_CCT,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B_C);CHKERRQ(ierr); 127316f15bc4SStefano Zampini ierr = MatDestroy(&S_CCT);CHKERRQ(ierr); 127406656605SStefano Zampini if (n_vertices) { 127516f15bc4SStefano Zampini Mat B_V,S_VCT; 127606656605SStefano Zampini 127706656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_vertices,work,&B_V);CHKERRQ(ierr); 127806656605SStefano Zampini ierr = MatTranspose(S_VC,MAT_INITIAL_MATRIX,&S_VCT);CHKERRQ(ierr); 127906656605SStefano Zampini ierr = MatTransposeMatMult(C_CR,S_VCT,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B_V);CHKERRQ(ierr); 128006656605SStefano Zampini ierr = MatDestroy(&B_V);CHKERRQ(ierr); 128116f15bc4SStefano Zampini ierr = MatDestroy(&S_VCT);CHKERRQ(ierr); 128206656605SStefano Zampini } 128306656605SStefano Zampini ierr = MatDestroy(&B_C);CHKERRQ(ierr); 128480677318SStefano Zampini } else { /* if there are no constraints, reset work */ 128580677318SStefano Zampini ierr = PetscMemzero(work,n_R*pcbddc->local_primal_size*sizeof(PetscScalar));CHKERRQ(ierr); 128606656605SStefano Zampini } 128716f15bc4SStefano Zampini if (n_vertices && n_R) { 128806656605SStefano Zampini Mat A_VRT; 128980677318SStefano Zampini PetscScalar *marray; 129006656605SStefano Zampini PetscBLASInt B_N,B_one = 1; 129106656605SStefano Zampini 129280677318SStefano Zampini ierr = MatTranspose(A_VR,MAT_INITIAL_MATRIX,&A_VRT);CHKERRQ(ierr); 129380677318SStefano Zampini ierr = MatConvert(A_VRT,impMatType,MAT_REUSE_MATRIX,&A_VRT);CHKERRQ(ierr); 129480677318SStefano Zampini ierr = MatDenseGetArray(A_VRT,&marray);CHKERRQ(ierr); 129506656605SStefano Zampini ierr = PetscBLASIntCast(n_vertices*n_R,&B_N);CHKERRQ(ierr); 129680677318SStefano Zampini PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&B_N,&m_one,marray,&B_one,work,&B_one)); 129780677318SStefano Zampini ierr = MatDenseRestoreArray(A_VRT,&marray);CHKERRQ(ierr); 129816f15bc4SStefano Zampini ierr = MatDestroy(&A_VRT);CHKERRQ(ierr); 129906656605SStefano Zampini } 130006656605SStefano Zampini 130106656605SStefano Zampini if (F) { /* currently there's no support for MatTransposeMatSolve(F,B,X) */ 130206656605SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 130306656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*n_R);CHKERRQ(ierr); 130406656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,work+(i+pcbddc->local_primal_size)*n_R);CHKERRQ(ierr); 130506656605SStefano Zampini ierr = MatSolveTranspose(F,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 130606656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 130706656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 130806656605SStefano Zampini } 130906656605SStefano Zampini } else { 131006656605SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 131106656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*n_R);CHKERRQ(ierr); 131206656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,work+(i+pcbddc->local_primal_size)*n_R);CHKERRQ(ierr); 131306656605SStefano Zampini ierr = KSPSolveTranspose(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 131406656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 131506656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 131606656605SStefano Zampini } 131706656605SStefano Zampini } 131806656605SStefano Zampini /* coarse basis functions */ 131906656605SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 132006656605SStefano Zampini PetscScalar *y; 132106656605SStefano Zampini 132206656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+n_R*(i+pcbddc->local_primal_size));CHKERRQ(ierr); 132306656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_psi_B,&y);CHKERRQ(ierr); 132406656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*i);CHKERRQ(ierr); 132506656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 132606656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 132706656605SStefano Zampini if (i<n_vertices) { 132806656605SStefano Zampini y[n_B*i+idx_V_B[i]] = 1.0; 132906656605SStefano Zampini } 133006656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_psi_B,&y);CHKERRQ(ierr); 133106656605SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 133206656605SStefano Zampini 133306656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 133406656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_psi_D,&y);CHKERRQ(ierr); 133506656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*i);CHKERRQ(ierr); 133606656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 133706656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 133806656605SStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 133906656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_psi_D,&y);CHKERRQ(ierr); 134006656605SStefano Zampini } 134106656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 134206656605SStefano Zampini } 134306656605SStefano Zampini } 1344d62866d3SStefano Zampini /* free memory */ 134588ebb749SStefano Zampini ierr = PetscFree(idx_V_B);CHKERRQ(ierr); 134606656605SStefano Zampini ierr = MatDestroy(&S_VV);CHKERRQ(ierr); 134706656605SStefano Zampini ierr = MatDestroy(&S_CV);CHKERRQ(ierr); 134806656605SStefano Zampini ierr = MatDestroy(&S_VC);CHKERRQ(ierr); 134906656605SStefano Zampini ierr = MatDestroy(&S_CC);CHKERRQ(ierr); 1350d62866d3SStefano Zampini ierr = PetscFree(work);CHKERRQ(ierr); 1351d62866d3SStefano Zampini if (n_vertices) { 1352d62866d3SStefano Zampini ierr = MatDestroy(&A_VR);CHKERRQ(ierr); 1353d62866d3SStefano Zampini } 1354d62866d3SStefano Zampini if (n_constraints) { 1355d62866d3SStefano Zampini ierr = MatDestroy(&C_CR);CHKERRQ(ierr); 1356d62866d3SStefano Zampini } 135788ebb749SStefano Zampini /* Checking coarse_sub_mat and coarse basis functios */ 135888ebb749SStefano Zampini /* Symmetric case : It should be \Phi^{(j)^T} A^{(j)} \Phi^{(j)}=coarse_sub_mat */ 135988ebb749SStefano Zampini /* Non-symmetric case : It should be \Psi^{(j)^T} A^{(j)} \Phi^{(j)}=coarse_sub_mat */ 1360d12edf2fSStefano Zampini if (pcbddc->dbg_flag) { 136188ebb749SStefano Zampini Mat coarse_sub_mat; 136225084f0cSStefano Zampini Mat AUXMAT,TM1,TM2,TM3,TM4; 136388ebb749SStefano Zampini Mat coarse_phi_D,coarse_phi_B; 136488ebb749SStefano Zampini Mat coarse_psi_D,coarse_psi_B; 136588ebb749SStefano Zampini Mat A_II,A_BB,A_IB,A_BI; 13668bec7fa6SStefano Zampini Mat C_B,CPHI; 13678bec7fa6SStefano Zampini IS is_dummy; 13688bec7fa6SStefano Zampini Vec mones; 136988ebb749SStefano Zampini MatType checkmattype=MATSEQAIJ; 137088ebb749SStefano Zampini PetscReal real_value; 137188ebb749SStefano Zampini 137288ebb749SStefano Zampini ierr = MatConvert(pcis->A_II,checkmattype,MAT_INITIAL_MATRIX,&A_II);CHKERRQ(ierr); 137388ebb749SStefano Zampini ierr = MatConvert(pcis->A_IB,checkmattype,MAT_INITIAL_MATRIX,&A_IB);CHKERRQ(ierr); 137488ebb749SStefano Zampini ierr = MatConvert(pcis->A_BI,checkmattype,MAT_INITIAL_MATRIX,&A_BI);CHKERRQ(ierr); 137588ebb749SStefano Zampini ierr = MatConvert(pcis->A_BB,checkmattype,MAT_INITIAL_MATRIX,&A_BB);CHKERRQ(ierr); 137688ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_phi_D,checkmattype,MAT_INITIAL_MATRIX,&coarse_phi_D);CHKERRQ(ierr); 137788ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_phi_B,checkmattype,MAT_INITIAL_MATRIX,&coarse_phi_B);CHKERRQ(ierr); 1378c0553b1fSStefano Zampini if (unsymmetric_check) { 137988ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_psi_D,checkmattype,MAT_INITIAL_MATRIX,&coarse_psi_D);CHKERRQ(ierr); 138088ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_psi_B,checkmattype,MAT_INITIAL_MATRIX,&coarse_psi_B);CHKERRQ(ierr); 138188ebb749SStefano Zampini } 138288ebb749SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,pcbddc->local_primal_size,pcbddc->local_primal_size,coarse_submat_vals,&coarse_sub_mat);CHKERRQ(ierr); 138388ebb749SStefano Zampini 138425084f0cSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 13853301b35fSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Check coarse sub mat computation (symmetric %d)\n",pcbddc->symmetric_primal);CHKERRQ(ierr); 138625084f0cSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1387c0553b1fSStefano Zampini if (unsymmetric_check) { 138888ebb749SStefano Zampini ierr = MatMatMult(A_II,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 138988ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM1);CHKERRQ(ierr); 139088ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 139188ebb749SStefano Zampini ierr = MatMatMult(A_BB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 139288ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM2);CHKERRQ(ierr); 139388ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 139488ebb749SStefano Zampini ierr = MatMatMult(A_IB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 139588ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM3);CHKERRQ(ierr); 139688ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 139788ebb749SStefano Zampini ierr = MatMatMult(A_BI,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 139888ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM4);CHKERRQ(ierr); 139988ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 140088ebb749SStefano Zampini } else { 140188ebb749SStefano Zampini ierr = MatPtAP(A_II,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&TM1);CHKERRQ(ierr); 140288ebb749SStefano Zampini ierr = MatPtAP(A_BB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&TM2);CHKERRQ(ierr); 140388ebb749SStefano Zampini ierr = MatMatMult(A_IB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 140488ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_phi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM3);CHKERRQ(ierr); 140588ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 140688ebb749SStefano Zampini ierr = MatMatMult(A_BI,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 140788ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_phi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM4);CHKERRQ(ierr); 140888ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 140988ebb749SStefano Zampini } 141088ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM2,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 141188ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM3,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 141288ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM4,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 141388ebb749SStefano Zampini ierr = MatConvert(TM1,MATSEQDENSE,MAT_REUSE_MATRIX,&TM1);CHKERRQ(ierr); 1414d12edf2fSStefano Zampini if (n_benign) { 1415d12edf2fSStefano Zampini Mat B0_I,B0_B,B0_BPHI,B0_IPHI; 1416d12edf2fSStefano Zampini PetscScalar *data,*data2; 1417d12edf2fSStefano Zampini 1418d12edf2fSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,1,0,1,&is_dummy);CHKERRQ(ierr); 1419d12edf2fSStefano Zampini ierr = MatGetSubMatrix(B0,is_dummy,pcis->is_B_local,MAT_INITIAL_MATRIX,&B0_B); 1420d12edf2fSStefano Zampini ierr = MatGetSubMatrix(B0,is_dummy,pcis->is_I_local,MAT_INITIAL_MATRIX,&B0_I); 1421d12edf2fSStefano Zampini ierr = MatMatMult(B0_B,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&B0_BPHI);CHKERRQ(ierr); 1422d12edf2fSStefano Zampini ierr = MatConvert(B0_BPHI,MATSEQDENSE,MAT_REUSE_MATRIX,&B0_BPHI);CHKERRQ(ierr); 1423d12edf2fSStefano Zampini ierr = MatDenseGetArray(TM1,&data);CHKERRQ(ierr); 1424d12edf2fSStefano Zampini ierr = MatDenseGetArray(B0_BPHI,&data2);CHKERRQ(ierr); 1425d12edf2fSStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 1426d16cbb6bSStefano Zampini data[(pcbddc->local_primal_size-1)*pcbddc->local_primal_size+i] += data2[i]; 1427d16cbb6bSStefano Zampini data[(i+1)*pcbddc->local_primal_size-1] += data2[i]; 1428d12edf2fSStefano Zampini } 1429d12edf2fSStefano Zampini ierr = MatDenseRestoreArray(TM1,&data);CHKERRQ(ierr); 1430d12edf2fSStefano Zampini ierr = MatDenseRestoreArray(B0_BPHI,&data2);CHKERRQ(ierr); 1431d12edf2fSStefano Zampini ierr = MatDestroy(&B0_B);CHKERRQ(ierr); 1432d12edf2fSStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 1433d12edf2fSStefano Zampini ierr = MatDestroy(&B0_BPHI);CHKERRQ(ierr); 1434d12edf2fSStefano Zampini ierr = MatMatMult(B0_I,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&B0_IPHI);CHKERRQ(ierr); 1435d12edf2fSStefano Zampini ierr = MatDestroy(&B0_I);CHKERRQ(ierr); 1436d12edf2fSStefano Zampini ierr = MatNorm(B0_IPHI,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 1437d12edf2fSStefano Zampini ierr = MatDestroy(&B0_IPHI);CHKERRQ(ierr); 1438d12edf2fSStefano Zampini } 1439d12edf2fSStefano Zampini #if 0 1440d12edf2fSStefano Zampini { 1441d12edf2fSStefano Zampini PetscViewer viewer; 1442d12edf2fSStefano Zampini char filename[256]; 1443d12edf2fSStefano Zampini sprintf(filename,"proj_local_coarse_mat%d.m",PetscGlobalRank); 1444d12edf2fSStefano Zampini ierr = PetscViewerASCIIOpen(PETSC_COMM_SELF,filename,&viewer);CHKERRQ(ierr); 1445d12edf2fSStefano Zampini ierr = PetscViewerSetFormat(viewer,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 1446d12edf2fSStefano Zampini ierr = MatView(TM1,viewer);CHKERRQ(ierr); 1447d12edf2fSStefano Zampini ierr = PetscViewerDestroy(&viewer);CHKERRQ(ierr); 1448d12edf2fSStefano Zampini } 1449d12edf2fSStefano Zampini #endif 145081d9aea3SBarry Smith ierr = MatAXPY(TM1,m_one,coarse_sub_mat,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 14518bec7fa6SStefano Zampini ierr = MatNorm(TM1,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 14520fccc4e9SStefano Zampini ierr = PetscViewerASCIISynchronizedAllow(pcbddc->dbg_viewer,PETSC_TRUE);CHKERRQ(ierr); 145306656605SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d matrix error % 1.14e\n",PetscGlobalRank,real_value);CHKERRQ(ierr); 14548bec7fa6SStefano Zampini 14558bec7fa6SStefano Zampini /* check constraints */ 1456d12edf2fSStefano Zampini if (!n_benign) { 14578bec7fa6SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcbddc->local_primal_size,0,1,&is_dummy);CHKERRQ(ierr); 14588bec7fa6SStefano Zampini ierr = MatGetSubMatrix(pcbddc->ConstraintMatrix,is_dummy,pcis->is_B_local,MAT_INITIAL_MATRIX,&C_B); 14598bec7fa6SStefano Zampini ierr = MatMatMult(C_B,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&CPHI);CHKERRQ(ierr); 14608bec7fa6SStefano Zampini ierr = MatCreateVecs(CPHI,&mones,NULL);CHKERRQ(ierr); 14618bec7fa6SStefano Zampini ierr = VecSet(mones,-1.0);CHKERRQ(ierr); 14628bec7fa6SStefano Zampini ierr = MatDiagonalSet(CPHI,mones,ADD_VALUES);CHKERRQ(ierr); 14638bec7fa6SStefano Zampini ierr = MatNorm(CPHI,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 1464bdae7319SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d phi constraints error % 1.14e\n",PetscGlobalRank,real_value);CHKERRQ(ierr); 1465c0553b1fSStefano Zampini if (unsymmetric_check) { 1466bdae7319SStefano Zampini ierr = MatMatMult(C_B,coarse_psi_B,MAT_REUSE_MATRIX,1.0,&CPHI);CHKERRQ(ierr); 1467bdae7319SStefano Zampini ierr = VecSet(mones,-1.0);CHKERRQ(ierr); 1468bdae7319SStefano Zampini ierr = MatDiagonalSet(CPHI,mones,ADD_VALUES);CHKERRQ(ierr); 1469bdae7319SStefano Zampini ierr = MatNorm(CPHI,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 1470bdae7319SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d psi constraints error % 1.14e\n",PetscGlobalRank,real_value);CHKERRQ(ierr); 147188ebb749SStefano Zampini } 14728bec7fa6SStefano Zampini ierr = MatDestroy(&C_B);CHKERRQ(ierr); 14738bec7fa6SStefano Zampini ierr = MatDestroy(&CPHI);CHKERRQ(ierr); 14748bec7fa6SStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 14758bec7fa6SStefano Zampini ierr = VecDestroy(&mones);CHKERRQ(ierr); 1476d12edf2fSStefano Zampini } 147725084f0cSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 147888ebb749SStefano Zampini ierr = MatDestroy(&A_II);CHKERRQ(ierr); 147988ebb749SStefano Zampini ierr = MatDestroy(&A_BB);CHKERRQ(ierr); 148088ebb749SStefano Zampini ierr = MatDestroy(&A_IB);CHKERRQ(ierr); 148188ebb749SStefano Zampini ierr = MatDestroy(&A_BI);CHKERRQ(ierr); 148288ebb749SStefano Zampini ierr = MatDestroy(&TM1);CHKERRQ(ierr); 148388ebb749SStefano Zampini ierr = MatDestroy(&TM2);CHKERRQ(ierr); 148488ebb749SStefano Zampini ierr = MatDestroy(&TM3);CHKERRQ(ierr); 148588ebb749SStefano Zampini ierr = MatDestroy(&TM4);CHKERRQ(ierr); 148688ebb749SStefano Zampini ierr = MatDestroy(&coarse_phi_D);CHKERRQ(ierr); 148788ebb749SStefano Zampini ierr = MatDestroy(&coarse_phi_B);CHKERRQ(ierr); 1488c0553b1fSStefano Zampini if (unsymmetric_check) { 148988ebb749SStefano Zampini ierr = MatDestroy(&coarse_psi_D);CHKERRQ(ierr); 149088ebb749SStefano Zampini ierr = MatDestroy(&coarse_psi_B);CHKERRQ(ierr); 149188ebb749SStefano Zampini } 149288ebb749SStefano Zampini ierr = MatDestroy(&coarse_sub_mat);CHKERRQ(ierr); 149388ebb749SStefano Zampini } 1494d12edf2fSStefano Zampini ierr = MatDestroy(&B0);CHKERRQ(ierr); 14958629588bSStefano Zampini /* get back data */ 14968629588bSStefano Zampini *coarse_submat_vals_n = coarse_submat_vals; 149788ebb749SStefano Zampini PetscFunctionReturn(0); 149888ebb749SStefano Zampini } 149988ebb749SStefano Zampini 150088ebb749SStefano Zampini #undef __FUNCT__ 1501d65f70fdSStefano Zampini #define __FUNCT__ "MatGetSubMatrixUnsorted" 1502d648f858SStefano Zampini PetscErrorCode MatGetSubMatrixUnsorted(Mat A, IS isrow, IS iscol, Mat* B) 1503aa0d41d4SStefano Zampini { 1504d65f70fdSStefano Zampini Mat *work_mat; 1505d65f70fdSStefano Zampini IS isrow_s,iscol_s; 1506d65f70fdSStefano Zampini PetscBool rsorted,csorted; 1507d65f70fdSStefano Zampini PetscInt rsize,*idxs_perm_r,csize,*idxs_perm_c; 1508aa0d41d4SStefano Zampini PetscErrorCode ierr; 1509aa0d41d4SStefano Zampini 1510aa0d41d4SStefano Zampini PetscFunctionBegin; 1511d65f70fdSStefano Zampini ierr = ISSorted(isrow,&rsorted);CHKERRQ(ierr); 1512d65f70fdSStefano Zampini ierr = ISSorted(iscol,&csorted);CHKERRQ(ierr); 1513d65f70fdSStefano Zampini ierr = ISGetLocalSize(isrow,&rsize);CHKERRQ(ierr); 1514d65f70fdSStefano Zampini ierr = ISGetLocalSize(iscol,&csize);CHKERRQ(ierr); 1515aa0d41d4SStefano Zampini 1516d65f70fdSStefano Zampini if (!rsorted) { 1517906d46d4SStefano Zampini const PetscInt *idxs; 1518906d46d4SStefano Zampini PetscInt *idxs_sorted,i; 1519aa0d41d4SStefano Zampini 1520d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_perm_r);CHKERRQ(ierr); 1521d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_sorted);CHKERRQ(ierr); 1522d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 1523d65f70fdSStefano Zampini idxs_perm_r[i] = i; 1524aa0d41d4SStefano Zampini } 1525d65f70fdSStefano Zampini ierr = ISGetIndices(isrow,&idxs);CHKERRQ(ierr); 1526d65f70fdSStefano Zampini ierr = PetscSortIntWithPermutation(rsize,idxs,idxs_perm_r);CHKERRQ(ierr); 1527d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 1528d65f70fdSStefano Zampini idxs_sorted[i] = idxs[idxs_perm_r[i]]; 1529aa0d41d4SStefano Zampini } 1530d65f70fdSStefano Zampini ierr = ISRestoreIndices(isrow,&idxs);CHKERRQ(ierr); 1531d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,rsize,idxs_sorted,PETSC_OWN_POINTER,&isrow_s);CHKERRQ(ierr); 1532d65f70fdSStefano Zampini } else { 1533d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)isrow);CHKERRQ(ierr); 1534d65f70fdSStefano Zampini isrow_s = isrow; 1535aa0d41d4SStefano Zampini } 1536906d46d4SStefano Zampini 1537d65f70fdSStefano Zampini if (!csorted) { 1538d65f70fdSStefano Zampini if (isrow == iscol) { 1539d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)isrow_s);CHKERRQ(ierr); 1540d65f70fdSStefano Zampini iscol_s = isrow_s; 1541d65f70fdSStefano Zampini } else { 1542d65f70fdSStefano Zampini const PetscInt *idxs; 1543d65f70fdSStefano Zampini PetscInt *idxs_sorted,i; 1544906d46d4SStefano Zampini 1545d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_perm_c);CHKERRQ(ierr); 1546d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_sorted);CHKERRQ(ierr); 1547d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 1548d65f70fdSStefano Zampini idxs_perm_c[i] = i; 1549d65f70fdSStefano Zampini } 1550d65f70fdSStefano Zampini ierr = ISGetIndices(iscol,&idxs);CHKERRQ(ierr); 1551d65f70fdSStefano Zampini ierr = PetscSortIntWithPermutation(csize,idxs,idxs_perm_c);CHKERRQ(ierr); 1552d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 1553d65f70fdSStefano Zampini idxs_sorted[i] = idxs[idxs_perm_c[i]]; 1554d65f70fdSStefano Zampini } 1555d65f70fdSStefano Zampini ierr = ISRestoreIndices(iscol,&idxs);CHKERRQ(ierr); 1556d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,csize,idxs_sorted,PETSC_OWN_POINTER,&iscol_s);CHKERRQ(ierr); 1557d65f70fdSStefano Zampini } 1558d65f70fdSStefano Zampini } else { 1559d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)iscol);CHKERRQ(ierr); 1560d65f70fdSStefano Zampini iscol_s = iscol; 1561d65f70fdSStefano Zampini } 1562d65f70fdSStefano Zampini 1563d648f858SStefano Zampini ierr = MatGetSubMatrices(A,1,&isrow_s,&iscol_s,MAT_INITIAL_MATRIX,&work_mat);CHKERRQ(ierr); 1564d65f70fdSStefano Zampini 1565d65f70fdSStefano Zampini if (!rsorted || !csorted) { 1566906d46d4SStefano Zampini Mat new_mat; 1567d65f70fdSStefano Zampini IS is_perm_r,is_perm_c; 1568906d46d4SStefano Zampini 1569d65f70fdSStefano Zampini if (!rsorted) { 1570d65f70fdSStefano Zampini PetscInt *idxs_r,i; 1571d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_r);CHKERRQ(ierr); 1572d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 1573d65f70fdSStefano Zampini idxs_r[idxs_perm_r[i]] = i; 1574906d46d4SStefano Zampini } 1575d65f70fdSStefano Zampini ierr = PetscFree(idxs_perm_r);CHKERRQ(ierr); 1576d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,rsize,idxs_r,PETSC_OWN_POINTER,&is_perm_r);CHKERRQ(ierr); 1577d65f70fdSStefano Zampini } else { 1578d65f70fdSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,rsize,0,1,&is_perm_r);CHKERRQ(ierr); 1579906d46d4SStefano Zampini } 1580d65f70fdSStefano Zampini ierr = ISSetPermutation(is_perm_r);CHKERRQ(ierr); 1581d65f70fdSStefano Zampini 1582d65f70fdSStefano Zampini if (!csorted) { 1583d65f70fdSStefano Zampini if (isrow_s == iscol_s) { 1584d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)is_perm_r);CHKERRQ(ierr); 1585d65f70fdSStefano Zampini is_perm_c = is_perm_r; 1586d65f70fdSStefano Zampini } else { 1587d65f70fdSStefano Zampini PetscInt *idxs_c,i; 1588d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_c);CHKERRQ(ierr); 1589d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 1590d65f70fdSStefano Zampini idxs_c[idxs_perm_c[i]] = i; 1591d65f70fdSStefano Zampini } 1592d65f70fdSStefano Zampini ierr = PetscFree(idxs_perm_c);CHKERRQ(ierr); 1593d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,csize,idxs_c,PETSC_OWN_POINTER,&is_perm_c);CHKERRQ(ierr); 1594d65f70fdSStefano Zampini } 1595d65f70fdSStefano Zampini } else { 1596d65f70fdSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,csize,0,1,&is_perm_c);CHKERRQ(ierr); 1597d65f70fdSStefano Zampini } 1598d65f70fdSStefano Zampini ierr = ISSetPermutation(is_perm_c);CHKERRQ(ierr); 1599d65f70fdSStefano Zampini 1600d65f70fdSStefano Zampini ierr = MatPermute(work_mat[0],is_perm_r,is_perm_c,&new_mat);CHKERRQ(ierr); 1601d65f70fdSStefano Zampini ierr = MatDestroy(&work_mat[0]);CHKERRQ(ierr); 1602d65f70fdSStefano Zampini work_mat[0] = new_mat; 1603d65f70fdSStefano Zampini ierr = ISDestroy(&is_perm_r);CHKERRQ(ierr); 1604d65f70fdSStefano Zampini ierr = ISDestroy(&is_perm_c);CHKERRQ(ierr); 1605d65f70fdSStefano Zampini } 1606d65f70fdSStefano Zampini 1607d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)work_mat[0]);CHKERRQ(ierr); 1608d65f70fdSStefano Zampini *B = work_mat[0]; 1609d65f70fdSStefano Zampini ierr = MatDestroyMatrices(1,&work_mat);CHKERRQ(ierr); 1610d65f70fdSStefano Zampini ierr = ISDestroy(&isrow_s);CHKERRQ(ierr); 1611d65f70fdSStefano Zampini ierr = ISDestroy(&iscol_s);CHKERRQ(ierr); 1612d65f70fdSStefano Zampini PetscFunctionReturn(0); 1613d65f70fdSStefano Zampini } 1614d65f70fdSStefano Zampini 1615d65f70fdSStefano Zampini #undef __FUNCT__ 16165e8657edSStefano Zampini #define __FUNCT__ "PCBDDCComputeLocalMatrix" 16175e8657edSStefano Zampini PetscErrorCode PCBDDCComputeLocalMatrix(PC pc, Mat ChangeOfBasisMatrix) 1618aa0d41d4SStefano Zampini { 1619aa0d41d4SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 16205e8657edSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 1621d65f70fdSStefano Zampini Mat new_mat; 16225e8657edSStefano Zampini IS is_local,is_global; 1623d65f70fdSStefano Zampini PetscInt local_size; 1624d65f70fdSStefano Zampini PetscBool isseqaij; 1625aa0d41d4SStefano Zampini PetscErrorCode ierr; 1626aa0d41d4SStefano Zampini 1627aa0d41d4SStefano Zampini PetscFunctionBegin; 1628aa0d41d4SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 16295e8657edSStefano Zampini ierr = MatGetSize(matis->A,&local_size,NULL);CHKERRQ(ierr); 16305e8657edSStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)matis->A),local_size,0,1,&is_local);CHKERRQ(ierr); 1631b087196eSStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(pc->pmat->rmap->mapping,is_local,&is_global);CHKERRQ(ierr); 1632aa0d41d4SStefano Zampini ierr = ISDestroy(&is_local);CHKERRQ(ierr); 1633d648f858SStefano Zampini ierr = MatGetSubMatrixUnsorted(ChangeOfBasisMatrix,is_global,is_global,&new_mat);CHKERRQ(ierr); 1634aa0d41d4SStefano Zampini ierr = ISDestroy(&is_global);CHKERRQ(ierr); 1635906d46d4SStefano Zampini 1636906d46d4SStefano Zampini /* check */ 1637906d46d4SStefano Zampini if (pcbddc->dbg_flag) { 1638906d46d4SStefano Zampini Vec x,x_change; 1639906d46d4SStefano Zampini PetscReal error; 1640906d46d4SStefano Zampini 16415e8657edSStefano Zampini ierr = MatCreateVecs(ChangeOfBasisMatrix,&x,&x_change);CHKERRQ(ierr); 1642906d46d4SStefano Zampini ierr = VecSetRandom(x,NULL);CHKERRQ(ierr); 16435e8657edSStefano Zampini ierr = MatMult(ChangeOfBasisMatrix,x,x_change);CHKERRQ(ierr); 1644e176bc59SStefano Zampini ierr = VecScatterBegin(matis->cctx,x,matis->x,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1645e176bc59SStefano Zampini ierr = VecScatterEnd(matis->cctx,x,matis->x,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1646d65f70fdSStefano Zampini ierr = MatMult(new_mat,matis->x,matis->y);CHKERRQ(ierr); 1647e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,matis->y,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1648e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,matis->y,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1649906d46d4SStefano Zampini ierr = VecAXPY(x,-1.0,x_change);CHKERRQ(ierr); 1650906d46d4SStefano Zampini ierr = VecNorm(x,NORM_INFINITY,&error);CHKERRQ(ierr); 1651906d46d4SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1652906d46d4SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Error global vs local change on N: %1.6e\n",error);CHKERRQ(ierr); 1653906d46d4SStefano Zampini ierr = VecDestroy(&x);CHKERRQ(ierr); 1654906d46d4SStefano Zampini ierr = VecDestroy(&x_change);CHKERRQ(ierr); 1655906d46d4SStefano Zampini } 1656906d46d4SStefano Zampini 165722d5777bSStefano Zampini /* TODO: HOW TO WORK WITH BAIJ and SBAIJ and SEQDENSE? */ 16589b28b3ffSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)matis->A,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 165922d5777bSStefano Zampini if (isseqaij) { 1660d65f70fdSStefano Zampini ierr = MatPtAP(matis->A,new_mat,MAT_INITIAL_MATRIX,2.0,&pcbddc->local_mat);CHKERRQ(ierr); 1661aa0d41d4SStefano Zampini } else { 1662aa0d41d4SStefano Zampini Mat work_mat; 1663aa0d41d4SStefano Zampini ierr = MatConvert(matis->A,MATSEQAIJ,MAT_INITIAL_MATRIX,&work_mat);CHKERRQ(ierr); 1664d65f70fdSStefano Zampini ierr = MatPtAP(work_mat,new_mat,MAT_INITIAL_MATRIX,2.0,&pcbddc->local_mat);CHKERRQ(ierr); 1665aa0d41d4SStefano Zampini ierr = MatDestroy(&work_mat);CHKERRQ(ierr); 1666aa0d41d4SStefano Zampini } 16673301b35fSStefano Zampini if (matis->A->symmetric_set) { 16683301b35fSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_SYMMETRIC,matis->A->symmetric);CHKERRQ(ierr); 1669e496cd5dSStefano Zampini #if !defined(PETSC_USE_COMPLEX) 16703301b35fSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_HERMITIAN,matis->A->symmetric);CHKERRQ(ierr); 1671e496cd5dSStefano Zampini #endif 16723301b35fSStefano Zampini } 167345a1bb75SStefano Zampini /* 167445a1bb75SStefano Zampini ierr = PetscViewerSetFormat(PETSC_VIEWER_STDOUT_SELF,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 1675d65f70fdSStefano Zampini ierr = MatView(new_mat,(PetscViewer)0);CHKERRQ(ierr); 167645a1bb75SStefano Zampini */ 1677d65f70fdSStefano Zampini ierr = MatDestroy(&new_mat);CHKERRQ(ierr); 1678aa0d41d4SStefano Zampini PetscFunctionReturn(0); 1679aa0d41d4SStefano Zampini } 1680aa0d41d4SStefano Zampini 1681aa0d41d4SStefano Zampini #undef __FUNCT__ 1682a64d13efSStefano Zampini #define __FUNCT__ "PCBDDCSetUpLocalScatters" 16838ce42a96SStefano Zampini PetscErrorCode PCBDDCSetUpLocalScatters(PC pc) 1684a64d13efSStefano Zampini { 1685a64d13efSStefano Zampini PC_IS* pcis = (PC_IS*)(pc->data); 1686a64d13efSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 1687d62866d3SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 168853892102SStefano Zampini PetscInt *idx_R_local=NULL; 16893a50541eSStefano Zampini PetscInt n_vertices,i,j,n_R,n_D,n_B; 16903a50541eSStefano Zampini PetscInt vbs,bs; 16916816873aSStefano Zampini PetscBT bitmask=NULL; 1692a64d13efSStefano Zampini PetscErrorCode ierr; 1693a64d13efSStefano Zampini 1694a64d13efSStefano Zampini PetscFunctionBegin; 1695b23d619eSStefano Zampini /* 1696b23d619eSStefano Zampini No need to setup local scatters if 1697b23d619eSStefano Zampini - primal space is unchanged 1698b23d619eSStefano Zampini AND 1699b23d619eSStefano Zampini - we actually have locally some primal dofs (could not be true in multilevel or for isolated subdomains) 1700b23d619eSStefano Zampini AND 1701b23d619eSStefano Zampini - we are not in debugging mode (this is needed since there are Synchronized prints at the end of the subroutine 1702b23d619eSStefano Zampini */ 1703b23d619eSStefano Zampini if (!pcbddc->new_primal_space_local && pcbddc->local_primal_size && !pcbddc->dbg_flag) { 1704f4ddd8eeSStefano Zampini PetscFunctionReturn(0); 1705f4ddd8eeSStefano Zampini } 1706f4ddd8eeSStefano Zampini /* destroy old objects */ 1707f4ddd8eeSStefano Zampini ierr = ISDestroy(&pcbddc->is_R_local);CHKERRQ(ierr); 1708f4ddd8eeSStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_B);CHKERRQ(ierr); 1709f4ddd8eeSStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_D);CHKERRQ(ierr); 1710a64d13efSStefano Zampini /* Set Non-overlapping dimensions */ 1711b371cd4fSStefano Zampini n_B = pcis->n_B; 1712b371cd4fSStefano Zampini n_D = pcis->n - n_B; 1713b371cd4fSStefano Zampini n_vertices = pcbddc->n_vertices; 17143a50541eSStefano Zampini 1715a64d13efSStefano Zampini /* Dohrmann's notation: dofs splitted in R (Remaining: all dofs but the vertices) and V (Vertices) */ 17166816873aSStefano Zampini 171753892102SStefano Zampini /* create auxiliary bitmask and allocate workspace */ 17186816873aSStefano Zampini if (!sub_schurs->reuse_mumps) { 1719854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n-n_vertices,&idx_R_local);CHKERRQ(ierr); 1720a64d13efSStefano Zampini ierr = PetscBTCreate(pcis->n,&bitmask);CHKERRQ(ierr); 1721a64d13efSStefano Zampini for (i=0;i<n_vertices;i++) { 17220e6343abSStefano Zampini ierr = PetscBTSet(bitmask,pcbddc->local_primal_ref_node[i]);CHKERRQ(ierr); 1723a64d13efSStefano Zampini } 1724a64d13efSStefano Zampini 1725a64d13efSStefano Zampini for (i=0, n_R=0; i<pcis->n; i++) { 17264641a718SStefano Zampini if (!PetscBTLookup(bitmask,i)) { 17276816873aSStefano Zampini idx_R_local[n_R++] = i; 1728a64d13efSStefano Zampini } 1729a64d13efSStefano Zampini } 173053892102SStefano Zampini } else { /* A different ordering (already computed) is present if we are reusing MUMPS Schur solver */ 17316816873aSStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 17326816873aSStefano Zampini 173353892102SStefano Zampini ierr = ISGetIndices(reuse_mumps->is_R,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 173453892102SStefano Zampini ierr = ISGetLocalSize(reuse_mumps->is_R,&n_R);CHKERRQ(ierr); 17356816873aSStefano Zampini } 17363a50541eSStefano Zampini 17373a50541eSStefano Zampini /* Block code */ 17383a50541eSStefano Zampini vbs = 1; 17393a50541eSStefano Zampini ierr = MatGetBlockSize(pcbddc->local_mat,&bs);CHKERRQ(ierr); 17403a50541eSStefano Zampini if (bs>1 && !(n_vertices%bs)) { 17413a50541eSStefano Zampini PetscBool is_blocked = PETSC_TRUE; 17423a50541eSStefano Zampini PetscInt *vary; 1743d3df7717SStefano Zampini if (!sub_schurs->reuse_mumps) { 1744785e854fSJed Brown ierr = PetscMalloc1(pcis->n/bs,&vary);CHKERRQ(ierr); 17453a50541eSStefano Zampini ierr = PetscMemzero(vary,pcis->n/bs*sizeof(PetscInt));CHKERRQ(ierr); 1746d3df7717SStefano Zampini /* Verify that the vertex indices correspond to each element in a block (code taken from sbaij2.c) */ 1747d3df7717SStefano 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 */ 17480e6343abSStefano Zampini for (i=0; i<n_vertices; i++) vary[pcbddc->local_primal_ref_node[i]/bs]++; 1749d3df7717SStefano Zampini for (i=0; i<pcis->n/bs; i++) { 17503a50541eSStefano Zampini if (vary[i]!=0 && vary[i]!=bs) { 17513a50541eSStefano Zampini is_blocked = PETSC_FALSE; 17523a50541eSStefano Zampini break; 17533a50541eSStefano Zampini } 17543a50541eSStefano Zampini } 1755d3df7717SStefano Zampini ierr = PetscFree(vary);CHKERRQ(ierr); 1756d3df7717SStefano Zampini } else { 1757d3df7717SStefano Zampini /* Verify directly the R set */ 1758d3df7717SStefano Zampini for (i=0; i<n_R/bs; i++) { 1759d3df7717SStefano Zampini PetscInt j,node=idx_R_local[bs*i]; 1760d3df7717SStefano Zampini for (j=1; j<bs; j++) { 1761d3df7717SStefano Zampini if (node != idx_R_local[bs*i+j]-j) { 1762d3df7717SStefano Zampini is_blocked = PETSC_FALSE; 1763d3df7717SStefano Zampini break; 1764d3df7717SStefano Zampini } 1765d3df7717SStefano Zampini } 1766d3df7717SStefano Zampini } 1767d3df7717SStefano Zampini } 17683a50541eSStefano Zampini if (is_blocked) { /* build compressed IS for R nodes (complement of vertices) */ 17693a50541eSStefano Zampini vbs = bs; 17703a50541eSStefano Zampini for (i=0;i<n_R/vbs;i++) { 17713a50541eSStefano Zampini idx_R_local[i] = idx_R_local[vbs*i]/vbs; 17723a50541eSStefano Zampini } 17733a50541eSStefano Zampini } 17743a50541eSStefano Zampini } 17753a50541eSStefano Zampini ierr = ISCreateBlock(PETSC_COMM_SELF,vbs,n_R/vbs,idx_R_local,PETSC_COPY_VALUES,&pcbddc->is_R_local);CHKERRQ(ierr); 177653892102SStefano Zampini if (sub_schurs->reuse_mumps) { 177753892102SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 177853892102SStefano Zampini 177953892102SStefano Zampini ierr = ISRestoreIndices(reuse_mumps->is_R,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 178053892102SStefano Zampini ierr = ISDestroy(&reuse_mumps->is_R);CHKERRQ(ierr); 178153892102SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->is_R_local);CHKERRQ(ierr); 178253892102SStefano Zampini reuse_mumps->is_R = pcbddc->is_R_local; 178353892102SStefano Zampini } else { 17843a50541eSStefano Zampini ierr = PetscFree(idx_R_local);CHKERRQ(ierr); 178553892102SStefano Zampini } 1786a64d13efSStefano Zampini 1787a64d13efSStefano Zampini /* print some info if requested */ 1788a64d13efSStefano Zampini if (pcbddc->dbg_flag) { 1789c9ed8603SStefano Zampini PetscInt benign = 0; 1790019a44ceSStefano Zampini 1791c9ed8603SStefano Zampini if (pcbddc->benign_p0_lidx >= 0) benign = 1; 1792a64d13efSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 1793a64d13efSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 17940fccc4e9SStefano Zampini ierr = PetscViewerASCIISynchronizedAllow(pcbddc->dbg_viewer,PETSC_TRUE);CHKERRQ(ierr); 1795a64d13efSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d local dimensions\n",PetscGlobalRank);CHKERRQ(ierr); 1796a64d13efSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"local_size = %d, dirichlet_size = %d, boundary_size = %d\n",pcis->n,n_D,n_B);CHKERRQ(ierr); 1797019a44ceSStefano 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-benign,pcbddc->local_primal_size);CHKERRQ(ierr); 1798a64d13efSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1799a64d13efSStefano Zampini } 1800a64d13efSStefano Zampini 1801a64d13efSStefano Zampini /* VecScatters pcbddc->R_to_B and (optionally) pcbddc->R_to_D */ 18026816873aSStefano Zampini if (!sub_schurs->reuse_mumps) { 18036816873aSStefano Zampini IS is_aux1,is_aux2; 18046816873aSStefano Zampini PetscInt *aux_array1,*aux_array2,*is_indices,*idx_R_local; 18056816873aSStefano Zampini 18063a50541eSStefano Zampini ierr = ISGetIndices(pcbddc->is_R_local,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 1807854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n_B-n_vertices,&aux_array1);CHKERRQ(ierr); 1808854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n_B-n_vertices,&aux_array2);CHKERRQ(ierr); 1809a64d13efSStefano Zampini ierr = ISGetIndices(pcis->is_I_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 18104641a718SStefano Zampini for (i=0; i<n_D; i++) { 18114641a718SStefano Zampini ierr = PetscBTSet(bitmask,is_indices[i]);CHKERRQ(ierr); 18124641a718SStefano Zampini } 1813a64d13efSStefano Zampini ierr = ISRestoreIndices(pcis->is_I_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 1814a64d13efSStefano Zampini for (i=0, j=0; i<n_R; i++) { 18154641a718SStefano Zampini if (!PetscBTLookup(bitmask,idx_R_local[i])) { 18164641a718SStefano Zampini aux_array1[j++] = i; 1817a64d13efSStefano Zampini } 1818a64d13efSStefano Zampini } 1819a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array1,PETSC_OWN_POINTER,&is_aux1);CHKERRQ(ierr); 1820a64d13efSStefano Zampini ierr = ISGetIndices(pcis->is_B_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 1821a64d13efSStefano Zampini for (i=0, j=0; i<n_B; i++) { 18224641a718SStefano Zampini if (!PetscBTLookup(bitmask,is_indices[i])) { 18234641a718SStefano Zampini aux_array2[j++] = i; 1824a64d13efSStefano Zampini } 1825a64d13efSStefano Zampini } 1826a64d13efSStefano Zampini ierr = ISRestoreIndices(pcis->is_B_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 1827a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array2,PETSC_OWN_POINTER,&is_aux2);CHKERRQ(ierr); 1828a64d13efSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,is_aux1,pcis->vec1_B,is_aux2,&pcbddc->R_to_B);CHKERRQ(ierr); 1829a64d13efSStefano Zampini ierr = ISDestroy(&is_aux1);CHKERRQ(ierr); 1830a64d13efSStefano Zampini ierr = ISDestroy(&is_aux2);CHKERRQ(ierr); 1831a64d13efSStefano Zampini 18328eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 1833785e854fSJed Brown ierr = PetscMalloc1(n_D,&aux_array1);CHKERRQ(ierr); 1834a64d13efSStefano Zampini for (i=0, j=0; i<n_R; i++) { 18354641a718SStefano Zampini if (PetscBTLookup(bitmask,idx_R_local[i])) { 18364641a718SStefano Zampini aux_array1[j++] = i; 1837a64d13efSStefano Zampini } 1838a64d13efSStefano Zampini } 1839a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array1,PETSC_OWN_POINTER,&is_aux1);CHKERRQ(ierr); 1840a64d13efSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,is_aux1,pcis->vec1_D,(IS)0,&pcbddc->R_to_D);CHKERRQ(ierr); 1841a64d13efSStefano Zampini ierr = ISDestroy(&is_aux1);CHKERRQ(ierr); 1842a64d13efSStefano Zampini } 18434641a718SStefano Zampini ierr = PetscBTDestroy(&bitmask);CHKERRQ(ierr); 18443a50541eSStefano Zampini ierr = ISRestoreIndices(pcbddc->is_R_local,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 1845d62866d3SStefano Zampini } else { 184653892102SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 18476816873aSStefano Zampini IS tis; 18486816873aSStefano Zampini PetscInt schur_size; 18496816873aSStefano Zampini 185053892102SStefano Zampini ierr = ISGetLocalSize(reuse_mumps->is_B,&schur_size);CHKERRQ(ierr); 18516816873aSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,schur_size,n_D,1,&tis);CHKERRQ(ierr); 185253892102SStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,tis,pcis->vec1_B,reuse_mumps->is_B,&pcbddc->R_to_B);CHKERRQ(ierr); 18536816873aSStefano Zampini ierr = ISDestroy(&tis);CHKERRQ(ierr); 18546816873aSStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 18556816873aSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,n_D,0,1,&tis);CHKERRQ(ierr); 18566816873aSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,tis,pcis->vec1_D,(IS)0,&pcbddc->R_to_D);CHKERRQ(ierr); 18576816873aSStefano Zampini ierr = ISDestroy(&tis);CHKERRQ(ierr); 1858d62866d3SStefano Zampini } 1859d62866d3SStefano Zampini } 1860a64d13efSStefano Zampini PetscFunctionReturn(0); 1861a64d13efSStefano Zampini } 1862a64d13efSStefano Zampini 1863304d26faSStefano Zampini 1864304d26faSStefano Zampini #undef __FUNCT__ 1865304d26faSStefano Zampini #define __FUNCT__ "PCBDDCSetUpLocalSolvers" 1866684f6988SStefano Zampini PetscErrorCode PCBDDCSetUpLocalSolvers(PC pc, PetscBool dirichlet, PetscBool neumann) 1867304d26faSStefano Zampini { 1868304d26faSStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 1869304d26faSStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 1870304d26faSStefano Zampini PC pc_temp; 1871304d26faSStefano Zampini Mat A_RR; 1872f4ddd8eeSStefano Zampini MatReuse reuse; 1873304d26faSStefano Zampini PetscScalar m_one = -1.0; 1874304d26faSStefano Zampini PetscReal value; 187504708bb6SStefano Zampini PetscInt n_D,n_R; 18769577ea80SStefano Zampini PetscBool use_exact,use_exact_reduced,issbaij; 1877304d26faSStefano Zampini PetscErrorCode ierr; 1878e604994aSStefano Zampini /* prefixes stuff */ 1879312be037SStefano Zampini char dir_prefix[256],neu_prefix[256],str_level[16]; 1880e604994aSStefano Zampini size_t len; 1881304d26faSStefano Zampini 1882304d26faSStefano Zampini PetscFunctionBegin; 1883304d26faSStefano Zampini 1884e604994aSStefano Zampini /* compute prefixes */ 1885e604994aSStefano Zampini ierr = PetscStrcpy(dir_prefix,"");CHKERRQ(ierr); 1886e604994aSStefano Zampini ierr = PetscStrcpy(neu_prefix,"");CHKERRQ(ierr); 1887e604994aSStefano Zampini if (!pcbddc->current_level) { 1888e604994aSStefano Zampini ierr = PetscStrcpy(dir_prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 1889e604994aSStefano Zampini ierr = PetscStrcpy(neu_prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 1890e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,"pc_bddc_dirichlet_");CHKERRQ(ierr); 1891e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,"pc_bddc_neumann_");CHKERRQ(ierr); 1892e604994aSStefano Zampini } else { 1893e604994aSStefano Zampini ierr = PetscStrcpy(str_level,"");CHKERRQ(ierr); 1894312be037SStefano Zampini sprintf(str_level,"l%d_",(int)(pcbddc->current_level)); 1895e604994aSStefano Zampini ierr = PetscStrlen(((PetscObject)pc)->prefix,&len);CHKERRQ(ierr); 1896e604994aSStefano Zampini len -= 15; /* remove "pc_bddc_coarse_" */ 1897312be037SStefano Zampini if (pcbddc->current_level>1) len -= 3; /* remove "lX_" with X level number */ 1898312be037SStefano Zampini if (pcbddc->current_level>10) len -= 1; /* remove another char from level number */ 189934d6797cSStefano Zampini ierr = PetscStrncpy(dir_prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 190034d6797cSStefano Zampini ierr = PetscStrncpy(neu_prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 1901e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,"pc_bddc_dirichlet_");CHKERRQ(ierr); 1902e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,"pc_bddc_neumann_");CHKERRQ(ierr); 1903e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,str_level);CHKERRQ(ierr); 1904e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,str_level);CHKERRQ(ierr); 1905e604994aSStefano Zampini } 1906e604994aSStefano Zampini 1907304d26faSStefano Zampini /* DIRICHLET PROBLEM */ 1908684f6988SStefano Zampini if (dirichlet) { 1909d5574798SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 19103301b35fSStefano Zampini if (pcbddc->local_mat->symmetric_set) { 19113301b35fSStefano Zampini ierr = MatSetOption(pcis->A_II,MAT_SYMMETRIC,pcbddc->local_mat->symmetric_set);CHKERRQ(ierr); 1912964fefecSStefano Zampini } 1913ac78edfcSStefano Zampini /* Matrix for Dirichlet problem is pcis->A_II */ 1914964fefecSStefano Zampini n_D = pcis->n - pcis->n_B; 1915304d26faSStefano Zampini if (!pcbddc->ksp_D) { /* create object if not yet build */ 1916304d26faSStefano Zampini ierr = KSPCreate(PETSC_COMM_SELF,&pcbddc->ksp_D);CHKERRQ(ierr); 1917304d26faSStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->ksp_D,(PetscObject)pc,1);CHKERRQ(ierr); 1918304d26faSStefano Zampini /* default */ 1919304d26faSStefano Zampini ierr = KSPSetType(pcbddc->ksp_D,KSPPREONLY);CHKERRQ(ierr); 1920e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->ksp_D,dir_prefix);CHKERRQ(ierr); 19219577ea80SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcis->A_II,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 1922304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_D,&pc_temp);CHKERRQ(ierr); 19239577ea80SStefano Zampini if (issbaij) { 19249577ea80SStefano Zampini ierr = PCSetType(pc_temp,PCCHOLESKY);CHKERRQ(ierr); 19259577ea80SStefano Zampini } else { 1926304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCLU);CHKERRQ(ierr); 19279577ea80SStefano Zampini } 1928304d26faSStefano Zampini /* Allow user's customization */ 1929304d26faSStefano Zampini ierr = KSPSetFromOptions(pcbddc->ksp_D);CHKERRQ(ierr); 1930304d26faSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 1931304d26faSStefano Zampini } 1932d1e9a80fSBarry Smith ierr = KSPSetOperators(pcbddc->ksp_D,pcis->A_II,pcis->A_II);CHKERRQ(ierr); 1933d62866d3SStefano Zampini if (sub_schurs->reuse_mumps) { 1934d62866d3SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 1935d62866d3SStefano Zampini 1936d62866d3SStefano Zampini ierr = KSPSetPC(pcbddc->ksp_D,reuse_mumps->interior_solver);CHKERRQ(ierr); 1937d5574798SStefano Zampini } 1938304d26faSStefano Zampini /* umfpack interface has a bug when matrix dimension is zero. TODO solve from umfpack interface */ 1939304d26faSStefano Zampini if (!n_D) { 1940304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_D,&pc_temp);CHKERRQ(ierr); 1941304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCNONE);CHKERRQ(ierr); 1942304d26faSStefano Zampini } 1943304d26faSStefano Zampini /* Set Up KSP for Dirichlet problem of BDDC */ 1944304d26faSStefano Zampini ierr = KSPSetUp(pcbddc->ksp_D);CHKERRQ(ierr); 1945304d26faSStefano Zampini /* set ksp_D into pcis data */ 1946304d26faSStefano Zampini ierr = KSPDestroy(&pcis->ksp_D);CHKERRQ(ierr); 1947304d26faSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->ksp_D);CHKERRQ(ierr); 1948304d26faSStefano Zampini pcis->ksp_D = pcbddc->ksp_D; 1949684f6988SStefano Zampini } 1950304d26faSStefano Zampini 1951304d26faSStefano Zampini /* NEUMANN PROBLEM */ 1952684f6988SStefano Zampini A_RR = 0; 1953684f6988SStefano Zampini if (neumann) { 1954d62866d3SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 195504708bb6SStefano Zampini PetscInt ibs,mbs; 195604708bb6SStefano Zampini PetscBool issbaij; 195704708bb6SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 1958f4ddd8eeSStefano Zampini /* Matrix for Neumann problem is A_RR -> we need to create/reuse it at this point */ 19598ce42a96SStefano Zampini ierr = ISGetSize(pcbddc->is_R_local,&n_R);CHKERRQ(ierr); 1960f4ddd8eeSStefano Zampini if (pcbddc->ksp_R) { /* already created ksp */ 1961f4ddd8eeSStefano Zampini PetscInt nn_R; 196281d9aea3SBarry Smith ierr = KSPGetOperators(pcbddc->ksp_R,NULL,&A_RR);CHKERRQ(ierr); 1963f4ddd8eeSStefano Zampini ierr = PetscObjectReference((PetscObject)A_RR);CHKERRQ(ierr); 1964f4ddd8eeSStefano Zampini ierr = MatGetSize(A_RR,&nn_R,NULL);CHKERRQ(ierr); 1965f4ddd8eeSStefano Zampini if (nn_R != n_R) { /* old ksp is not reusable, so reset it */ 1966f4ddd8eeSStefano Zampini ierr = KSPReset(pcbddc->ksp_R);CHKERRQ(ierr); 1967f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 1968f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 1969f4ddd8eeSStefano Zampini } else { /* same sizes, but nonzero pattern depend on primal vertices so it can be changed */ 1970727cdba6SStefano Zampini if (pcbddc->new_primal_space_local) { /* we are not sure the matrix will have the same nonzero pattern */ 1971f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 1972f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 1973f4ddd8eeSStefano Zampini } else { /* safe to reuse the matrix */ 1974f4ddd8eeSStefano Zampini reuse = MAT_REUSE_MATRIX; 1975f4ddd8eeSStefano Zampini } 1976f4ddd8eeSStefano Zampini } 1977f4ddd8eeSStefano Zampini /* last check */ 1978d1e9a80fSBarry Smith if (pc->flag == DIFFERENT_NONZERO_PATTERN) { 1979f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 1980f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 1981f4ddd8eeSStefano Zampini } 1982f4ddd8eeSStefano Zampini } else { /* first time, so we need to create the matrix */ 1983f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 1984f4ddd8eeSStefano Zampini } 1985f4ddd8eeSStefano Zampini /* extract A_RR */ 1986af732b37SStefano Zampini ierr = MatGetBlockSize(pcbddc->local_mat,&mbs);CHKERRQ(ierr); 1987af732b37SStefano Zampini ierr = ISGetBlockSize(pcbddc->is_R_local,&ibs);CHKERRQ(ierr); 198804708bb6SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcbddc->local_mat,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 198904708bb6SStefano Zampini if (ibs != mbs) { /* need to convert to SEQAIJ to extract any submatrix with is_R_local */ 199004708bb6SStefano Zampini if (matis->A == pcbddc->local_mat) { 199104708bb6SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 199204708bb6SStefano Zampini ierr = MatConvert(matis->A,MATSEQAIJ,MAT_INITIAL_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 1993af732b37SStefano Zampini } else { 199404708bb6SStefano Zampini ierr = MatConvert(pcbddc->local_mat,MATSEQAIJ,MAT_REUSE_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 19956816873aSStefano Zampini } 199604708bb6SStefano Zampini } else if (issbaij) { /* need to convert to BAIJ to get offdiagonal blocks */ 199704708bb6SStefano Zampini if (matis->A == pcbddc->local_mat) { 199804708bb6SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 199904708bb6SStefano Zampini ierr = MatConvert(matis->A,MATSEQBAIJ,MAT_INITIAL_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 200004708bb6SStefano Zampini } else { 200104708bb6SStefano Zampini ierr = MatConvert(pcbddc->local_mat,MATSEQBAIJ,MAT_REUSE_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 200204708bb6SStefano Zampini } 200304708bb6SStefano Zampini } 200404708bb6SStefano Zampini if (!sub_schurs->reuse_mumps) { 2005f4ddd8eeSStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,pcbddc->is_R_local,pcbddc->is_R_local,reuse,&A_RR);CHKERRQ(ierr); 20063301b35fSStefano Zampini if (pcbddc->local_mat->symmetric_set) { 20073301b35fSStefano Zampini ierr = MatSetOption(A_RR,MAT_SYMMETRIC,pcbddc->local_mat->symmetric_set);CHKERRQ(ierr); 20086816873aSStefano Zampini } 20096816873aSStefano Zampini } else { 20106816873aSStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 20116816873aSStefano Zampini 20126816873aSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 20136816873aSStefano Zampini ierr = PCGetOperators(reuse_mumps->correction_solver,&A_RR,NULL);CHKERRQ(ierr); 20146816873aSStefano Zampini ierr = PetscObjectReference((PetscObject)A_RR);CHKERRQ(ierr); 2015af732b37SStefano Zampini } 2016f4ddd8eeSStefano Zampini if (!pcbddc->ksp_R) { /* create object if not present */ 2017304d26faSStefano Zampini ierr = KSPCreate(PETSC_COMM_SELF,&pcbddc->ksp_R);CHKERRQ(ierr); 2018304d26faSStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->ksp_R,(PetscObject)pc,1);CHKERRQ(ierr); 2019304d26faSStefano Zampini /* default */ 2020304d26faSStefano Zampini ierr = KSPSetType(pcbddc->ksp_R,KSPPREONLY);CHKERRQ(ierr); 2021e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->ksp_R,neu_prefix);CHKERRQ(ierr); 2022304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_temp);CHKERRQ(ierr); 20239577ea80SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)A_RR,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 20249577ea80SStefano Zampini if (issbaij) { 20259577ea80SStefano Zampini ierr = PCSetType(pc_temp,PCCHOLESKY);CHKERRQ(ierr); 20269577ea80SStefano Zampini } else { 2027304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCLU);CHKERRQ(ierr); 20289577ea80SStefano Zampini } 2029304d26faSStefano Zampini /* Allow user's customization */ 2030304d26faSStefano Zampini ierr = KSPSetFromOptions(pcbddc->ksp_R);CHKERRQ(ierr); 2031304d26faSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 2032304d26faSStefano Zampini } 2033d1e9a80fSBarry Smith ierr = KSPSetOperators(pcbddc->ksp_R,A_RR,A_RR);CHKERRQ(ierr); 2034304d26faSStefano Zampini /* umfpack interface has a bug when matrix dimension is zero. TODO solve from umfpack interface */ 2035304d26faSStefano Zampini if (!n_R) { 2036304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_temp);CHKERRQ(ierr); 2037304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCNONE);CHKERRQ(ierr); 2038304d26faSStefano Zampini } 2039d62866d3SStefano Zampini /* Reuse MUMPS solver if it is present */ 2040d62866d3SStefano Zampini if (sub_schurs->reuse_mumps) { 2041d62866d3SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 2042d62866d3SStefano Zampini 2043d62866d3SStefano Zampini ierr = KSPSetPC(pcbddc->ksp_R,reuse_mumps->correction_solver);CHKERRQ(ierr); 2044d62866d3SStefano Zampini } 2045304d26faSStefano Zampini /* Set Up KSP for Neumann problem of BDDC */ 2046304d26faSStefano Zampini ierr = KSPSetUp(pcbddc->ksp_R);CHKERRQ(ierr); 2047684f6988SStefano Zampini } 20486816873aSStefano Zampini /* free Neumann problem's matrix */ 20496816873aSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 2050304d26faSStefano Zampini 2051304d26faSStefano Zampini /* check Dirichlet and Neumann solvers and adapt them if a nullspace correction is needed */ 20520fccc4e9SStefano Zampini if (pcbddc->NullSpace || pcbddc->dbg_flag) { 2053684f6988SStefano Zampini if (pcbddc->dbg_flag) { 2054684f6988SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 2055684f6988SStefano Zampini ierr = PetscViewerASCIISynchronizedAllow(pcbddc->dbg_viewer,PETSC_TRUE);CHKERRQ(ierr); 2056684f6988SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 2057684f6988SStefano Zampini } 2058684f6988SStefano Zampini if (dirichlet) { /* Dirichlet */ 20590fccc4e9SStefano Zampini ierr = VecSetRandom(pcis->vec1_D,NULL);CHKERRQ(ierr); 20600fccc4e9SStefano Zampini ierr = MatMult(pcis->A_II,pcis->vec1_D,pcis->vec2_D);CHKERRQ(ierr); 20610fccc4e9SStefano Zampini ierr = KSPSolve(pcbddc->ksp_D,pcis->vec2_D,pcis->vec2_D);CHKERRQ(ierr); 20620fccc4e9SStefano Zampini ierr = VecAXPY(pcis->vec1_D,m_one,pcis->vec2_D);CHKERRQ(ierr); 20630fccc4e9SStefano Zampini ierr = VecNorm(pcis->vec1_D,NORM_INFINITY,&value);CHKERRQ(ierr); 2064304d26faSStefano Zampini /* need to be adapted? */ 2065b8ffe317SStefano Zampini use_exact = (PetscAbsReal(value) > 1.e-4 ? PETSC_FALSE : PETSC_TRUE); 2066b8ffe317SStefano Zampini ierr = MPI_Allreduce(&use_exact,&use_exact_reduced,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 2067b8ffe317SStefano Zampini ierr = PCBDDCSetUseExactDirichlet(pc,use_exact_reduced);CHKERRQ(ierr); 2068304d26faSStefano Zampini /* print info */ 2069304d26faSStefano Zampini if (pcbddc->dbg_flag) { 2070e604994aSStefano 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); 2071304d26faSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 2072304d26faSStefano Zampini } 2073b8ffe317SStefano Zampini if (pcbddc->NullSpace && !use_exact_reduced && !pcbddc->switch_static) { 2074298c0119SStefano Zampini ierr = PCBDDCNullSpaceAssembleCorrection(pc,PETSC_TRUE,pcis->is_I_local);CHKERRQ(ierr); 2075304d26faSStefano Zampini } 2076684f6988SStefano Zampini } 2077684f6988SStefano Zampini if (neumann) { /* Neumann */ 20786816873aSStefano Zampini ierr = KSPGetOperators(pcbddc->ksp_R,&A_RR,NULL);CHKERRQ(ierr); 20790fccc4e9SStefano Zampini ierr = VecSetRandom(pcbddc->vec1_R,NULL);CHKERRQ(ierr); 20800fccc4e9SStefano Zampini ierr = MatMult(A_RR,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 20810fccc4e9SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec2_R,pcbddc->vec2_R);CHKERRQ(ierr); 20820fccc4e9SStefano Zampini ierr = VecAXPY(pcbddc->vec1_R,m_one,pcbddc->vec2_R);CHKERRQ(ierr); 20830fccc4e9SStefano Zampini ierr = VecNorm(pcbddc->vec1_R,NORM_INFINITY,&value);CHKERRQ(ierr); 2084304d26faSStefano Zampini /* need to be adapted? */ 2085b8ffe317SStefano Zampini use_exact = (PetscAbsReal(value) > 1.e-4 ? PETSC_FALSE : PETSC_TRUE); 2086b8ffe317SStefano Zampini ierr = MPI_Allreduce(&use_exact,&use_exact_reduced,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 2087304d26faSStefano Zampini /* print info */ 2088304d26faSStefano Zampini if (pcbddc->dbg_flag) { 2089e604994aSStefano 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); 2090304d26faSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 2091304d26faSStefano Zampini } 2092b8ffe317SStefano Zampini if (pcbddc->NullSpace && !use_exact_reduced) { /* is it the right logic? */ 2093298c0119SStefano Zampini ierr = PCBDDCNullSpaceAssembleCorrection(pc,PETSC_FALSE,pcbddc->is_R_local);CHKERRQ(ierr); 2094304d26faSStefano Zampini } 20950fccc4e9SStefano Zampini } 2096684f6988SStefano Zampini } 2097304d26faSStefano Zampini PetscFunctionReturn(0); 2098304d26faSStefano Zampini } 2099304d26faSStefano Zampini 2100304d26faSStefano Zampini #undef __FUNCT__ 2101ba15a52eSStefano Zampini #define __FUNCT__ "PCBDDCSolveSubstructureCorrection" 210280677318SStefano Zampini static PetscErrorCode PCBDDCSolveSubstructureCorrection(PC pc, Vec inout_B, Vec inout_D, PetscBool applytranspose) 2103674ae819SStefano Zampini { 2104674ae819SStefano Zampini PetscErrorCode ierr; 2105674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 2106be83ff47SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 2107674ae819SStefano Zampini 2108674ae819SStefano Zampini PetscFunctionBegin; 2109be83ff47SStefano Zampini if (!sub_schurs->reuse_mumps) { 211080677318SStefano Zampini ierr = VecSet(pcbddc->vec1_R,0.);CHKERRQ(ierr); 211120c7b377SStefano Zampini } 211280677318SStefano Zampini if (!pcbddc->switch_static) { 211380677318SStefano Zampini if (applytranspose && pcbddc->local_auxmat1) { 211480677318SStefano Zampini ierr = MatMultTranspose(pcbddc->local_auxmat2,inout_B,pcbddc->vec1_C);CHKERRQ(ierr); 211580677318SStefano Zampini ierr = MatMultTransposeAdd(pcbddc->local_auxmat1,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr); 211620c7b377SStefano Zampini } 2117be83ff47SStefano Zampini if (!sub_schurs->reuse_mumps) { 211880677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 211980677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 212020c7b377SStefano Zampini } else { 2121be83ff47SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 2122be83ff47SStefano Zampini 212353892102SStefano Zampini ierr = VecScatterBegin(reuse_mumps->correction_scatter_B,inout_B,reuse_mumps->rhs_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 212453892102SStefano Zampini ierr = VecScatterEnd(reuse_mumps->correction_scatter_B,inout_B,reuse_mumps->rhs_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 212520c7b377SStefano Zampini } 2126be83ff47SStefano Zampini } else { 212780677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 212880677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 212980677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,inout_D,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 213080677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,inout_D,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 213180677318SStefano Zampini if (applytranspose && pcbddc->local_auxmat1) { 213280677318SStefano Zampini ierr = MatMultTranspose(pcbddc->local_auxmat2,pcbddc->vec1_R,pcbddc->vec1_C);CHKERRQ(ierr); 213380677318SStefano Zampini ierr = MatMultTransposeAdd(pcbddc->local_auxmat1,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr); 213480677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 213580677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2136674ae819SStefano Zampini } 2137674ae819SStefano Zampini } 2138be83ff47SStefano Zampini if (!sub_schurs->reuse_mumps) { 213980677318SStefano Zampini if (applytranspose) { 214080677318SStefano Zampini ierr = KSPSolveTranspose(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 214180677318SStefano Zampini } else { 214280677318SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 214380677318SStefano Zampini } 214453892102SStefano Zampini #if defined(PETSC_HAVE_MUMPS) 2145be83ff47SStefano Zampini } else { 2146be83ff47SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 2147be83ff47SStefano Zampini 2148be83ff47SStefano Zampini if (applytranspose) { 214953892102SStefano Zampini ierr = MatMumpsSolveSchurComplementTranspose(reuse_mumps->F,reuse_mumps->rhs_B,reuse_mumps->sol_B);CHKERRQ(ierr); 2150be83ff47SStefano Zampini } else { 215153892102SStefano Zampini ierr = MatMumpsSolveSchurComplement(reuse_mumps->F,reuse_mumps->rhs_B,reuse_mumps->sol_B);CHKERRQ(ierr); 2152be83ff47SStefano Zampini } 215353892102SStefano Zampini #endif 2154be83ff47SStefano Zampini } 215580677318SStefano Zampini ierr = VecSet(inout_B,0.);CHKERRQ(ierr); 215680677318SStefano Zampini if (!pcbddc->switch_static) { 2157be83ff47SStefano Zampini if (!sub_schurs->reuse_mumps) { 215880677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 215980677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2160be83ff47SStefano Zampini } else { 2161be83ff47SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 2162be83ff47SStefano Zampini 216353892102SStefano Zampini ierr = VecScatterBegin(reuse_mumps->correction_scatter_B,reuse_mumps->sol_B,inout_B,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 216453892102SStefano Zampini ierr = VecScatterEnd(reuse_mumps->correction_scatter_B,reuse_mumps->sol_B,inout_B,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2165be83ff47SStefano Zampini } 216680677318SStefano Zampini if (!applytranspose && pcbddc->local_auxmat1) { 216780677318SStefano Zampini ierr = MatMult(pcbddc->local_auxmat1,inout_B,pcbddc->vec1_C);CHKERRQ(ierr); 216880677318SStefano Zampini ierr = MatMultAdd(pcbddc->local_auxmat2,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr); 216980677318SStefano Zampini } 217080677318SStefano Zampini } else { 217180677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 217280677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 217380677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 217480677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 217580677318SStefano Zampini if (!applytranspose && pcbddc->local_auxmat1) { 217680677318SStefano Zampini ierr = MatMult(pcbddc->local_auxmat1,inout_B,pcbddc->vec1_C);CHKERRQ(ierr); 217780677318SStefano Zampini ierr = MatMultAdd(pcbddc->local_auxmat2,pcbddc->vec1_C,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 217880677318SStefano Zampini } 217980677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 218080677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 218180677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 218280677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2183674ae819SStefano Zampini } 2184674ae819SStefano Zampini PetscFunctionReturn(0); 2185674ae819SStefano Zampini } 2186674ae819SStefano Zampini 2187dc359a40SStefano Zampini /* parameter apply transpose determines if the interface preconditioner should be applied transposed or not */ 2188674ae819SStefano Zampini #undef __FUNCT__ 2189674ae819SStefano Zampini #define __FUNCT__ "PCBDDCApplyInterfacePreconditioner" 2190dc359a40SStefano Zampini PetscErrorCode PCBDDCApplyInterfacePreconditioner(PC pc, PetscBool applytranspose) 2191674ae819SStefano Zampini { 2192674ae819SStefano Zampini PetscErrorCode ierr; 2193674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 2194674ae819SStefano Zampini PC_IS* pcis = (PC_IS*) (pc->data); 2195674ae819SStefano Zampini const PetscScalar zero = 0.0; 2196674ae819SStefano Zampini 2197674ae819SStefano Zampini PetscFunctionBegin; 2198dc359a40SStefano Zampini /* Application of PSI^T or PHI^T (depending on applytranspose, see comment above) */ 2199dc359a40SStefano Zampini if (applytranspose) { 2200674ae819SStefano Zampini ierr = MatMultTranspose(pcbddc->coarse_phi_B,pcis->vec1_B,pcbddc->vec1_P);CHKERRQ(ierr); 22018eeda7d8SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultTransposeAdd(pcbddc->coarse_phi_D,pcis->vec1_D,pcbddc->vec1_P,pcbddc->vec1_P);CHKERRQ(ierr); } 2202dc359a40SStefano Zampini } else { 2203674ae819SStefano Zampini ierr = MatMultTranspose(pcbddc->coarse_psi_B,pcis->vec1_B,pcbddc->vec1_P);CHKERRQ(ierr); 2204674ae819SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultTransposeAdd(pcbddc->coarse_psi_D,pcis->vec1_D,pcbddc->vec1_P,pcbddc->vec1_P);CHKERRQ(ierr); } 220515aaf578SStefano Zampini } 2206efc2fbd9SStefano Zampini 2207efc2fbd9SStefano Zampini /* add p0 to the last value of vec1_P holding the coarse dof relative to p0 */ 2208c9ed8603SStefano Zampini if (pcbddc->benign_p0_lidx >= 0) { 2209efc2fbd9SStefano Zampini PetscScalar *array; 2210efc2fbd9SStefano Zampini 2211efc2fbd9SStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 2212efc2fbd9SStefano Zampini array[pcbddc->local_primal_size-1] += pcbddc->benign_p0; 2213efc2fbd9SStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 2214efc2fbd9SStefano Zampini } 2215efc2fbd9SStefano Zampini 221612edc857SStefano Zampini /* start communications from local primal nodes to rhs of coarse solver */ 221712edc857SStefano Zampini ierr = VecSet(pcbddc->coarse_vec,zero);CHKERRQ(ierr); 221812edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataBegin(pc,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 221912edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataEnd(pc,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 222012edc857SStefano Zampini 22219f00e9b4SStefano Zampini /* Coarse solution -> rhs and sol updated inside PCBDDCScattarCoarseDataBegin/End */ 222212edc857SStefano Zampini /* TODO remove null space when doing multilevel */ 222312edc857SStefano Zampini if (pcbddc->coarse_ksp) { 2224964fefecSStefano Zampini Vec rhs,sol; 2225964fefecSStefano Zampini 2226964fefecSStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&rhs);CHKERRQ(ierr); 2227964fefecSStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&sol);CHKERRQ(ierr); 222812edc857SStefano Zampini if (applytranspose) { 2229964fefecSStefano Zampini ierr = KSPSolveTranspose(pcbddc->coarse_ksp,rhs,sol);CHKERRQ(ierr); 223012edc857SStefano Zampini } else { 2231964fefecSStefano Zampini ierr = KSPSolve(pcbddc->coarse_ksp,rhs,sol);CHKERRQ(ierr); 223212edc857SStefano Zampini } 223312edc857SStefano Zampini } 2234674ae819SStefano Zampini 2235674ae819SStefano Zampini /* Local solution on R nodes */ 223680677318SStefano Zampini if (pcis->n) { /* in/out pcbddc->vec1_B,pcbddc->vec1_D */ 223780677318SStefano Zampini ierr = PCBDDCSolveSubstructureCorrection(pc,pcis->vec1_B,pcis->vec1_D,applytranspose);CHKERRQ(ierr); 22389f00e9b4SStefano Zampini } 2239674ae819SStefano Zampini 22409f00e9b4SStefano Zampini /* communications from coarse sol to local primal nodes */ 22419f00e9b4SStefano Zampini ierr = PCBDDCScatterCoarseDataBegin(pc,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 224212edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataEnd(pc,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2243674ae819SStefano Zampini 2244674ae819SStefano Zampini /* Sum contributions from two levels */ 2245dc359a40SStefano Zampini if (applytranspose) { 2246dc359a40SStefano Zampini ierr = MatMultAdd(pcbddc->coarse_psi_B,pcbddc->vec1_P,pcis->vec1_B,pcis->vec1_B);CHKERRQ(ierr); 2247dc359a40SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultAdd(pcbddc->coarse_psi_D,pcbddc->vec1_P,pcis->vec1_D,pcis->vec1_D);CHKERRQ(ierr); } 2248dc359a40SStefano Zampini } else { 2249674ae819SStefano Zampini ierr = MatMultAdd(pcbddc->coarse_phi_B,pcbddc->vec1_P,pcis->vec1_B,pcis->vec1_B);CHKERRQ(ierr); 22508eeda7d8SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultAdd(pcbddc->coarse_phi_D,pcbddc->vec1_P,pcis->vec1_D,pcis->vec1_D);CHKERRQ(ierr); } 2251dc359a40SStefano Zampini } 2252efc2fbd9SStefano Zampini /* store p0 */ 2253c9ed8603SStefano Zampini if (pcbddc->benign_p0_lidx >= 0) { 2254efc2fbd9SStefano Zampini PetscScalar *array; 2255efc2fbd9SStefano Zampini 2256efc2fbd9SStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 2257efc2fbd9SStefano Zampini pcbddc->benign_p0 = array[pcbddc->local_primal_size-1]; 2258efc2fbd9SStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 2259efc2fbd9SStefano Zampini } 2260674ae819SStefano Zampini PetscFunctionReturn(0); 2261674ae819SStefano Zampini } 2262674ae819SStefano Zampini 2263674ae819SStefano Zampini #undef __FUNCT__ 2264674ae819SStefano Zampini #define __FUNCT__ "PCBDDCScatterCoarseDataBegin" 226512edc857SStefano Zampini PetscErrorCode PCBDDCScatterCoarseDataBegin(PC pc,InsertMode imode, ScatterMode smode) 2266674ae819SStefano Zampini { 2267674ae819SStefano Zampini PetscErrorCode ierr; 2268674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 226958da7f69SStefano Zampini PetscScalar *array; 227012edc857SStefano Zampini Vec from,to; 2271674ae819SStefano Zampini 2272674ae819SStefano Zampini PetscFunctionBegin; 227312edc857SStefano Zampini if (smode == SCATTER_REVERSE) { /* from global to local -> get data from coarse solution */ 227412edc857SStefano Zampini from = pcbddc->coarse_vec; 227512edc857SStefano Zampini to = pcbddc->vec1_P; 227612edc857SStefano Zampini if (pcbddc->coarse_ksp) { /* get array from coarse processes */ 227712edc857SStefano Zampini Vec tvec; 227858da7f69SStefano Zampini 227958da7f69SStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 228058da7f69SStefano Zampini ierr = VecResetArray(tvec);CHKERRQ(ierr); 228112edc857SStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 228258da7f69SStefano Zampini ierr = VecGetArray(tvec,&array);CHKERRQ(ierr); 228358da7f69SStefano Zampini ierr = VecPlaceArray(from,array);CHKERRQ(ierr); 228458da7f69SStefano Zampini ierr = VecRestoreArray(tvec,&array);CHKERRQ(ierr); 228512edc857SStefano Zampini } 228612edc857SStefano Zampini } else { /* from local to global -> put data in coarse right hand side */ 228712edc857SStefano Zampini from = pcbddc->vec1_P; 228812edc857SStefano Zampini to = pcbddc->coarse_vec; 228912edc857SStefano Zampini } 229012edc857SStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,from,to,imode,smode);CHKERRQ(ierr); 2291674ae819SStefano Zampini PetscFunctionReturn(0); 2292674ae819SStefano Zampini } 2293674ae819SStefano Zampini 2294674ae819SStefano Zampini #undef __FUNCT__ 2295674ae819SStefano Zampini #define __FUNCT__ "PCBDDCScatterCoarseDataEnd" 229612edc857SStefano Zampini PetscErrorCode PCBDDCScatterCoarseDataEnd(PC pc, InsertMode imode, ScatterMode smode) 2297674ae819SStefano Zampini { 2298674ae819SStefano Zampini PetscErrorCode ierr; 2299674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 230058da7f69SStefano Zampini PetscScalar *array; 230112edc857SStefano Zampini Vec from,to; 2302674ae819SStefano Zampini 2303674ae819SStefano Zampini PetscFunctionBegin; 230412edc857SStefano Zampini if (smode == SCATTER_REVERSE) { /* from global to local -> get data from coarse solution */ 230512edc857SStefano Zampini from = pcbddc->coarse_vec; 230612edc857SStefano Zampini to = pcbddc->vec1_P; 230712edc857SStefano Zampini } else { /* from local to global -> put data in coarse right hand side */ 230812edc857SStefano Zampini from = pcbddc->vec1_P; 230912edc857SStefano Zampini to = pcbddc->coarse_vec; 231012edc857SStefano Zampini } 231112edc857SStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,from,to,imode,smode);CHKERRQ(ierr); 231212edc857SStefano Zampini if (smode == SCATTER_FORWARD) { 231312edc857SStefano Zampini if (pcbddc->coarse_ksp) { /* get array from coarse processes */ 231412edc857SStefano Zampini Vec tvec; 231558da7f69SStefano Zampini 231612edc857SStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 231758da7f69SStefano Zampini ierr = VecGetArray(to,&array);CHKERRQ(ierr); 231858da7f69SStefano Zampini ierr = VecPlaceArray(tvec,array);CHKERRQ(ierr); 231958da7f69SStefano Zampini ierr = VecRestoreArray(to,&array);CHKERRQ(ierr); 232058da7f69SStefano Zampini } 232158da7f69SStefano Zampini } else { 232258da7f69SStefano Zampini if (pcbddc->coarse_ksp) { /* restore array of pcbddc->coarse_vec */ 232358da7f69SStefano Zampini ierr = VecResetArray(from);CHKERRQ(ierr); 232412edc857SStefano Zampini } 232512edc857SStefano Zampini } 2326674ae819SStefano Zampini PetscFunctionReturn(0); 2327674ae819SStefano Zampini } 2328674ae819SStefano Zampini 2329984c4197SStefano Zampini /* uncomment for testing purposes */ 2330984c4197SStefano Zampini /* #define PETSC_MISSING_LAPACK_GESVD 1 */ 2331674ae819SStefano Zampini #undef __FUNCT__ 2332674ae819SStefano Zampini #define __FUNCT__ "PCBDDCConstraintsSetUp" 2333674ae819SStefano Zampini PetscErrorCode PCBDDCConstraintsSetUp(PC pc) 2334674ae819SStefano Zampini { 2335674ae819SStefano Zampini PetscErrorCode ierr; 2336674ae819SStefano Zampini PC_IS* pcis = (PC_IS*)(pc->data); 2337674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 2338674ae819SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 2339984c4197SStefano Zampini /* one and zero */ 2340984c4197SStefano Zampini PetscScalar one=1.0,zero=0.0; 2341984c4197SStefano Zampini /* space to store constraints and their local indices */ 23429162d606SStefano Zampini PetscScalar *constraints_data; 23439162d606SStefano Zampini PetscInt *constraints_idxs,*constraints_idxs_B; 23449162d606SStefano Zampini PetscInt *constraints_idxs_ptr,*constraints_data_ptr; 23459162d606SStefano Zampini PetscInt *constraints_n; 2346984c4197SStefano Zampini /* iterators */ 2347b3d85658SStefano Zampini PetscInt i,j,k,total_counts,total_counts_cc,cum; 2348984c4197SStefano Zampini /* BLAS integers */ 2349e310c8b4SStefano Zampini PetscBLASInt lwork,lierr; 2350e310c8b4SStefano Zampini PetscBLASInt Blas_N,Blas_M,Blas_K,Blas_one=1; 2351c4303822SStefano Zampini PetscBLASInt Blas_LDA,Blas_LDB,Blas_LDC; 2352727cdba6SStefano Zampini /* reuse */ 23530e6343abSStefano Zampini PetscInt olocal_primal_size,olocal_primal_size_cc; 23540e6343abSStefano Zampini PetscInt *olocal_primal_ref_node,*olocal_primal_ref_mult; 2355984c4197SStefano Zampini /* change of basis */ 2356b3d85658SStefano Zampini PetscBool qr_needed; 23579162d606SStefano Zampini PetscBT change_basis,qr_needed_idx; 2358984c4197SStefano Zampini /* auxiliary stuff */ 235964efe560SStefano Zampini PetscInt *nnz,*is_indices; 23608a0068c3SStefano Zampini PetscInt ncc; 2361984c4197SStefano Zampini /* some quantities */ 236245a1bb75SStefano Zampini PetscInt n_vertices,total_primal_vertices,valid_constraints; 2363a58a30b4SStefano Zampini PetscInt size_of_constraint,max_size_of_constraint=0,max_constraints,temp_constraints; 2364984c4197SStefano Zampini 2365674ae819SStefano Zampini PetscFunctionBegin; 23668e61c736SStefano Zampini /* Destroy Mat objects computed previously */ 23678e61c736SStefano Zampini ierr = MatDestroy(&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 23688e61c736SStefano Zampini ierr = MatDestroy(&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 2369088faed8SStefano Zampini /* save info on constraints from previous setup (if any) */ 2370088faed8SStefano Zampini olocal_primal_size = pcbddc->local_primal_size; 23710e6343abSStefano Zampini olocal_primal_size_cc = pcbddc->local_primal_size_cc; 23720e6343abSStefano Zampini ierr = PetscMalloc2(olocal_primal_size_cc,&olocal_primal_ref_node,olocal_primal_size_cc,&olocal_primal_ref_mult);CHKERRQ(ierr); 23730e6343abSStefano Zampini ierr = PetscMemcpy(olocal_primal_ref_node,pcbddc->local_primal_ref_node,olocal_primal_size_cc*sizeof(PetscInt));CHKERRQ(ierr); 23740e6343abSStefano Zampini ierr = PetscMemcpy(olocal_primal_ref_mult,pcbddc->local_primal_ref_mult,olocal_primal_size_cc*sizeof(PetscInt));CHKERRQ(ierr); 23750e6343abSStefano Zampini ierr = PetscFree2(pcbddc->local_primal_ref_node,pcbddc->local_primal_ref_mult);CHKERRQ(ierr); 2376088faed8SStefano Zampini ierr = PetscFree(pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 2377cf5a6209SStefano Zampini 2378cf5a6209SStefano Zampini /* print some info */ 2379cf5a6209SStefano Zampini if (pcbddc->dbg_flag) { 2380cf5a6209SStefano Zampini IS vertices; 2381cf5a6209SStefano Zampini PetscInt nv,nedges,nfaces; 2382cf5a6209SStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,&nfaces,NULL,&nedges,NULL,&vertices);CHKERRQ(ierr); 2383cf5a6209SStefano Zampini ierr = ISGetSize(vertices,&nv);CHKERRQ(ierr); 2384cf5a6209SStefano Zampini ierr = ISDestroy(&vertices);CHKERRQ(ierr); 2385cf5a6209SStefano Zampini ierr = PetscViewerASCIISynchronizedAllow(pcbddc->dbg_viewer,PETSC_TRUE);CHKERRQ(ierr); 2386cf5a6209SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"--------------------------------------------------------------\n");CHKERRQ(ierr); 2387cf5a6209SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate vertices (%d)\n",PetscGlobalRank,nv,pcbddc->use_vertices);CHKERRQ(ierr); 2388fd14bc51SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate edges (%d)\n",PetscGlobalRank,nedges,pcbddc->use_edges);CHKERRQ(ierr); 2389fd14bc51SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate faces (%d)\n",PetscGlobalRank,nfaces,pcbddc->use_faces);CHKERRQ(ierr); 2390cf5a6209SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 2391cf5a6209SStefano Zampini } 2392cf5a6209SStefano Zampini 2393cf5a6209SStefano Zampini if (!pcbddc->adaptive_selection) { 23949162d606SStefano Zampini IS ISForVertices,*ISForFaces,*ISForEdges; 2395cf5a6209SStefano Zampini MatNullSpace nearnullsp; 2396cf5a6209SStefano Zampini const Vec *nearnullvecs; 2397cf5a6209SStefano Zampini Vec *localnearnullsp; 2398cf5a6209SStefano Zampini PetscScalar *array; 2399cf5a6209SStefano Zampini PetscInt n_ISForFaces,n_ISForEdges,nnsp_size; 2400cf5a6209SStefano Zampini PetscBool nnsp_has_cnst; 2401674ae819SStefano Zampini /* LAPACK working arrays for SVD or POD */ 2402b3d85658SStefano Zampini PetscBool skip_lapack,boolforchange; 2403674ae819SStefano Zampini PetscScalar *work; 2404674ae819SStefano Zampini PetscReal *singular_vals; 2405674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX) 2406674ae819SStefano Zampini PetscReal *rwork; 2407674ae819SStefano Zampini #endif 2408674ae819SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 2409674ae819SStefano Zampini PetscScalar *temp_basis,*correlation_mat; 2410674ae819SStefano Zampini #else 2411964fefecSStefano Zampini PetscBLASInt dummy_int=1; 2412964fefecSStefano Zampini PetscScalar dummy_scalar=1.; 2413674ae819SStefano Zampini #endif 2414674ae819SStefano Zampini 2415674ae819SStefano Zampini /* Get index sets for faces, edges and vertices from graph */ 2416d06fc5fdSStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,&n_ISForFaces,&ISForFaces,&n_ISForEdges,&ISForEdges,&ISForVertices);CHKERRQ(ierr); 2417d06fc5fdSStefano Zampini /* free unneeded index sets */ 2418d06fc5fdSStefano Zampini if (!pcbddc->use_vertices) { 2419d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForVertices);CHKERRQ(ierr); 2420674ae819SStefano Zampini } 2421d06fc5fdSStefano Zampini if (!pcbddc->use_edges) { 2422d06fc5fdSStefano Zampini for (i=0;i<n_ISForEdges;i++) { 2423d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForEdges[i]);CHKERRQ(ierr); 2424d06fc5fdSStefano Zampini } 2425d06fc5fdSStefano Zampini ierr = PetscFree(ISForEdges);CHKERRQ(ierr); 2426d06fc5fdSStefano Zampini n_ISForEdges = 0; 2427d06fc5fdSStefano Zampini } 2428d06fc5fdSStefano Zampini if (!pcbddc->use_faces) { 2429d06fc5fdSStefano Zampini for (i=0;i<n_ISForFaces;i++) { 2430d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForFaces[i]);CHKERRQ(ierr); 2431d06fc5fdSStefano Zampini } 2432d06fc5fdSStefano Zampini ierr = PetscFree(ISForFaces);CHKERRQ(ierr); 2433d06fc5fdSStefano Zampini n_ISForFaces = 0; 2434d06fc5fdSStefano Zampini } 243570022509SStefano Zampini 243670022509SStefano Zampini #if defined(PETSC_USE_DEBUG) 243770022509SStefano Zampini /* HACK: when solving singular problems not using vertices, a change of basis is mandatory. 243870022509SStefano Zampini Also use_change_of_basis should be consistent among processors */ 243970022509SStefano Zampini if (pcbddc->NullSpace) { 244070022509SStefano Zampini PetscBool tbool[2],gbool[2]; 244170022509SStefano Zampini 244270022509SStefano Zampini if (!ISForVertices && !pcbddc->user_ChangeOfBasisMatrix) { 2443b8ffe317SStefano Zampini pcbddc->use_change_of_basis = PETSC_TRUE; 2444d06fc5fdSStefano Zampini if (!ISForEdges) { 2445d06fc5fdSStefano Zampini pcbddc->use_change_on_faces = PETSC_TRUE; 2446d06fc5fdSStefano Zampini } 2447b8ffe317SStefano Zampini } 2448d06fc5fdSStefano Zampini tbool[0] = pcbddc->use_change_of_basis; 2449d06fc5fdSStefano Zampini tbool[1] = pcbddc->use_change_on_faces; 2450d06fc5fdSStefano Zampini ierr = MPI_Allreduce(tbool,gbool,2,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 2451d06fc5fdSStefano Zampini pcbddc->use_change_of_basis = gbool[0]; 2452d06fc5fdSStefano Zampini pcbddc->use_change_on_faces = gbool[1]; 245398a51de6SStefano Zampini } 245470022509SStefano Zampini #endif 245508122e43SStefano Zampini 2456674ae819SStefano Zampini /* check if near null space is attached to global mat */ 2457674ae819SStefano Zampini ierr = MatGetNearNullSpace(pc->pmat,&nearnullsp);CHKERRQ(ierr); 2458674ae819SStefano Zampini if (nearnullsp) { 2459674ae819SStefano Zampini ierr = MatNullSpaceGetVecs(nearnullsp,&nnsp_has_cnst,&nnsp_size,&nearnullvecs);CHKERRQ(ierr); 2460f4ddd8eeSStefano Zampini /* remove any stored info */ 2461f4ddd8eeSStefano Zampini ierr = MatNullSpaceDestroy(&pcbddc->onearnullspace);CHKERRQ(ierr); 2462f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->onearnullvecs_state);CHKERRQ(ierr); 2463f4ddd8eeSStefano Zampini /* store information for BDDC solver reuse */ 2464f4ddd8eeSStefano Zampini ierr = PetscObjectReference((PetscObject)nearnullsp);CHKERRQ(ierr); 2465f4ddd8eeSStefano Zampini pcbddc->onearnullspace = nearnullsp; 2466473ba861SJed Brown ierr = PetscMalloc1(nnsp_size,&pcbddc->onearnullvecs_state);CHKERRQ(ierr); 2467f4ddd8eeSStefano Zampini for (i=0;i<nnsp_size;i++) { 2468f4ddd8eeSStefano Zampini ierr = PetscObjectStateGet((PetscObject)nearnullvecs[i],&pcbddc->onearnullvecs_state[i]);CHKERRQ(ierr); 2469f4ddd8eeSStefano Zampini } 2470984c4197SStefano Zampini } else { /* if near null space is not provided BDDC uses constants by default */ 2471984c4197SStefano Zampini nnsp_size = 0; 2472674ae819SStefano Zampini nnsp_has_cnst = PETSC_TRUE; 2473674ae819SStefano Zampini } 2474984c4197SStefano Zampini /* get max number of constraints on a single cc */ 2475984c4197SStefano Zampini max_constraints = nnsp_size; 2476984c4197SStefano Zampini if (nnsp_has_cnst) max_constraints++; 2477984c4197SStefano Zampini 2478674ae819SStefano Zampini /* 2479674ae819SStefano Zampini Evaluate maximum storage size needed by the procedure 24809162d606SStefano Zampini - Indices for connected component i stored at "constraints_idxs + constraints_idxs_ptr[i]" 24819162d606SStefano Zampini - Values for constraints on connected component i stored at "constraints_data + constraints_data_ptr[i]" 24829162d606SStefano Zampini There can be multiple constraints per connected component 2483674ae819SStefano Zampini */ 2484674ae819SStefano Zampini n_vertices = 0; 2485674ae819SStefano Zampini if (ISForVertices) { 2486674ae819SStefano Zampini ierr = ISGetSize(ISForVertices,&n_vertices);CHKERRQ(ierr); 2487674ae819SStefano Zampini } 24889162d606SStefano Zampini ncc = n_vertices+n_ISForFaces+n_ISForEdges; 24899162d606SStefano Zampini ierr = PetscMalloc3(ncc+1,&constraints_idxs_ptr,ncc+1,&constraints_data_ptr,ncc,&constraints_n);CHKERRQ(ierr); 24909162d606SStefano Zampini 24919162d606SStefano Zampini total_counts = n_ISForFaces+n_ISForEdges; 24929162d606SStefano Zampini total_counts *= max_constraints; 2493674ae819SStefano Zampini total_counts += n_vertices; 24944641a718SStefano Zampini ierr = PetscBTCreate(total_counts,&change_basis);CHKERRQ(ierr); 24959162d606SStefano Zampini 2496674ae819SStefano Zampini total_counts = 0; 2497674ae819SStefano Zampini max_size_of_constraint = 0; 2498674ae819SStefano Zampini for (i=0;i<n_ISForEdges+n_ISForFaces;i++) { 24999162d606SStefano Zampini IS used_is; 2500674ae819SStefano Zampini if (i<n_ISForEdges) { 25019162d606SStefano Zampini used_is = ISForEdges[i]; 2502674ae819SStefano Zampini } else { 25039162d606SStefano Zampini used_is = ISForFaces[i-n_ISForEdges]; 2504674ae819SStefano Zampini } 25059162d606SStefano Zampini ierr = ISGetSize(used_is,&j);CHKERRQ(ierr); 2506674ae819SStefano Zampini total_counts += j; 2507674ae819SStefano Zampini max_size_of_constraint = PetscMax(j,max_size_of_constraint); 2508674ae819SStefano Zampini } 25099162d606SStefano 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); 25109162d606SStefano Zampini 2511984c4197SStefano Zampini /* get local part of global near null space vectors */ 2512785e854fSJed Brown ierr = PetscMalloc1(nnsp_size,&localnearnullsp);CHKERRQ(ierr); 2513984c4197SStefano Zampini for (k=0;k<nnsp_size;k++) { 2514984c4197SStefano Zampini ierr = VecDuplicate(pcis->vec1_N,&localnearnullsp[k]);CHKERRQ(ierr); 2515e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,nearnullvecs[k],localnearnullsp[k],INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2516e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,nearnullvecs[k],localnearnullsp[k],INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2517984c4197SStefano Zampini } 2518674ae819SStefano Zampini 2519242a89d7SStefano Zampini /* whether or not to skip lapack calls */ 2520242a89d7SStefano Zampini skip_lapack = PETSC_TRUE; 2521a773dcb8SStefano Zampini if (n_ISForFaces+n_ISForEdges && max_constraints > 1 && !pcbddc->use_nnsp_true) skip_lapack = PETSC_FALSE; 2522242a89d7SStefano Zampini 2523984c4197SStefano Zampini /* First we issue queries to allocate optimal workspace for LAPACKgesvd (or LAPACKsyev if SVD is missing) */ 2524a773dcb8SStefano Zampini if (!skip_lapack) { 2525674ae819SStefano Zampini PetscScalar temp_work; 2526911cabfeSStefano Zampini 2527674ae819SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 2528984c4197SStefano Zampini /* Proper Orthogonal Decomposition (POD) using the snapshot method */ 2529785e854fSJed Brown ierr = PetscMalloc1(max_constraints*max_constraints,&correlation_mat);CHKERRQ(ierr); 2530785e854fSJed Brown ierr = PetscMalloc1(max_constraints,&singular_vals);CHKERRQ(ierr); 2531785e854fSJed Brown ierr = PetscMalloc1(max_size_of_constraint*max_constraints,&temp_basis);CHKERRQ(ierr); 2532674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX) 2533785e854fSJed Brown ierr = PetscMalloc1(3*max_constraints,&rwork);CHKERRQ(ierr); 2534674ae819SStefano Zampini #endif 2535674ae819SStefano Zampini /* now we evaluate the optimal workspace using query with lwork=-1 */ 2536c8244a33SStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_N);CHKERRQ(ierr); 2537c8244a33SStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_LDA);CHKERRQ(ierr); 2538674ae819SStefano Zampini lwork = -1; 2539674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 2540674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 2541c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,&temp_work,&lwork,&lierr)); 2542674ae819SStefano Zampini #else 2543c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,&temp_work,&lwork,rwork,&lierr)); 2544674ae819SStefano Zampini #endif 2545674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2546984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to SYEV Lapack routine %d",(int)lierr); 2547674ae819SStefano Zampini #else /* on missing GESVD */ 2548674ae819SStefano Zampini /* SVD */ 2549674ae819SStefano Zampini PetscInt max_n,min_n; 2550674ae819SStefano Zampini max_n = max_size_of_constraint; 2551984c4197SStefano Zampini min_n = max_constraints; 2552984c4197SStefano Zampini if (max_size_of_constraint < max_constraints) { 2553674ae819SStefano Zampini min_n = max_size_of_constraint; 2554984c4197SStefano Zampini max_n = max_constraints; 2555674ae819SStefano Zampini } 2556785e854fSJed Brown ierr = PetscMalloc1(min_n,&singular_vals);CHKERRQ(ierr); 2557674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX) 2558785e854fSJed Brown ierr = PetscMalloc1(5*min_n,&rwork);CHKERRQ(ierr); 2559674ae819SStefano Zampini #endif 2560674ae819SStefano Zampini /* now we evaluate the optimal workspace using query with lwork=-1 */ 2561674ae819SStefano Zampini lwork = -1; 2562e310c8b4SStefano Zampini ierr = PetscBLASIntCast(max_n,&Blas_M);CHKERRQ(ierr); 2563e310c8b4SStefano Zampini ierr = PetscBLASIntCast(min_n,&Blas_N);CHKERRQ(ierr); 2564b7d8b9f8SStefano Zampini ierr = PetscBLASIntCast(max_n,&Blas_LDA);CHKERRQ(ierr); 2565674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 2566674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 25679162d606SStefano 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)); 2568674ae819SStefano Zampini #else 25699162d606SStefano 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)); 2570674ae819SStefano Zampini #endif 2571674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2572984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to GESVD Lapack routine %d",(int)lierr); 2573984c4197SStefano Zampini #endif /* on missing GESVD */ 2574674ae819SStefano Zampini /* Allocate optimal workspace */ 2575674ae819SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(temp_work),&lwork);CHKERRQ(ierr); 2576854ce69bSBarry Smith ierr = PetscMalloc1(lwork,&work);CHKERRQ(ierr); 2577674ae819SStefano Zampini } 2578674ae819SStefano Zampini /* Now we can loop on constraining sets */ 2579674ae819SStefano Zampini total_counts = 0; 25809162d606SStefano Zampini constraints_idxs_ptr[0] = 0; 25819162d606SStefano Zampini constraints_data_ptr[0] = 0; 2582674ae819SStefano Zampini /* vertices */ 25839162d606SStefano Zampini if (n_vertices) { 2584674ae819SStefano Zampini ierr = ISGetIndices(ISForVertices,(const PetscInt**)&is_indices);CHKERRQ(ierr); 25859162d606SStefano Zampini ierr = PetscMemcpy(constraints_idxs,is_indices,n_vertices*sizeof(PetscInt));CHKERRQ(ierr); 2586674ae819SStefano Zampini for (i=0;i<n_vertices;i++) { 25879162d606SStefano Zampini constraints_n[total_counts] = 1; 25889162d606SStefano Zampini constraints_data[total_counts] = 1.0; 25899162d606SStefano Zampini constraints_idxs_ptr[total_counts+1] = constraints_idxs_ptr[total_counts]+1; 25909162d606SStefano Zampini constraints_data_ptr[total_counts+1] = constraints_data_ptr[total_counts]+1; 2591674ae819SStefano Zampini total_counts++; 2592674ae819SStefano Zampini } 2593674ae819SStefano Zampini ierr = ISRestoreIndices(ISForVertices,(const PetscInt**)&is_indices);CHKERRQ(ierr); 2594674ae819SStefano Zampini n_vertices = total_counts; 2595674ae819SStefano Zampini } 2596984c4197SStefano Zampini 2597674ae819SStefano Zampini /* edges and faces */ 25989162d606SStefano Zampini total_counts_cc = total_counts; 2599911cabfeSStefano Zampini for (ncc=0;ncc<n_ISForEdges+n_ISForFaces;ncc++) { 26009162d606SStefano Zampini IS used_is; 26019162d606SStefano Zampini PetscBool idxs_copied = PETSC_FALSE; 26029162d606SStefano Zampini 2603911cabfeSStefano Zampini if (ncc<n_ISForEdges) { 26049162d606SStefano Zampini used_is = ISForEdges[ncc]; 2605984c4197SStefano Zampini boolforchange = pcbddc->use_change_of_basis; /* change or not the basis on the edge */ 2606674ae819SStefano Zampini } else { 26079162d606SStefano Zampini used_is = ISForFaces[ncc-n_ISForEdges]; 2608984c4197SStefano Zampini boolforchange = (PetscBool)(pcbddc->use_change_of_basis && pcbddc->use_change_on_faces); /* change or not the basis on the face */ 2609674ae819SStefano Zampini } 2610674ae819SStefano Zampini temp_constraints = 0; /* zero the number of constraints I have on this conn comp */ 26119162d606SStefano Zampini 26129162d606SStefano Zampini ierr = ISGetSize(used_is,&size_of_constraint);CHKERRQ(ierr); 26139162d606SStefano Zampini ierr = ISGetIndices(used_is,(const PetscInt**)&is_indices);CHKERRQ(ierr); 2614984c4197SStefano Zampini /* change of basis should not be performed on local periodic nodes */ 2615984c4197SStefano Zampini if (pcbddc->mat_graph->mirrors && pcbddc->mat_graph->mirrors[is_indices[0]]) boolforchange = PETSC_FALSE; 2616674ae819SStefano Zampini if (nnsp_has_cnst) { 26175b08dc53SStefano Zampini PetscScalar quad_value; 26189162d606SStefano Zampini 26199162d606SStefano Zampini ierr = PetscMemcpy(constraints_idxs + constraints_idxs_ptr[total_counts_cc],is_indices,size_of_constraint*sizeof(PetscInt));CHKERRQ(ierr); 26209162d606SStefano Zampini idxs_copied = PETSC_TRUE; 26219162d606SStefano Zampini 2622a773dcb8SStefano Zampini if (!pcbddc->use_nnsp_true) { 2623674ae819SStefano Zampini quad_value = (PetscScalar)(1.0/PetscSqrtReal((PetscReal)size_of_constraint)); 2624a773dcb8SStefano Zampini } else { 2625a773dcb8SStefano Zampini quad_value = 1.0; 2626a773dcb8SStefano Zampini } 2627674ae819SStefano Zampini for (j=0;j<size_of_constraint;j++) { 26289162d606SStefano Zampini constraints_data[constraints_data_ptr[total_counts_cc]+j] = quad_value; 2629674ae819SStefano Zampini } 26309162d606SStefano Zampini temp_constraints++; 2631674ae819SStefano Zampini total_counts++; 2632674ae819SStefano Zampini } 2633674ae819SStefano Zampini for (k=0;k<nnsp_size;k++) { 2634984c4197SStefano Zampini PetscReal real_value; 26359162d606SStefano Zampini PetscScalar *ptr_to_data; 26369162d606SStefano Zampini 2637984c4197SStefano Zampini ierr = VecGetArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 26389162d606SStefano Zampini ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]+temp_constraints*size_of_constraint]; 2639674ae819SStefano Zampini for (j=0;j<size_of_constraint;j++) { 26409162d606SStefano Zampini ptr_to_data[j] = array[is_indices[j]]; 2641674ae819SStefano Zampini } 2642984c4197SStefano Zampini ierr = VecRestoreArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 2643984c4197SStefano Zampini /* check if array is null on the connected component */ 2644e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 26459162d606SStefano Zampini PetscStackCallBLAS("BLASasum",real_value = BLASasum_(&Blas_N,ptr_to_data,&Blas_one)); 26465b08dc53SStefano Zampini if (real_value > 0.0) { /* keep indices and values */ 2647674ae819SStefano Zampini temp_constraints++; 2648674ae819SStefano Zampini total_counts++; 26499162d606SStefano Zampini if (!idxs_copied) { 26509162d606SStefano Zampini ierr = PetscMemcpy(constraints_idxs + constraints_idxs_ptr[total_counts_cc],is_indices,size_of_constraint*sizeof(PetscInt));CHKERRQ(ierr); 26519162d606SStefano Zampini idxs_copied = PETSC_TRUE; 2652674ae819SStefano Zampini } 2653674ae819SStefano Zampini } 26549162d606SStefano Zampini } 26559162d606SStefano Zampini ierr = ISRestoreIndices(used_is,(const PetscInt**)&is_indices);CHKERRQ(ierr); 265645a1bb75SStefano Zampini valid_constraints = temp_constraints; 2657eb97c9d2SStefano Zampini if (!pcbddc->use_nnsp_true && temp_constraints) { 2658a773dcb8SStefano Zampini if (temp_constraints == 1) { /* just normalize the constraint */ 26599162d606SStefano Zampini PetscScalar norm,*ptr_to_data; 26609162d606SStefano Zampini 26619162d606SStefano Zampini ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]]; 2662a773dcb8SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 26639162d606SStefano Zampini PetscStackCallBLAS("BLASdot",norm = BLASdot_(&Blas_N,ptr_to_data,&Blas_one,ptr_to_data,&Blas_one)); 2664a773dcb8SStefano Zampini norm = 1.0/PetscSqrtReal(PetscRealPart(norm)); 26659162d606SStefano Zampini PetscStackCallBLAS("BLASscal",BLASscal_(&Blas_N,&norm,ptr_to_data,&Blas_one)); 2666a773dcb8SStefano Zampini } else { /* perform SVD */ 2667984c4197SStefano Zampini PetscReal tol = 1.0e-8; /* tolerance for retaining eigenmodes */ 26689162d606SStefano Zampini PetscScalar *ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]]; 2669674ae819SStefano Zampini 2670674ae819SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 2671984c4197SStefano Zampini /* SVD: Y = U*S*V^H -> U (eigenvectors of Y*Y^H) = Y*V*(S)^\dag 2672984c4197SStefano Zampini POD: Y^H*Y = V*D*V^H, D = S^H*S -> U = Y*V*D^(-1/2) 2673984c4197SStefano Zampini -> When PETSC_USE_COMPLEX and PETSC_MISSING_LAPACK_GESVD are defined 2674984c4197SStefano Zampini the constraints basis will differ (by a complex factor with absolute value equal to 1) 2675984c4197SStefano Zampini from that computed using LAPACKgesvd 2676984c4197SStefano Zampini -> This is due to a different computation of eigenvectors in LAPACKheev 2677984c4197SStefano Zampini -> The quality of the POD-computed basis will be the same */ 2678984c4197SStefano Zampini ierr = PetscMemzero(correlation_mat,temp_constraints*temp_constraints*sizeof(PetscScalar));CHKERRQ(ierr); 2679674ae819SStefano Zampini /* Store upper triangular part of correlation matrix */ 2680e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 2681984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 2682674ae819SStefano Zampini for (j=0;j<temp_constraints;j++) { 2683674ae819SStefano Zampini for (k=0;k<j+1;k++) { 26849162d606SStefano 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)); 2685674ae819SStefano Zampini } 2686674ae819SStefano Zampini } 2687e310c8b4SStefano Zampini /* compute eigenvalues and eigenvectors of correlation matrix */ 2688e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 2689e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_LDA);CHKERRQ(ierr); 2690674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 2691c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,work,&lwork,&lierr)); 2692674ae819SStefano Zampini #else 2693c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,work,&lwork,rwork,&lierr)); 2694674ae819SStefano Zampini #endif 2695674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2696984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYEV Lapack routine %d",(int)lierr); 2697984c4197SStefano Zampini /* retain eigenvalues greater than tol: note that LAPACKsyev gives eigs in ascending order */ 2698674ae819SStefano Zampini j = 0; 2699984c4197SStefano Zampini while (j < temp_constraints && singular_vals[j] < tol) j++; 2700674ae819SStefano Zampini total_counts = total_counts-j; 270145a1bb75SStefano Zampini valid_constraints = temp_constraints-j; 2702e310c8b4SStefano Zampini /* scale and copy POD basis into used quadrature memory */ 2703c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 2704c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 2705c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_K);CHKERRQ(ierr); 2706c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 2707c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_LDB);CHKERRQ(ierr); 2708c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDC);CHKERRQ(ierr); 2709674ae819SStefano Zampini if (j<temp_constraints) { 2710984c4197SStefano Zampini PetscInt ii; 2711984c4197SStefano Zampini for (k=j;k<temp_constraints;k++) singular_vals[k] = 1.0/PetscSqrtReal(singular_vals[k]); 2712674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 27139162d606SStefano 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)); 2714674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2715984c4197SStefano Zampini for (k=0;k<temp_constraints-j;k++) { 2716674ae819SStefano Zampini for (ii=0;ii<size_of_constraint;ii++) { 27179162d606SStefano 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]; 2718674ae819SStefano Zampini } 2719674ae819SStefano Zampini } 2720674ae819SStefano Zampini } 2721674ae819SStefano Zampini #else /* on missing GESVD */ 2722e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 2723e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 2724b7d8b9f8SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 2725674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 2726674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 27279162d606SStefano 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)); 2728674ae819SStefano Zampini #else 27299162d606SStefano 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)); 2730674ae819SStefano Zampini #endif 2731984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GESVD Lapack routine %d",(int)lierr); 2732674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2733984c4197SStefano Zampini /* retain eigenvalues greater than tol: note that LAPACKgesvd gives eigs in descending order */ 2734e310c8b4SStefano Zampini k = temp_constraints; 2735e310c8b4SStefano Zampini if (k > size_of_constraint) k = size_of_constraint; 2736674ae819SStefano Zampini j = 0; 2737e310c8b4SStefano Zampini while (j < k && singular_vals[k-j-1] < tol) j++; 273845a1bb75SStefano Zampini valid_constraints = k-j; 2739911cabfeSStefano Zampini total_counts = total_counts-temp_constraints+valid_constraints; 2740984c4197SStefano Zampini #endif /* on missing GESVD */ 2741674ae819SStefano Zampini } 2742a773dcb8SStefano Zampini } 27439162d606SStefano Zampini /* update pointers information */ 27449162d606SStefano Zampini if (valid_constraints) { 27459162d606SStefano Zampini constraints_n[total_counts_cc] = valid_constraints; 27469162d606SStefano Zampini constraints_idxs_ptr[total_counts_cc+1] = constraints_idxs_ptr[total_counts_cc]+size_of_constraint; 27479162d606SStefano Zampini constraints_data_ptr[total_counts_cc+1] = constraints_data_ptr[total_counts_cc]+size_of_constraint*valid_constraints; 27489162d606SStefano Zampini /* set change_of_basis flag */ 274945a1bb75SStefano Zampini if (boolforchange) { 2750b3d85658SStefano Zampini PetscBTSet(change_basis,total_counts_cc); 27519162d606SStefano Zampini } 2752b3d85658SStefano Zampini total_counts_cc++; 275345a1bb75SStefano Zampini } 275445a1bb75SStefano Zampini } 2755984c4197SStefano Zampini /* free workspace */ 27568f1c130eSStefano Zampini if (!skip_lapack) { 2757984c4197SStefano Zampini ierr = PetscFree(work);CHKERRQ(ierr); 2758984c4197SStefano Zampini #if defined(PETSC_USE_COMPLEX) 2759984c4197SStefano Zampini ierr = PetscFree(rwork);CHKERRQ(ierr); 2760984c4197SStefano Zampini #endif 2761984c4197SStefano Zampini ierr = PetscFree(singular_vals);CHKERRQ(ierr); 2762984c4197SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 2763984c4197SStefano Zampini ierr = PetscFree(correlation_mat);CHKERRQ(ierr); 2764984c4197SStefano Zampini ierr = PetscFree(temp_basis);CHKERRQ(ierr); 2765984c4197SStefano Zampini #endif 2766984c4197SStefano Zampini } 2767984c4197SStefano Zampini for (k=0;k<nnsp_size;k++) { 2768984c4197SStefano Zampini ierr = VecDestroy(&localnearnullsp[k]);CHKERRQ(ierr); 2769984c4197SStefano Zampini } 2770984c4197SStefano Zampini ierr = PetscFree(localnearnullsp);CHKERRQ(ierr); 2771cf5a6209SStefano Zampini /* free index sets of faces, edges and vertices */ 2772cf5a6209SStefano Zampini for (i=0;i<n_ISForFaces;i++) { 2773cf5a6209SStefano Zampini ierr = ISDestroy(&ISForFaces[i]);CHKERRQ(ierr); 2774cf5a6209SStefano Zampini } 2775cf5a6209SStefano Zampini if (n_ISForFaces) { 2776cf5a6209SStefano Zampini ierr = PetscFree(ISForFaces);CHKERRQ(ierr); 2777cf5a6209SStefano Zampini } 2778cf5a6209SStefano Zampini for (i=0;i<n_ISForEdges;i++) { 2779cf5a6209SStefano Zampini ierr = ISDestroy(&ISForEdges[i]);CHKERRQ(ierr); 2780cf5a6209SStefano Zampini } 2781cf5a6209SStefano Zampini if (n_ISForEdges) { 2782cf5a6209SStefano Zampini ierr = PetscFree(ISForEdges);CHKERRQ(ierr); 2783cf5a6209SStefano Zampini } 2784cf5a6209SStefano Zampini ierr = ISDestroy(&ISForVertices);CHKERRQ(ierr); 278508122e43SStefano Zampini } else { 278608122e43SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 2787984c4197SStefano Zampini 278808122e43SStefano Zampini total_counts = 0; 278908122e43SStefano Zampini n_vertices = 0; 2790d62866d3SStefano Zampini if (sub_schurs->is_vertices && pcbddc->use_vertices) { 2791d62866d3SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_vertices,&n_vertices);CHKERRQ(ierr); 279208122e43SStefano Zampini } 279308122e43SStefano Zampini max_constraints = 0; 27949162d606SStefano Zampini total_counts_cc = 0; 279508122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs+n_vertices;i++) { 279608122e43SStefano Zampini total_counts += pcbddc->adaptive_constraints_n[i]; 27979162d606SStefano Zampini if (pcbddc->adaptive_constraints_n[i]) total_counts_cc++; 279808122e43SStefano Zampini max_constraints = PetscMax(max_constraints,pcbddc->adaptive_constraints_n[i]); 279908122e43SStefano Zampini } 28009162d606SStefano Zampini constraints_idxs_ptr = pcbddc->adaptive_constraints_idxs_ptr; 28019162d606SStefano Zampini constraints_data_ptr = pcbddc->adaptive_constraints_data_ptr; 28029162d606SStefano Zampini constraints_idxs = pcbddc->adaptive_constraints_idxs; 28039162d606SStefano Zampini constraints_data = pcbddc->adaptive_constraints_data; 280474d5cdf7SStefano Zampini /* constraints_n differs from pcbddc->adaptive_constraints_n */ 28059162d606SStefano Zampini ierr = PetscMalloc1(total_counts_cc,&constraints_n);CHKERRQ(ierr); 28069162d606SStefano Zampini total_counts_cc = 0; 28079162d606SStefano Zampini for (i=0;i<sub_schurs->n_subs+n_vertices;i++) { 28089162d606SStefano Zampini if (pcbddc->adaptive_constraints_n[i]) { 28099162d606SStefano Zampini constraints_n[total_counts_cc++] = pcbddc->adaptive_constraints_n[i]; 281008122e43SStefano Zampini } 281108122e43SStefano Zampini } 28129162d606SStefano Zampini #if 0 28139162d606SStefano Zampini printf("Found %d totals (%d)\n",total_counts_cc,total_counts); 28149162d606SStefano Zampini for (i=0;i<total_counts_cc;i++) { 28159162d606SStefano Zampini printf("const %d, start %d",i,constraints_idxs_ptr[i]); 28169162d606SStefano Zampini printf(" end %d:\n",constraints_idxs_ptr[i+1]); 28179162d606SStefano Zampini for (j=constraints_idxs_ptr[i];j<constraints_idxs_ptr[i+1];j++) { 28189162d606SStefano Zampini printf(" %d",constraints_idxs[j]); 28199162d606SStefano Zampini } 28209162d606SStefano Zampini printf("\n"); 28219162d606SStefano Zampini printf("number of cc: %d\n",constraints_n[i]); 28229162d606SStefano Zampini } 28231b968477SStefano Zampini for (i=0;i<n_vertices;i++) { 28248bec7fa6SStefano Zampini PetscPrintf(PETSC_COMM_SELF,"[%d] vertex %d, n %d\n",PetscGlobalRank,i,pcbddc->adaptive_constraints_n[i]); 28251b968477SStefano Zampini } 28261b968477SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 28278bec7fa6SStefano 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]); 28281b968477SStefano Zampini } 282908122e43SStefano Zampini #endif 283008122e43SStefano Zampini 28318bec7fa6SStefano Zampini max_size_of_constraint = 0; 28329162d606SStefano 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]); 28339162d606SStefano Zampini ierr = PetscMalloc1(constraints_idxs_ptr[total_counts_cc],&constraints_idxs_B);CHKERRQ(ierr); 283408122e43SStefano Zampini /* Change of basis */ 2835b3d85658SStefano Zampini ierr = PetscBTCreate(total_counts_cc,&change_basis);CHKERRQ(ierr); 283608122e43SStefano Zampini if (pcbddc->use_change_of_basis) { 283708122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 283808122e43SStefano Zampini if (PetscBTLookup(sub_schurs->is_edge,i) || pcbddc->use_change_on_faces) { 2839b3d85658SStefano Zampini ierr = PetscBTSet(change_basis,i+n_vertices);CHKERRQ(ierr); 284008122e43SStefano Zampini } 284108122e43SStefano Zampini } 284208122e43SStefano Zampini } 284308122e43SStefano Zampini } 2844984c4197SStefano Zampini pcbddc->local_primal_size = total_counts; 2845019a44ceSStefano Zampini /* allocating one extra space (in case an extra primal dof should be stored for the benign trick */ 2846019a44ceSStefano Zampini ierr = PetscMalloc1(pcbddc->local_primal_size+1,&pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 284708122e43SStefano Zampini 28489162d606SStefano Zampini /* map constraints_idxs in boundary numbering */ 28499162d606SStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,constraints_idxs_ptr[total_counts_cc],constraints_idxs,&i,constraints_idxs_B);CHKERRQ(ierr); 28509162d606SStefano Zampini if (i != constraints_idxs_ptr[total_counts_cc]) { 28519162d606SStefano Zampini SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_SUP,"Error in boundary numbering for constraints indices %d != %d\n",constraints_idxs_ptr[total_counts_cc],i); 285208122e43SStefano Zampini } 2853674ae819SStefano Zampini 2854674ae819SStefano Zampini /* Create constraint matrix */ 2855674ae819SStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 285616f15bc4SStefano Zampini ierr = MatSetType(pcbddc->ConstraintMatrix,MATAIJ);CHKERRQ(ierr); 2857984c4197SStefano Zampini ierr = MatSetSizes(pcbddc->ConstraintMatrix,pcbddc->local_primal_size,pcis->n,pcbddc->local_primal_size,pcis->n);CHKERRQ(ierr); 2858984c4197SStefano Zampini 2859984c4197SStefano Zampini /* find primal_dofs: subdomain corners plus dofs selected as primal after change of basis */ 2860a717540cSStefano Zampini /* determine if a QR strategy is needed for change of basis */ 2861a717540cSStefano Zampini qr_needed = PETSC_FALSE; 286274d5cdf7SStefano Zampini ierr = PetscBTCreate(total_counts_cc,&qr_needed_idx);CHKERRQ(ierr); 2863984c4197SStefano Zampini total_primal_vertices=0; 2864b3d85658SStefano Zampini pcbddc->local_primal_size_cc = 0; 28659162d606SStefano Zampini for (i=0;i<total_counts_cc;i++) { 28669162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 2867984c4197SStefano Zampini if (size_of_constraint == 1) { 28689162d606SStefano Zampini pcbddc->primal_indices_local_idxs[total_primal_vertices++] = constraints_idxs[constraints_idxs_ptr[i]]; 2869b3d85658SStefano Zampini pcbddc->local_primal_size_cc += 1; 287064efe560SStefano Zampini } else if (PetscBTLookup(change_basis,i)) { 28719162d606SStefano Zampini for (k=0;k<constraints_n[i];k++) { 28729162d606SStefano Zampini pcbddc->primal_indices_local_idxs[total_primal_vertices++] = constraints_idxs[constraints_idxs_ptr[i]+k]; 2873a717540cSStefano Zampini } 2874b3d85658SStefano Zampini pcbddc->local_primal_size_cc += constraints_n[i]; 287574d5cdf7SStefano Zampini if (constraints_n[i] > 1 || pcbddc->use_qr_single || pcbddc->faster_deluxe) { 2876a717540cSStefano Zampini PetscBTSet(qr_needed_idx,i); 2877a717540cSStefano Zampini qr_needed = PETSC_TRUE; 2878a717540cSStefano Zampini } 2879fa434743SStefano Zampini } else { 2880b3d85658SStefano Zampini pcbddc->local_primal_size_cc += 1; 2881fa434743SStefano Zampini } 2882a717540cSStefano Zampini } 2883b371cd4fSStefano Zampini /* note that the local variable n_vertices used below stores the number of pointwise constraints */ 2884b371cd4fSStefano Zampini pcbddc->n_vertices = total_primal_vertices; 2885674ae819SStefano Zampini /* permute indices in order to have a sorted set of vertices */ 288670022509SStefano Zampini ierr = PetscSortInt(total_primal_vertices,pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 2887b3d85658SStefano Zampini 2888019a44ceSStefano Zampini /* allocating one extra space (in case an extra primal dof should be stored for the benign trick */ 2889019a44ceSStefano Zampini ierr = PetscMalloc2(pcbddc->local_primal_size_cc+1,&pcbddc->local_primal_ref_node,pcbddc->local_primal_size_cc+1,&pcbddc->local_primal_ref_mult);CHKERRQ(ierr); 28900e6343abSStefano Zampini ierr = PetscMemcpy(pcbddc->local_primal_ref_node,pcbddc->primal_indices_local_idxs,total_primal_vertices*sizeof(PetscInt));CHKERRQ(ierr); 28910e6343abSStefano Zampini for (i=0;i<total_primal_vertices;i++) pcbddc->local_primal_ref_mult[i] = 1; 2892984c4197SStefano Zampini 2893984c4197SStefano Zampini /* nonzero structure of constraint matrix */ 289474d5cdf7SStefano Zampini /* and get reference dof for local constraints */ 2895785e854fSJed Brown ierr = PetscMalloc1(pcbddc->local_primal_size,&nnz);CHKERRQ(ierr); 2896984c4197SStefano Zampini for (i=0;i<total_primal_vertices;i++) nnz[i] = 1; 289774d5cdf7SStefano Zampini 2898984c4197SStefano Zampini j = total_primal_vertices; 289974d5cdf7SStefano Zampini total_counts = total_primal_vertices; 2900b3d85658SStefano Zampini cum = total_primal_vertices; 29019162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 29024641a718SStefano Zampini if (!PetscBTLookup(change_basis,i)) { 2903b3d85658SStefano Zampini pcbddc->local_primal_ref_node[cum] = constraints_idxs[constraints_idxs_ptr[i]]; 2904b3d85658SStefano Zampini pcbddc->local_primal_ref_mult[cum] = constraints_n[i]; 2905b3d85658SStefano Zampini cum++; 29069162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 290774d5cdf7SStefano Zampini for (k=0;k<constraints_n[i];k++) { 290874d5cdf7SStefano Zampini pcbddc->primal_indices_local_idxs[total_counts++] = constraints_idxs[constraints_idxs_ptr[i]+k]; 290974d5cdf7SStefano Zampini nnz[j+k] = size_of_constraint; 291074d5cdf7SStefano Zampini } 29119162d606SStefano Zampini j += constraints_n[i]; 2912674ae819SStefano Zampini } 2913674ae819SStefano Zampini } 2914674ae819SStefano Zampini ierr = MatSeqAIJSetPreallocation(pcbddc->ConstraintMatrix,0,nnz);CHKERRQ(ierr); 2915674ae819SStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 2916088faed8SStefano Zampini 2917674ae819SStefano Zampini /* set values in constraint matrix */ 2918984c4197SStefano Zampini for (i=0;i<total_primal_vertices;i++) { 29190e6343abSStefano Zampini ierr = MatSetValue(pcbddc->ConstraintMatrix,i,pcbddc->local_primal_ref_node[i],1.0,INSERT_VALUES);CHKERRQ(ierr); 2920674ae819SStefano Zampini } 2921984c4197SStefano Zampini total_counts = total_primal_vertices; 29229162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 29234641a718SStefano Zampini if (!PetscBTLookup(change_basis,i)) { 29249162d606SStefano Zampini PetscInt *cols; 29259162d606SStefano Zampini 29269162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 29279162d606SStefano Zampini cols = constraints_idxs+constraints_idxs_ptr[i]; 29289162d606SStefano Zampini for (k=0;k<constraints_n[i];k++) { 29299162d606SStefano Zampini PetscInt row = total_counts+k; 29309162d606SStefano Zampini PetscScalar *vals; 29319162d606SStefano Zampini 29329162d606SStefano Zampini vals = constraints_data+constraints_data_ptr[i]+k*size_of_constraint; 29339162d606SStefano Zampini ierr = MatSetValues(pcbddc->ConstraintMatrix,1,&row,size_of_constraint,cols,vals,INSERT_VALUES);CHKERRQ(ierr); 29349162d606SStefano Zampini } 29359162d606SStefano Zampini total_counts += constraints_n[i]; 2936674ae819SStefano Zampini } 2937674ae819SStefano Zampini } 2938674ae819SStefano Zampini /* assembling */ 2939674ae819SStefano Zampini ierr = MatAssemblyBegin(pcbddc->ConstraintMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2940674ae819SStefano Zampini ierr = MatAssemblyEnd(pcbddc->ConstraintMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2941088faed8SStefano Zampini 2942984c4197SStefano Zampini /* 294345a1bb75SStefano Zampini ierr = PetscViewerSetFormat(PETSC_VIEWER_STDOUT_SELF,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 2944984c4197SStefano Zampini ierr = MatView(pcbddc->ConstraintMatrix,(PetscViewer)0);CHKERRQ(ierr); 2945984c4197SStefano Zampini */ 2946674ae819SStefano Zampini /* Create matrix for change of basis. We don't need it in case pcbddc->use_change_of_basis is FALSE */ 2947674ae819SStefano Zampini if (pcbddc->use_change_of_basis) { 2948026de310SStefano Zampini /* dual and primal dofs on a single cc */ 2949984c4197SStefano Zampini PetscInt dual_dofs,primal_dofs; 2950984c4197SStefano Zampini /* working stuff for GEQRF */ 295181d9aea3SBarry Smith PetscScalar *qr_basis,*qr_tau = NULL,*qr_work,lqr_work_t; 2952984c4197SStefano Zampini PetscBLASInt lqr_work; 2953984c4197SStefano Zampini /* working stuff for UNGQR */ 2954984c4197SStefano Zampini PetscScalar *gqr_work,lgqr_work_t; 2955984c4197SStefano Zampini PetscBLASInt lgqr_work; 2956984c4197SStefano Zampini /* working stuff for TRTRS */ 2957984c4197SStefano Zampini PetscScalar *trs_rhs; 29583f08241aSStefano Zampini PetscBLASInt Blas_NRHS; 2959984c4197SStefano Zampini /* pointers for values insertion into change of basis matrix */ 2960984c4197SStefano Zampini PetscInt *start_rows,*start_cols; 2961984c4197SStefano Zampini PetscScalar *start_vals; 2962984c4197SStefano Zampini /* working stuff for values insertion */ 29634641a718SStefano Zampini PetscBT is_primal; 296464efe560SStefano Zampini PetscInt *aux_primal_numbering_B; 2965906d46d4SStefano Zampini /* matrix sizes */ 2966906d46d4SStefano Zampini PetscInt global_size,local_size; 2967906d46d4SStefano Zampini /* temporary change of basis */ 2968906d46d4SStefano Zampini Mat localChangeOfBasisMatrix; 2969cf5a6209SStefano Zampini /* extra space for debugging */ 2970cf5a6209SStefano Zampini PetscScalar *dbg_work; 2971984c4197SStefano Zampini 2972906d46d4SStefano Zampini /* local temporary change of basis acts on local interfaces -> dimension is n_B x n_B */ 2973906d46d4SStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&localChangeOfBasisMatrix);CHKERRQ(ierr); 297416f15bc4SStefano Zampini ierr = MatSetType(localChangeOfBasisMatrix,MATAIJ);CHKERRQ(ierr); 2975bbb9e6c6SStefano Zampini ierr = MatSetSizes(localChangeOfBasisMatrix,pcis->n,pcis->n,pcis->n,pcis->n);CHKERRQ(ierr); 2976906d46d4SStefano Zampini /* nonzeros for local mat */ 2977bbb9e6c6SStefano Zampini ierr = PetscMalloc1(pcis->n,&nnz);CHKERRQ(ierr); 2978bbb9e6c6SStefano Zampini for (i=0;i<pcis->n;i++) nnz[i]=1; 29799162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 2980a717540cSStefano Zampini if (PetscBTLookup(change_basis,i)) { 29819162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 2982a717540cSStefano Zampini if (PetscBTLookup(qr_needed_idx,i)) { 29839162d606SStefano Zampini for (j=0;j<size_of_constraint;j++) nnz[constraints_idxs[constraints_idxs_ptr[i]+j]] = size_of_constraint; 2984a717540cSStefano Zampini } else { 29859162d606SStefano Zampini nnz[constraints_idxs[constraints_idxs_ptr[i]]] = size_of_constraint; 29869162d606SStefano Zampini for (j=1;j<size_of_constraint;j++) nnz[constraints_idxs[constraints_idxs_ptr[i]+j]] = 2; 2987a717540cSStefano Zampini } 2988a717540cSStefano Zampini } 2989a717540cSStefano Zampini } 2990906d46d4SStefano Zampini ierr = MatSeqAIJSetPreallocation(localChangeOfBasisMatrix,0,nnz);CHKERRQ(ierr); 2991bbb9e6c6SStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 2992a717540cSStefano Zampini /* Set initial identity in the matrix */ 2993bbb9e6c6SStefano Zampini for (i=0;i<pcis->n;i++) { 2994906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,i,i,1.0,INSERT_VALUES);CHKERRQ(ierr); 2995a717540cSStefano Zampini } 2996a717540cSStefano Zampini 2997a717540cSStefano Zampini if (pcbddc->dbg_flag) { 2998a717540cSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"--------------------------------------------------------------\n");CHKERRQ(ierr); 2999a717540cSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Checking change of basis computation for subdomain %04d\n",PetscGlobalRank);CHKERRQ(ierr); 3000a717540cSStefano Zampini } 3001a717540cSStefano Zampini 3002a717540cSStefano Zampini 3003a717540cSStefano Zampini /* Now we loop on the constraints which need a change of basis */ 3004a717540cSStefano Zampini /* 3005a717540cSStefano Zampini Change of basis matrix is evaluated similarly to the FIRST APPROACH in 3006a717540cSStefano Zampini Klawonn and Widlund, Dual-primal FETI-DP methods for linear elasticity, (see Sect 6.2.1) 3007a717540cSStefano Zampini 3008a6b551f4SStefano Zampini Basic blocks of change of basis matrix T computed by 3009a717540cSStefano Zampini 3010a6b551f4SStefano 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) 3011a6b551f4SStefano Zampini 3012a6b551f4SStefano Zampini | 1 0 ... 0 s_1/S | 3013a6b551f4SStefano Zampini | 0 1 ... 0 s_2/S | 3014a717540cSStefano Zampini | ... | 3015a6b551f4SStefano Zampini | 0 ... 1 s_{n-1}/S | 3016a6b551f4SStefano Zampini | -s_1/s_n ... -s_{n-1}/s_n s_n/S | 3017a717540cSStefano Zampini 3018a6b551f4SStefano Zampini with S = \sum_{i=1}^n s_i^2 3019a6b551f4SStefano Zampini NOTE: in the above example, the primal dof is the last one of the edge in LOCAL ordering 3020a6b551f4SStefano Zampini in the current implementation, the primal dof is the first one of the edge in GLOBAL ordering 3021a6b551f4SStefano Zampini 3022a6b551f4SStefano Zampini - QR decomposition of constraints otherwise 3023a717540cSStefano Zampini */ 3024a717540cSStefano Zampini if (qr_needed) { 3025984c4197SStefano Zampini /* space to store Q */ 3026854ce69bSBarry Smith ierr = PetscMalloc1(max_size_of_constraint*max_size_of_constraint,&qr_basis);CHKERRQ(ierr); 3027984c4197SStefano Zampini /* first we issue queries for optimal work */ 30283f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_M);CHKERRQ(ierr); 30293f08241aSStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_N);CHKERRQ(ierr); 30303f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 3031984c4197SStefano Zampini lqr_work = -1; 30323f08241aSStefano Zampini PetscStackCallBLAS("LAPACKgeqrf",LAPACKgeqrf_(&Blas_M,&Blas_N,qr_basis,&Blas_LDA,qr_tau,&lqr_work_t,&lqr_work,&lierr)); 3033984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to GEQRF Lapack routine %d",(int)lierr); 3034984c4197SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lqr_work_t),&lqr_work);CHKERRQ(ierr); 3035785e854fSJed Brown ierr = PetscMalloc1((PetscInt)PetscRealPart(lqr_work_t),&qr_work);CHKERRQ(ierr); 3036984c4197SStefano Zampini lgqr_work = -1; 30373f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_M);CHKERRQ(ierr); 30383f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_N);CHKERRQ(ierr); 30393f08241aSStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_K);CHKERRQ(ierr); 30403f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 30413f08241aSStefano Zampini if (Blas_K>Blas_M) Blas_K=Blas_M; /* adjust just for computing optimal work */ 30423f08241aSStefano Zampini PetscStackCallBLAS("LAPACKungqr",LAPACKungqr_(&Blas_M,&Blas_N,&Blas_K,qr_basis,&Blas_LDA,qr_tau,&lgqr_work_t,&lgqr_work,&lierr)); 3043984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to UNGQR Lapack routine %d",(int)lierr); 3044984c4197SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lgqr_work_t),&lgqr_work);CHKERRQ(ierr); 3045785e854fSJed Brown ierr = PetscMalloc1((PetscInt)PetscRealPart(lgqr_work_t),&gqr_work);CHKERRQ(ierr); 3046984c4197SStefano Zampini /* array to store scaling factors for reflectors */ 3047785e854fSJed Brown ierr = PetscMalloc1(max_constraints,&qr_tau);CHKERRQ(ierr); 3048984c4197SStefano Zampini /* array to store rhs and solution of triangular solver */ 3049785e854fSJed Brown ierr = PetscMalloc1(max_constraints*max_constraints,&trs_rhs);CHKERRQ(ierr); 3050a717540cSStefano Zampini /* allocating workspace for check */ 3051a717540cSStefano Zampini if (pcbddc->dbg_flag) { 3052cf5a6209SStefano Zampini ierr = PetscMalloc1(max_size_of_constraint*(max_constraints+max_size_of_constraint),&dbg_work);CHKERRQ(ierr); 3053a717540cSStefano Zampini } 3054a717540cSStefano Zampini } 3055984c4197SStefano Zampini /* array to store whether a node is primal or not */ 30564641a718SStefano Zampini ierr = PetscBTCreate(pcis->n_B,&is_primal);CHKERRQ(ierr); 3057473ba861SJed Brown ierr = PetscMalloc1(total_primal_vertices,&aux_primal_numbering_B);CHKERRQ(ierr); 30580e6343abSStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,total_primal_vertices,pcbddc->local_primal_ref_node,&i,aux_primal_numbering_B);CHKERRQ(ierr); 305939e2fb2aSStefano Zampini if (i != total_primal_vertices) { 306039e2fb2aSStefano Zampini SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_SUP,"Error in boundary numbering for BDDC vertices! %d != %d\n",total_primal_vertices,i); 30614641a718SStefano Zampini } 306239e2fb2aSStefano Zampini for (i=0;i<total_primal_vertices;i++) { 306339e2fb2aSStefano Zampini ierr = PetscBTSet(is_primal,aux_primal_numbering_B[i]);CHKERRQ(ierr); 306439e2fb2aSStefano Zampini } 306539e2fb2aSStefano Zampini ierr = PetscFree(aux_primal_numbering_B);CHKERRQ(ierr); 3066984c4197SStefano Zampini 3067a717540cSStefano Zampini /* loop on constraints and see whether or not they need a change of basis and compute it */ 30689162d606SStefano Zampini for (total_counts=n_vertices;total_counts<total_counts_cc;total_counts++) { 30699162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[total_counts+1]-constraints_idxs_ptr[total_counts]; 30704641a718SStefano Zampini if (PetscBTLookup(change_basis,total_counts)) { 3071984c4197SStefano Zampini /* get constraint info */ 30729162d606SStefano Zampini primal_dofs = constraints_n[total_counts]; 3073984c4197SStefano Zampini dual_dofs = size_of_constraint-primal_dofs; 3074984c4197SStefano Zampini 3075984c4197SStefano Zampini if (pcbddc->dbg_flag) { 30769162d606SStefano 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); 3077674ae819SStefano Zampini } 3078984c4197SStefano Zampini 3079fa434743SStefano Zampini if (PetscBTLookup(qr_needed_idx,total_counts)) { /* QR */ 3080a717540cSStefano Zampini 3081a717540cSStefano Zampini /* copy quadrature constraints for change of basis check */ 3082a717540cSStefano Zampini if (pcbddc->dbg_flag) { 30839162d606SStefano Zampini ierr = PetscMemcpy(dbg_work,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 3084a717540cSStefano Zampini } 3085984c4197SStefano Zampini /* copy temporary constraints into larger work vector (in order to store all columns of Q) */ 30869162d606SStefano Zampini ierr = PetscMemcpy(qr_basis,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 3087984c4197SStefano Zampini 3088984c4197SStefano Zampini /* compute QR decomposition of constraints */ 30893f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 30903f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 30913f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 3092674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 30933f08241aSStefano Zampini PetscStackCallBLAS("LAPACKgeqrf",LAPACKgeqrf_(&Blas_M,&Blas_N,qr_basis,&Blas_LDA,qr_tau,qr_work,&lqr_work,&lierr)); 3094984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GEQRF Lapack routine %d",(int)lierr); 3095674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 3096984c4197SStefano Zampini 3097984c4197SStefano Zampini /* explictly compute R^-T */ 3098984c4197SStefano Zampini ierr = PetscMemzero(trs_rhs,primal_dofs*primal_dofs*sizeof(*trs_rhs));CHKERRQ(ierr); 3099984c4197SStefano Zampini for (j=0;j<primal_dofs;j++) trs_rhs[j*(primal_dofs+1)] = 1.0; 31003f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 31013f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_NRHS);CHKERRQ(ierr); 31023f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 31033f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDB);CHKERRQ(ierr); 3104984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 31053f08241aSStefano Zampini PetscStackCallBLAS("LAPACKtrtrs",LAPACKtrtrs_("U","T","N",&Blas_N,&Blas_NRHS,qr_basis,&Blas_LDA,trs_rhs,&Blas_LDB,&lierr)); 3106984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in TRTRS Lapack routine %d",(int)lierr); 3107984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 3108984c4197SStefano Zampini 3109a717540cSStefano Zampini /* explicitly compute all columns of Q (Q = [Q1 | Q2] ) overwriting QR factorization in qr_basis */ 31103f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 31113f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 31123f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_K);CHKERRQ(ierr); 31133f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 3114984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 31153f08241aSStefano Zampini PetscStackCallBLAS("LAPACKungqr",LAPACKungqr_(&Blas_M,&Blas_N,&Blas_K,qr_basis,&Blas_LDA,qr_tau,gqr_work,&lgqr_work,&lierr)); 3116984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in UNGQR Lapack routine %d",(int)lierr); 3117984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 3118984c4197SStefano Zampini 3119984c4197SStefano Zampini /* first primal_dofs columns of Q need to be re-scaled in order to be unitary w.r.t constraints 3120984c4197SStefano Zampini i.e. C_{pxn}*Q_{nxn} should be equal to [I_pxp | 0_pxd] (see check below) 3121984c4197SStefano Zampini where n=size_of_constraint, p=primal_dofs, d=dual_dofs (n=p+d), I and 0 identity and null matrix resp. */ 31223f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 31233f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 31243f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_K);CHKERRQ(ierr); 31253f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 31263f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDB);CHKERRQ(ierr); 31273f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDC);CHKERRQ(ierr); 3128984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 31299162d606SStefano 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)); 3130984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 31319162d606SStefano Zampini ierr = PetscMemcpy(qr_basis,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 3132984c4197SStefano Zampini 3133984c4197SStefano Zampini /* insert values in change of basis matrix respecting global ordering of new primal dofs */ 31349162d606SStefano Zampini start_rows = &constraints_idxs[constraints_idxs_ptr[total_counts]]; 3135984c4197SStefano Zampini /* insert cols for primal dofs */ 3136984c4197SStefano Zampini for (j=0;j<primal_dofs;j++) { 3137984c4197SStefano Zampini start_vals = &qr_basis[j*size_of_constraint]; 31389162d606SStefano Zampini start_cols = &constraints_idxs[constraints_idxs_ptr[total_counts]+j]; 3139906d46d4SStefano Zampini ierr = MatSetValues(localChangeOfBasisMatrix,size_of_constraint,start_rows,1,start_cols,start_vals,INSERT_VALUES);CHKERRQ(ierr); 3140984c4197SStefano Zampini } 3141984c4197SStefano Zampini /* insert cols for dual dofs */ 3142984c4197SStefano Zampini for (j=0,k=0;j<dual_dofs;k++) { 31439162d606SStefano Zampini if (!PetscBTLookup(is_primal,constraints_idxs_B[constraints_idxs_ptr[total_counts]+k])) { 3144984c4197SStefano Zampini start_vals = &qr_basis[(primal_dofs+j)*size_of_constraint]; 31459162d606SStefano Zampini start_cols = &constraints_idxs[constraints_idxs_ptr[total_counts]+k]; 3146906d46d4SStefano Zampini ierr = MatSetValues(localChangeOfBasisMatrix,size_of_constraint,start_rows,1,start_cols,start_vals,INSERT_VALUES);CHKERRQ(ierr); 3147984c4197SStefano Zampini j++; 3148674ae819SStefano Zampini } 3149674ae819SStefano Zampini } 3150984c4197SStefano Zampini 3151984c4197SStefano Zampini /* check change of basis */ 3152984c4197SStefano Zampini if (pcbddc->dbg_flag) { 3153984c4197SStefano Zampini PetscInt ii,jj; 3154984c4197SStefano Zampini PetscBool valid_qr=PETSC_TRUE; 3155c4303822SStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_M);CHKERRQ(ierr); 3156c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 3157c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_K);CHKERRQ(ierr); 3158c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 3159c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDB);CHKERRQ(ierr); 3160c4303822SStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDC);CHKERRQ(ierr); 3161984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 3162cf5a6209SStefano 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)); 3163984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 3164984c4197SStefano Zampini for (jj=0;jj<size_of_constraint;jj++) { 3165984c4197SStefano Zampini for (ii=0;ii<primal_dofs;ii++) { 3166cf5a6209SStefano Zampini if (ii != jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]) > 1.e-12) valid_qr = PETSC_FALSE; 3167cf5a6209SStefano 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; 3168674ae819SStefano Zampini } 3169674ae819SStefano Zampini } 3170984c4197SStefano Zampini if (!valid_qr) { 317122d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> wrong change of basis!\n");CHKERRQ(ierr); 3172984c4197SStefano Zampini for (jj=0;jj<size_of_constraint;jj++) { 3173984c4197SStefano Zampini for (ii=0;ii<primal_dofs;ii++) { 3174cf5a6209SStefano Zampini if (ii != jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]) > 1.e-12) { 3175cf5a6209SStefano 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])); 3176674ae819SStefano Zampini } 3177cf5a6209SStefano Zampini if (ii == jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]-1.0) > 1.e-12) { 3178cf5a6209SStefano 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])); 3179984c4197SStefano Zampini } 3180984c4197SStefano Zampini } 3181984c4197SStefano Zampini } 3182674ae819SStefano Zampini } else { 318322d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> right change of basis!\n");CHKERRQ(ierr); 3184674ae819SStefano Zampini } 3185674ae819SStefano Zampini } 3186a717540cSStefano Zampini } else { /* simple transformation block */ 3187a717540cSStefano Zampini PetscInt row,col; 3188a6b551f4SStefano Zampini PetscScalar val,norm; 3189a6b551f4SStefano Zampini 3190a6b551f4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 31919162d606SStefano 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)); 3192a717540cSStefano Zampini for (j=0;j<size_of_constraint;j++) { 31939162d606SStefano Zampini PetscInt row_B = constraints_idxs_B[constraints_idxs_ptr[total_counts]+j]; 31949162d606SStefano Zampini row = constraints_idxs[constraints_idxs_ptr[total_counts]+j]; 3195bbb9e6c6SStefano Zampini if (!PetscBTLookup(is_primal,row_B)) { 31969162d606SStefano Zampini col = constraints_idxs[constraints_idxs_ptr[total_counts]]; 3197906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,row,1.0,INSERT_VALUES);CHKERRQ(ierr); 31989162d606SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,col,constraints_data[constraints_data_ptr[total_counts]+j]/norm,INSERT_VALUES);CHKERRQ(ierr); 3199a717540cSStefano Zampini } else { 3200a717540cSStefano Zampini for (k=0;k<size_of_constraint;k++) { 32019162d606SStefano Zampini col = constraints_idxs[constraints_idxs_ptr[total_counts]+k]; 3202a717540cSStefano Zampini if (row != col) { 32039162d606SStefano Zampini val = -constraints_data[constraints_data_ptr[total_counts]+k]/constraints_data[constraints_data_ptr[total_counts]]; 3204a717540cSStefano Zampini } else { 32059162d606SStefano Zampini val = constraints_data[constraints_data_ptr[total_counts]]/norm; 3206a717540cSStefano Zampini } 3207906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,col,val,INSERT_VALUES);CHKERRQ(ierr); 3208a717540cSStefano Zampini } 3209a717540cSStefano Zampini } 3210a717540cSStefano Zampini } 321198a51de6SStefano Zampini if (pcbddc->dbg_flag) { 321222d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> using standard change of basis\n");CHKERRQ(ierr); 3213a717540cSStefano Zampini } 3214674ae819SStefano Zampini } 3215984c4197SStefano Zampini } else { 3216984c4197SStefano Zampini if (pcbddc->dbg_flag) { 32179162d606SStefano 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); 3218674ae819SStefano Zampini } 3219674ae819SStefano Zampini } 3220674ae819SStefano Zampini } 3221a717540cSStefano Zampini 3222a717540cSStefano Zampini /* free workspace */ 3223a717540cSStefano Zampini if (qr_needed) { 3224984c4197SStefano Zampini if (pcbddc->dbg_flag) { 3225cf5a6209SStefano Zampini ierr = PetscFree(dbg_work);CHKERRQ(ierr); 3226984c4197SStefano Zampini } 3227984c4197SStefano Zampini ierr = PetscFree(trs_rhs);CHKERRQ(ierr); 3228984c4197SStefano Zampini ierr = PetscFree(qr_tau);CHKERRQ(ierr); 3229984c4197SStefano Zampini ierr = PetscFree(qr_work);CHKERRQ(ierr); 3230984c4197SStefano Zampini ierr = PetscFree(gqr_work);CHKERRQ(ierr); 3231984c4197SStefano Zampini ierr = PetscFree(qr_basis);CHKERRQ(ierr); 3232674ae819SStefano Zampini } 3233a717540cSStefano Zampini ierr = PetscBTDestroy(&is_primal);CHKERRQ(ierr); 3234906d46d4SStefano Zampini ierr = MatAssemblyBegin(localChangeOfBasisMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3235906d46d4SStefano Zampini ierr = MatAssemblyEnd(localChangeOfBasisMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3236906d46d4SStefano Zampini 3237906d46d4SStefano Zampini /* assembling of global change of variable */ 3238bbb9e6c6SStefano Zampini { 3239bbb9e6c6SStefano Zampini Mat tmat; 324016f15bc4SStefano Zampini PetscInt bs; 324116f15bc4SStefano Zampini 3242906d46d4SStefano Zampini ierr = VecGetSize(pcis->vec1_global,&global_size);CHKERRQ(ierr); 3243906d46d4SStefano Zampini ierr = VecGetLocalSize(pcis->vec1_global,&local_size);CHKERRQ(ierr); 3244bbb9e6c6SStefano Zampini ierr = MatDuplicate(pc->pmat,MAT_DO_NOT_COPY_VALUES,&tmat);CHKERRQ(ierr); 3245bbb9e6c6SStefano Zampini ierr = MatISSetLocalMat(tmat,localChangeOfBasisMatrix);CHKERRQ(ierr); 3246bbb9e6c6SStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)pc),&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 3247bbb9e6c6SStefano Zampini ierr = MatSetType(pcbddc->ChangeOfBasisMatrix,MATAIJ);CHKERRQ(ierr); 324816f15bc4SStefano Zampini ierr = MatGetBlockSize(pc->pmat,&bs);CHKERRQ(ierr); 324916f15bc4SStefano Zampini ierr = MatSetBlockSize(pcbddc->ChangeOfBasisMatrix,bs);CHKERRQ(ierr); 3250906d46d4SStefano Zampini ierr = MatSetSizes(pcbddc->ChangeOfBasisMatrix,local_size,local_size,global_size,global_size);CHKERRQ(ierr); 3251bbb9e6c6SStefano Zampini ierr = MatISSetMPIXAIJPreallocation_Private(tmat,pcbddc->ChangeOfBasisMatrix,PETSC_TRUE);CHKERRQ(ierr); 3252bbb9e6c6SStefano Zampini ierr = MatISGetMPIXAIJ(tmat,MAT_REUSE_MATRIX,&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 3253bbb9e6c6SStefano Zampini ierr = MatDestroy(&tmat);CHKERRQ(ierr); 3254bbb9e6c6SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 3255bbb9e6c6SStefano Zampini ierr = VecSet(pcis->vec1_N,1.0);CHKERRQ(ierr); 3256e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 3257e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 3258bbb9e6c6SStefano Zampini ierr = VecReciprocal(pcis->vec1_global);CHKERRQ(ierr); 3259bbb9e6c6SStefano Zampini ierr = MatDiagonalScale(pcbddc->ChangeOfBasisMatrix,pcis->vec1_global,NULL);CHKERRQ(ierr); 3260906d46d4SStefano Zampini } 3261906d46d4SStefano Zampini /* check */ 3262906d46d4SStefano Zampini if (pcbddc->dbg_flag) { 3263906d46d4SStefano Zampini PetscReal error; 3264906d46d4SStefano Zampini Vec x,x_change; 3265906d46d4SStefano Zampini 3266906d46d4SStefano Zampini ierr = VecDuplicate(pcis->vec1_global,&x);CHKERRQ(ierr); 3267906d46d4SStefano Zampini ierr = VecDuplicate(pcis->vec1_global,&x_change);CHKERRQ(ierr); 3268906d46d4SStefano Zampini ierr = VecSetRandom(x,NULL);CHKERRQ(ierr); 3269906d46d4SStefano Zampini ierr = VecCopy(x,pcis->vec1_global);CHKERRQ(ierr); 3270e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,x,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 3271e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,x,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 3272bbb9e6c6SStefano Zampini ierr = MatMult(localChangeOfBasisMatrix,pcis->vec1_N,pcis->vec2_N);CHKERRQ(ierr); 3273e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec2_N,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 3274e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec2_N,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 3275906d46d4SStefano Zampini ierr = MatMult(pcbddc->ChangeOfBasisMatrix,pcis->vec1_global,x_change);CHKERRQ(ierr); 3276906d46d4SStefano Zampini ierr = VecAXPY(x,-1.0,x_change);CHKERRQ(ierr); 3277906d46d4SStefano Zampini ierr = VecNorm(x,NORM_INFINITY,&error);CHKERRQ(ierr); 3278906d46d4SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 3279bbb9e6c6SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Error global vs local change: %1.6e\n",error);CHKERRQ(ierr); 3280906d46d4SStefano Zampini ierr = VecDestroy(&x);CHKERRQ(ierr); 3281906d46d4SStefano Zampini ierr = VecDestroy(&x_change);CHKERRQ(ierr); 3282906d46d4SStefano Zampini } 3283b96c3477SStefano Zampini 3284b96c3477SStefano Zampini /* adapt sub_schurs computed (if any) */ 3285b96c3477SStefano Zampini if (pcbddc->use_deluxe_scaling) { 3286b96c3477SStefano Zampini PCBDDCSubSchurs sub_schurs=pcbddc->sub_schurs; 3287b96c3477SStefano Zampini if (sub_schurs->S_Ej_all) { 3288ac632422SStefano Zampini Mat S_new,tmat; 3289b087196eSStefano Zampini ISLocalToGlobalMapping NtoSall; 3290b087196eSStefano Zampini IS is_all_N,is_V,is_V_Sall; 3291b087196eSStefano Zampini const PetscScalar *array; 3292b087196eSStefano Zampini const PetscInt *idxs_V,*idxs_all; 3293b087196eSStefano Zampini PetscInt i,n_V; 3294bbb9e6c6SStefano Zampini 3295bbb9e6c6SStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(pcis->BtoNmap,sub_schurs->is_Ej_all,&is_all_N);CHKERRQ(ierr); 32966816873aSStefano Zampini ierr = MatGetSubMatrix(localChangeOfBasisMatrix,is_all_N,is_all_N,MAT_INITIAL_MATRIX,&tmat);CHKERRQ(ierr); 3297b087196eSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,pcbddc->n_vertices,pcbddc->local_primal_ref_node,PETSC_COPY_VALUES,&is_V);CHKERRQ(ierr); 3298b087196eSStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(is_all_N,&NtoSall);CHKERRQ(ierr); 3299b087196eSStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(NtoSall,IS_GTOLM_DROP,is_V,&is_V_Sall);CHKERRQ(ierr); 3300b087196eSStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&NtoSall);CHKERRQ(ierr); 3301bbb9e6c6SStefano Zampini ierr = ISDestroy(&is_all_N);CHKERRQ(ierr); 3302b087196eSStefano Zampini ierr = ISDestroy(&is_V);CHKERRQ(ierr); 3303ac632422SStefano Zampini ierr = MatPtAP(sub_schurs->S_Ej_all,tmat,MAT_INITIAL_MATRIX,1.0,&S_new);CHKERRQ(ierr); 3304b96c3477SStefano Zampini ierr = MatDestroy(&sub_schurs->S_Ej_all);CHKERRQ(ierr); 3305ac632422SStefano Zampini ierr = PetscObjectReference((PetscObject)S_new);CHKERRQ(ierr); 3306b087196eSStefano Zampini ierr = MatZeroRowsColumnsIS(S_new,is_V_Sall,1.,NULL,NULL);CHKERRQ(ierr); 3307b087196eSStefano Zampini ierr = ISGetLocalSize(is_V_Sall,&n_V);CHKERRQ(ierr); 3308b087196eSStefano Zampini ierr = ISGetIndices(is_V_Sall,&idxs_V);CHKERRQ(ierr); 3309b087196eSStefano Zampini ierr = ISGetIndices(sub_schurs->is_Ej_all,&idxs_all);CHKERRQ(ierr); 3310b087196eSStefano Zampini ierr = VecGetArrayRead(pcis->D,&array);CHKERRQ(ierr); 3311b087196eSStefano Zampini for (i=0;i<n_V;i++) { 3312b087196eSStefano Zampini PetscScalar val; 3313b087196eSStefano Zampini PetscInt idx; 3314b087196eSStefano Zampini 3315b087196eSStefano Zampini idx = idxs_V[i]; 3316b087196eSStefano Zampini val = array[idxs_all[idxs_V[i]]]; 3317b087196eSStefano Zampini ierr = MatSetValue(S_new,idx,idx,val,INSERT_VALUES);CHKERRQ(ierr); 3318b087196eSStefano Zampini } 3319b087196eSStefano Zampini ierr = MatAssemblyBegin(S_new,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3320b087196eSStefano Zampini ierr = MatAssemblyEnd(S_new,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3321ac632422SStefano Zampini sub_schurs->S_Ej_all = S_new; 3322ac632422SStefano Zampini ierr = MatDestroy(&S_new);CHKERRQ(ierr); 3323ac632422SStefano Zampini if (sub_schurs->sum_S_Ej_all) { 3324ac632422SStefano Zampini ierr = MatPtAP(sub_schurs->sum_S_Ej_all,tmat,MAT_INITIAL_MATRIX,1.0,&S_new);CHKERRQ(ierr); 3325b96c3477SStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_all);CHKERRQ(ierr); 3326ac632422SStefano Zampini ierr = PetscObjectReference((PetscObject)S_new);CHKERRQ(ierr); 3327b087196eSStefano Zampini ierr = MatZeroRowsColumnsIS(S_new,is_V_Sall,1.,NULL,NULL);CHKERRQ(ierr); 3328ac632422SStefano Zampini sub_schurs->sum_S_Ej_all = S_new; 3329ac632422SStefano Zampini ierr = MatDestroy(&S_new);CHKERRQ(ierr); 3330ac632422SStefano Zampini } 3331b087196eSStefano Zampini ierr = VecRestoreArrayRead(pcis->D,&array);CHKERRQ(ierr); 3332b087196eSStefano Zampini ierr = ISRestoreIndices(sub_schurs->is_Ej_all,&idxs_all);CHKERRQ(ierr); 3333b087196eSStefano Zampini ierr = ISRestoreIndices(is_V_Sall,&idxs_V);CHKERRQ(ierr); 3334b96c3477SStefano Zampini ierr = MatDestroy(&tmat);CHKERRQ(ierr); 3335b087196eSStefano Zampini ierr = ISDestroy(&is_V_Sall);CHKERRQ(ierr); 3336b96c3477SStefano Zampini } 3337b96c3477SStefano Zampini } 3338906d46d4SStefano Zampini ierr = MatDestroy(&localChangeOfBasisMatrix);CHKERRQ(ierr); 3339906d46d4SStefano Zampini } else if (pcbddc->user_ChangeOfBasisMatrix) { 3340b9b85e73SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->user_ChangeOfBasisMatrix);CHKERRQ(ierr); 3341b9b85e73SStefano Zampini pcbddc->ChangeOfBasisMatrix = pcbddc->user_ChangeOfBasisMatrix; 3342b9b85e73SStefano Zampini } 3343906d46d4SStefano Zampini 3344906d46d4SStefano Zampini /* set up change of basis context */ 3345906d46d4SStefano Zampini if (pcbddc->ChangeOfBasisMatrix) { 3346906d46d4SStefano Zampini PCBDDCChange_ctx change_ctx; 3347906d46d4SStefano Zampini 3348906d46d4SStefano Zampini if (!pcbddc->new_global_mat) { 3349906d46d4SStefano Zampini PetscInt global_size,local_size; 3350906d46d4SStefano Zampini 3351906d46d4SStefano Zampini ierr = VecGetSize(pcis->vec1_global,&global_size);CHKERRQ(ierr); 3352906d46d4SStefano Zampini ierr = VecGetLocalSize(pcis->vec1_global,&local_size);CHKERRQ(ierr); 3353906d46d4SStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)pc),&pcbddc->new_global_mat);CHKERRQ(ierr); 3354906d46d4SStefano Zampini ierr = MatSetSizes(pcbddc->new_global_mat,local_size,local_size,global_size,global_size);CHKERRQ(ierr); 3355906d46d4SStefano Zampini ierr = MatSetType(pcbddc->new_global_mat,MATSHELL);CHKERRQ(ierr); 3356906d46d4SStefano Zampini ierr = MatShellSetOperation(pcbddc->new_global_mat,MATOP_MULT,(void (*)(void))PCBDDCMatMult_Private);CHKERRQ(ierr); 3357906d46d4SStefano Zampini ierr = MatShellSetOperation(pcbddc->new_global_mat,MATOP_MULT_TRANSPOSE,(void (*)(void))PCBDDCMatMultTranspose_Private);CHKERRQ(ierr); 3358906d46d4SStefano Zampini ierr = PetscNew(&change_ctx);CHKERRQ(ierr); 3359906d46d4SStefano Zampini ierr = MatShellSetContext(pcbddc->new_global_mat,change_ctx);CHKERRQ(ierr); 3360906d46d4SStefano Zampini } else { 3361906d46d4SStefano Zampini ierr = MatShellGetContext(pcbddc->new_global_mat,&change_ctx);CHKERRQ(ierr); 3362906d46d4SStefano Zampini ierr = MatDestroy(&change_ctx->global_change);CHKERRQ(ierr); 3363906d46d4SStefano Zampini ierr = VecDestroyVecs(2,&change_ctx->work);CHKERRQ(ierr); 3364906d46d4SStefano Zampini } 3365906d46d4SStefano Zampini if (!pcbddc->user_ChangeOfBasisMatrix) { 3366906d46d4SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 3367906d46d4SStefano Zampini change_ctx->global_change = pcbddc->ChangeOfBasisMatrix; 3368906d46d4SStefano Zampini } else { 3369906d46d4SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->user_ChangeOfBasisMatrix);CHKERRQ(ierr); 3370906d46d4SStefano Zampini change_ctx->global_change = pcbddc->user_ChangeOfBasisMatrix; 3371906d46d4SStefano Zampini } 3372906d46d4SStefano Zampini ierr = VecDuplicateVecs(pcis->vec1_global,2,&change_ctx->work);CHKERRQ(ierr); 3373906d46d4SStefano Zampini ierr = MatSetUp(pcbddc->new_global_mat);CHKERRQ(ierr); 3374906d46d4SStefano Zampini ierr = MatAssemblyBegin(pcbddc->new_global_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3375906d46d4SStefano Zampini ierr = MatAssemblyEnd(pcbddc->new_global_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3376b9b85e73SStefano Zampini } 3377a717540cSStefano Zampini 3378019a44ceSStefano Zampini /* add pressure dof to set of primal nodes for numbering purposes */ 3379c9ed8603SStefano Zampini if (pcbddc->benign_p0_lidx >= 0) { 3380c9ed8603SStefano Zampini pcbddc->local_primal_ref_node[pcbddc->local_primal_size_cc] = pcbddc->benign_p0_lidx; 3381d16cbb6bSStefano Zampini pcbddc->primal_indices_local_idxs[pcbddc->local_primal_size] = pcbddc->benign_p0_lidx; 3382019a44ceSStefano Zampini pcbddc->local_primal_ref_mult[pcbddc->local_primal_size_cc] = 1; 3383019a44ceSStefano Zampini pcbddc->local_primal_size_cc++; 3384019a44ceSStefano Zampini pcbddc->local_primal_size++; 3385019a44ceSStefano Zampini } 3386019a44ceSStefano Zampini 3387019a44ceSStefano Zampini /* check if a new primal space has been introduced (also take into account benign trick) */ 3388727cdba6SStefano Zampini pcbddc->new_primal_space_local = PETSC_TRUE; 3389727cdba6SStefano Zampini if (olocal_primal_size == pcbddc->local_primal_size) { 3390aff50787SStefano Zampini ierr = PetscMemcmp(pcbddc->local_primal_ref_node,olocal_primal_ref_node,olocal_primal_size_cc*sizeof(PetscScalar),&pcbddc->new_primal_space_local);CHKERRQ(ierr); 3391c1c8e736SStefano Zampini pcbddc->new_primal_space_local = (PetscBool)(!pcbddc->new_primal_space_local); 33920e6343abSStefano Zampini if (!pcbddc->new_primal_space_local) { 3393aff50787SStefano Zampini ierr = PetscMemcmp(pcbddc->local_primal_ref_mult,olocal_primal_ref_mult,olocal_primal_size_cc*sizeof(PetscScalar),&pcbddc->new_primal_space_local);CHKERRQ(ierr); 3394727cdba6SStefano Zampini pcbddc->new_primal_space_local = (PetscBool)(!pcbddc->new_primal_space_local); 3395727cdba6SStefano Zampini } 33960e6343abSStefano Zampini } 33970e6343abSStefano Zampini ierr = PetscFree2(olocal_primal_ref_node,olocal_primal_ref_mult);CHKERRQ(ierr); 3398727cdba6SStefano Zampini /* new_primal_space will be used for numbering of coarse dofs, so it should be the same across all subdomains */ 3399727cdba6SStefano Zampini ierr = MPI_Allreduce(&pcbddc->new_primal_space_local,&pcbddc->new_primal_space,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 3400727cdba6SStefano Zampini 3401a717540cSStefano Zampini /* flush dbg viewer */ 3402b8ffe317SStefano Zampini if (pcbddc->dbg_flag) { 3403b8ffe317SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 3404b8ffe317SStefano Zampini } 3405a717540cSStefano Zampini 3406e310c8b4SStefano Zampini /* free workspace */ 3407a717540cSStefano Zampini ierr = PetscBTDestroy(&qr_needed_idx);CHKERRQ(ierr); 34084641a718SStefano Zampini ierr = PetscBTDestroy(&change_basis);CHKERRQ(ierr); 340908122e43SStefano Zampini if (!pcbddc->adaptive_selection) { 34109162d606SStefano Zampini ierr = PetscFree3(constraints_idxs_ptr,constraints_data_ptr,constraints_n);CHKERRQ(ierr); 34119162d606SStefano Zampini ierr = PetscFree3(constraints_data,constraints_idxs,constraints_idxs_B);CHKERRQ(ierr); 341208122e43SStefano Zampini } else { 34139162d606SStefano Zampini ierr = PetscFree5(pcbddc->adaptive_constraints_n, 34149162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr, 34159162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr, 341608122e43SStefano Zampini pcbddc->adaptive_constraints_idxs, 341708122e43SStefano Zampini pcbddc->adaptive_constraints_data);CHKERRQ(ierr); 34189162d606SStefano Zampini ierr = PetscFree(constraints_n);CHKERRQ(ierr); 34199162d606SStefano Zampini ierr = PetscFree(constraints_idxs_B);CHKERRQ(ierr); 342008122e43SStefano Zampini } 3421674ae819SStefano Zampini PetscFunctionReturn(0); 3422674ae819SStefano Zampini } 3423674ae819SStefano Zampini 3424674ae819SStefano Zampini #undef __FUNCT__ 3425674ae819SStefano Zampini #define __FUNCT__ "PCBDDCAnalyzeInterface" 3426674ae819SStefano Zampini PetscErrorCode PCBDDCAnalyzeInterface(PC pc) 3427674ae819SStefano Zampini { 3428674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 3429674ae819SStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 3430674ae819SStefano Zampini Mat_IS *matis = (Mat_IS*)pc->pmat->data; 34317fb0e2dbSStefano Zampini PetscInt ierr,i,vertex_size,N; 3432674ae819SStefano Zampini PetscViewer viewer=pcbddc->dbg_viewer; 3433674ae819SStefano Zampini 3434674ae819SStefano Zampini PetscFunctionBegin; 34358e61c736SStefano Zampini /* Reset previously computed graph */ 34368e61c736SStefano Zampini ierr = PCBDDCGraphReset(pcbddc->mat_graph);CHKERRQ(ierr); 3437674ae819SStefano Zampini /* Init local Graph struct */ 34387fb0e2dbSStefano Zampini ierr = MatGetSize(pc->pmat,&N,NULL);CHKERRQ(ierr); 34393bbff08aSStefano Zampini ierr = PCBDDCGraphInit(pcbddc->mat_graph,pcis->mapping,N);CHKERRQ(ierr); 3440674ae819SStefano Zampini 3441575ad6abSStefano Zampini /* Check validity of the csr graph passed in by the user */ 34425099eff2SStefano Zampini if (pcbddc->mat_graph->nvtxs_csr && pcbddc->mat_graph->nvtxs_csr != pcbddc->mat_graph->nvtxs) { 34435099eff2SStefano 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); 3444575ad6abSStefano Zampini } 34459577ea80SStefano Zampini 3446674ae819SStefano Zampini /* Set default CSR adjacency of local dofs if not provided by the user with PCBDDCSetLocalAdjacencyGraph */ 34474d379d7bSStefano Zampini if (!pcbddc->mat_graph->xadj || !pcbddc->mat_graph->adjncy) { 34484d379d7bSStefano Zampini PetscInt *xadj,*adjncy; 34494d379d7bSStefano Zampini PetscInt nvtxs; 3450e496cd5dSStefano Zampini PetscBool flg_row=PETSC_FALSE; 3451674ae819SStefano Zampini 34524d379d7bSStefano Zampini if (pcbddc->use_local_adj) { 34532fffb893SStefano Zampini 34542fffb893SStefano Zampini ierr = MatGetRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 34552fffb893SStefano Zampini if (flg_row) { 34564d379d7bSStefano Zampini ierr = PCBDDCSetLocalAdjacencyGraph(pc,nvtxs,xadj,adjncy,PETSC_COPY_VALUES);CHKERRQ(ierr); 3457b96c3477SStefano Zampini pcbddc->computed_rowadj = PETSC_TRUE; 34582fffb893SStefano Zampini } 34592fffb893SStefano Zampini ierr = MatRestoreRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 34609b28b941SStefano Zampini } else if (pcbddc->current_level && pcis->n_B) { /* just compute subdomain's connected components for coarser levels when the local boundary is not empty */ 34614d379d7bSStefano Zampini IS is_dummy; 34624d379d7bSStefano Zampini ISLocalToGlobalMapping l2gmap_dummy; 34634d379d7bSStefano Zampini PetscInt j,sum; 34644d379d7bSStefano Zampini PetscInt *cxadj,*cadjncy; 34654d379d7bSStefano Zampini const PetscInt *idxs; 34664d379d7bSStefano Zampini PCBDDCGraph graph; 34674d379d7bSStefano Zampini PetscBT is_on_boundary; 34684d379d7bSStefano Zampini 34694d379d7bSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcis->n,0,1,&is_dummy);CHKERRQ(ierr); 34704d379d7bSStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(is_dummy,&l2gmap_dummy);CHKERRQ(ierr); 34714d379d7bSStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 34724d379d7bSStefano Zampini ierr = PCBDDCGraphCreate(&graph);CHKERRQ(ierr); 34737fb0e2dbSStefano Zampini ierr = PCBDDCGraphInit(graph,l2gmap_dummy,pcis->n);CHKERRQ(ierr); 34744d379d7bSStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&l2gmap_dummy);CHKERRQ(ierr); 3475e496cd5dSStefano Zampini ierr = MatGetRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 3476e496cd5dSStefano Zampini if (flg_row) { 34774d379d7bSStefano Zampini graph->xadj = xadj; 34784d379d7bSStefano Zampini graph->adjncy = adjncy; 3479e496cd5dSStefano Zampini } 34804d379d7bSStefano Zampini ierr = PCBDDCGraphSetUp(graph,1,NULL,NULL,0,NULL,NULL);CHKERRQ(ierr); 34814d379d7bSStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(graph);CHKERRQ(ierr); 3482e496cd5dSStefano Zampini ierr = MatRestoreRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 34834d379d7bSStefano Zampini 34844d379d7bSStefano Zampini if (pcbddc->dbg_flag) { 34859b28b941SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"[%d] Found %d subdomains (local size %d)\n",PetscGlobalRank,graph->ncc,pcis->n);CHKERRQ(ierr); 34864d379d7bSStefano Zampini for (i=0;i<graph->ncc;i++) { 34874d379d7bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"[%d] %d cc size %d\n",PetscGlobalRank,i,graph->cptr[i+1]-graph->cptr[i]);CHKERRQ(ierr); 34884d379d7bSStefano Zampini } 34894d379d7bSStefano Zampini } 34904d379d7bSStefano Zampini 3491e496cd5dSStefano Zampini ierr = PetscBTCreate(pcis->n,&is_on_boundary);CHKERRQ(ierr); 34924d379d7bSStefano Zampini ierr = ISGetIndices(pcis->is_B_local,&idxs);CHKERRQ(ierr); 34934d379d7bSStefano Zampini for (i=0;i<pcis->n_B;i++) { 34944d379d7bSStefano Zampini ierr = PetscBTSet(is_on_boundary,idxs[i]);CHKERRQ(ierr); 34954d379d7bSStefano Zampini } 34964d379d7bSStefano Zampini ierr = ISRestoreIndices(pcis->is_B_local,&idxs);CHKERRQ(ierr); 34974d379d7bSStefano Zampini 3498e496cd5dSStefano Zampini ierr = PetscCalloc1(pcis->n+1,&cxadj);CHKERRQ(ierr); 34994d379d7bSStefano Zampini sum = 0; 35004d379d7bSStefano Zampini for (i=0;i<graph->ncc;i++) { 35014d379d7bSStefano Zampini PetscInt sizecc = 0; 35024d379d7bSStefano Zampini for (j=graph->cptr[i];j<graph->cptr[i+1];j++) { 35034d379d7bSStefano Zampini if (PetscBTLookup(is_on_boundary,graph->queue[j])) { 35044d379d7bSStefano Zampini sizecc++; 35054d379d7bSStefano Zampini } 35064d379d7bSStefano Zampini } 35074d379d7bSStefano Zampini for (j=graph->cptr[i];j<graph->cptr[i+1];j++) { 35084d379d7bSStefano Zampini if (PetscBTLookup(is_on_boundary,graph->queue[j])) { 35094d379d7bSStefano Zampini cxadj[graph->queue[j]] = sizecc; 35104d379d7bSStefano Zampini } 35114d379d7bSStefano Zampini } 35124d379d7bSStefano Zampini sum += sizecc*sizecc; 35134d379d7bSStefano Zampini } 35144d379d7bSStefano Zampini ierr = PetscMalloc1(sum,&cadjncy);CHKERRQ(ierr); 35154d379d7bSStefano Zampini sum = 0; 3516e496cd5dSStefano Zampini for (i=0;i<pcis->n;i++) { 35174d379d7bSStefano Zampini PetscInt temp = cxadj[i]; 35184d379d7bSStefano Zampini cxadj[i] = sum; 35194d379d7bSStefano Zampini sum += temp; 35204d379d7bSStefano Zampini } 3521e496cd5dSStefano Zampini cxadj[pcis->n] = sum; 35224d379d7bSStefano Zampini for (i=0;i<graph->ncc;i++) { 35234d379d7bSStefano Zampini for (j=graph->cptr[i];j<graph->cptr[i+1];j++) { 35244d379d7bSStefano Zampini if (PetscBTLookup(is_on_boundary,graph->queue[j])) { 35254d379d7bSStefano Zampini PetscInt k,sizecc = 0; 35264d379d7bSStefano Zampini for (k=graph->cptr[i];k<graph->cptr[i+1];k++) { 35274d379d7bSStefano Zampini if (PetscBTLookup(is_on_boundary,graph->queue[k])) { 35284d379d7bSStefano Zampini cadjncy[cxadj[graph->queue[j]]+sizecc]=graph->queue[k]; 35294d379d7bSStefano Zampini sizecc++; 35304d379d7bSStefano Zampini } 35314d379d7bSStefano Zampini } 35324d379d7bSStefano Zampini } 35334d379d7bSStefano Zampini } 35344d379d7bSStefano Zampini } 35359b28b941SStefano Zampini if (sum) { 3536e496cd5dSStefano Zampini ierr = PCBDDCSetLocalAdjacencyGraph(pc,pcis->n,cxadj,cadjncy,PETSC_OWN_POINTER);CHKERRQ(ierr); 35374d379d7bSStefano Zampini } else { 35384d379d7bSStefano Zampini ierr = PetscFree(cxadj);CHKERRQ(ierr); 35394d379d7bSStefano Zampini ierr = PetscFree(cadjncy);CHKERRQ(ierr); 35404d379d7bSStefano Zampini } 35414d379d7bSStefano Zampini graph->xadj = 0; 35424d379d7bSStefano Zampini graph->adjncy = 0; 35434d379d7bSStefano Zampini ierr = PCBDDCGraphDestroy(&graph);CHKERRQ(ierr); 35444d379d7bSStefano Zampini ierr = PetscBTDestroy(&is_on_boundary);CHKERRQ(ierr); 35454d379d7bSStefano Zampini } 3546674ae819SStefano Zampini } 35479b28b941SStefano Zampini if (pcbddc->dbg_flag) { 35489b28b941SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 3549674ae819SStefano Zampini } 3550674ae819SStefano Zampini 355163602bcaSStefano Zampini /* Set default dofs' splitting if no information has been provided by the user with PCBDDCSetDofsSplitting or PCBDDCSetDofsSplittingLocal */ 3552674ae819SStefano Zampini vertex_size = 1; 355363602bcaSStefano Zampini if (pcbddc->user_provided_isfordofs) { 355463602bcaSStefano Zampini if (pcbddc->n_ISForDofs) { /* need to convert from global to local and remove references to global dofs splitting */ 355595ecbf38SStefano Zampini ierr = PetscMalloc1(pcbddc->n_ISForDofs,&pcbddc->ISForDofsLocal);CHKERRQ(ierr); 355663602bcaSStefano Zampini for (i=0;i<pcbddc->n_ISForDofs;i++) { 3557e176bc59SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->rctx,pcis->vec1_global,pcis->vec1_N,pcbddc->ISForDofs[i],&pcbddc->ISForDofsLocal[i]);CHKERRQ(ierr); 355863602bcaSStefano Zampini ierr = ISDestroy(&pcbddc->ISForDofs[i]);CHKERRQ(ierr); 3559674ae819SStefano Zampini } 356063602bcaSStefano Zampini pcbddc->n_ISForDofsLocal = pcbddc->n_ISForDofs; 356163602bcaSStefano Zampini pcbddc->n_ISForDofs = 0; 356263602bcaSStefano Zampini ierr = PetscFree(pcbddc->ISForDofs);CHKERRQ(ierr); 3563674ae819SStefano Zampini } 356463602bcaSStefano Zampini /* mat block size as vertex size (used for elasticity with rigid body modes as nearnullspace) */ 3565674ae819SStefano Zampini ierr = MatGetBlockSize(matis->A,&vertex_size);CHKERRQ(ierr); 356663602bcaSStefano Zampini } else { 356763602bcaSStefano Zampini if (!pcbddc->n_ISForDofsLocal) { /* field split not present, create it in local ordering */ 356863602bcaSStefano Zampini ierr = MatGetBlockSize(pc->pmat,&pcbddc->n_ISForDofsLocal);CHKERRQ(ierr); 3569854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->n_ISForDofsLocal,&pcbddc->ISForDofsLocal);CHKERRQ(ierr); 357063602bcaSStefano Zampini for (i=0;i<pcbddc->n_ISForDofsLocal;i++) { 357163602bcaSStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)pc),pcis->n/pcbddc->n_ISForDofsLocal,i,pcbddc->n_ISForDofsLocal,&pcbddc->ISForDofsLocal[i]);CHKERRQ(ierr); 357263602bcaSStefano Zampini } 357363602bcaSStefano Zampini } 3574674ae819SStefano Zampini } 3575674ae819SStefano Zampini 3576674ae819SStefano Zampini /* Setup of Graph */ 3577785d1243SStefano Zampini if (!pcbddc->DirichletBoundariesLocal && pcbddc->DirichletBoundaries) { /* need to convert from global to local */ 3578e176bc59SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->rctx,pcis->vec1_global,pcis->vec1_N,pcbddc->DirichletBoundaries,&pcbddc->DirichletBoundariesLocal);CHKERRQ(ierr); 3579785d1243SStefano Zampini } 3580785d1243SStefano Zampini if (!pcbddc->NeumannBoundariesLocal && pcbddc->NeumannBoundaries) { /* need to convert from global to local */ 3581e176bc59SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->rctx,pcis->vec1_global,pcis->vec1_N,pcbddc->NeumannBoundaries,&pcbddc->NeumannBoundariesLocal);CHKERRQ(ierr); 3582785d1243SStefano Zampini } 3583302440fdSBarry Smith ierr = PCBDDCGraphSetUp(pcbddc->mat_graph,vertex_size,pcbddc->NeumannBoundariesLocal,pcbddc->DirichletBoundariesLocal,pcbddc->n_ISForDofsLocal,pcbddc->ISForDofsLocal,pcbddc->user_primal_vertices);CHKERRQ(ierr); 3584674ae819SStefano Zampini 3585674ae819SStefano Zampini /* Graph's connected components analysis */ 3586674ae819SStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(pcbddc->mat_graph);CHKERRQ(ierr); 3587674ae819SStefano Zampini 3588674ae819SStefano Zampini /* print some info to stdout */ 3589674ae819SStefano Zampini if (pcbddc->dbg_flag) { 3590302440fdSBarry Smith ierr = PCBDDCGraphASCIIView(pcbddc->mat_graph,pcbddc->dbg_flag,viewer);CHKERRQ(ierr); 3591674ae819SStefano Zampini } 3592fb180af4SStefano Zampini 3593fb180af4SStefano Zampini /* mark topography has done */ 3594fb180af4SStefano Zampini pcbddc->recompute_topography = PETSC_FALSE; 3595674ae819SStefano Zampini PetscFunctionReturn(0); 3596674ae819SStefano Zampini } 3597674ae819SStefano Zampini 3598dc456d91SStefano Zampini /* given an index sets possibly with holes, renumbers the indexes removing the holes */ 3599674ae819SStefano Zampini #undef __FUNCT__ 3600674ae819SStefano Zampini #define __FUNCT__ "PCBDDCSubsetNumbering" 3601dc456d91SStefano Zampini PetscErrorCode PCBDDCSubsetNumbering(IS subset, IS subset_mult, PetscInt *N_n, IS *subset_n) 3602674ae819SStefano Zampini { 3603dc456d91SStefano Zampini PetscSF sf; 3604dc456d91SStefano Zampini PetscLayout map; 3605dc456d91SStefano Zampini const PetscInt *idxs; 3606dc456d91SStefano Zampini PetscInt *leaf_data,*root_data,*gidxs; 3607dc456d91SStefano Zampini PetscInt N,n,i,lbounds[2],gbounds[2],Nl; 3608dc456d91SStefano Zampini PetscInt n_n,nlocals,start,first_index; 3609dc456d91SStefano Zampini PetscMPIInt commsize; 3610674ae819SStefano Zampini PetscBool first_found; 3611674ae819SStefano Zampini PetscErrorCode ierr; 3612674ae819SStefano Zampini 3613674ae819SStefano Zampini PetscFunctionBegin; 3614dc456d91SStefano Zampini ierr = ISGetLocalSize(subset,&n);CHKERRQ(ierr); 3615dc456d91SStefano Zampini if (subset_mult) { 3616dc456d91SStefano Zampini PetscCheckSameComm(subset,1,subset_mult,2); 3617dc456d91SStefano Zampini ierr = ISGetLocalSize(subset,&i);CHKERRQ(ierr); 3618dc456d91SStefano Zampini if (i != n) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Local subset and multiplicity sizes don't match! %d != %d",n,i); 3619674ae819SStefano Zampini } 3620dc456d91SStefano Zampini /* create workspace layout for computing global indices of subset */ 3621dc456d91SStefano Zampini ierr = ISGetIndices(subset,&idxs);CHKERRQ(ierr); 3622dc456d91SStefano Zampini lbounds[0] = lbounds[1] = 0; 3623dc456d91SStefano Zampini for (i=0;i<n;i++) { 3624dc456d91SStefano Zampini if (idxs[i] < lbounds[0]) lbounds[0] = idxs[i]; 3625dc456d91SStefano Zampini else if (idxs[i] > lbounds[1]) lbounds[1] = idxs[i]; 3626674ae819SStefano Zampini } 3627dc456d91SStefano Zampini lbounds[0] = -lbounds[0]; 3628dc456d91SStefano Zampini ierr = MPI_Allreduce(lbounds,gbounds,2,MPIU_INT,MPI_MAX,PetscObjectComm((PetscObject)subset));CHKERRQ(ierr); 3629dc456d91SStefano Zampini gbounds[0] = -gbounds[0]; 3630dc456d91SStefano Zampini N = gbounds[1] - gbounds[0] + 1; 3631dc456d91SStefano Zampini ierr = PetscLayoutCreate(PetscObjectComm((PetscObject)subset),&map);CHKERRQ(ierr); 3632dc456d91SStefano Zampini ierr = PetscLayoutSetBlockSize(map,1);CHKERRQ(ierr); 3633dc456d91SStefano Zampini ierr = PetscLayoutSetSize(map,N);CHKERRQ(ierr); 3634dc456d91SStefano Zampini ierr = PetscLayoutSetUp(map);CHKERRQ(ierr); 3635dc456d91SStefano Zampini ierr = PetscLayoutGetLocalSize(map,&Nl);CHKERRQ(ierr); 3636dc456d91SStefano Zampini 3637dc456d91SStefano Zampini /* create sf : leaf_data == multiplicity of indexes, root data == global index in layout */ 3638dc456d91SStefano Zampini ierr = PetscMalloc2(n,&leaf_data,Nl,&root_data);CHKERRQ(ierr); 3639dc456d91SStefano Zampini if (subset_mult) { 3640dc456d91SStefano Zampini const PetscInt* idxs_mult; 3641dc456d91SStefano Zampini 3642dc456d91SStefano Zampini ierr = ISGetIndices(subset_mult,&idxs_mult);CHKERRQ(ierr); 3643dc456d91SStefano Zampini ierr = PetscMemcpy(leaf_data,idxs_mult,n*sizeof(PetscInt));CHKERRQ(ierr); 3644dc456d91SStefano Zampini ierr = ISRestoreIndices(subset_mult,&idxs_mult);CHKERRQ(ierr); 3645674ae819SStefano Zampini } else { 3646dc456d91SStefano Zampini for (i=0;i<n;i++) leaf_data[i] = 1; 3647674ae819SStefano Zampini } 3648dc456d91SStefano Zampini /* local size of new subset */ 3649dc456d91SStefano Zampini n_n = 0; 3650dc456d91SStefano Zampini for (i=0;i<n;i++) n_n += leaf_data[i]; 3651dc456d91SStefano Zampini 3652dc456d91SStefano Zampini /* global indexes in layout */ 3653dc456d91SStefano Zampini ierr = PetscMalloc1(n_n,&gidxs);CHKERRQ(ierr); /* allocating possibly extra space in gidxs which will be used later */ 3654dc456d91SStefano Zampini for (i=0;i<n;i++) gidxs[i] = idxs[i] - gbounds[0]; 3655dc456d91SStefano Zampini ierr = ISRestoreIndices(subset,&idxs);CHKERRQ(ierr); 3656dc456d91SStefano Zampini ierr = PetscSFCreate(PetscObjectComm((PetscObject)subset),&sf);CHKERRQ(ierr); 3657dc456d91SStefano Zampini ierr = PetscSFSetGraphLayout(sf,map,n,NULL,PETSC_COPY_VALUES,gidxs);CHKERRQ(ierr); 3658dc456d91SStefano Zampini ierr = PetscLayoutDestroy(&map);CHKERRQ(ierr); 3659dc456d91SStefano Zampini 3660dc456d91SStefano Zampini /* reduce from leaves to roots */ 3661dc456d91SStefano Zampini ierr = PetscMemzero(root_data,Nl*sizeof(PetscInt));CHKERRQ(ierr); 366264a8e5bcSStefano Zampini ierr = PetscSFReduceBegin(sf,MPIU_INT,leaf_data,root_data,MPI_MAX);CHKERRQ(ierr); 366364a8e5bcSStefano Zampini ierr = PetscSFReduceEnd(sf,MPIU_INT,leaf_data,root_data,MPI_MAX);CHKERRQ(ierr); 3664dc456d91SStefano Zampini 3665dc456d91SStefano Zampini /* count indexes in local part of layout */ 3666674ae819SStefano Zampini nlocals = 0; 3667674ae819SStefano Zampini first_index = -1; 3668674ae819SStefano Zampini first_found = PETSC_FALSE; 3669dc456d91SStefano Zampini for (i=0;i<Nl;i++) { 3670dc456d91SStefano Zampini if (!first_found && root_data[i]) { 3671674ae819SStefano Zampini first_found = PETSC_TRUE; 3672674ae819SStefano Zampini first_index = i; 3673674ae819SStefano Zampini } 3674dc456d91SStefano Zampini nlocals += root_data[i]; 3675674ae819SStefano Zampini } 3676dc456d91SStefano Zampini 3677dc456d91SStefano Zampini /* cumulative of number of indexes and size of subset without holes */ 36785fa240b1SStefano Zampini #if defined(PETSC_HAVE_MPI_EXSCAN) 3679dc456d91SStefano Zampini start = 0; 368064a8e5bcSStefano Zampini ierr = MPI_Exscan(&nlocals,&start,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)subset));CHKERRQ(ierr); 36815fa240b1SStefano Zampini #else 368264a8e5bcSStefano Zampini ierr = MPI_Scan(&nlocals,&start,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)subset));CHKERRQ(ierr); 36835fa240b1SStefano Zampini start = start-nlocals; 36845fa240b1SStefano Zampini #endif 36855fa240b1SStefano Zampini 3686dc456d91SStefano Zampini if (N_n) { /* compute total size of new subset if requested */ 3687dc456d91SStefano Zampini *N_n = start + nlocals; 3688dc456d91SStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)subset),&commsize);CHKERRQ(ierr); 3689dc456d91SStefano Zampini ierr = MPI_Bcast(N_n,1,MPIU_INT,commsize-1,PetscObjectComm((PetscObject)subset));CHKERRQ(ierr); 3690674ae819SStefano Zampini } 36915fa240b1SStefano Zampini 36925fa240b1SStefano Zampini /* adapt root data with cumulative */ 3693674ae819SStefano Zampini if (first_found) { 3694dc456d91SStefano Zampini PetscInt old_index; 3695dc456d91SStefano Zampini 3696dc456d91SStefano Zampini root_data[first_index] += start; 3697674ae819SStefano Zampini old_index = first_index; 3698dc456d91SStefano Zampini for (i=first_index+1;i<Nl;i++) { 3699dc456d91SStefano Zampini if (root_data[i]) { 3700dc456d91SStefano Zampini root_data[i] += root_data[old_index]; 3701674ae819SStefano Zampini old_index = i; 3702674ae819SStefano Zampini } 3703674ae819SStefano Zampini } 3704674ae819SStefano Zampini } 3705dc456d91SStefano Zampini 3706dc456d91SStefano Zampini /* from roots to leaves */ 3707dc456d91SStefano Zampini ierr = PetscSFBcastBegin(sf,MPIU_INT,root_data,leaf_data);CHKERRQ(ierr); 3708dc456d91SStefano Zampini ierr = PetscSFBcastEnd(sf,MPIU_INT,root_data,leaf_data);CHKERRQ(ierr); 3709dc456d91SStefano Zampini ierr = PetscSFDestroy(&sf);CHKERRQ(ierr); 3710dc456d91SStefano Zampini 3711dc456d91SStefano Zampini /* create new IS with global indexes without holes */ 3712dc456d91SStefano Zampini if (subset_mult) { 3713dc456d91SStefano Zampini const PetscInt* idxs_mult; 3714dc456d91SStefano Zampini PetscInt cum; 3715dc456d91SStefano Zampini 3716dc456d91SStefano Zampini cum = 0; 3717dc456d91SStefano Zampini ierr = ISGetIndices(subset_mult,&idxs_mult);CHKERRQ(ierr); 3718dc456d91SStefano Zampini for (i=0;i<n;i++) { 3719dc456d91SStefano Zampini PetscInt j; 3720dc456d91SStefano Zampini for (j=0;j<idxs_mult[i];j++) gidxs[cum++] = leaf_data[i] - idxs_mult[i] + j; 3721674ae819SStefano Zampini } 3722dc456d91SStefano Zampini ierr = ISRestoreIndices(subset_mult,&idxs_mult);CHKERRQ(ierr); 3723674ae819SStefano Zampini } else { 3724dc456d91SStefano Zampini for (i=0;i<n;i++) { 3725dc456d91SStefano Zampini gidxs[i] = leaf_data[i]-1; 3726674ae819SStefano Zampini } 3727674ae819SStefano Zampini } 3728dc456d91SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)subset),n_n,gidxs,PETSC_OWN_POINTER,subset_n);CHKERRQ(ierr); 3729dc456d91SStefano Zampini ierr = PetscFree2(leaf_data,root_data);CHKERRQ(ierr); 3730674ae819SStefano Zampini PetscFunctionReturn(0); 3731674ae819SStefano Zampini } 37329a7d3425SStefano Zampini 37339a7d3425SStefano Zampini #undef __FUNCT__ 37349a7d3425SStefano Zampini #define __FUNCT__ "PCBDDCOrthonormalizeVecs" 37359a7d3425SStefano Zampini PetscErrorCode PCBDDCOrthonormalizeVecs(PetscInt n, Vec vecs[]) 37369a7d3425SStefano Zampini { 37379a7d3425SStefano Zampini PetscInt i,j; 37389a7d3425SStefano Zampini PetscScalar *alphas; 37399a7d3425SStefano Zampini PetscErrorCode ierr; 37409a7d3425SStefano Zampini 37419a7d3425SStefano Zampini PetscFunctionBegin; 37429a7d3425SStefano Zampini /* this implements stabilized Gram-Schmidt */ 3743785e854fSJed Brown ierr = PetscMalloc1(n,&alphas);CHKERRQ(ierr); 37449a7d3425SStefano Zampini for (i=0;i<n;i++) { 37459a7d3425SStefano Zampini ierr = VecNormalize(vecs[i],NULL);CHKERRQ(ierr); 37469a7d3425SStefano Zampini if (i<n) { ierr = VecMDot(vecs[i],n-i-1,&vecs[i+1],&alphas[i+1]);CHKERRQ(ierr); } 37479a7d3425SStefano Zampini for (j=i+1;j<n;j++) { ierr = VecAXPY(vecs[j],PetscConj(-alphas[j]),vecs[i]);CHKERRQ(ierr); } 37489a7d3425SStefano Zampini } 37499a7d3425SStefano Zampini ierr = PetscFree(alphas);CHKERRQ(ierr); 37509a7d3425SStefano Zampini PetscFunctionReturn(0); 37519a7d3425SStefano Zampini } 37529a7d3425SStefano Zampini 3753e7931f94SStefano Zampini #undef __FUNCT__ 375470cf5478SStefano Zampini #define __FUNCT__ "MatISGetSubassemblingPattern" 3755b0c7d250SStefano Zampini PetscErrorCode MatISGetSubassemblingPattern(Mat mat, PetscInt n_subdomains, PetscInt redprocs, IS* is_sends) 3756e7931f94SStefano Zampini { 375752e5ac9dSStefano Zampini IS ranks_send_to; 3758e7931f94SStefano Zampini PetscInt n_neighs,*neighs,*n_shared,**shared; 3759e7931f94SStefano Zampini PetscMPIInt size,rank,color; 376052e5ac9dSStefano Zampini PetscInt *xadj,*adjncy; 376152e5ac9dSStefano Zampini PetscInt *adjncy_wgt,*v_wgt,*ranks_send_to_idx; 37623837a79fSStefano Zampini PetscInt i,local_size,threshold=0; 37632b510759SStefano Zampini PetscBool use_vwgt=PETSC_FALSE,use_square=PETSC_FALSE; 3764e7931f94SStefano Zampini PetscSubcomm subcomm; 376552e5ac9dSStefano Zampini PetscErrorCode ierr; 3766a57a6d2fSStefano Zampini 3767e7931f94SStefano Zampini PetscFunctionBegin; 37682b510759SStefano Zampini ierr = PetscOptionsGetBool(NULL,"-matis_partitioning_use_square",&use_square,NULL);CHKERRQ(ierr); 37692b510759SStefano Zampini ierr = PetscOptionsGetBool(NULL,"-matis_partitioning_use_vwgt",&use_vwgt,NULL);CHKERRQ(ierr); 37702b510759SStefano Zampini ierr = PetscOptionsGetInt(NULL,"-matis_partitioning_threshold",&threshold,NULL);CHKERRQ(ierr); 3771e7931f94SStefano Zampini 3772e7931f94SStefano Zampini /* Get info on mapping */ 37733bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetSize(mat->rmap->mapping,&local_size);CHKERRQ(ierr); 37743bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetInfo(mat->rmap->mapping,&n_neighs,&neighs,&n_shared,&shared);CHKERRQ(ierr); 3775e7931f94SStefano Zampini 3776e7931f94SStefano Zampini /* build local CSR graph of subdomains' connectivity */ 3777785e854fSJed Brown ierr = PetscMalloc1(2,&xadj);CHKERRQ(ierr); 3778e7931f94SStefano Zampini xadj[0] = 0; 3779e7931f94SStefano Zampini xadj[1] = PetscMax(n_neighs-1,0); 3780785e854fSJed Brown ierr = PetscMalloc1(xadj[1],&adjncy);CHKERRQ(ierr); 3781785e854fSJed Brown ierr = PetscMalloc1(xadj[1],&adjncy_wgt);CHKERRQ(ierr); 3782e7931f94SStefano Zampini 37832b510759SStefano Zampini if (threshold) { 3784d023bfaeSStefano Zampini PetscInt xadj_count = 0; 37852b510759SStefano Zampini for (i=1;i<n_neighs;i++) { 3786d023bfaeSStefano Zampini if (n_shared[i] > threshold) { 3787d023bfaeSStefano Zampini adjncy[xadj_count] = neighs[i]; 3788d023bfaeSStefano Zampini adjncy_wgt[xadj_count] = n_shared[i]; 3789d023bfaeSStefano Zampini xadj_count++; 3790e7931f94SStefano Zampini } 3791e7931f94SStefano Zampini } 3792d023bfaeSStefano Zampini xadj[1] = xadj_count; 3793c8587f34SStefano Zampini } else { 3794e7931f94SStefano Zampini if (xadj[1]) { 3795e7931f94SStefano Zampini ierr = PetscMemcpy(adjncy,&neighs[1],xadj[1]*sizeof(*adjncy));CHKERRQ(ierr); 3796e7931f94SStefano Zampini ierr = PetscMemcpy(adjncy_wgt,&n_shared[1],xadj[1]*sizeof(*adjncy_wgt));CHKERRQ(ierr); 3797c8587f34SStefano Zampini } 3798e7931f94SStefano Zampini } 37993bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingRestoreInfo(mat->rmap->mapping,&n_neighs,&neighs,&n_shared,&shared);CHKERRQ(ierr); 3800e7931f94SStefano Zampini if (use_square) { 3801e7931f94SStefano Zampini for (i=0;i<xadj[1];i++) { 3802e7931f94SStefano Zampini adjncy_wgt[i] = adjncy_wgt[i]*adjncy_wgt[i]; 3803e7931f94SStefano Zampini } 3804e7931f94SStefano Zampini } 3805e7931f94SStefano Zampini ierr = PetscSortIntWithArray(xadj[1],adjncy,adjncy_wgt);CHKERRQ(ierr); 3806e7931f94SStefano Zampini 38073837a79fSStefano Zampini ierr = PetscMalloc1(1,&ranks_send_to_idx);CHKERRQ(ierr); 3808e7931f94SStefano Zampini 3809e7931f94SStefano Zampini /* 3810e7931f94SStefano Zampini Restrict work on active processes only. 3811e7931f94SStefano Zampini */ 3812e7931f94SStefano Zampini ierr = PetscSubcommCreate(PetscObjectComm((PetscObject)mat),&subcomm);CHKERRQ(ierr); 3813e7931f94SStefano Zampini ierr = PetscSubcommSetNumber(subcomm,2);CHKERRQ(ierr); /* 2 groups, active process and not active processes */ 3814e7931f94SStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)mat),&rank);CHKERRQ(ierr); 38152b510759SStefano Zampini ierr = PetscMPIIntCast(!local_size,&color);CHKERRQ(ierr); 3816d3531aaaSJed Brown ierr = PetscSubcommSetTypeGeneral(subcomm,color,rank);CHKERRQ(ierr); 3817e7931f94SStefano Zampini if (color) { 3818e7931f94SStefano Zampini ierr = PetscFree(xadj);CHKERRQ(ierr); 3819e7931f94SStefano Zampini ierr = PetscFree(adjncy);CHKERRQ(ierr); 3820e7931f94SStefano Zampini ierr = PetscFree(adjncy_wgt);CHKERRQ(ierr); 3821c8587f34SStefano Zampini } else { 382252e5ac9dSStefano Zampini Mat subdomain_adj; 382352e5ac9dSStefano Zampini IS new_ranks,new_ranks_contig; 382452e5ac9dSStefano Zampini MatPartitioning partitioner; 382552e5ac9dSStefano Zampini PetscInt prank,rstart=0,rend=0; 382652e5ac9dSStefano Zampini PetscInt *is_indices,*oldranks; 3827b0c7d250SStefano Zampini PetscBool aggregate; 3828b0c7d250SStefano Zampini 3829306c2d5bSBarry Smith ierr = MPI_Comm_size(PetscSubcommChild(subcomm),&size);CHKERRQ(ierr); 3830785e854fSJed Brown ierr = PetscMalloc1(size,&oldranks);CHKERRQ(ierr); 3831e7931f94SStefano Zampini prank = rank; 3832306c2d5bSBarry Smith ierr = MPI_Allgather(&prank,1,MPIU_INT,oldranks,1,MPIU_INT,PetscSubcommChild(subcomm));CHKERRQ(ierr); 38338002ef2cSStefano Zampini /* 3834e7931f94SStefano Zampini for (i=0;i<size;i++) { 3835e7931f94SStefano Zampini PetscPrintf(subcomm->comm,"oldranks[%d] = %d\n",i,oldranks[i]); 3836c8587f34SStefano Zampini } 38378002ef2cSStefano Zampini */ 3838e7931f94SStefano Zampini for (i=0;i<xadj[1];i++) { 3839e7931f94SStefano Zampini ierr = PetscFindInt(adjncy[i],size,oldranks,&adjncy[i]);CHKERRQ(ierr); 3840c8587f34SStefano Zampini } 3841e7931f94SStefano Zampini ierr = PetscSortIntWithArray(xadj[1],adjncy,adjncy_wgt);CHKERRQ(ierr); 3842b0c7d250SStefano Zampini aggregate = ((redprocs > 0 && redprocs < size) ? PETSC_TRUE : PETSC_FALSE); 3843b0c7d250SStefano Zampini if (aggregate) { 3844b0c7d250SStefano Zampini PetscInt lrows,row,ncols,*cols; 3845b0c7d250SStefano Zampini PetscMPIInt nrank; 3846b0c7d250SStefano Zampini PetscScalar *vals; 3847b0c7d250SStefano Zampini 3848b0c7d250SStefano Zampini ierr = MPI_Comm_rank(PetscSubcommChild(subcomm),&nrank);CHKERRQ(ierr); 3849b0c7d250SStefano Zampini lrows = 0; 3850b0c7d250SStefano Zampini if (nrank<redprocs) { 3851b0c7d250SStefano Zampini lrows = size/redprocs; 3852b0c7d250SStefano Zampini if (nrank<size%redprocs) lrows++; 3853b0c7d250SStefano Zampini } 38545fa240b1SStefano Zampini ierr = MatCreateAIJ(PetscSubcommChild(subcomm),lrows,lrows,size,size,50,NULL,50,NULL,&subdomain_adj);CHKERRQ(ierr); 3855b0c7d250SStefano Zampini ierr = MatGetOwnershipRange(subdomain_adj,&rstart,&rend);CHKERRQ(ierr); 3856b0c7d250SStefano Zampini ierr = MatSetOption(subdomain_adj,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_FALSE);CHKERRQ(ierr); 3857b0c7d250SStefano Zampini ierr = MatSetOption(subdomain_adj,MAT_NEW_NONZERO_ALLOCATION_ERR,PETSC_FALSE);CHKERRQ(ierr); 3858b0c7d250SStefano Zampini row = nrank; 3859b0c7d250SStefano Zampini ncols = xadj[1]-xadj[0]; 3860b0c7d250SStefano Zampini cols = adjncy; 3861b0c7d250SStefano Zampini ierr = PetscMalloc1(ncols,&vals);CHKERRQ(ierr); 3862b0c7d250SStefano Zampini for (i=0;i<ncols;i++) vals[i] = adjncy_wgt[i]; 3863b0c7d250SStefano Zampini ierr = MatSetValues(subdomain_adj,1,&row,ncols,cols,vals,INSERT_VALUES);CHKERRQ(ierr); 3864b0c7d250SStefano Zampini ierr = MatAssemblyBegin(subdomain_adj,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3865b0c7d250SStefano Zampini ierr = MatAssemblyEnd(subdomain_adj,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 386652e5ac9dSStefano Zampini ierr = PetscFree(xadj);CHKERRQ(ierr); 386752e5ac9dSStefano Zampini ierr = PetscFree(adjncy);CHKERRQ(ierr); 386852e5ac9dSStefano Zampini ierr = PetscFree(adjncy_wgt);CHKERRQ(ierr); 3869b0c7d250SStefano Zampini ierr = PetscFree(vals);CHKERRQ(ierr); 3870b0c7d250SStefano Zampini } else { 3871306c2d5bSBarry Smith ierr = MatCreateMPIAdj(PetscSubcommChild(subcomm),1,(PetscInt)size,xadj,adjncy,adjncy_wgt,&subdomain_adj);CHKERRQ(ierr); 3872b0c7d250SStefano Zampini } 387322b6e8a2SStefano Zampini /* ierr = MatView(subdomain_adj,0);CHKERRQ(ierr); */ 3874e7931f94SStefano Zampini 3875e7931f94SStefano Zampini /* Partition */ 3876306c2d5bSBarry Smith ierr = MatPartitioningCreate(PetscSubcommChild(subcomm),&partitioner);CHKERRQ(ierr); 3877e7931f94SStefano Zampini ierr = MatPartitioningSetAdjacency(partitioner,subdomain_adj);CHKERRQ(ierr); 3878e7931f94SStefano Zampini if (use_vwgt) { 38793837a79fSStefano Zampini ierr = PetscMalloc1(1,&v_wgt);CHKERRQ(ierr); 3880e7931f94SStefano Zampini v_wgt[0] = local_size; 3881e7931f94SStefano Zampini ierr = MatPartitioningSetVertexWeights(partitioner,v_wgt);CHKERRQ(ierr); 3882c8587f34SStefano Zampini } 388328143c3dSStefano Zampini n_subdomains = PetscMin((PetscInt)size,n_subdomains); 388428143c3dSStefano Zampini ierr = MatPartitioningSetNParts(partitioner,n_subdomains);CHKERRQ(ierr); 3885e7931f94SStefano Zampini ierr = MatPartitioningSetFromOptions(partitioner);CHKERRQ(ierr); 3886e7931f94SStefano Zampini ierr = MatPartitioningApply(partitioner,&new_ranks);CHKERRQ(ierr); 388722b6e8a2SStefano Zampini /* ierr = MatPartitioningView(partitioner,0);CHKERRQ(ierr); */ 3888e7931f94SStefano Zampini 388952e5ac9dSStefano Zampini /* renumber new_ranks to avoid "holes" in new set of processors */ 389052e5ac9dSStefano Zampini ierr = PCBDDCSubsetNumbering(new_ranks,NULL,NULL,&new_ranks_contig);CHKERRQ(ierr); 389152e5ac9dSStefano Zampini ierr = ISDestroy(&new_ranks);CHKERRQ(ierr); 389252e5ac9dSStefano Zampini ierr = ISGetIndices(new_ranks_contig,(const PetscInt**)&is_indices);CHKERRQ(ierr); 3893b0c7d250SStefano Zampini if (!redprocs) { 3894b0c7d250SStefano Zampini ranks_send_to_idx[0] = oldranks[is_indices[0]]; 389528143c3dSStefano Zampini } else { 3896b0c7d250SStefano Zampini PetscInt idxs[1]; 3897b0c7d250SStefano Zampini PetscMPIInt tag; 3898b0c7d250SStefano Zampini MPI_Request *reqs; 3899b0c7d250SStefano Zampini 3900b0c7d250SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)subdomain_adj,&tag);CHKERRQ(ierr); 3901b0c7d250SStefano Zampini ierr = PetscMalloc1(rend-rstart,&reqs);CHKERRQ(ierr); 3902b0c7d250SStefano Zampini for (i=rstart;i<rend;i++) { 3903b0c7d250SStefano Zampini ierr = MPI_Isend(is_indices+i-rstart,1,MPIU_INT,i,tag,PetscSubcommChild(subcomm),&reqs[i-rstart]);CHKERRQ(ierr); 390428143c3dSStefano Zampini } 3905b0c7d250SStefano Zampini ierr = MPI_Recv(idxs,1,MPIU_INT,MPI_ANY_SOURCE,tag,PetscSubcommChild(subcomm),MPI_STATUS_IGNORE);CHKERRQ(ierr); 3906b0c7d250SStefano Zampini ierr = MPI_Waitall(rend-rstart,reqs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 3907b0c7d250SStefano Zampini ierr = PetscFree(reqs);CHKERRQ(ierr); 3908b0c7d250SStefano Zampini ranks_send_to_idx[0] = oldranks[idxs[0]]; 3909e7931f94SStefano Zampini } 391052e5ac9dSStefano Zampini ierr = ISRestoreIndices(new_ranks_contig,(const PetscInt**)&is_indices);CHKERRQ(ierr); 3911e7931f94SStefano Zampini /* clean up */ 3912e7931f94SStefano Zampini ierr = PetscFree(oldranks);CHKERRQ(ierr); 391352e5ac9dSStefano Zampini ierr = ISDestroy(&new_ranks_contig);CHKERRQ(ierr); 3914e7931f94SStefano Zampini ierr = MatDestroy(&subdomain_adj);CHKERRQ(ierr); 3915e7931f94SStefano Zampini ierr = MatPartitioningDestroy(&partitioner);CHKERRQ(ierr); 3916e7931f94SStefano Zampini } 3917e7931f94SStefano Zampini ierr = PetscSubcommDestroy(&subcomm);CHKERRQ(ierr); 3918e7931f94SStefano Zampini 3919e7931f94SStefano Zampini /* assemble parallel IS for sends */ 3920e7931f94SStefano Zampini i = 1; 3921e7931f94SStefano Zampini if (color) i=0; 3922e7931f94SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)mat),i,ranks_send_to_idx,PETSC_OWN_POINTER,&ranks_send_to);CHKERRQ(ierr); 3923e7931f94SStefano Zampini /* get back IS */ 3924e7931f94SStefano Zampini *is_sends = ranks_send_to; 3925e7931f94SStefano Zampini PetscFunctionReturn(0); 3926e7931f94SStefano Zampini } 3927e7931f94SStefano Zampini 3928e7931f94SStefano Zampini typedef enum {MATDENSE_PRIVATE=0,MATAIJ_PRIVATE,MATBAIJ_PRIVATE,MATSBAIJ_PRIVATE}MatTypePrivate; 3929e7931f94SStefano Zampini 3930e7931f94SStefano Zampini #undef __FUNCT__ 3931e7931f94SStefano Zampini #define __FUNCT__ "MatISSubassemble" 393253a05cb3SStefano Zampini PetscErrorCode MatISSubassemble(Mat mat, IS is_sends, PetscInt n_subdomains, PetscBool restrict_comm, PetscBool restrict_full, MatReuse reuse, Mat *mat_n, PetscInt nis, IS isarray[]) 3933e7931f94SStefano Zampini { 393470cf5478SStefano Zampini Mat local_mat; 3935e7931f94SStefano Zampini IS is_sends_internal; 39369d30be91SStefano Zampini PetscInt rows,cols,new_local_rows; 393728143c3dSStefano Zampini PetscInt i,bs,buf_size_idxs,buf_size_idxs_is,buf_size_vals; 39389d30be91SStefano Zampini PetscBool ismatis,isdense,newisdense,destroy_mat; 3939e7931f94SStefano Zampini ISLocalToGlobalMapping l2gmap; 3940e7931f94SStefano Zampini PetscInt* l2gmap_indices; 3941e7931f94SStefano Zampini const PetscInt* is_indices; 3942e7931f94SStefano Zampini MatType new_local_type; 3943e7931f94SStefano Zampini /* buffers */ 3944e7931f94SStefano Zampini PetscInt *ptr_idxs,*send_buffer_idxs,*recv_buffer_idxs; 394528143c3dSStefano Zampini PetscInt *ptr_idxs_is,*send_buffer_idxs_is,*recv_buffer_idxs_is; 39469d30be91SStefano Zampini PetscInt *recv_buffer_idxs_local; 3947e7931f94SStefano Zampini PetscScalar *ptr_vals,*send_buffer_vals,*recv_buffer_vals; 3948e7931f94SStefano Zampini /* MPI */ 394928143c3dSStefano Zampini MPI_Comm comm,comm_n; 395028143c3dSStefano Zampini PetscSubcomm subcomm; 3951e7931f94SStefano Zampini PetscMPIInt n_sends,n_recvs,commsize; 395228143c3dSStefano Zampini PetscMPIInt *iflags,*ilengths_idxs,*ilengths_vals,*ilengths_idxs_is; 395328143c3dSStefano Zampini PetscMPIInt *onodes,*onodes_is,*olengths_idxs,*olengths_idxs_is,*olengths_vals; 395428143c3dSStefano Zampini PetscMPIInt len,tag_idxs,tag_idxs_is,tag_vals,source_dest; 395528143c3dSStefano Zampini MPI_Request *send_req_idxs,*send_req_idxs_is,*send_req_vals; 395628143c3dSStefano Zampini MPI_Request *recv_req_idxs,*recv_req_idxs_is,*recv_req_vals; 3957e7931f94SStefano Zampini PetscErrorCode ierr; 3958e7931f94SStefano Zampini 3959e7931f94SStefano Zampini PetscFunctionBegin; 396028143c3dSStefano Zampini /* TODO: add missing checks */ 396128143c3dSStefano Zampini PetscValidLogicalCollectiveInt(mat,n_subdomains,3); 396228143c3dSStefano Zampini PetscValidLogicalCollectiveBool(mat,restrict_comm,4); 396328143c3dSStefano Zampini PetscValidLogicalCollectiveEnum(mat,reuse,5); 396428143c3dSStefano Zampini PetscValidLogicalCollectiveInt(mat,nis,7); 3965e7931f94SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)mat,MATIS,&ismatis);CHKERRQ(ierr); 396628143c3dSStefano Zampini if (!ismatis) SETERRQ1(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot use %s on a matrix object which is not of type MATIS",__FUNCT__); 3967e7931f94SStefano Zampini ierr = MatISGetLocalMat(mat,&local_mat);CHKERRQ(ierr); 3968e7931f94SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)local_mat,MATSEQDENSE,&isdense);CHKERRQ(ierr); 3969e7931f94SStefano Zampini if (!isdense) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Currently cannot subassemble MATIS when local matrix type is not of type SEQDENSE"); 3970e7931f94SStefano Zampini ierr = MatGetSize(local_mat,&rows,&cols);CHKERRQ(ierr); 3971e7931f94SStefano Zampini if (rows != cols) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Local MATIS matrices should be square"); 397228143c3dSStefano Zampini if (reuse == MAT_REUSE_MATRIX && *mat_n) { 397370cf5478SStefano Zampini PetscInt mrows,mcols,mnrows,mncols; 397470cf5478SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)*mat_n,MATIS,&ismatis);CHKERRQ(ierr); 397528143c3dSStefano Zampini if (!ismatis) SETERRQ(PetscObjectComm((PetscObject)*mat_n),PETSC_ERR_SUP,"Cannot reuse a matrix which is not of type MATIS"); 397670cf5478SStefano Zampini ierr = MatGetSize(mat,&mrows,&mcols);CHKERRQ(ierr); 397770cf5478SStefano Zampini ierr = MatGetSize(*mat_n,&mnrows,&mncols);CHKERRQ(ierr); 397870cf5478SStefano Zampini if (mrows != mnrows) SETERRQ2(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot reuse matrix! Wrong number of rows %D != %D",mrows,mnrows); 397970cf5478SStefano Zampini if (mcols != mncols) SETERRQ2(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot reuse matrix! Wrong number of cols %D != %D",mcols,mncols); 398070cf5478SStefano Zampini } 3981e7931f94SStefano Zampini ierr = MatGetBlockSize(local_mat,&bs);CHKERRQ(ierr); 3982e7931f94SStefano Zampini PetscValidLogicalCollectiveInt(mat,bs,0); 3983e7931f94SStefano Zampini /* prepare IS for sending if not provided */ 3984e7931f94SStefano Zampini if (!is_sends) { 398528143c3dSStefano Zampini if (!n_subdomains) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"You should specify either an IS or a target number of subdomains"); 3986b0c7d250SStefano Zampini ierr = MatISGetSubassemblingPattern(mat,n_subdomains,0,&is_sends_internal);CHKERRQ(ierr); 3987c8587f34SStefano Zampini } else { 3988e7931f94SStefano Zampini ierr = PetscObjectReference((PetscObject)is_sends);CHKERRQ(ierr); 3989e7931f94SStefano Zampini is_sends_internal = is_sends; 3990c8587f34SStefano Zampini } 3991e7931f94SStefano Zampini 3992e7931f94SStefano Zampini /* get comm */ 3993a316fed8SStefano Zampini ierr = PetscObjectGetComm((PetscObject)mat,&comm);CHKERRQ(ierr); 3994e7931f94SStefano Zampini 3995e7931f94SStefano Zampini /* compute number of sends */ 3996e7931f94SStefano Zampini ierr = ISGetLocalSize(is_sends_internal,&i);CHKERRQ(ierr); 3997e7931f94SStefano Zampini ierr = PetscMPIIntCast(i,&n_sends);CHKERRQ(ierr); 3998e7931f94SStefano Zampini 3999e7931f94SStefano Zampini /* compute number of receives */ 4000e7931f94SStefano Zampini ierr = MPI_Comm_size(comm,&commsize);CHKERRQ(ierr); 4001785e854fSJed Brown ierr = PetscMalloc1(commsize,&iflags);CHKERRQ(ierr); 4002e7931f94SStefano Zampini ierr = PetscMemzero(iflags,commsize*sizeof(*iflags));CHKERRQ(ierr); 4003e7931f94SStefano Zampini ierr = ISGetIndices(is_sends_internal,&is_indices);CHKERRQ(ierr); 4004e7931f94SStefano Zampini for (i=0;i<n_sends;i++) iflags[is_indices[i]] = 1; 4005e7931f94SStefano Zampini ierr = PetscGatherNumberOfMessages(comm,iflags,NULL,&n_recvs);CHKERRQ(ierr); 4006e7931f94SStefano Zampini ierr = PetscFree(iflags);CHKERRQ(ierr); 4007e7931f94SStefano Zampini 400828143c3dSStefano Zampini /* restrict comm if requested */ 400928143c3dSStefano Zampini subcomm = 0; 401028143c3dSStefano Zampini destroy_mat = PETSC_FALSE; 401128143c3dSStefano Zampini if (restrict_comm) { 4012779c1cceSStefano Zampini PetscMPIInt color,subcommsize; 4013779c1cceSStefano Zampini 401428143c3dSStefano Zampini color = 0; 401553a05cb3SStefano Zampini if (restrict_full) { 401653a05cb3SStefano Zampini if (!n_recvs) color = 1; /* processes not receiving anything will not partecipate in new comm (full restriction) */ 401753a05cb3SStefano Zampini } else { 401853a05cb3SStefano Zampini if (!n_recvs && n_sends) color = 1; /* just those processes that are sending but not receiving anything will not partecipate in new comm */ 401953a05cb3SStefano Zampini } 402028143c3dSStefano Zampini ierr = MPI_Allreduce(&color,&subcommsize,1,MPI_INT,MPI_SUM,comm);CHKERRQ(ierr); 402128143c3dSStefano Zampini subcommsize = commsize - subcommsize; 402228143c3dSStefano Zampini /* check if reuse has been requested */ 402328143c3dSStefano Zampini if (reuse == MAT_REUSE_MATRIX) { 402428143c3dSStefano Zampini if (*mat_n) { 402528143c3dSStefano Zampini PetscMPIInt subcommsize2; 402628143c3dSStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)*mat_n),&subcommsize2);CHKERRQ(ierr); 402728143c3dSStefano Zampini if (subcommsize != subcommsize2) SETERRQ2(PetscObjectComm((PetscObject)*mat_n),PETSC_ERR_PLIB,"Cannot reuse matrix! wrong subcomm size %d != %d",subcommsize,subcommsize2); 402828143c3dSStefano Zampini comm_n = PetscObjectComm((PetscObject)*mat_n); 402928143c3dSStefano Zampini } else { 403028143c3dSStefano Zampini comm_n = PETSC_COMM_SELF; 403128143c3dSStefano Zampini } 403228143c3dSStefano Zampini } else { /* MAT_INITIAL_MATRIX */ 4033779c1cceSStefano Zampini PetscMPIInt rank; 4034779c1cceSStefano Zampini 4035779c1cceSStefano Zampini ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 403628143c3dSStefano Zampini ierr = PetscSubcommCreate(comm,&subcomm);CHKERRQ(ierr); 403728143c3dSStefano Zampini ierr = PetscSubcommSetNumber(subcomm,2);CHKERRQ(ierr); 403828143c3dSStefano Zampini ierr = PetscSubcommSetTypeGeneral(subcomm,color,rank);CHKERRQ(ierr); 4039306c2d5bSBarry Smith comm_n = PetscSubcommChild(subcomm); 404028143c3dSStefano Zampini } 404128143c3dSStefano Zampini /* flag to destroy *mat_n if not significative */ 404228143c3dSStefano Zampini if (color) destroy_mat = PETSC_TRUE; 404328143c3dSStefano Zampini } else { 404428143c3dSStefano Zampini comm_n = comm; 404528143c3dSStefano Zampini } 404628143c3dSStefano Zampini 4047e7931f94SStefano Zampini /* prepare send/receive buffers */ 4048785e854fSJed Brown ierr = PetscMalloc1(commsize,&ilengths_idxs);CHKERRQ(ierr); 4049e7931f94SStefano Zampini ierr = PetscMemzero(ilengths_idxs,commsize*sizeof(*ilengths_idxs));CHKERRQ(ierr); 4050785e854fSJed Brown ierr = PetscMalloc1(commsize,&ilengths_vals);CHKERRQ(ierr); 4051e7931f94SStefano Zampini ierr = PetscMemzero(ilengths_vals,commsize*sizeof(*ilengths_vals));CHKERRQ(ierr); 405228143c3dSStefano Zampini if (nis) { 4053854ce69bSBarry Smith ierr = PetscCalloc1(commsize,&ilengths_idxs_is);CHKERRQ(ierr); 405428143c3dSStefano Zampini } 4055e7931f94SStefano Zampini 405628143c3dSStefano Zampini /* Get data from local matrices */ 4057e7931f94SStefano Zampini if (!isdense) { 4058a26c9d0eSStefano Zampini SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Subassembling of AIJ local matrices not yet implemented"); 4059e7931f94SStefano Zampini /* TODO: See below some guidelines on how to prepare the local buffers */ 4060e7931f94SStefano Zampini /* 4061e7931f94SStefano Zampini send_buffer_vals should contain the raw values of the local matrix 4062e7931f94SStefano Zampini send_buffer_idxs should contain: 4063e7931f94SStefano Zampini - MatType_PRIVATE type 4064e7931f94SStefano Zampini - PetscInt size_of_l2gmap 4065e7931f94SStefano Zampini - PetscInt global_row_indices[size_of_l2gmap] 4066e7931f94SStefano Zampini - PetscInt all_other_info_which_is_needed_to_compute_preallocation_and_set_values 4067e7931f94SStefano Zampini */ 4068e7931f94SStefano Zampini } else { 4069e7931f94SStefano Zampini ierr = MatDenseGetArray(local_mat,&send_buffer_vals);CHKERRQ(ierr); 40703bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetSize(mat->rmap->mapping,&i);CHKERRQ(ierr); 4071854ce69bSBarry Smith ierr = PetscMalloc1(i+2,&send_buffer_idxs);CHKERRQ(ierr); 4072e7931f94SStefano Zampini send_buffer_idxs[0] = (PetscInt)MATDENSE_PRIVATE; 4073e7931f94SStefano Zampini send_buffer_idxs[1] = i; 40743bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetIndices(mat->rmap->mapping,(const PetscInt**)&ptr_idxs);CHKERRQ(ierr); 4075e7931f94SStefano Zampini ierr = PetscMemcpy(&send_buffer_idxs[2],ptr_idxs,i*sizeof(PetscInt));CHKERRQ(ierr); 40763bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingRestoreIndices(mat->rmap->mapping,(const PetscInt**)&ptr_idxs);CHKERRQ(ierr); 4077e7931f94SStefano Zampini ierr = PetscMPIIntCast(i,&len);CHKERRQ(ierr); 4078e7931f94SStefano Zampini for (i=0;i<n_sends;i++) { 4079e7931f94SStefano Zampini ilengths_vals[is_indices[i]] = len*len; 4080e7931f94SStefano Zampini ilengths_idxs[is_indices[i]] = len+2; 4081c8587f34SStefano Zampini } 4082c8587f34SStefano Zampini } 4083e7931f94SStefano Zampini ierr = PetscGatherMessageLengths2(comm,n_sends,n_recvs,ilengths_idxs,ilengths_vals,&onodes,&olengths_idxs,&olengths_vals);CHKERRQ(ierr); 408428143c3dSStefano Zampini /* additional is (if any) */ 408528143c3dSStefano Zampini if (nis) { 408628143c3dSStefano Zampini PetscMPIInt psum; 408728143c3dSStefano Zampini PetscInt j; 408828143c3dSStefano Zampini for (j=0,psum=0;j<nis;j++) { 408928143c3dSStefano Zampini PetscInt plen; 409028143c3dSStefano Zampini ierr = ISGetLocalSize(isarray[j],&plen);CHKERRQ(ierr); 409128143c3dSStefano Zampini ierr = PetscMPIIntCast(plen,&len);CHKERRQ(ierr); 409228143c3dSStefano Zampini psum += len+1; /* indices + lenght */ 409328143c3dSStefano Zampini } 4094854ce69bSBarry Smith ierr = PetscMalloc1(psum,&send_buffer_idxs_is);CHKERRQ(ierr); 409528143c3dSStefano Zampini for (j=0,psum=0;j<nis;j++) { 409628143c3dSStefano Zampini PetscInt plen; 409728143c3dSStefano Zampini const PetscInt *is_array_idxs; 409828143c3dSStefano Zampini ierr = ISGetLocalSize(isarray[j],&plen);CHKERRQ(ierr); 409928143c3dSStefano Zampini send_buffer_idxs_is[psum] = plen; 410028143c3dSStefano Zampini ierr = ISGetIndices(isarray[j],&is_array_idxs);CHKERRQ(ierr); 410128143c3dSStefano Zampini ierr = PetscMemcpy(&send_buffer_idxs_is[psum+1],is_array_idxs,plen*sizeof(PetscInt));CHKERRQ(ierr); 410228143c3dSStefano Zampini ierr = ISRestoreIndices(isarray[j],&is_array_idxs);CHKERRQ(ierr); 410328143c3dSStefano Zampini psum += plen+1; /* indices + lenght */ 410428143c3dSStefano Zampini } 410528143c3dSStefano Zampini for (i=0;i<n_sends;i++) { 410628143c3dSStefano Zampini ilengths_idxs_is[is_indices[i]] = psum; 410728143c3dSStefano Zampini } 410828143c3dSStefano Zampini ierr = PetscGatherMessageLengths(comm,n_sends,n_recvs,ilengths_idxs_is,&onodes_is,&olengths_idxs_is);CHKERRQ(ierr); 410928143c3dSStefano Zampini } 411028143c3dSStefano Zampini 4111e7931f94SStefano Zampini buf_size_idxs = 0; 4112e7931f94SStefano Zampini buf_size_vals = 0; 411328143c3dSStefano Zampini buf_size_idxs_is = 0; 4114e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 4115e7931f94SStefano Zampini buf_size_idxs += (PetscInt)olengths_idxs[i]; 4116e7931f94SStefano Zampini buf_size_vals += (PetscInt)olengths_vals[i]; 411728143c3dSStefano Zampini if (nis) buf_size_idxs_is += (PetscInt)olengths_idxs_is[i]; 4118e7931f94SStefano Zampini } 4119785e854fSJed Brown ierr = PetscMalloc1(buf_size_idxs,&recv_buffer_idxs);CHKERRQ(ierr); 4120785e854fSJed Brown ierr = PetscMalloc1(buf_size_vals,&recv_buffer_vals);CHKERRQ(ierr); 412195ecbf38SStefano Zampini ierr = PetscMalloc1(buf_size_idxs_is,&recv_buffer_idxs_is);CHKERRQ(ierr); 4122e7931f94SStefano Zampini 4123e7931f94SStefano Zampini /* get new tags for clean communications */ 4124e7931f94SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_idxs);CHKERRQ(ierr); 4125e7931f94SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_vals);CHKERRQ(ierr); 412628143c3dSStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_idxs_is);CHKERRQ(ierr); 4127e7931f94SStefano Zampini 4128e7931f94SStefano Zampini /* allocate for requests */ 4129785e854fSJed Brown ierr = PetscMalloc1(n_sends,&send_req_idxs);CHKERRQ(ierr); 4130785e854fSJed Brown ierr = PetscMalloc1(n_sends,&send_req_vals);CHKERRQ(ierr); 413195ecbf38SStefano Zampini ierr = PetscMalloc1(n_sends,&send_req_idxs_is);CHKERRQ(ierr); 4132785e854fSJed Brown ierr = PetscMalloc1(n_recvs,&recv_req_idxs);CHKERRQ(ierr); 4133785e854fSJed Brown ierr = PetscMalloc1(n_recvs,&recv_req_vals);CHKERRQ(ierr); 413495ecbf38SStefano Zampini ierr = PetscMalloc1(n_recvs,&recv_req_idxs_is);CHKERRQ(ierr); 4135e7931f94SStefano Zampini 4136e7931f94SStefano Zampini /* communications */ 4137e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 4138e7931f94SStefano Zampini ptr_vals = recv_buffer_vals; 413928143c3dSStefano Zampini ptr_idxs_is = recv_buffer_idxs_is; 4140e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 4141e7931f94SStefano Zampini source_dest = onodes[i]; 4142e7931f94SStefano Zampini ierr = MPI_Irecv(ptr_idxs,olengths_idxs[i],MPIU_INT,source_dest,tag_idxs,comm,&recv_req_idxs[i]);CHKERRQ(ierr); 4143e7931f94SStefano Zampini ierr = MPI_Irecv(ptr_vals,olengths_vals[i],MPIU_SCALAR,source_dest,tag_vals,comm,&recv_req_vals[i]);CHKERRQ(ierr); 4144e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 4145e7931f94SStefano Zampini ptr_vals += olengths_vals[i]; 414628143c3dSStefano Zampini if (nis) { 414728143c3dSStefano 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); 414828143c3dSStefano Zampini ptr_idxs_is += olengths_idxs_is[i]; 414928143c3dSStefano Zampini } 4150e7931f94SStefano Zampini } 4151e7931f94SStefano Zampini for (i=0;i<n_sends;i++) { 4152e7931f94SStefano Zampini ierr = PetscMPIIntCast(is_indices[i],&source_dest);CHKERRQ(ierr); 4153e7931f94SStefano Zampini ierr = MPI_Isend(send_buffer_idxs,ilengths_idxs[source_dest],MPIU_INT,source_dest,tag_idxs,comm,&send_req_idxs[i]);CHKERRQ(ierr); 4154e7931f94SStefano Zampini ierr = MPI_Isend(send_buffer_vals,ilengths_vals[source_dest],MPIU_SCALAR,source_dest,tag_vals,comm,&send_req_vals[i]);CHKERRQ(ierr); 415528143c3dSStefano Zampini if (nis) { 415628143c3dSStefano 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); 415728143c3dSStefano Zampini } 4158e7931f94SStefano Zampini } 4159e7931f94SStefano Zampini ierr = ISRestoreIndices(is_sends_internal,&is_indices);CHKERRQ(ierr); 4160e7931f94SStefano Zampini ierr = ISDestroy(&is_sends_internal);CHKERRQ(ierr); 4161e7931f94SStefano Zampini 4162e7931f94SStefano Zampini /* assemble new l2g map */ 4163e7931f94SStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_idxs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 4164e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 41659d30be91SStefano Zampini new_local_rows = 0; 4166e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 41679d30be91SStefano Zampini new_local_rows += *(ptr_idxs+1); /* second element is the local size of the l2gmap */ 4168e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 4169e7931f94SStefano Zampini } 41709d30be91SStefano Zampini ierr = PetscMalloc1(new_local_rows,&l2gmap_indices);CHKERRQ(ierr); 4171e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 41729d30be91SStefano Zampini new_local_rows = 0; 4173e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 41749d30be91SStefano Zampini ierr = PetscMemcpy(&l2gmap_indices[new_local_rows],ptr_idxs+2,(*(ptr_idxs+1))*sizeof(PetscInt));CHKERRQ(ierr); 41759d30be91SStefano Zampini new_local_rows += *(ptr_idxs+1); /* second element is the local size of the l2gmap */ 4176e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 4177e7931f94SStefano Zampini } 41789d30be91SStefano Zampini ierr = PetscSortRemoveDupsInt(&new_local_rows,l2gmap_indices);CHKERRQ(ierr); 41799d30be91SStefano Zampini ierr = ISLocalToGlobalMappingCreate(comm_n,1,new_local_rows,l2gmap_indices,PETSC_COPY_VALUES,&l2gmap);CHKERRQ(ierr); 4180e7931f94SStefano Zampini ierr = PetscFree(l2gmap_indices);CHKERRQ(ierr); 4181e7931f94SStefano Zampini 4182e7931f94SStefano Zampini /* infer new local matrix type from received local matrices type */ 4183e7931f94SStefano Zampini /* currently if all local matrices are of type X, then the resulting matrix will be of type X, except for the dense case */ 4184e7931f94SStefano 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) */ 4185e7931f94SStefano Zampini if (n_recvs) { 418628143c3dSStefano Zampini MatTypePrivate new_local_type_private = (MatTypePrivate)send_buffer_idxs[0]; 4187e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 4188e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 4189e7931f94SStefano Zampini if ((PetscInt)new_local_type_private != *ptr_idxs) { 4190e7931f94SStefano Zampini new_local_type_private = MATAIJ_PRIVATE; 4191e7931f94SStefano Zampini break; 4192e7931f94SStefano Zampini } 4193e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 4194e7931f94SStefano Zampini } 4195e7931f94SStefano Zampini switch (new_local_type_private) { 419628143c3dSStefano Zampini case MATDENSE_PRIVATE: 419728143c3dSStefano Zampini if (n_recvs>1) { /* subassembling of dense matrices does not give a dense matrix! */ 4198e7931f94SStefano Zampini new_local_type = MATSEQAIJ; 4199e7931f94SStefano Zampini bs = 1; 420028143c3dSStefano Zampini } else { /* if I receive only 1 dense matrix */ 420128143c3dSStefano Zampini new_local_type = MATSEQDENSE; 420228143c3dSStefano Zampini bs = 1; 420328143c3dSStefano Zampini } 4204e7931f94SStefano Zampini break; 4205e7931f94SStefano Zampini case MATAIJ_PRIVATE: 4206e7931f94SStefano Zampini new_local_type = MATSEQAIJ; 4207e7931f94SStefano Zampini bs = 1; 4208e7931f94SStefano Zampini break; 4209e7931f94SStefano Zampini case MATBAIJ_PRIVATE: 4210e7931f94SStefano Zampini new_local_type = MATSEQBAIJ; 4211e7931f94SStefano Zampini break; 4212e7931f94SStefano Zampini case MATSBAIJ_PRIVATE: 4213e7931f94SStefano Zampini new_local_type = MATSEQSBAIJ; 4214e7931f94SStefano Zampini break; 4215e7931f94SStefano Zampini default: 42169d30be91SStefano Zampini SETERRQ2(comm,PETSC_ERR_SUP,"Unsupported private type %d in %s",new_local_type_private,__FUNCT__); 4217e7931f94SStefano Zampini break; 4218e7931f94SStefano Zampini } 421928143c3dSStefano Zampini } else { /* by default, new_local_type is seqdense */ 422028143c3dSStefano Zampini new_local_type = MATSEQDENSE; 422128143c3dSStefano Zampini bs = 1; 4222e7931f94SStefano Zampini } 4223e7931f94SStefano Zampini 422470cf5478SStefano Zampini /* create MATIS object if needed */ 422570cf5478SStefano Zampini if (reuse == MAT_INITIAL_MATRIX) { 4226e7931f94SStefano Zampini ierr = MatGetSize(mat,&rows,&cols);CHKERRQ(ierr); 4227e176bc59SStefano Zampini ierr = MatCreateIS(comm_n,bs,PETSC_DECIDE,PETSC_DECIDE,rows,cols,l2gmap,NULL,mat_n);CHKERRQ(ierr); 422870cf5478SStefano Zampini } else { 422970cf5478SStefano Zampini /* it also destroys the local matrices */ 423070cf5478SStefano Zampini ierr = MatSetLocalToGlobalMapping(*mat_n,l2gmap,l2gmap);CHKERRQ(ierr); 423170cf5478SStefano Zampini } 423270cf5478SStefano Zampini ierr = MatISGetLocalMat(*mat_n,&local_mat);CHKERRQ(ierr); 4233e7931f94SStefano Zampini ierr = MatSetType(local_mat,new_local_type);CHKERRQ(ierr); 42349d30be91SStefano Zampini 42359d30be91SStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_vals,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 42369d30be91SStefano Zampini 42379d30be91SStefano Zampini /* Global to local map of received indices */ 42389d30be91SStefano Zampini ierr = PetscMalloc1(buf_size_idxs,&recv_buffer_idxs_local);CHKERRQ(ierr); /* needed for values insertion */ 42399d30be91SStefano Zampini ierr = ISGlobalToLocalMappingApply(l2gmap,IS_GTOLM_MASK,buf_size_idxs,recv_buffer_idxs,&i,recv_buffer_idxs_local);CHKERRQ(ierr); 42409d30be91SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&l2gmap);CHKERRQ(ierr); 42419d30be91SStefano Zampini 42429d30be91SStefano Zampini /* restore attributes -> type of incoming data and its size */ 42439d30be91SStefano Zampini buf_size_idxs = 0; 42449d30be91SStefano Zampini for (i=0;i<n_recvs;i++) { 42459d30be91SStefano Zampini recv_buffer_idxs_local[buf_size_idxs] = recv_buffer_idxs[buf_size_idxs]; 42469d30be91SStefano Zampini recv_buffer_idxs_local[buf_size_idxs+1] = recv_buffer_idxs[buf_size_idxs+1]; 42479d30be91SStefano Zampini buf_size_idxs += (PetscInt)olengths_idxs[i]; 42489d30be91SStefano Zampini } 42499d30be91SStefano Zampini ierr = PetscFree(recv_buffer_idxs);CHKERRQ(ierr); 42509d30be91SStefano Zampini 42519d30be91SStefano Zampini /* set preallocation */ 42529d30be91SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)local_mat,MATSEQDENSE,&newisdense);CHKERRQ(ierr); 42539d30be91SStefano Zampini if (!newisdense) { 42549d30be91SStefano Zampini PetscInt *new_local_nnz=0; 42559d30be91SStefano Zampini 42569d30be91SStefano Zampini ptr_vals = recv_buffer_vals; 42579d30be91SStefano Zampini ptr_idxs = recv_buffer_idxs_local; 42589d30be91SStefano Zampini if (n_recvs) { 42599d30be91SStefano Zampini ierr = PetscCalloc1(new_local_rows,&new_local_nnz);CHKERRQ(ierr); 42609d30be91SStefano Zampini } 42619d30be91SStefano Zampini for (i=0;i<n_recvs;i++) { 42629d30be91SStefano Zampini PetscInt j; 42639d30be91SStefano Zampini if (*ptr_idxs == (PetscInt)MATDENSE_PRIVATE) { /* preallocation provided for dense case only */ 42649d30be91SStefano Zampini for (j=0;j<*(ptr_idxs+1);j++) { 42659d30be91SStefano Zampini new_local_nnz[*(ptr_idxs+2+j)] += *(ptr_idxs+1); 42669d30be91SStefano Zampini } 42679d30be91SStefano Zampini } else { 42689d30be91SStefano Zampini /* TODO */ 42699d30be91SStefano Zampini } 42709d30be91SStefano Zampini ptr_idxs += olengths_idxs[i]; 42719d30be91SStefano Zampini } 42729d30be91SStefano Zampini if (new_local_nnz) { 42739d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] = PetscMin(new_local_nnz[i],new_local_rows); 42749d30be91SStefano Zampini ierr = MatSeqAIJSetPreallocation(local_mat,0,new_local_nnz);CHKERRQ(ierr); 42759d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] /= bs; 42769d30be91SStefano Zampini ierr = MatSeqBAIJSetPreallocation(local_mat,bs,0,new_local_nnz);CHKERRQ(ierr); 42779d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] = PetscMax(new_local_nnz[i]-i,0); 42789d30be91SStefano Zampini ierr = MatSeqSBAIJSetPreallocation(local_mat,bs,0,new_local_nnz);CHKERRQ(ierr); 42799d30be91SStefano Zampini } else { 42809d30be91SStefano Zampini ierr = MatSetUp(local_mat);CHKERRQ(ierr); 42819d30be91SStefano Zampini } 42829d30be91SStefano Zampini ierr = PetscFree(new_local_nnz);CHKERRQ(ierr); 42839d30be91SStefano Zampini } else { 42849d30be91SStefano Zampini ierr = MatSetUp(local_mat);CHKERRQ(ierr); 42859d30be91SStefano Zampini } 4286e7931f94SStefano Zampini 4287e7931f94SStefano Zampini /* set values */ 4288e7931f94SStefano Zampini ptr_vals = recv_buffer_vals; 42899d30be91SStefano Zampini ptr_idxs = recv_buffer_idxs_local; 4290e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 4291e7931f94SStefano Zampini if (*ptr_idxs == (PetscInt)MATDENSE_PRIVATE) { /* values insertion provided for dense case only */ 4292e7931f94SStefano Zampini ierr = MatSetOption(local_mat,MAT_ROW_ORIENTED,PETSC_FALSE);CHKERRQ(ierr); 42939d30be91SStefano Zampini ierr = MatSetValues(local_mat,*(ptr_idxs+1),ptr_idxs+2,*(ptr_idxs+1),ptr_idxs+2,ptr_vals,ADD_VALUES);CHKERRQ(ierr); 4294e7931f94SStefano Zampini ierr = MatAssemblyBegin(local_mat,MAT_FLUSH_ASSEMBLY);CHKERRQ(ierr); 4295e7931f94SStefano Zampini ierr = MatAssemblyEnd(local_mat,MAT_FLUSH_ASSEMBLY);CHKERRQ(ierr); 4296e7931f94SStefano Zampini ierr = MatSetOption(local_mat,MAT_ROW_ORIENTED,PETSC_TRUE);CHKERRQ(ierr); 429728143c3dSStefano Zampini } else { 429828143c3dSStefano Zampini /* TODO */ 4299e7931f94SStefano Zampini } 4300e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 4301e7931f94SStefano Zampini ptr_vals += olengths_vals[i]; 4302e7931f94SStefano Zampini } 4303e7931f94SStefano Zampini ierr = MatAssemblyBegin(local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4304e7931f94SStefano Zampini ierr = MatAssemblyEnd(local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 430570cf5478SStefano Zampini ierr = MatAssemblyBegin(*mat_n,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 430670cf5478SStefano Zampini ierr = MatAssemblyEnd(*mat_n,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 43079d30be91SStefano Zampini ierr = PetscFree(recv_buffer_vals);CHKERRQ(ierr); 43089d30be91SStefano Zampini ierr = PetscFree(recv_buffer_idxs_local);CHKERRQ(ierr); 4309e7931f94SStefano Zampini 4310dfd14d43SStefano Zampini #if 0 431128143c3dSStefano Zampini if (!restrict_comm) { /* check */ 4312e7931f94SStefano Zampini Vec lvec,rvec; 4313e7931f94SStefano Zampini PetscReal infty_error; 4314e7931f94SStefano Zampini 43152a7a6963SBarry Smith ierr = MatCreateVecs(mat,&rvec,&lvec);CHKERRQ(ierr); 4316e7931f94SStefano Zampini ierr = VecSetRandom(rvec,NULL);CHKERRQ(ierr); 4317e7931f94SStefano Zampini ierr = MatMult(mat,rvec,lvec);CHKERRQ(ierr); 4318e7931f94SStefano Zampini ierr = VecScale(lvec,-1.0);CHKERRQ(ierr); 431970cf5478SStefano Zampini ierr = MatMultAdd(*mat_n,rvec,lvec,lvec);CHKERRQ(ierr); 4320e7931f94SStefano Zampini ierr = VecNorm(lvec,NORM_INFINITY,&infty_error);CHKERRQ(ierr); 4321e7931f94SStefano Zampini ierr = PetscPrintf(PetscObjectComm((PetscObject)mat),"Infinity error subassembling %1.6e\n",infty_error); 4322e7931f94SStefano Zampini ierr = VecDestroy(&rvec);CHKERRQ(ierr); 4323e7931f94SStefano Zampini ierr = VecDestroy(&lvec);CHKERRQ(ierr); 4324e7931f94SStefano Zampini } 432528143c3dSStefano Zampini #endif 4326e7931f94SStefano Zampini 432728143c3dSStefano Zampini /* assemble new additional is (if any) */ 432828143c3dSStefano Zampini if (nis) { 432928143c3dSStefano Zampini PetscInt **temp_idxs,*count_is,j,psum; 433028143c3dSStefano Zampini 433128143c3dSStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_idxs_is,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 4332854ce69bSBarry Smith ierr = PetscCalloc1(nis,&count_is);CHKERRQ(ierr); 433328143c3dSStefano Zampini ptr_idxs = recv_buffer_idxs_is; 433428143c3dSStefano Zampini psum = 0; 433528143c3dSStefano Zampini for (i=0;i<n_recvs;i++) { 433628143c3dSStefano Zampini for (j=0;j<nis;j++) { 433728143c3dSStefano Zampini PetscInt plen = *(ptr_idxs); /* first element is the local size of IS's indices */ 433828143c3dSStefano Zampini count_is[j] += plen; /* increment counting of buffer for j-th IS */ 433928143c3dSStefano Zampini psum += plen; 434028143c3dSStefano Zampini ptr_idxs += plen+1; /* shift pointer to received data */ 434128143c3dSStefano Zampini } 434228143c3dSStefano Zampini } 4343854ce69bSBarry Smith ierr = PetscMalloc1(nis,&temp_idxs);CHKERRQ(ierr); 4344854ce69bSBarry Smith ierr = PetscMalloc1(psum,&temp_idxs[0]);CHKERRQ(ierr); 434528143c3dSStefano Zampini for (i=1;i<nis;i++) { 434628143c3dSStefano Zampini temp_idxs[i] = temp_idxs[i-1]+count_is[i-1]; 434728143c3dSStefano Zampini } 434828143c3dSStefano Zampini ierr = PetscMemzero(count_is,nis*sizeof(PetscInt));CHKERRQ(ierr); 434928143c3dSStefano Zampini ptr_idxs = recv_buffer_idxs_is; 435028143c3dSStefano Zampini for (i=0;i<n_recvs;i++) { 435128143c3dSStefano Zampini for (j=0;j<nis;j++) { 435228143c3dSStefano Zampini PetscInt plen = *(ptr_idxs); /* first element is the local size of IS's indices */ 435328143c3dSStefano Zampini ierr = PetscMemcpy(&temp_idxs[j][count_is[j]],ptr_idxs+1,plen*sizeof(PetscInt));CHKERRQ(ierr); 435428143c3dSStefano Zampini count_is[j] += plen; /* increment starting point of buffer for j-th IS */ 435528143c3dSStefano Zampini ptr_idxs += plen+1; /* shift pointer to received data */ 435628143c3dSStefano Zampini } 435728143c3dSStefano Zampini } 435828143c3dSStefano Zampini for (i=0;i<nis;i++) { 435928143c3dSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 436028143c3dSStefano Zampini ierr = PetscSortRemoveDupsInt(&count_is[i],temp_idxs[i]);CHKERRQ(ierr);CHKERRQ(ierr); 436128143c3dSStefano Zampini ierr = ISCreateGeneral(comm_n,count_is[i],temp_idxs[i],PETSC_COPY_VALUES,&isarray[i]);CHKERRQ(ierr); 436228143c3dSStefano Zampini } 436328143c3dSStefano Zampini ierr = PetscFree(count_is);CHKERRQ(ierr); 436428143c3dSStefano Zampini ierr = PetscFree(temp_idxs[0]);CHKERRQ(ierr); 436528143c3dSStefano Zampini ierr = PetscFree(temp_idxs);CHKERRQ(ierr); 436628143c3dSStefano Zampini } 4367e7931f94SStefano Zampini /* free workspace */ 436828143c3dSStefano Zampini ierr = PetscFree(recv_buffer_idxs_is);CHKERRQ(ierr); 4369e7931f94SStefano Zampini ierr = MPI_Waitall(n_sends,send_req_idxs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 4370e7931f94SStefano Zampini ierr = PetscFree(send_buffer_idxs);CHKERRQ(ierr); 4371e7931f94SStefano Zampini ierr = MPI_Waitall(n_sends,send_req_vals,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 4372e7931f94SStefano Zampini if (isdense) { 4373e7931f94SStefano Zampini ierr = MatISGetLocalMat(mat,&local_mat);CHKERRQ(ierr); 4374e7931f94SStefano Zampini ierr = MatDenseRestoreArray(local_mat,&send_buffer_vals);CHKERRQ(ierr); 4375e7931f94SStefano Zampini } else { 4376e7931f94SStefano Zampini /* ierr = PetscFree(send_buffer_vals);CHKERRQ(ierr); */ 4377e7931f94SStefano Zampini } 437828143c3dSStefano Zampini if (nis) { 437928143c3dSStefano Zampini ierr = MPI_Waitall(n_sends,send_req_idxs_is,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 438028143c3dSStefano Zampini ierr = PetscFree(send_buffer_idxs_is);CHKERRQ(ierr); 438128143c3dSStefano Zampini } 4382e7931f94SStefano Zampini ierr = PetscFree(recv_req_idxs);CHKERRQ(ierr); 4383e7931f94SStefano Zampini ierr = PetscFree(recv_req_vals);CHKERRQ(ierr); 438428143c3dSStefano Zampini ierr = PetscFree(recv_req_idxs_is);CHKERRQ(ierr); 4385e7931f94SStefano Zampini ierr = PetscFree(send_req_idxs);CHKERRQ(ierr); 4386e7931f94SStefano Zampini ierr = PetscFree(send_req_vals);CHKERRQ(ierr); 438728143c3dSStefano Zampini ierr = PetscFree(send_req_idxs_is);CHKERRQ(ierr); 4388e7931f94SStefano Zampini ierr = PetscFree(ilengths_vals);CHKERRQ(ierr); 4389e7931f94SStefano Zampini ierr = PetscFree(ilengths_idxs);CHKERRQ(ierr); 4390e7931f94SStefano Zampini ierr = PetscFree(olengths_vals);CHKERRQ(ierr); 4391e7931f94SStefano Zampini ierr = PetscFree(olengths_idxs);CHKERRQ(ierr); 4392e7931f94SStefano Zampini ierr = PetscFree(onodes);CHKERRQ(ierr); 439328143c3dSStefano Zampini if (nis) { 439428143c3dSStefano Zampini ierr = PetscFree(ilengths_idxs_is);CHKERRQ(ierr); 439528143c3dSStefano Zampini ierr = PetscFree(olengths_idxs_is);CHKERRQ(ierr); 439628143c3dSStefano Zampini ierr = PetscFree(onodes_is);CHKERRQ(ierr); 439728143c3dSStefano Zampini } 439828143c3dSStefano Zampini ierr = PetscSubcommDestroy(&subcomm);CHKERRQ(ierr); 439928143c3dSStefano Zampini if (destroy_mat) { /* destroy mat is true only if restrict comm is true and process will not partecipate */ 440028143c3dSStefano Zampini ierr = MatDestroy(mat_n);CHKERRQ(ierr); 440128143c3dSStefano Zampini for (i=0;i<nis;i++) { 440228143c3dSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 440328143c3dSStefano Zampini } 440453a05cb3SStefano Zampini *mat_n = NULL; 440528143c3dSStefano Zampini } 4406e7931f94SStefano Zampini PetscFunctionReturn(0); 4407e7931f94SStefano Zampini } 4408a57a6d2fSStefano Zampini 440912edc857SStefano Zampini /* temporary hack into ksp private data structure */ 4410af0996ceSBarry Smith #include <petsc/private/kspimpl.h> 441112edc857SStefano Zampini 4412c8587f34SStefano Zampini #undef __FUNCT__ 4413c8587f34SStefano Zampini #define __FUNCT__ "PCBDDCSetUpCoarseSolver" 4414c8587f34SStefano Zampini PetscErrorCode PCBDDCSetUpCoarseSolver(PC pc,PetscScalar* coarse_submat_vals) 4415c8587f34SStefano Zampini { 4416c8587f34SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 4417c8587f34SStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 441820a2ab83SStefano Zampini Mat coarse_mat,coarse_mat_is,coarse_submat_dense; 44199881197aSStefano Zampini MatNullSpace CoarseNullSpace=NULL; 442020a2ab83SStefano Zampini ISLocalToGlobalMapping coarse_islg; 44216e683305SStefano Zampini IS coarse_is,*isarray; 44226e683305SStefano Zampini PetscInt i,im_active=-1,active_procs=-1; 44236e683305SStefano Zampini PetscInt nis,nisdofs,nisneu; 4424f9eb5b7dSStefano Zampini PC pc_temp; 4425c8587f34SStefano Zampini PCType coarse_pc_type; 4426c8587f34SStefano Zampini KSPType coarse_ksp_type; 4427f9eb5b7dSStefano Zampini PetscBool multilevel_requested,multilevel_allowed; 44284f3a063dSStefano Zampini PetscBool isredundant,isbddc,isnn,coarse_reuse; 44296e683305SStefano Zampini Mat t_coarse_mat_is; 44306e683305SStefano Zampini PetscInt void_procs,ncoarse_ml,ncoarse_ds,ncoarse; 44316e683305SStefano Zampini PetscMPIInt all_procs; 443274e2c79eSStefano Zampini PetscBool csin_ml,csin_ds,csin,csin_type_simple,redist; 443368457ee5SStefano Zampini PetscBool compute_vecs = PETSC_FALSE; 443422bc73bbSStefano Zampini PetscScalar *array; 44359881197aSStefano Zampini PetscErrorCode ierr; 4436fdc09c96SStefano Zampini 4437c8587f34SStefano Zampini PetscFunctionBegin; 4438c8587f34SStefano Zampini /* Assign global numbering to coarse dofs */ 443968457ee5SStefano 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 */ 4440fa7f1dd8SStefano Zampini PetscInt ocoarse_size; 44415a75c04eSSatish Balay compute_vecs = PETSC_TRUE; 4442fa7f1dd8SStefano Zampini ocoarse_size = pcbddc->coarse_size; 4443f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->global_primal_indices);CHKERRQ(ierr); 4444f4ddd8eeSStefano Zampini ierr = PCBDDCComputePrimalNumbering(pc,&pcbddc->coarse_size,&pcbddc->global_primal_indices);CHKERRQ(ierr); 4445f4ddd8eeSStefano Zampini /* see if we can avoid some work */ 4446fa7f1dd8SStefano Zampini if (pcbddc->coarse_ksp) { /* coarse ksp has already been created */ 444751bea450SStefano Zampini /* if the coarse size is different or we are using adaptive selection, better to not reuse the coarse matrix */ 444851bea450SStefano Zampini if (ocoarse_size != pcbddc->coarse_size || pcbddc->adaptive_selection) { 4449dc4bcba2SStefano Zampini PC pc; 4450dc4bcba2SStefano Zampini PetscBool isbddc; 4451dc4bcba2SStefano Zampini 4452dc4bcba2SStefano Zampini /* temporary workaround since PCBDDC does not have a reset method so far */ 4453dc4bcba2SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&pc);CHKERRQ(ierr); 4454dc4bcba2SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc,PCBDDC,&isbddc);CHKERRQ(ierr); 4455dc4bcba2SStefano Zampini if (isbddc) { 4456dc4bcba2SStefano Zampini ierr = PCDestroy(&pc);CHKERRQ(ierr); 4457dc4bcba2SStefano Zampini } 4458727cdba6SStefano Zampini ierr = KSPReset(pcbddc->coarse_ksp);CHKERRQ(ierr); 4459fa7f1dd8SStefano Zampini coarse_reuse = PETSC_FALSE; 4460fa7f1dd8SStefano Zampini } else { /* we can safely reuse already computed coarse matrix */ 4461fa7f1dd8SStefano Zampini coarse_reuse = PETSC_TRUE; 4462f4ddd8eeSStefano Zampini } 4463fa7f1dd8SStefano Zampini } else { /* there's no coarse ksp, so we need to create the coarse matrix too */ 4464fa7f1dd8SStefano Zampini coarse_reuse = PETSC_FALSE; 4465f4ddd8eeSStefano Zampini } 446670cf5478SStefano Zampini /* reset any subassembling information */ 446770cf5478SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling);CHKERRQ(ierr); 44686e683305SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 44696e683305SStefano Zampini } else { /* primal space is unchanged, so we can reuse coarse matrix */ 4470fa7f1dd8SStefano Zampini coarse_reuse = PETSC_TRUE; 4471f4ddd8eeSStefano Zampini } 4472c8587f34SStefano Zampini 44736e683305SStefano Zampini /* count "active" (i.e. with positive local size) and "void" processes */ 44742b510759SStefano Zampini im_active = !!(pcis->n); 44752b510759SStefano Zampini ierr = MPI_Allreduce(&im_active,&active_procs,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 44766e683305SStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)pc),&all_procs);CHKERRQ(ierr); 44776e683305SStefano Zampini void_procs = all_procs-active_procs; 44786e683305SStefano Zampini csin_type_simple = PETSC_TRUE; 447974e2c79eSStefano Zampini redist = PETSC_FALSE; 448022bc73bbSStefano Zampini if (pcbddc->current_level && void_procs) { 44816e683305SStefano Zampini csin_ml = PETSC_TRUE; 44826e683305SStefano Zampini ncoarse_ml = void_procs; 4483779c1cceSStefano Zampini /* it has no sense to redistribute on a set of processors larger than the number of active processes */ 4484779c1cceSStefano Zampini if (pcbddc->redistribute_coarse > 0 && pcbddc->redistribute_coarse < active_procs) { 44856e683305SStefano Zampini csin_ds = PETSC_TRUE; 448618a45a71SStefano Zampini ncoarse_ds = pcbddc->redistribute_coarse; 448718a45a71SStefano Zampini redist = PETSC_TRUE; 448818a45a71SStefano Zampini } else { 44896e683305SStefano Zampini csin_ds = PETSC_TRUE; 4490779c1cceSStefano Zampini ncoarse_ds = active_procs; 4491779c1cceSStefano Zampini redist = PETSC_TRUE; 449218a45a71SStefano Zampini } 44936e683305SStefano Zampini } else { 44946e683305SStefano Zampini csin_ml = PETSC_FALSE; 44956e683305SStefano Zampini ncoarse_ml = all_procs; 44966e683305SStefano Zampini if (void_procs) { 44976e683305SStefano Zampini csin_ds = PETSC_TRUE; 44986e683305SStefano Zampini ncoarse_ds = void_procs; 44996e683305SStefano Zampini csin_type_simple = PETSC_FALSE; 45006e683305SStefano Zampini } else { 4501779c1cceSStefano Zampini if (pcbddc->redistribute_coarse > 0 && pcbddc->redistribute_coarse < all_procs) { 450274e2c79eSStefano Zampini csin_ds = PETSC_TRUE; 450374e2c79eSStefano Zampini ncoarse_ds = pcbddc->redistribute_coarse; 450474e2c79eSStefano Zampini redist = PETSC_TRUE; 450574e2c79eSStefano Zampini } else { 45066e683305SStefano Zampini csin_ds = PETSC_FALSE; 45076e683305SStefano Zampini ncoarse_ds = all_procs; 45086e683305SStefano Zampini } 45096e683305SStefano Zampini } 451074e2c79eSStefano Zampini } 45116e683305SStefano Zampini 45126e683305SStefano Zampini /* 45136e683305SStefano Zampini test if we can go multilevel: three conditions must be satisfied: 45146e683305SStefano Zampini - we have not exceeded the number of levels requested 45156e683305SStefano Zampini - we can actually subassemble the active processes 45166e683305SStefano Zampini - we can find a suitable number of MPI processes where we can place the subassembled problem 45176e683305SStefano Zampini */ 45186e683305SStefano Zampini multilevel_allowed = PETSC_FALSE; 45196e683305SStefano Zampini multilevel_requested = PETSC_FALSE; 45206e683305SStefano Zampini if (pcbddc->current_level < pcbddc->max_levels) { 45216e683305SStefano Zampini multilevel_requested = PETSC_TRUE; 45226e683305SStefano Zampini if (active_procs/pcbddc->coarsening_ratio < 2 || ncoarse_ml/pcbddc->coarsening_ratio < 2) { 4523f9eb5b7dSStefano Zampini multilevel_allowed = PETSC_FALSE; 45242b510759SStefano Zampini } else { 4525f9eb5b7dSStefano Zampini multilevel_allowed = PETSC_TRUE; 4526c8587f34SStefano Zampini } 4527c8587f34SStefano Zampini } 45286e683305SStefano Zampini /* determine number of process partecipating to coarse solver */ 45296e683305SStefano Zampini if (multilevel_allowed) { 45306e683305SStefano Zampini ncoarse = ncoarse_ml; 45316e683305SStefano Zampini csin = csin_ml; 453258da7f69SStefano Zampini redist = PETSC_FALSE; 45336e683305SStefano Zampini } else { 45346e683305SStefano Zampini ncoarse = ncoarse_ds; 45356e683305SStefano Zampini csin = csin_ds; 45366e683305SStefano Zampini } 4537e7931f94SStefano Zampini 4538abbbba34SStefano Zampini /* creates temporary l2gmap and IS for coarse indexes */ 4539abbbba34SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),pcbddc->local_primal_size,pcbddc->global_primal_indices,PETSC_COPY_VALUES,&coarse_is);CHKERRQ(ierr); 4540abbbba34SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(coarse_is,&coarse_islg);CHKERRQ(ierr); 4541abbbba34SStefano Zampini 4542abbbba34SStefano Zampini /* creates temporary MATIS object for coarse matrix */ 454322bc73bbSStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,pcbddc->local_primal_size,pcbddc->local_primal_size,NULL,&coarse_submat_dense);CHKERRQ(ierr); 454422bc73bbSStefano Zampini ierr = MatDenseGetArray(coarse_submat_dense,&array);CHKERRQ(ierr); 454522bc73bbSStefano Zampini ierr = PetscMemcpy(array,coarse_submat_vals,sizeof(*coarse_submat_vals)*pcbddc->local_primal_size*pcbddc->local_primal_size);CHKERRQ(ierr); 454622bc73bbSStefano Zampini ierr = MatDenseRestoreArray(coarse_submat_dense,&array);CHKERRQ(ierr); 4547b1ecc7b1SStefano Zampini #if 0 4548b9b85e73SStefano Zampini { 4549b9b85e73SStefano Zampini PetscViewer viewer; 4550b9b85e73SStefano Zampini char filename[256]; 4551b1ecc7b1SStefano Zampini sprintf(filename,"local_coarse_mat%d_level%d.m",PetscGlobalRank,pcbddc->current_level); 4552b9b85e73SStefano Zampini ierr = PetscViewerASCIIOpen(PETSC_COMM_SELF,filename,&viewer);CHKERRQ(ierr); 4553b9b85e73SStefano Zampini ierr = PetscViewerSetFormat(viewer,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 4554b9b85e73SStefano Zampini ierr = MatView(coarse_submat_dense,viewer);CHKERRQ(ierr); 4555b9b85e73SStefano Zampini ierr = PetscViewerDestroy(&viewer);CHKERRQ(ierr); 4556b9b85e73SStefano Zampini } 4557b9b85e73SStefano Zampini #endif 4558e176bc59SStefano 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); 45596e683305SStefano Zampini ierr = MatISSetLocalMat(t_coarse_mat_is,coarse_submat_dense);CHKERRQ(ierr); 45606e683305SStefano Zampini ierr = MatAssemblyBegin(t_coarse_mat_is,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 45616e683305SStefano Zampini ierr = MatAssemblyEnd(t_coarse_mat_is,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4562abbbba34SStefano Zampini ierr = MatDestroy(&coarse_submat_dense);CHKERRQ(ierr); 4563abbbba34SStefano Zampini 45646e683305SStefano Zampini /* compute dofs splitting and neumann boundaries for coarse dofs */ 45656e683305SStefano Zampini if (multilevel_allowed && (pcbddc->n_ISForDofsLocal || pcbddc->NeumannBoundariesLocal) ) { /* protects from unneded computations */ 45666e683305SStefano Zampini PetscInt *tidxs,*tidxs2,nout,tsize,i; 45676e683305SStefano Zampini const PetscInt *idxs; 45686e683305SStefano Zampini ISLocalToGlobalMapping tmap; 45696e683305SStefano Zampini 45706e683305SStefano Zampini /* create map between primal indices (in local representative ordering) and local primal numbering */ 45710be93d54SStefano Zampini ierr = ISLocalToGlobalMappingCreate(PETSC_COMM_SELF,1,pcbddc->local_primal_size,pcbddc->primal_indices_local_idxs,PETSC_COPY_VALUES,&tmap);CHKERRQ(ierr); 45726e683305SStefano Zampini /* allocate space for temporary storage */ 4573854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->local_primal_size,&tidxs);CHKERRQ(ierr); 4574854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->local_primal_size,&tidxs2);CHKERRQ(ierr); 45756e683305SStefano Zampini /* allocate for IS array */ 45766e683305SStefano Zampini nisdofs = pcbddc->n_ISForDofsLocal; 45776e683305SStefano Zampini nisneu = !!pcbddc->NeumannBoundariesLocal; 45786e683305SStefano Zampini nis = nisdofs + nisneu; 4579854ce69bSBarry Smith ierr = PetscMalloc1(nis,&isarray);CHKERRQ(ierr); 45806e683305SStefano Zampini /* dofs splitting */ 45816e683305SStefano Zampini for (i=0;i<nisdofs;i++) { 45826e683305SStefano Zampini /* ierr = ISView(pcbddc->ISForDofsLocal[i],0);CHKERRQ(ierr); */ 45836e683305SStefano Zampini ierr = ISGetLocalSize(pcbddc->ISForDofsLocal[i],&tsize);CHKERRQ(ierr); 45846e683305SStefano Zampini ierr = ISGetIndices(pcbddc->ISForDofsLocal[i],&idxs);CHKERRQ(ierr); 45856e683305SStefano Zampini ierr = ISGlobalToLocalMappingApply(tmap,IS_GTOLM_DROP,tsize,idxs,&nout,tidxs);CHKERRQ(ierr); 45866e683305SStefano Zampini ierr = ISRestoreIndices(pcbddc->ISForDofsLocal[i],&idxs);CHKERRQ(ierr); 45876e683305SStefano Zampini ierr = ISLocalToGlobalMappingApply(coarse_islg,nout,tidxs,tidxs2);CHKERRQ(ierr); 45886e683305SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pcbddc->ISForDofsLocal[i]),nout,tidxs2,PETSC_COPY_VALUES,&isarray[i]);CHKERRQ(ierr); 45896e683305SStefano Zampini /* ierr = ISView(isarray[i],0);CHKERRQ(ierr); */ 45906e683305SStefano Zampini } 45916e683305SStefano Zampini /* neumann boundaries */ 45926e683305SStefano Zampini if (pcbddc->NeumannBoundariesLocal) { 45936e683305SStefano Zampini /* ierr = ISView(pcbddc->NeumannBoundariesLocal,0);CHKERRQ(ierr); */ 45946e683305SStefano Zampini ierr = ISGetLocalSize(pcbddc->NeumannBoundariesLocal,&tsize);CHKERRQ(ierr); 45956e683305SStefano Zampini ierr = ISGetIndices(pcbddc->NeumannBoundariesLocal,&idxs);CHKERRQ(ierr); 45966e683305SStefano Zampini ierr = ISGlobalToLocalMappingApply(tmap,IS_GTOLM_DROP,tsize,idxs,&nout,tidxs);CHKERRQ(ierr); 45976e683305SStefano Zampini ierr = ISRestoreIndices(pcbddc->NeumannBoundariesLocal,&idxs);CHKERRQ(ierr); 45986e683305SStefano Zampini ierr = ISLocalToGlobalMappingApply(coarse_islg,nout,tidxs,tidxs2);CHKERRQ(ierr); 45996e683305SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pcbddc->NeumannBoundariesLocal),nout,tidxs2,PETSC_COPY_VALUES,&isarray[nisdofs]);CHKERRQ(ierr); 46006e683305SStefano Zampini /* ierr = ISView(isarray[nisdofs],0);CHKERRQ(ierr); */ 46016e683305SStefano Zampini } 46026e683305SStefano Zampini /* free memory */ 46036e683305SStefano Zampini ierr = PetscFree(tidxs);CHKERRQ(ierr); 46046e683305SStefano Zampini ierr = PetscFree(tidxs2);CHKERRQ(ierr); 46056e683305SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&tmap);CHKERRQ(ierr); 46066e683305SStefano Zampini } else { 46076e683305SStefano Zampini nis = 0; 46086e683305SStefano Zampini nisdofs = 0; 46096e683305SStefano Zampini nisneu = 0; 46106e683305SStefano Zampini isarray = NULL; 46116e683305SStefano Zampini } 46126e683305SStefano Zampini /* destroy no longer needed map */ 46136e683305SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&coarse_islg);CHKERRQ(ierr); 46146e683305SStefano Zampini 46156e683305SStefano Zampini /* restrict on coarse candidates (if needed) */ 46166e683305SStefano Zampini coarse_mat_is = NULL; 46176e683305SStefano Zampini if (csin) { 46186e683305SStefano Zampini if (!pcbddc->coarse_subassembling_init ) { /* creates subassembling init pattern if not present */ 461974e2c79eSStefano Zampini if (redist) { 462074e2c79eSStefano Zampini PetscMPIInt rank; 4621779c1cceSStefano Zampini PetscInt spc,n_spc_p1,dest[1],destsize; 462274e2c79eSStefano Zampini 462374e2c79eSStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)pc),&rank);CHKERRQ(ierr); 462458da7f69SStefano Zampini spc = active_procs/ncoarse; 462558da7f69SStefano Zampini n_spc_p1 = active_procs%ncoarse; 4626779c1cceSStefano Zampini if (im_active) { 4627779c1cceSStefano Zampini destsize = 1; 462874e2c79eSStefano Zampini if (rank > n_spc_p1*(spc+1)-1) { 462974e2c79eSStefano Zampini dest[0] = n_spc_p1+(rank-(n_spc_p1*(spc+1)))/spc; 463074e2c79eSStefano Zampini } else { 463174e2c79eSStefano Zampini dest[0] = rank/(spc+1); 463274e2c79eSStefano Zampini } 463374e2c79eSStefano Zampini } else { 4634779c1cceSStefano Zampini destsize = 0; 46356e683305SStefano Zampini } 4636779c1cceSStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),destsize,dest,PETSC_COPY_VALUES,&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 4637779c1cceSStefano Zampini } else if (csin_type_simple) { 46386e683305SStefano Zampini PetscMPIInt rank; 46396e683305SStefano Zampini PetscInt issize,isidx; 4640779c1cceSStefano Zampini 46416e683305SStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)pc),&rank);CHKERRQ(ierr); 46426e683305SStefano Zampini if (im_active) { 46436e683305SStefano Zampini issize = 1; 46446e683305SStefano Zampini isidx = (PetscInt)rank; 46456e683305SStefano Zampini } else { 46466e683305SStefano Zampini issize = 0; 46476e683305SStefano Zampini isidx = -1; 46486e683305SStefano Zampini } 46496e683305SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),issize,&isidx,PETSC_COPY_VALUES,&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 4650779c1cceSStefano Zampini } else { /* get a suitable subassembling pattern from MATIS code */ 4651b0c7d250SStefano Zampini ierr = MatISGetSubassemblingPattern(t_coarse_mat_is,ncoarse,pcbddc->coarse_adj_red,&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 46526e683305SStefano Zampini } 4653779c1cceSStefano Zampini 4654779c1cceSStefano Zampini /* we need to shift on coarse candidates either if we are not redistributing or we are redistributing and we have enough void processes */ 4655779c1cceSStefano Zampini if (!redist || ncoarse <= void_procs) { 4656779c1cceSStefano Zampini PetscInt ncoarse_cand,tissize,*nisindices; 4657779c1cceSStefano Zampini PetscInt *coarse_candidates; 4658779c1cceSStefano Zampini const PetscInt* tisindices; 4659779c1cceSStefano Zampini 4660779c1cceSStefano Zampini /* get coarse candidates' ranks in pc communicator */ 4661779c1cceSStefano Zampini ierr = PetscMalloc1(all_procs,&coarse_candidates);CHKERRQ(ierr); 4662779c1cceSStefano Zampini ierr = MPI_Allgather(&im_active,1,MPIU_INT,coarse_candidates,1,MPIU_INT,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 4663779c1cceSStefano Zampini for (i=0,ncoarse_cand=0;i<all_procs;i++) { 4664779c1cceSStefano Zampini if (!coarse_candidates[i]) { 4665779c1cceSStefano Zampini coarse_candidates[ncoarse_cand++]=i; 4666779c1cceSStefano Zampini } 4667779c1cceSStefano Zampini } 4668779c1cceSStefano Zampini if (ncoarse_cand < ncoarse) SETERRQ2(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"This should not happen! %d < %d",ncoarse_cand,ncoarse); 4669779c1cceSStefano Zampini 4670779c1cceSStefano Zampini 46716e683305SStefano Zampini if (pcbddc->dbg_flag) { 46726e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 46736e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Subassembling pattern init (before shift)\n");CHKERRQ(ierr); 46746e683305SStefano Zampini ierr = ISView(pcbddc->coarse_subassembling_init,pcbddc->dbg_viewer);CHKERRQ(ierr); 46756e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Coarse candidates\n");CHKERRQ(ierr); 4676779c1cceSStefano Zampini for (i=0;i<ncoarse_cand;i++) { 46776e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"%d ",coarse_candidates[i]);CHKERRQ(ierr); 46786e683305SStefano Zampini } 46796e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"\n");CHKERRQ(ierr); 46806e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 46816e683305SStefano Zampini } 46826e683305SStefano Zampini /* shift the pattern on coarse candidates */ 46836e683305SStefano Zampini ierr = ISGetLocalSize(pcbddc->coarse_subassembling_init,&tissize);CHKERRQ(ierr); 46846e683305SStefano Zampini ierr = ISGetIndices(pcbddc->coarse_subassembling_init,&tisindices);CHKERRQ(ierr); 4685854ce69bSBarry Smith ierr = PetscMalloc1(tissize,&nisindices);CHKERRQ(ierr); 46866e683305SStefano Zampini for (i=0;i<tissize;i++) nisindices[i] = coarse_candidates[tisindices[i]]; 46876e683305SStefano Zampini ierr = ISRestoreIndices(pcbddc->coarse_subassembling_init,&tisindices);CHKERRQ(ierr); 46886e683305SStefano Zampini ierr = ISGeneralSetIndices(pcbddc->coarse_subassembling_init,tissize,nisindices,PETSC_OWN_POINTER);CHKERRQ(ierr); 46896e683305SStefano Zampini ierr = PetscFree(coarse_candidates);CHKERRQ(ierr); 46906e683305SStefano Zampini } 46916e683305SStefano Zampini if (pcbddc->dbg_flag) { 46926e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 46936e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Subassembling pattern init\n");CHKERRQ(ierr); 46946e683305SStefano Zampini ierr = ISView(pcbddc->coarse_subassembling_init,pcbddc->dbg_viewer);CHKERRQ(ierr); 46956e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 46966e683305SStefano Zampini } 4697779c1cceSStefano Zampini } 46986e683305SStefano Zampini /* get temporary coarse mat in IS format restricted on coarse procs (plus additional index sets of isarray) */ 469953a05cb3SStefano Zampini if (multilevel_allowed) { /* we need to keep tracking of void processes for future placements */ 470053a05cb3SStefano Zampini ierr = MatISSubassemble(t_coarse_mat_is,pcbddc->coarse_subassembling_init,0,PETSC_TRUE,PETSC_FALSE,MAT_INITIAL_MATRIX,&coarse_mat_is,nis,isarray);CHKERRQ(ierr); 470153a05cb3SStefano Zampini } else { /* this is the last level, so use just receiving processes in subcomm */ 470253a05cb3SStefano Zampini ierr = MatISSubassemble(t_coarse_mat_is,pcbddc->coarse_subassembling_init,0,PETSC_TRUE,PETSC_TRUE,MAT_INITIAL_MATRIX,&coarse_mat_is,nis,isarray);CHKERRQ(ierr); 470353a05cb3SStefano Zampini } 47046e683305SStefano Zampini } else { 47056e683305SStefano Zampini if (pcbddc->dbg_flag) { 47066e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 47076e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Subassembling pattern init not needed\n");CHKERRQ(ierr); 47086e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 47096e683305SStefano Zampini } 47106e683305SStefano Zampini ierr = PetscObjectReference((PetscObject)t_coarse_mat_is);CHKERRQ(ierr); 47116e683305SStefano Zampini coarse_mat_is = t_coarse_mat_is; 47126e683305SStefano Zampini } 47136e683305SStefano Zampini 47146e683305SStefano Zampini /* create local to global scatters for coarse problem */ 471568457ee5SStefano Zampini if (compute_vecs) { 47166e683305SStefano Zampini PetscInt lrows; 47176e683305SStefano Zampini ierr = VecDestroy(&pcbddc->coarse_vec);CHKERRQ(ierr); 47186e683305SStefano Zampini if (coarse_mat_is) { 47196e683305SStefano Zampini ierr = MatGetLocalSize(coarse_mat_is,&lrows,NULL);CHKERRQ(ierr); 47206e683305SStefano Zampini } else { 47216e683305SStefano Zampini lrows = 0; 47226e683305SStefano Zampini } 47236e683305SStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pc),&pcbddc->coarse_vec);CHKERRQ(ierr); 47246e683305SStefano Zampini ierr = VecSetSizes(pcbddc->coarse_vec,lrows,PETSC_DECIDE);CHKERRQ(ierr); 47256e683305SStefano Zampini ierr = VecSetType(pcbddc->coarse_vec,VECSTANDARD);CHKERRQ(ierr); 47266e683305SStefano Zampini ierr = VecScatterDestroy(&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 47276e683305SStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_P,NULL,pcbddc->coarse_vec,coarse_is,&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 47286e683305SStefano Zampini } 47296e683305SStefano Zampini ierr = ISDestroy(&coarse_is);CHKERRQ(ierr); 47306e683305SStefano Zampini ierr = MatDestroy(&t_coarse_mat_is);CHKERRQ(ierr); 4731c8587f34SStefano Zampini 4732f9eb5b7dSStefano Zampini /* set defaults for coarse KSP and PC */ 4733f9eb5b7dSStefano Zampini if (multilevel_allowed) { 4734f9eb5b7dSStefano Zampini coarse_ksp_type = KSPRICHARDSON; 4735f9eb5b7dSStefano Zampini coarse_pc_type = PCBDDC; 4736f9eb5b7dSStefano Zampini } else { 4737f9eb5b7dSStefano Zampini coarse_ksp_type = KSPPREONLY; 4738f9eb5b7dSStefano Zampini coarse_pc_type = PCREDUNDANT; 4739c8587f34SStefano Zampini } 4740c8587f34SStefano Zampini 47416e683305SStefano Zampini /* print some info if requested */ 47426e683305SStefano Zampini if (pcbddc->dbg_flag) { 47436e683305SStefano Zampini if (!multilevel_allowed) { 47446e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 47456e683305SStefano Zampini if (multilevel_requested) { 47466e683305SStefano 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); 47476e683305SStefano Zampini } else if (pcbddc->max_levels) { 47486e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Maximum number of requested levels reached (%d)\n",pcbddc->max_levels);CHKERRQ(ierr); 47496e683305SStefano Zampini } 47506e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 47516e683305SStefano Zampini } 47526e683305SStefano Zampini } 47536e683305SStefano Zampini 4754f9eb5b7dSStefano Zampini /* create the coarse KSP object only once with defaults */ 47556e683305SStefano Zampini if (coarse_mat_is) { 47566e683305SStefano Zampini MatReuse coarse_mat_reuse; 47576a1308c2SStefano Zampini PetscViewer dbg_viewer = NULL; 47586e683305SStefano Zampini if (pcbddc->dbg_flag) { 47596e683305SStefano Zampini dbg_viewer = PETSC_VIEWER_STDOUT_(PetscObjectComm((PetscObject)coarse_mat_is)); 47606e683305SStefano Zampini ierr = PetscViewerASCIIAddTab(dbg_viewer,2*pcbddc->current_level);CHKERRQ(ierr); 47616e683305SStefano Zampini } 4762f9eb5b7dSStefano Zampini if (!pcbddc->coarse_ksp) { 4763312be037SStefano Zampini char prefix[256],str_level[16]; 4764e604994aSStefano Zampini size_t len; 47656e683305SStefano Zampini ierr = KSPCreate(PetscObjectComm((PetscObject)coarse_mat_is),&pcbddc->coarse_ksp);CHKERRQ(ierr); 4766422a814eSBarry Smith ierr = KSPSetErrorIfNotConverged(pcbddc->coarse_ksp,pc->erroriffailure);CHKERRQ(ierr); 4767c8587f34SStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->coarse_ksp,(PetscObject)pc,1);CHKERRQ(ierr); 4768f9eb5b7dSStefano Zampini ierr = KSPSetTolerances(pcbddc->coarse_ksp,PETSC_DEFAULT,PETSC_DEFAULT,PETSC_DEFAULT,1);CHKERRQ(ierr); 47695f76c7aeSStefano Zampini ierr = KSPSetOperators(pcbddc->coarse_ksp,coarse_mat_is,coarse_mat_is);CHKERRQ(ierr); 4770c8587f34SStefano Zampini ierr = KSPSetType(pcbddc->coarse_ksp,coarse_ksp_type);CHKERRQ(ierr); 47716e683305SStefano Zampini ierr = KSPSetNormType(pcbddc->coarse_ksp,KSP_NORM_NONE);CHKERRQ(ierr); 4772c8587f34SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&pc_temp);CHKERRQ(ierr); 4773c8587f34SStefano Zampini ierr = PCSetType(pc_temp,coarse_pc_type);CHKERRQ(ierr); 4774e604994aSStefano Zampini /* prefix */ 4775e604994aSStefano Zampini ierr = PetscStrcpy(prefix,"");CHKERRQ(ierr); 4776e604994aSStefano Zampini ierr = PetscStrcpy(str_level,"");CHKERRQ(ierr); 4777e604994aSStefano Zampini if (!pcbddc->current_level) { 4778e604994aSStefano Zampini ierr = PetscStrcpy(prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 4779e604994aSStefano Zampini ierr = PetscStrcat(prefix,"pc_bddc_coarse_");CHKERRQ(ierr); 4780c8587f34SStefano Zampini } else { 4781e604994aSStefano Zampini ierr = PetscStrlen(((PetscObject)pc)->prefix,&len);CHKERRQ(ierr); 4782312be037SStefano Zampini if (pcbddc->current_level>1) len -= 3; /* remove "lX_" with X level number */ 4783312be037SStefano Zampini if (pcbddc->current_level>10) len -= 1; /* remove another char from level number */ 478434d6797cSStefano Zampini ierr = PetscStrncpy(prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 4785312be037SStefano Zampini sprintf(str_level,"l%d_",(int)(pcbddc->current_level)); 4786e604994aSStefano Zampini ierr = PetscStrcat(prefix,str_level);CHKERRQ(ierr); 4787e604994aSStefano Zampini } 4788e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->coarse_ksp,prefix);CHKERRQ(ierr); 47893e3c6dadSStefano Zampini /* propagate BDDC info to the next level (these are dummy calls if pc_temp is not of type PCBDDC) */ 47903e3c6dadSStefano Zampini ierr = PCBDDCSetLevel(pc_temp,pcbddc->current_level+1);CHKERRQ(ierr); 47913e3c6dadSStefano Zampini ierr = PCBDDCSetCoarseningRatio(pc_temp,pcbddc->coarsening_ratio);CHKERRQ(ierr); 47923e3c6dadSStefano Zampini ierr = PCBDDCSetLevels(pc_temp,pcbddc->max_levels);CHKERRQ(ierr); 4793f9eb5b7dSStefano Zampini /* allow user customization */ 4794f9eb5b7dSStefano Zampini ierr = KSPSetFromOptions(pcbddc->coarse_ksp);CHKERRQ(ierr); 47953e3c6dadSStefano Zampini } 47963e3c6dadSStefano Zampini /* propagate BDDC info to the next level (these are dummy calls if pc_temp is not of type PCBDDC) */ 479751bea450SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&pc_temp);CHKERRQ(ierr); 47983e3c6dadSStefano Zampini if (nisdofs) { 47993e3c6dadSStefano Zampini ierr = PCBDDCSetDofsSplitting(pc_temp,nisdofs,isarray);CHKERRQ(ierr); 48003e3c6dadSStefano Zampini for (i=0;i<nisdofs;i++) { 48013e3c6dadSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 48023e3c6dadSStefano Zampini } 48033e3c6dadSStefano Zampini } 48043e3c6dadSStefano Zampini if (nisneu) { 48053e3c6dadSStefano Zampini ierr = PCBDDCSetNeumannBoundaries(pc_temp,isarray[nisdofs]);CHKERRQ(ierr); 48063e3c6dadSStefano Zampini ierr = ISDestroy(&isarray[nisdofs]);CHKERRQ(ierr); 4807312be037SStefano Zampini } 4808f9eb5b7dSStefano Zampini 4809f9eb5b7dSStefano Zampini /* get some info after set from options */ 4810f9eb5b7dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCNN,&isnn);CHKERRQ(ierr); 4811f9eb5b7dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCBDDC,&isbddc);CHKERRQ(ierr); 48124f3a063dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCREDUNDANT,&isredundant);CHKERRQ(ierr); 48136e683305SStefano Zampini if (isbddc && !multilevel_allowed) { /* multilevel can only be requested via pc_bddc_set_levels */ 4814f9eb5b7dSStefano Zampini ierr = PCSetType(pc_temp,coarse_pc_type);CHKERRQ(ierr); 4815f9eb5b7dSStefano Zampini isbddc = PETSC_FALSE; 4816f9eb5b7dSStefano Zampini } 481739f0f02cSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 48184f3a063dSStefano Zampini if (isredundant) { 48194f3a063dSStefano Zampini KSP inner_ksp; 48204f3a063dSStefano Zampini PC inner_pc; 48214f3a063dSStefano Zampini ierr = PCRedundantGetKSP(pc_temp,&inner_ksp);CHKERRQ(ierr); 48224f3a063dSStefano Zampini ierr = KSPGetPC(inner_ksp,&inner_pc);CHKERRQ(ierr); 48234f3a063dSStefano Zampini ierr = PCFactorSetReuseFill(inner_pc,PETSC_TRUE);CHKERRQ(ierr); 48244f3a063dSStefano Zampini } 4825f9eb5b7dSStefano Zampini 4826f9eb5b7dSStefano Zampini /* assemble coarse matrix */ 4827fa7f1dd8SStefano Zampini if (coarse_reuse) { 482881d9aea3SBarry Smith ierr = KSPGetOperators(pcbddc->coarse_ksp,&coarse_mat,NULL);CHKERRQ(ierr); 4829fa7f1dd8SStefano Zampini ierr = PetscObjectReference((PetscObject)coarse_mat);CHKERRQ(ierr); 48306e683305SStefano Zampini coarse_mat_reuse = MAT_REUSE_MATRIX; 4831fa7f1dd8SStefano Zampini } else { 48326e683305SStefano Zampini coarse_mat_reuse = MAT_INITIAL_MATRIX; 4833fa7f1dd8SStefano Zampini } 4834c8587f34SStefano Zampini if (isbddc || isnn) { 483522bc73bbSStefano Zampini if (pcbddc->coarsening_ratio > 1) { 483670cf5478SStefano Zampini if (!pcbddc->coarse_subassembling) { /* subassembling info is not present */ 4837b0c7d250SStefano Zampini ierr = MatISGetSubassemblingPattern(coarse_mat_is,active_procs/pcbddc->coarsening_ratio,pcbddc->coarse_adj_red,&pcbddc->coarse_subassembling);CHKERRQ(ierr); 483822b6e8a2SStefano Zampini if (pcbddc->dbg_flag) { 48396e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 48406e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Subassembling pattern\n");CHKERRQ(ierr); 48416e683305SStefano Zampini ierr = ISView(pcbddc->coarse_subassembling,dbg_viewer);CHKERRQ(ierr); 48426e683305SStefano Zampini ierr = PetscViewerFlush(dbg_viewer);CHKERRQ(ierr); 484322b6e8a2SStefano Zampini } 484470cf5478SStefano Zampini } 484553a05cb3SStefano Zampini ierr = MatISSubassemble(coarse_mat_is,pcbddc->coarse_subassembling,0,PETSC_FALSE,PETSC_FALSE,coarse_mat_reuse,&coarse_mat,0,NULL);CHKERRQ(ierr); 484670cf5478SStefano Zampini } else { 484722bc73bbSStefano Zampini ierr = PetscObjectReference((PetscObject)coarse_mat_is);CHKERRQ(ierr); 484822bc73bbSStefano Zampini coarse_mat = coarse_mat_is; 484922bc73bbSStefano Zampini } 485022bc73bbSStefano Zampini } else { 48512e1e5fa4SStefano Zampini ierr = MatISGetMPIXAIJ(coarse_mat_is,coarse_mat_reuse,&coarse_mat);CHKERRQ(ierr); 4852c8587f34SStefano Zampini } 4853c8587f34SStefano Zampini ierr = MatDestroy(&coarse_mat_is);CHKERRQ(ierr); 4854c8587f34SStefano Zampini 48553301b35fSStefano Zampini /* propagate symmetry info of coarse matrix */ 48565a16e3a0SStefano Zampini ierr = MatSetOption(coarse_mat,MAT_STRUCTURALLY_SYMMETRIC,PETSC_TRUE);CHKERRQ(ierr); 48573301b35fSStefano Zampini if (pc->pmat->symmetric_set) { 48583301b35fSStefano Zampini ierr = MatSetOption(coarse_mat,MAT_SYMMETRIC,pc->pmat->symmetric);CHKERRQ(ierr); 48593301b35fSStefano Zampini } 48603301b35fSStefano Zampini if (pc->pmat->hermitian_set) { 48613301b35fSStefano Zampini ierr = MatSetOption(coarse_mat,MAT_HERMITIAN,pc->pmat->hermitian);CHKERRQ(ierr); 48623301b35fSStefano Zampini } 48633301b35fSStefano Zampini if (pc->pmat->spd_set) { 48643301b35fSStefano Zampini ierr = MatSetOption(coarse_mat,MAT_SPD,pc->pmat->spd);CHKERRQ(ierr); 48653301b35fSStefano Zampini } 48666e683305SStefano Zampini /* set operators */ 48675f76c7aeSStefano Zampini ierr = KSPSetOperators(pcbddc->coarse_ksp,coarse_mat,coarse_mat);CHKERRQ(ierr); 48686e683305SStefano Zampini if (pcbddc->dbg_flag) { 48696e683305SStefano Zampini ierr = PetscViewerASCIISubtractTab(dbg_viewer,2*pcbddc->current_level);CHKERRQ(ierr); 48706e683305SStefano Zampini } 48716e683305SStefano Zampini } else { /* processes non partecipating to coarse solver (if any) */ 48726e683305SStefano Zampini coarse_mat = 0; 48736e683305SStefano Zampini } 48746e683305SStefano Zampini ierr = PetscFree(isarray);CHKERRQ(ierr); 4875b1ecc7b1SStefano Zampini #if 0 4876b9b85e73SStefano Zampini { 4877b9b85e73SStefano Zampini PetscViewer viewer; 4878b9b85e73SStefano Zampini char filename[256]; 4879b1ecc7b1SStefano Zampini sprintf(filename,"coarse_mat_level%d.m",pcbddc->current_level); 4880b1ecc7b1SStefano Zampini ierr = PetscViewerASCIIOpen(PetscObjectComm((PetscObject)coarse_mat),filename,&viewer);CHKERRQ(ierr); 4881b9b85e73SStefano Zampini ierr = PetscViewerSetFormat(viewer,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 4882b9b85e73SStefano Zampini ierr = MatView(coarse_mat,viewer);CHKERRQ(ierr); 4883b9b85e73SStefano Zampini ierr = PetscViewerDestroy(&viewer);CHKERRQ(ierr); 4884b9b85e73SStefano Zampini } 4885b9b85e73SStefano Zampini #endif 4886c8587f34SStefano Zampini 4887c8587f34SStefano Zampini /* Compute coarse null space (special handling by BDDC only) */ 4888298c0119SStefano Zampini #if 0 4889c8587f34SStefano Zampini if (pcbddc->NullSpace) { 4890c8587f34SStefano Zampini ierr = PCBDDCNullSpaceAssembleCoarse(pc,coarse_mat,&CoarseNullSpace);CHKERRQ(ierr); 489198a51de6SStefano Zampini } 4892298c0119SStefano Zampini #endif 489398a51de6SStefano Zampini 489498a51de6SStefano Zampini if (pcbddc->coarse_ksp) { 489598a51de6SStefano Zampini Vec crhs,csol; 489698a51de6SStefano Zampini PetscBool ispreonly; 489704708bb6SStefano Zampini 489898a51de6SStefano Zampini if (CoarseNullSpace) { 4899c8587f34SStefano Zampini if (isbddc) { 4900c8587f34SStefano Zampini ierr = PCBDDCSetNullSpace(pc_temp,CoarseNullSpace);CHKERRQ(ierr); 4901c8587f34SStefano Zampini } else { 49025fa7ec2dSBarry Smith ierr = MatSetNullSpace(coarse_mat,CoarseNullSpace);CHKERRQ(ierr); 4903c8587f34SStefano Zampini } 4904c8587f34SStefano Zampini } 4905f9eb5b7dSStefano Zampini /* setup coarse ksp */ 4906f9eb5b7dSStefano Zampini ierr = KSPSetUp(pcbddc->coarse_ksp);CHKERRQ(ierr); 4907f347579bSStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&csol);CHKERRQ(ierr); 4908f347579bSStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&crhs);CHKERRQ(ierr); 49096e683305SStefano Zampini /* hack */ 4910f347579bSStefano Zampini if (!csol) { 49112a7a6963SBarry Smith ierr = MatCreateVecs(coarse_mat,&((pcbddc->coarse_ksp)->vec_sol),NULL);CHKERRQ(ierr); 4912f9eb5b7dSStefano Zampini } 4913f347579bSStefano Zampini if (!crhs) { 49142a7a6963SBarry Smith ierr = MatCreateVecs(coarse_mat,NULL,&((pcbddc->coarse_ksp)->vec_rhs));CHKERRQ(ierr); 4915f347579bSStefano Zampini } 4916cbcc2c2aSStefano Zampini /* Check coarse problem if in debug mode or if solving with an iterative method */ 4917cbcc2c2aSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcbddc->coarse_ksp,KSPPREONLY,&ispreonly);CHKERRQ(ierr); 49186e683305SStefano Zampini if (pcbddc->dbg_flag || (!ispreonly && pcbddc->use_coarse_estimates) ) { 4919c8587f34SStefano Zampini KSP check_ksp; 49202b510759SStefano Zampini KSPType check_ksp_type; 4921c8587f34SStefano Zampini PC check_pc; 49226e683305SStefano Zampini Vec check_vec,coarse_vec; 49236a1308c2SStefano Zampini PetscReal abs_infty_error,infty_error,lambda_min=1.0,lambda_max=1.0; 49242b510759SStefano Zampini PetscInt its; 49256e683305SStefano Zampini PetscBool compute_eigs; 49266e683305SStefano Zampini PetscReal *eigs_r,*eigs_c; 49276e683305SStefano Zampini PetscInt neigs; 49288e185a42SStefano Zampini const char *prefix; 4929c8587f34SStefano Zampini 49302b510759SStefano Zampini /* Create ksp object suitable for estimation of extreme eigenvalues */ 49316e683305SStefano Zampini ierr = KSPCreate(PetscObjectComm((PetscObject)pcbddc->coarse_ksp),&check_ksp);CHKERRQ(ierr); 4932422a814eSBarry Smith ierr = KSPSetErrorIfNotConverged(pcbddc->coarse_ksp,pc->erroriffailure);CHKERRQ(ierr); 493323ee1639SBarry Smith ierr = KSPSetOperators(check_ksp,coarse_mat,coarse_mat);CHKERRQ(ierr); 4934f4ddd8eeSStefano Zampini ierr = KSPSetTolerances(check_ksp,1.e-12,1.e-12,PETSC_DEFAULT,pcbddc->coarse_size);CHKERRQ(ierr); 49352b510759SStefano Zampini if (ispreonly) { 49362b510759SStefano Zampini check_ksp_type = KSPPREONLY; 49376e683305SStefano Zampini compute_eigs = PETSC_FALSE; 49382b510759SStefano Zampini } else { 4939cbcc2c2aSStefano Zampini check_ksp_type = KSPGMRES; 49406e683305SStefano Zampini compute_eigs = PETSC_TRUE; 4941c8587f34SStefano Zampini } 4942c8587f34SStefano Zampini ierr = KSPSetType(check_ksp,check_ksp_type);CHKERRQ(ierr); 49436e683305SStefano Zampini ierr = KSPSetComputeSingularValues(check_ksp,compute_eigs);CHKERRQ(ierr); 49446e683305SStefano Zampini ierr = KSPSetComputeEigenvalues(check_ksp,compute_eigs);CHKERRQ(ierr); 49456e683305SStefano Zampini ierr = KSPGMRESSetRestart(check_ksp,pcbddc->coarse_size+1);CHKERRQ(ierr); 4946a7dc3881SStefano Zampini ierr = KSPGetOptionsPrefix(pcbddc->coarse_ksp,&prefix);CHKERRQ(ierr); 4947a7dc3881SStefano Zampini ierr = KSPSetOptionsPrefix(check_ksp,prefix);CHKERRQ(ierr); 4948a7dc3881SStefano Zampini ierr = KSPAppendOptionsPrefix(check_ksp,"check_");CHKERRQ(ierr); 4949a7dc3881SStefano Zampini ierr = KSPSetFromOptions(check_ksp);CHKERRQ(ierr); 4950c8587f34SStefano Zampini ierr = KSPSetUp(check_ksp);CHKERRQ(ierr); 4951c8587f34SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&check_pc);CHKERRQ(ierr); 4952c8587f34SStefano Zampini ierr = KSPSetPC(check_ksp,check_pc);CHKERRQ(ierr); 4953c8587f34SStefano Zampini /* create random vec */ 49546e683305SStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&coarse_vec);CHKERRQ(ierr); 49556e683305SStefano Zampini ierr = VecDuplicate(coarse_vec,&check_vec);CHKERRQ(ierr); 4956c8587f34SStefano Zampini ierr = VecSetRandom(check_vec,NULL);CHKERRQ(ierr); 4957c8587f34SStefano Zampini if (CoarseNullSpace) { 4958c8587f34SStefano Zampini ierr = MatNullSpaceRemove(CoarseNullSpace,check_vec);CHKERRQ(ierr); 4959c8587f34SStefano Zampini } 49606e683305SStefano Zampini ierr = MatMult(coarse_mat,check_vec,coarse_vec);CHKERRQ(ierr); 4961c8587f34SStefano Zampini /* solve coarse problem */ 49626e683305SStefano Zampini ierr = KSPSolve(check_ksp,coarse_vec,coarse_vec);CHKERRQ(ierr); 4963c8587f34SStefano Zampini if (CoarseNullSpace) { 49646e683305SStefano Zampini ierr = MatNullSpaceRemove(CoarseNullSpace,coarse_vec);CHKERRQ(ierr); 4965c8587f34SStefano Zampini } 4966cbcc2c2aSStefano Zampini /* set eigenvalue estimation if preonly has not been requested */ 49676e683305SStefano Zampini if (compute_eigs) { 4968854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->coarse_size+1,&eigs_r);CHKERRQ(ierr); 4969854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->coarse_size+1,&eigs_c);CHKERRQ(ierr); 49706e683305SStefano Zampini ierr = KSPComputeEigenvalues(check_ksp,pcbddc->coarse_size+1,eigs_r,eigs_c,&neigs);CHKERRQ(ierr); 49716e683305SStefano Zampini lambda_max = eigs_r[neigs-1]; 49726e683305SStefano Zampini lambda_min = eigs_r[0]; 49736e683305SStefano Zampini if (pcbddc->use_coarse_estimates) { 49746e683305SStefano Zampini if (lambda_max>lambda_min) { 4975cbcc2c2aSStefano Zampini ierr = KSPChebyshevSetEigenvalues(pcbddc->coarse_ksp,lambda_max,lambda_min);CHKERRQ(ierr); 4976cbcc2c2aSStefano Zampini ierr = KSPRichardsonSetScale(pcbddc->coarse_ksp,2.0/(lambda_max+lambda_min));CHKERRQ(ierr); 4977cbcc2c2aSStefano Zampini } 4978c8587f34SStefano Zampini } 4979c8587f34SStefano Zampini } 4980cbcc2c2aSStefano Zampini 4981c8587f34SStefano Zampini /* check coarse problem residual error */ 49826e683305SStefano Zampini if (pcbddc->dbg_flag) { 49836e683305SStefano Zampini PetscViewer dbg_viewer = PETSC_VIEWER_STDOUT_(PetscObjectComm((PetscObject)pcbddc->coarse_ksp)); 49846e683305SStefano Zampini ierr = PetscViewerASCIIAddTab(dbg_viewer,2*(pcbddc->current_level+1));CHKERRQ(ierr); 49856e683305SStefano Zampini ierr = VecAXPY(check_vec,-1.0,coarse_vec);CHKERRQ(ierr); 4986c8587f34SStefano Zampini ierr = VecNorm(check_vec,NORM_INFINITY,&infty_error);CHKERRQ(ierr); 49876e683305SStefano Zampini ierr = MatMult(coarse_mat,check_vec,coarse_vec);CHKERRQ(ierr); 49886e683305SStefano Zampini ierr = VecNorm(coarse_vec,NORM_INFINITY,&abs_infty_error);CHKERRQ(ierr); 4989c8587f34SStefano Zampini ierr = VecDestroy(&check_vec);CHKERRQ(ierr); 4990779c1cceSStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem details (use estimates %d)\n",pcbddc->use_coarse_estimates);CHKERRQ(ierr); 49916e683305SStefano Zampini ierr = PetscObjectPrintClassNamePrefixType((PetscObject)(pcbddc->coarse_ksp),dbg_viewer);CHKERRQ(ierr); 49926e683305SStefano Zampini ierr = PetscObjectPrintClassNamePrefixType((PetscObject)(check_pc),dbg_viewer);CHKERRQ(ierr); 49936e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem exact infty_error : %1.6e\n",infty_error);CHKERRQ(ierr); 49946e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem residual infty_error: %1.6e\n",abs_infty_error);CHKERRQ(ierr); 49956e683305SStefano Zampini if (compute_eigs) { 49966e683305SStefano Zampini PetscReal lambda_max_s,lambda_min_s; 4997deec49d1SStefano Zampini ierr = KSPGetType(check_ksp,&check_ksp_type);CHKERRQ(ierr); 4998c8587f34SStefano Zampini ierr = KSPGetIterationNumber(check_ksp,&its);CHKERRQ(ierr); 49996e683305SStefano Zampini ierr = KSPComputeExtremeSingularValues(check_ksp,&lambda_max_s,&lambda_min_s);CHKERRQ(ierr); 50006e683305SStefano 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); 50016e683305SStefano Zampini for (i=0;i<neigs;i++) { 50026e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"%1.6e %1.6ei\n",eigs_r[i],eigs_c[i]);CHKERRQ(ierr); 5003c8587f34SStefano Zampini } 50046e683305SStefano Zampini } 50056e683305SStefano Zampini ierr = PetscViewerFlush(dbg_viewer);CHKERRQ(ierr); 50066e683305SStefano Zampini ierr = PetscViewerASCIISubtractTab(dbg_viewer,2*(pcbddc->current_level+1));CHKERRQ(ierr); 50076e683305SStefano Zampini } 5008c8587f34SStefano Zampini ierr = KSPDestroy(&check_ksp);CHKERRQ(ierr); 50096e683305SStefano Zampini if (compute_eigs) { 50106e683305SStefano Zampini ierr = PetscFree(eigs_r);CHKERRQ(ierr); 50116e683305SStefano Zampini ierr = PetscFree(eigs_c);CHKERRQ(ierr); 5012c8587f34SStefano Zampini } 50136e683305SStefano Zampini } 50146e683305SStefano Zampini } 5015cbcc2c2aSStefano Zampini /* print additional info */ 5016cbcc2c2aSStefano Zampini if (pcbddc->dbg_flag) { 50176e683305SStefano Zampini /* waits until all processes reaches this point */ 50186e683305SStefano Zampini ierr = PetscBarrier((PetscObject)pc);CHKERRQ(ierr); 5019cbcc2c2aSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Coarse solver setup completed at level %d\n",pcbddc->current_level);CHKERRQ(ierr); 5020cbcc2c2aSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 5021cbcc2c2aSStefano Zampini } 5022cbcc2c2aSStefano Zampini 50232b510759SStefano Zampini /* free memory */ 5024c8587f34SStefano Zampini ierr = MatNullSpaceDestroy(&CoarseNullSpace);CHKERRQ(ierr); 5025fdc635d7SStefano Zampini ierr = MatDestroy(&coarse_mat);CHKERRQ(ierr); 5026c8587f34SStefano Zampini PetscFunctionReturn(0); 5027c8587f34SStefano Zampini } 5028674ae819SStefano Zampini 5029f34684f1SStefano Zampini #undef __FUNCT__ 5030f34684f1SStefano Zampini #define __FUNCT__ "PCBDDCComputePrimalNumbering" 5031f34684f1SStefano Zampini PetscErrorCode PCBDDCComputePrimalNumbering(PC pc,PetscInt* coarse_size_n,PetscInt** local_primal_indices_n) 5032f34684f1SStefano Zampini { 5033f34684f1SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 5034f34684f1SStefano Zampini PC_IS* pcis = (PC_IS*)pc->data; 5035f34684f1SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 5036dc456d91SStefano Zampini IS subset,subset_mult,subset_n; 5037dc456d91SStefano Zampini PetscInt local_size,coarse_size=0; 503873be2a3aSStefano Zampini PetscInt *local_primal_indices=NULL; 5039dc456d91SStefano Zampini const PetscInt *t_local_primal_indices; 5040f34684f1SStefano Zampini PetscErrorCode ierr; 5041f34684f1SStefano Zampini 5042f34684f1SStefano Zampini PetscFunctionBegin; 5043f34684f1SStefano Zampini /* Compute global number of coarse dofs */ 50440e6343abSStefano Zampini if (pcbddc->local_primal_size && !pcbddc->local_primal_ref_node) { 50450e6343abSStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"BDDC ConstraintsSetUp should be called first"); 5046727cdba6SStefano Zampini } 5047dc456d91SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)(pc->pmat)),pcbddc->local_primal_size_cc,pcbddc->local_primal_ref_node,PETSC_COPY_VALUES,&subset_n);CHKERRQ(ierr); 50483bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(pcis->mapping,subset_n,&subset);CHKERRQ(ierr); 5049dc456d91SStefano Zampini ierr = ISDestroy(&subset_n);CHKERRQ(ierr); 5050dc456d91SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)(pc->pmat)),pcbddc->local_primal_size_cc,pcbddc->local_primal_ref_mult,PETSC_COPY_VALUES,&subset_mult);CHKERRQ(ierr); 5051dc456d91SStefano Zampini ierr = PCBDDCSubsetNumbering(subset,subset_mult,&coarse_size,&subset_n);CHKERRQ(ierr); 5052dc456d91SStefano Zampini ierr = ISDestroy(&subset);CHKERRQ(ierr); 5053dc456d91SStefano Zampini ierr = ISDestroy(&subset_mult);CHKERRQ(ierr); 5054dc456d91SStefano Zampini ierr = ISGetLocalSize(subset_n,&local_size);CHKERRQ(ierr); 50550e6343abSStefano Zampini if (local_size != pcbddc->local_primal_size) { 50560e6343abSStefano Zampini SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Invalid number of local primal indices computed %d != %d",local_size,pcbddc->local_primal_size); 50570e6343abSStefano Zampini } 5058dc456d91SStefano Zampini ierr = PetscMalloc1(local_size,&local_primal_indices);CHKERRQ(ierr); 5059dc456d91SStefano Zampini ierr = ISGetIndices(subset_n,&t_local_primal_indices);CHKERRQ(ierr); 5060dc456d91SStefano Zampini ierr = PetscMemcpy(local_primal_indices,t_local_primal_indices,local_size*sizeof(PetscInt));CHKERRQ(ierr); 5061dc456d91SStefano Zampini ierr = ISRestoreIndices(subset_n,&t_local_primal_indices);CHKERRQ(ierr); 5062dc456d91SStefano Zampini ierr = ISDestroy(&subset_n);CHKERRQ(ierr); 5063f34684f1SStefano Zampini 5064f34684f1SStefano Zampini /* check numbering */ 5065f34684f1SStefano Zampini if (pcbddc->dbg_flag) { 5066019a44ceSStefano Zampini PetscScalar coarsesum,*array,*array2; 5067dc456d91SStefano Zampini PetscInt i; 5068b9b85e73SStefano Zampini PetscBool set_error = PETSC_FALSE,set_error_reduced = PETSC_FALSE; 5069f34684f1SStefano Zampini 5070f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 5071f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 5072f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Check coarse indices\n");CHKERRQ(ierr); 50730fccc4e9SStefano Zampini ierr = PetscViewerASCIISynchronizedAllow(pcbddc->dbg_viewer,PETSC_TRUE);CHKERRQ(ierr); 5074019a44ceSStefano Zampini /* counter */ 5075019a44ceSStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 5076019a44ceSStefano Zampini ierr = VecSet(pcis->vec1_N,1.0);CHKERRQ(ierr); 5077019a44ceSStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 5078019a44ceSStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 5079019a44ceSStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_global,pcis->vec2_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 5080019a44ceSStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_global,pcis->vec2_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 5081019a44ceSStefano Zampini 5082f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_N,0.0);CHKERRQ(ierr); 5083f34684f1SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 5084727cdba6SStefano Zampini ierr = VecSetValue(pcis->vec1_N,pcbddc->primal_indices_local_idxs[i],1.0,INSERT_VALUES);CHKERRQ(ierr); 5085f34684f1SStefano Zampini } 5086f34684f1SStefano Zampini ierr = VecAssemblyBegin(pcis->vec1_N);CHKERRQ(ierr); 5087f34684f1SStefano Zampini ierr = VecAssemblyEnd(pcis->vec1_N);CHKERRQ(ierr); 5088f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 5089e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 5090e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 5091e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_global,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 5092e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_global,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 5093f34684f1SStefano Zampini ierr = VecGetArray(pcis->vec1_N,&array);CHKERRQ(ierr); 5094019a44ceSStefano Zampini ierr = VecGetArray(pcis->vec2_N,&array2);CHKERRQ(ierr); 5095f34684f1SStefano Zampini for (i=0;i<pcis->n;i++) { 5096019a44ceSStefano Zampini if (array[i] != 0.0 && array[i] != array2[i]) { 5097019a44ceSStefano Zampini PetscInt owned = (PetscInt)(array[i]); 5098019a44ceSStefano Zampini PetscInt neigh = (PetscInt)(array2[i]); 5099b9b85e73SStefano Zampini set_error = PETSC_TRUE; 5100019a44ceSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d: local index %d owned by a %d processes instead of %d!\n",PetscGlobalRank,i,owned,neigh);CHKERRQ(ierr); 5101f34684f1SStefano Zampini } 5102f34684f1SStefano Zampini } 5103019a44ceSStefano Zampini ierr = VecRestoreArray(pcis->vec2_N,&array2);CHKERRQ(ierr); 5104b9b85e73SStefano Zampini ierr = MPI_Allreduce(&set_error,&set_error_reduced,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 5105f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 5106f34684f1SStefano Zampini for (i=0;i<pcis->n;i++) { 5107f34684f1SStefano Zampini if (PetscRealPart(array[i]) > 0.0) array[i] = 1.0/PetscRealPart(array[i]); 5108f34684f1SStefano Zampini } 5109f34684f1SStefano Zampini ierr = VecRestoreArray(pcis->vec1_N,&array);CHKERRQ(ierr); 5110f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 5111e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 5112e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 5113f34684f1SStefano Zampini ierr = VecSum(pcis->vec1_global,&coarsesum);CHKERRQ(ierr); 5114f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Size of coarse problem is %d (%lf)\n",coarse_size,PetscRealPart(coarsesum));CHKERRQ(ierr); 5115b9b85e73SStefano Zampini if (pcbddc->dbg_flag > 1 || set_error_reduced) { 5116ca8b9ea9SStefano Zampini PetscInt *gidxs; 5117ca8b9ea9SStefano Zampini 5118ca8b9ea9SStefano Zampini ierr = PetscMalloc1(pcbddc->local_primal_size,&gidxs);CHKERRQ(ierr); 51193bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingApply(pcis->mapping,pcbddc->local_primal_size,pcbddc->primal_indices_local_idxs,gidxs);CHKERRQ(ierr); 5120f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Distribution of local primal indices\n");CHKERRQ(ierr); 5121f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 5122f34684f1SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d\n",PetscGlobalRank);CHKERRQ(ierr); 5123f34684f1SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 51244bc2dc4bSStefano 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); 5125f34684f1SStefano Zampini } 5126f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 5127ca8b9ea9SStefano Zampini ierr = PetscFree(gidxs);CHKERRQ(ierr); 5128f34684f1SStefano Zampini } 5129f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 5130302440fdSBarry Smith if (set_error_reduced) SETERRQ(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"BDDC Numbering of coarse dofs failed"); 5131f34684f1SStefano Zampini } 51328bec7fa6SStefano Zampini /* ierr = PetscPrintf(PetscObjectComm((PetscObject)pc),"Size of coarse problem is %d\n",coarse_size);CHKERRQ(ierr); */ 5133f34684f1SStefano Zampini /* get back data */ 5134f34684f1SStefano Zampini *coarse_size_n = coarse_size; 5135f34684f1SStefano Zampini *local_primal_indices_n = local_primal_indices; 5136674ae819SStefano Zampini PetscFunctionReturn(0); 5137674ae819SStefano Zampini } 5138674ae819SStefano Zampini 5139e456f2a8SStefano Zampini #undef __FUNCT__ 5140e456f2a8SStefano Zampini #define __FUNCT__ "PCBDDCGlobalToLocal" 5141a7dc3881SStefano Zampini PetscErrorCode PCBDDCGlobalToLocal(VecScatter g2l_ctx,Vec gwork, Vec lwork, IS globalis, IS* localis) 5142e456f2a8SStefano Zampini { 5143e456f2a8SStefano Zampini IS localis_t; 5144a7dc3881SStefano Zampini PetscInt i,lsize,*idxs,n; 5145e456f2a8SStefano Zampini PetscScalar *vals; 5146e456f2a8SStefano Zampini PetscErrorCode ierr; 5147e456f2a8SStefano Zampini 5148e456f2a8SStefano Zampini PetscFunctionBegin; 5149a7dc3881SStefano Zampini /* get indices in local ordering exploiting local to global map */ 5150e456f2a8SStefano Zampini ierr = ISGetLocalSize(globalis,&lsize);CHKERRQ(ierr); 5151854ce69bSBarry Smith ierr = PetscMalloc1(lsize,&vals);CHKERRQ(ierr); 5152e456f2a8SStefano Zampini for (i=0;i<lsize;i++) vals[i] = 1.0; 5153e456f2a8SStefano Zampini ierr = ISGetIndices(globalis,(const PetscInt**)&idxs);CHKERRQ(ierr); 5154a7dc3881SStefano Zampini ierr = VecSet(gwork,0.0);CHKERRQ(ierr); 5155a7dc3881SStefano Zampini ierr = VecSet(lwork,0.0);CHKERRQ(ierr); 51561035eff8SStefano Zampini if (idxs) { /* multilevel guard */ 5157a7dc3881SStefano Zampini ierr = VecSetValues(gwork,lsize,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 51581035eff8SStefano Zampini } 5159a7dc3881SStefano Zampini ierr = VecAssemblyBegin(gwork);CHKERRQ(ierr); 5160e456f2a8SStefano Zampini ierr = ISRestoreIndices(globalis,(const PetscInt**)&idxs);CHKERRQ(ierr); 5161e456f2a8SStefano Zampini ierr = PetscFree(vals);CHKERRQ(ierr); 5162a7dc3881SStefano Zampini ierr = VecAssemblyEnd(gwork);CHKERRQ(ierr); 5163a7dc3881SStefano Zampini /* now compute set in local ordering */ 5164a7dc3881SStefano Zampini ierr = VecScatterBegin(g2l_ctx,gwork,lwork,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 5165a7dc3881SStefano Zampini ierr = VecScatterEnd(g2l_ctx,gwork,lwork,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 5166a7dc3881SStefano Zampini ierr = VecGetArrayRead(lwork,(const PetscScalar**)&vals);CHKERRQ(ierr); 5167a7dc3881SStefano Zampini ierr = VecGetSize(lwork,&n);CHKERRQ(ierr); 5168a7dc3881SStefano Zampini for (i=0,lsize=0;i<n;i++) { 5169ff92baa0SMatthew G. Knepley if (PetscRealPart(vals[i]) > 0.5) { 5170e456f2a8SStefano Zampini lsize++; 5171e456f2a8SStefano Zampini } 5172e456f2a8SStefano Zampini } 5173854ce69bSBarry Smith ierr = PetscMalloc1(lsize,&idxs);CHKERRQ(ierr); 5174a7dc3881SStefano Zampini for (i=0,lsize=0;i<n;i++) { 5175ff92baa0SMatthew G. Knepley if (PetscRealPart(vals[i]) > 0.5) { 5176e456f2a8SStefano Zampini idxs[lsize++] = i; 5177e456f2a8SStefano Zampini } 5178e456f2a8SStefano Zampini } 5179a7dc3881SStefano Zampini ierr = VecRestoreArrayRead(lwork,(const PetscScalar**)&vals);CHKERRQ(ierr); 5180a7dc3881SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)gwork),lsize,idxs,PETSC_OWN_POINTER,&localis_t);CHKERRQ(ierr); 5181e456f2a8SStefano Zampini *localis = localis_t; 5182e456f2a8SStefano Zampini PetscFunctionReturn(0); 5183e456f2a8SStefano Zampini } 5184906d46d4SStefano Zampini 5185906d46d4SStefano Zampini /* the next two functions will be called in KSPMatMult if a change of basis has been requested */ 5186906d46d4SStefano Zampini #undef __FUNCT__ 5187906d46d4SStefano Zampini #define __FUNCT__ "PCBDDCMatMult_Private" 5188906d46d4SStefano Zampini static PetscErrorCode PCBDDCMatMult_Private(Mat A, Vec x, Vec y) 5189906d46d4SStefano Zampini { 5190906d46d4SStefano Zampini PCBDDCChange_ctx change_ctx; 5191906d46d4SStefano Zampini PetscErrorCode ierr; 5192906d46d4SStefano Zampini 5193906d46d4SStefano Zampini PetscFunctionBegin; 5194906d46d4SStefano Zampini ierr = MatShellGetContext(A,&change_ctx);CHKERRQ(ierr); 5195906d46d4SStefano Zampini ierr = MatMult(change_ctx->global_change,x,change_ctx->work[0]);CHKERRQ(ierr); 5196906d46d4SStefano Zampini ierr = MatMult(change_ctx->original_mat,change_ctx->work[0],change_ctx->work[1]);CHKERRQ(ierr); 5197906d46d4SStefano Zampini ierr = MatMultTranspose(change_ctx->global_change,change_ctx->work[1],y);CHKERRQ(ierr); 5198906d46d4SStefano Zampini PetscFunctionReturn(0); 5199906d46d4SStefano Zampini } 5200906d46d4SStefano Zampini 5201906d46d4SStefano Zampini #undef __FUNCT__ 5202906d46d4SStefano Zampini #define __FUNCT__ "PCBDDCMatMultTranspose_Private" 5203906d46d4SStefano Zampini static PetscErrorCode PCBDDCMatMultTranspose_Private(Mat A, Vec x, Vec y) 5204906d46d4SStefano Zampini { 5205906d46d4SStefano Zampini PCBDDCChange_ctx change_ctx; 5206906d46d4SStefano Zampini PetscErrorCode ierr; 5207906d46d4SStefano Zampini 5208906d46d4SStefano Zampini PetscFunctionBegin; 5209906d46d4SStefano Zampini ierr = MatShellGetContext(A,&change_ctx);CHKERRQ(ierr); 5210906d46d4SStefano Zampini ierr = MatMult(change_ctx->global_change,x,change_ctx->work[0]);CHKERRQ(ierr); 5211906d46d4SStefano Zampini ierr = MatMultTranspose(change_ctx->original_mat,change_ctx->work[0],change_ctx->work[1]);CHKERRQ(ierr); 5212906d46d4SStefano Zampini ierr = MatMultTranspose(change_ctx->global_change,change_ctx->work[1],y);CHKERRQ(ierr); 5213906d46d4SStefano Zampini PetscFunctionReturn(0); 5214906d46d4SStefano Zampini } 5215b96c3477SStefano Zampini 5216b96c3477SStefano Zampini #undef __FUNCT__ 5217b96c3477SStefano Zampini #define __FUNCT__ "PCBDDCSetUpSubSchurs" 521808122e43SStefano Zampini PetscErrorCode PCBDDCSetUpSubSchurs(PC pc) 5219b96c3477SStefano Zampini { 5220a64f4aa4SStefano Zampini PC_IS *pcis=(PC_IS*)pc->data; 5221b96c3477SStefano Zampini PC_BDDC *pcbddc=(PC_BDDC*)pc->data; 5222b96c3477SStefano Zampini PCBDDCSubSchurs sub_schurs=pcbddc->sub_schurs; 5223a64f4aa4SStefano Zampini Mat S_j; 5224b96c3477SStefano Zampini PetscInt *used_xadj,*used_adjncy; 5225b96c3477SStefano Zampini PetscBool free_used_adj; 5226b96c3477SStefano Zampini PetscErrorCode ierr; 5227b96c3477SStefano Zampini 5228b96c3477SStefano Zampini PetscFunctionBegin; 5229b96c3477SStefano Zampini /* decide the adjacency to be used for determining internal problems for local schur on subsets */ 5230b96c3477SStefano Zampini free_used_adj = PETSC_FALSE; 523108122e43SStefano Zampini if (pcbddc->sub_schurs_layers == -1) { 5232b96c3477SStefano Zampini used_xadj = NULL; 5233b96c3477SStefano Zampini used_adjncy = NULL; 5234b96c3477SStefano Zampini } else { 523508122e43SStefano Zampini if (pcbddc->sub_schurs_use_useradj && pcbddc->mat_graph->xadj) { 523608122e43SStefano Zampini used_xadj = pcbddc->mat_graph->xadj; 523708122e43SStefano Zampini used_adjncy = pcbddc->mat_graph->adjncy; 523808122e43SStefano Zampini } else if (pcbddc->computed_rowadj) { 5239b96c3477SStefano Zampini used_xadj = pcbddc->mat_graph->xadj; 5240b96c3477SStefano Zampini used_adjncy = pcbddc->mat_graph->adjncy; 5241b96c3477SStefano Zampini } else { 52422fffb893SStefano Zampini PetscBool flg_row=PETSC_FALSE; 5243b96c3477SStefano Zampini const PetscInt *xadj,*adjncy; 5244b96c3477SStefano Zampini PetscInt nvtxs; 5245b96c3477SStefano Zampini 52462fffb893SStefano Zampini ierr = MatGetRowIJ(pcbddc->local_mat,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 52472fffb893SStefano Zampini if (flg_row) { 5248b96c3477SStefano Zampini ierr = PetscMalloc2(nvtxs+1,&used_xadj,xadj[nvtxs],&used_adjncy);CHKERRQ(ierr); 5249b96c3477SStefano Zampini ierr = PetscMemcpy(used_xadj,xadj,(nvtxs+1)*sizeof(*xadj));CHKERRQ(ierr); 5250b96c3477SStefano Zampini ierr = PetscMemcpy(used_adjncy,adjncy,(xadj[nvtxs])*sizeof(*adjncy));CHKERRQ(ierr); 5251b96c3477SStefano Zampini free_used_adj = PETSC_TRUE; 52522fffb893SStefano Zampini } else { 52532fffb893SStefano Zampini pcbddc->sub_schurs_layers = -1; 52542fffb893SStefano Zampini used_xadj = NULL; 52552fffb893SStefano Zampini used_adjncy = NULL; 52562fffb893SStefano Zampini } 52572fffb893SStefano Zampini ierr = MatRestoreRowIJ(pcbddc->local_mat,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 5258b96c3477SStefano Zampini } 5259b96c3477SStefano Zampini } 5260d5574798SStefano Zampini 5261d5574798SStefano Zampini /* setup sub_schurs data */ 5262a64f4aa4SStefano Zampini ierr = MatCreateSchurComplement(pcis->A_II,pcis->A_II,pcis->A_IB,pcis->A_BI,pcis->A_BB,&S_j);CHKERRQ(ierr); 5263a64f4aa4SStefano Zampini if (!sub_schurs->use_mumps) { 5264a64f4aa4SStefano Zampini /* pcbddc->ksp_D up to date only if not using MUMPS */ 5265a64f4aa4SStefano Zampini ierr = MatSchurComplementSetKSP(S_j,pcbddc->ksp_D);CHKERRQ(ierr); 526606a4e24aSStefano Zampini ierr = PCBDDCSubSchursSetUp(sub_schurs,NULL,S_j,used_xadj,used_adjncy,pcbddc->sub_schurs_layers,pcbddc->faster_deluxe,pcbddc->adaptive_selection,PETSC_FALSE,PETSC_FALSE);CHKERRQ(ierr); 5267a64f4aa4SStefano Zampini } else { 52686816873aSStefano Zampini PetscBool reuse_solvers = (PetscBool)!pcbddc->use_change_of_basis; 526904708bb6SStefano Zampini PetscBool isseqaij; 52705feab87aSStefano Zampini if (!pcbddc->use_vertices && reuse_solvers) { 52715feab87aSStefano Zampini PetscInt n_vertices; 52725feab87aSStefano Zampini 52735feab87aSStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_vertices,&n_vertices);CHKERRQ(ierr); 52742034aafcSStefano Zampini reuse_solvers = (PetscBool)!n_vertices; 52755feab87aSStefano Zampini } 527604708bb6SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcbddc->local_mat,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 527704708bb6SStefano Zampini if (!isseqaij) { 527804708bb6SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 527904708bb6SStefano Zampini if (matis->A == pcbddc->local_mat) { 528004708bb6SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 528104708bb6SStefano Zampini ierr = MatConvert(matis->A,MATSEQAIJ,MAT_INITIAL_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 528204708bb6SStefano Zampini } else { 528304708bb6SStefano Zampini ierr = MatConvert(pcbddc->local_mat,MATSEQAIJ,MAT_REUSE_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 528404708bb6SStefano Zampini } 528504708bb6SStefano Zampini } 528606a4e24aSStefano Zampini ierr = PCBDDCSubSchursSetUp(sub_schurs,pcbddc->local_mat,S_j,used_xadj,used_adjncy,pcbddc->sub_schurs_layers,pcbddc->faster_deluxe,pcbddc->adaptive_selection,reuse_solvers,pcbddc->benign_saddle_point);CHKERRQ(ierr); 5287a64f4aa4SStefano Zampini } 5288a64f4aa4SStefano Zampini ierr = MatDestroy(&S_j);CHKERRQ(ierr); 5289b96c3477SStefano Zampini 5290b96c3477SStefano Zampini /* free adjacency */ 5291b96c3477SStefano Zampini if (free_used_adj) { 5292b96c3477SStefano Zampini ierr = PetscFree2(used_xadj,used_adjncy);CHKERRQ(ierr); 5293b96c3477SStefano Zampini } 5294b96c3477SStefano Zampini PetscFunctionReturn(0); 5295b96c3477SStefano Zampini } 5296b96c3477SStefano Zampini 5297b96c3477SStefano Zampini #undef __FUNCT__ 5298b96c3477SStefano Zampini #define __FUNCT__ "PCBDDCInitSubSchurs" 529908122e43SStefano Zampini PetscErrorCode PCBDDCInitSubSchurs(PC pc) 5300b96c3477SStefano Zampini { 5301b96c3477SStefano Zampini PC_IS *pcis=(PC_IS*)pc->data; 5302b96c3477SStefano Zampini PC_BDDC *pcbddc=(PC_BDDC*)pc->data; 5303b96c3477SStefano Zampini PCBDDCSubSchurs sub_schurs=pcbddc->sub_schurs; 5304b96c3477SStefano Zampini PCBDDCGraph graph; 5305b96c3477SStefano Zampini PetscErrorCode ierr; 5306b96c3477SStefano Zampini 5307b96c3477SStefano Zampini PetscFunctionBegin; 5308b96c3477SStefano Zampini /* attach interface graph for determining subsets */ 530908122e43SStefano Zampini if (pcbddc->sub_schurs_rebuild) { /* in case rebuild has been requested, it uses a graph generated only by the neighbouring information */ 53103301b35fSStefano Zampini IS verticesIS,verticescomm; 53113301b35fSStefano Zampini PetscInt vsize,*idxs; 5312b96c3477SStefano Zampini 5313b96c3477SStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,NULL,NULL,NULL,NULL,&verticesIS);CHKERRQ(ierr); 53143301b35fSStefano Zampini ierr = ISGetSize(verticesIS,&vsize);CHKERRQ(ierr); 53153301b35fSStefano Zampini ierr = ISGetIndices(verticesIS,(const PetscInt**)&idxs);CHKERRQ(ierr); 53163301b35fSStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),vsize,idxs,PETSC_COPY_VALUES,&verticescomm);CHKERRQ(ierr); 53173301b35fSStefano Zampini ierr = ISRestoreIndices(verticesIS,(const PetscInt**)&idxs);CHKERRQ(ierr); 53183301b35fSStefano Zampini ierr = ISDestroy(&verticesIS);CHKERRQ(ierr); 5319b96c3477SStefano Zampini ierr = PCBDDCGraphCreate(&graph);CHKERRQ(ierr); 53207fb0e2dbSStefano Zampini ierr = PCBDDCGraphInit(graph,pcbddc->mat_graph->l2gmap,pcbddc->mat_graph->nvtxs_global);CHKERRQ(ierr); 53213301b35fSStefano Zampini ierr = PCBDDCGraphSetUp(graph,0,NULL,pcbddc->DirichletBoundariesLocal,0,NULL,verticescomm);CHKERRQ(ierr); 53223301b35fSStefano Zampini ierr = ISDestroy(&verticescomm);CHKERRQ(ierr); 5323b96c3477SStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(graph);CHKERRQ(ierr); 5324b96c3477SStefano Zampini /* 5325b96c3477SStefano Zampini if (pcbddc->dbg_flag) { 5326b96c3477SStefano Zampini ierr = PCBDDCGraphASCIIView(graph,pcbddc->dbg_flag,pcbddc->dbg_viewer);CHKERRQ(ierr); 5327b96c3477SStefano Zampini } 5328b96c3477SStefano Zampini */ 5329b96c3477SStefano Zampini } else { 5330b96c3477SStefano Zampini graph = pcbddc->mat_graph; 5331b96c3477SStefano Zampini } 5332b96c3477SStefano Zampini 5333b96c3477SStefano Zampini /* sub_schurs init */ 5334a64f4aa4SStefano Zampini ierr = PCBDDCSubSchursInit(sub_schurs,pcis->is_I_local,pcis->is_B_local,graph,pcis->BtoNmap);CHKERRQ(ierr); 5335a64f4aa4SStefano Zampini 5336b96c3477SStefano Zampini /* free graph struct */ 533708122e43SStefano Zampini if (pcbddc->sub_schurs_rebuild) { 5338b96c3477SStefano Zampini ierr = PCBDDCGraphDestroy(&graph);CHKERRQ(ierr); 5339b96c3477SStefano Zampini } 5340b96c3477SStefano Zampini PetscFunctionReturn(0); 5341b96c3477SStefano Zampini } 5342fa34dd3eSStefano Zampini 5343fa34dd3eSStefano Zampini #undef __FUNCT__ 5344fa34dd3eSStefano Zampini #define __FUNCT__ "PCBDDCCheckOperator" 5345fa34dd3eSStefano Zampini PetscErrorCode PCBDDCCheckOperator(PC pc) 5346fa34dd3eSStefano Zampini { 5347fa34dd3eSStefano Zampini PC_IS *pcis=(PC_IS*)pc->data; 5348fa34dd3eSStefano Zampini PC_BDDC *pcbddc=(PC_BDDC*)pc->data; 5349fa34dd3eSStefano Zampini PetscErrorCode ierr; 5350fa34dd3eSStefano Zampini 5351fa34dd3eSStefano Zampini PetscFunctionBegin; 5352fa34dd3eSStefano Zampini if (pcbddc->n_vertices == pcbddc->local_primal_size) { 5353fa34dd3eSStefano Zampini IS zerodiag = NULL; 5354fa34dd3eSStefano Zampini Mat S_j,B0=NULL,B0_B=NULL; 5355fa34dd3eSStefano Zampini Vec dummy_vec=NULL,vec_check_B,vec_scale_P; 5356fa34dd3eSStefano Zampini PetscScalar norm,p0_check,*array,*array2; 5357fa34dd3eSStefano Zampini PetscInt i; 5358fa34dd3eSStefano Zampini 5359fa34dd3eSStefano Zampini /* B0 and B0_B */ 5360fa34dd3eSStefano Zampini if (zerodiag) { 5361fa34dd3eSStefano Zampini IS dummy; 5362fa34dd3eSStefano Zampini PetscInt ii[2]; 5363fa34dd3eSStefano Zampini 5364fa34dd3eSStefano Zampini ii[0] = 0; 5365fa34dd3eSStefano Zampini ii[1] = pcbddc->B0_ncol; 5366fa34dd3eSStefano Zampini ierr = MatCreateSeqAIJWithArrays(PETSC_COMM_SELF,1,pcis->n,ii,pcbddc->B0_cols,pcbddc->B0_vals,&B0);CHKERRQ(ierr); 5367fa34dd3eSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,1,0,1,&dummy);CHKERRQ(ierr); 5368fa34dd3eSStefano Zampini ierr = MatGetSubMatrix(B0,dummy,pcis->is_B_local,MAT_INITIAL_MATRIX,&B0_B);CHKERRQ(ierr); 5369fa34dd3eSStefano Zampini ierr = MatCreateVecs(B0_B,NULL,&dummy_vec);CHKERRQ(ierr); 5370fa34dd3eSStefano Zampini ierr = ISDestroy(&dummy);CHKERRQ(ierr); 5371fa34dd3eSStefano Zampini } 5372fa34dd3eSStefano Zampini /* I need a primal vector to scale primal nodes since BDDC sums contibutions */ 5373fa34dd3eSStefano Zampini ierr = VecDuplicate(pcbddc->vec1_P,&vec_scale_P);CHKERRQ(ierr); 5374fa34dd3eSStefano Zampini ierr = VecSet(pcbddc->vec1_P,1.0);CHKERRQ(ierr); 5375fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 5376fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 5377fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,vec_scale_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 5378fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,vec_scale_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 5379fa34dd3eSStefano Zampini ierr = VecReciprocal(vec_scale_P);CHKERRQ(ierr); 5380fa34dd3eSStefano Zampini /* S_j */ 5381fa34dd3eSStefano Zampini ierr = MatCreateSchurComplement(pcis->A_II,pcis->A_II,pcis->A_IB,pcis->A_BI,pcis->A_BB,&S_j);CHKERRQ(ierr); 5382fa34dd3eSStefano Zampini ierr = MatSchurComplementSetKSP(S_j,pcbddc->ksp_D);CHKERRQ(ierr); 5383fa34dd3eSStefano Zampini 5384fa34dd3eSStefano Zampini /* mimic vector in \widetilde{W}_\Gamma */ 5385fa34dd3eSStefano Zampini ierr = VecSetRandom(pcis->vec1_N,NULL);CHKERRQ(ierr); 5386fa34dd3eSStefano Zampini /* continuous in primal space */ 5387fa34dd3eSStefano Zampini ierr = VecSetRandom(pcbddc->coarse_vec,NULL);CHKERRQ(ierr); 5388fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,pcbddc->vec1_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 5389fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,pcbddc->vec1_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 5390fa34dd3eSStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 5391fa34dd3eSStefano Zampini if (zerodiag) { 5392fa34dd3eSStefano Zampini p0_check = array[pcbddc->local_primal_size-1]; 5393fa34dd3eSStefano Zampini } else { 5394fa34dd3eSStefano Zampini p0_check = 0; 5395fa34dd3eSStefano Zampini } 5396fa34dd3eSStefano Zampini ierr = VecSetValues(pcis->vec1_N,pcbddc->local_primal_size,pcbddc->local_primal_ref_node,array,INSERT_VALUES);CHKERRQ(ierr); 5397fa34dd3eSStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 5398fa34dd3eSStefano Zampini ierr = VecAssemblyBegin(pcis->vec1_N);CHKERRQ(ierr); 5399fa34dd3eSStefano Zampini ierr = VecAssemblyEnd(pcis->vec1_N);CHKERRQ(ierr); 5400fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcis->N_to_B,pcis->vec1_N,pcis->vec2_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 5401fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcis->N_to_B,pcis->vec1_N,pcis->vec2_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 5402fa34dd3eSStefano Zampini ierr = VecDuplicate(pcis->vec2_B,&vec_check_B);CHKERRQ(ierr); 5403fa34dd3eSStefano Zampini ierr = VecCopy(pcis->vec2_B,vec_check_B);CHKERRQ(ierr); 5404fa34dd3eSStefano Zampini 5405fa34dd3eSStefano Zampini /* assemble rhs for coarse problem */ 5406fa34dd3eSStefano Zampini /* widetilde{S}_\Gamma w_\Gamma + \widetilde{B0}^T_B p0 */ 5407fa34dd3eSStefano Zampini /* local with Schur */ 5408fa34dd3eSStefano Zampini ierr = MatMult(S_j,pcis->vec2_B,pcis->vec1_B);CHKERRQ(ierr); 5409fa34dd3eSStefano Zampini if (zerodiag) { 5410fa34dd3eSStefano Zampini ierr = VecGetArray(dummy_vec,&array);CHKERRQ(ierr); 5411fa34dd3eSStefano Zampini array[0] = p0_check; 5412fa34dd3eSStefano Zampini ierr = VecRestoreArray(dummy_vec,&array);CHKERRQ(ierr); 5413fa34dd3eSStefano Zampini ierr = MatMultTransposeAdd(B0_B,dummy_vec,pcis->vec1_B,pcis->vec1_B);CHKERRQ(ierr); 5414fa34dd3eSStefano Zampini } 5415fa34dd3eSStefano Zampini /* sum on primal nodes the local contributions */ 5416fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcis->N_to_B,pcis->vec1_B,pcis->vec1_N,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 5417fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcis->N_to_B,pcis->vec1_B,pcis->vec1_N,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 5418fa34dd3eSStefano Zampini ierr = VecGetArray(pcis->vec1_N,&array);CHKERRQ(ierr); 5419fa34dd3eSStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array2);CHKERRQ(ierr); 5420fa34dd3eSStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) array2[i] = array[pcbddc->local_primal_ref_node[i]]; 5421fa34dd3eSStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array2);CHKERRQ(ierr); 5422fa34dd3eSStefano Zampini ierr = VecRestoreArray(pcis->vec1_N,&array);CHKERRQ(ierr); 5423fa34dd3eSStefano Zampini ierr = VecSet(pcbddc->coarse_vec,0.);CHKERRQ(ierr); 5424fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 5425fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 5426fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,pcbddc->vec1_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 5427fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,pcbddc->vec1_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 5428fa34dd3eSStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 5429fa34dd3eSStefano Zampini /* scale primal nodes (BDDC sums contibutions) */ 5430fa34dd3eSStefano Zampini ierr = VecPointwiseMult(pcbddc->vec1_P,vec_scale_P,pcbddc->vec1_P);CHKERRQ(ierr); 5431fa34dd3eSStefano Zampini ierr = VecSetValues(pcis->vec1_N,pcbddc->local_primal_size,pcbddc->local_primal_ref_node,array,INSERT_VALUES);CHKERRQ(ierr); 5432fa34dd3eSStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 5433fa34dd3eSStefano Zampini ierr = VecAssemblyBegin(pcis->vec1_N);CHKERRQ(ierr); 5434fa34dd3eSStefano Zampini ierr = VecAssemblyEnd(pcis->vec1_N);CHKERRQ(ierr); 5435fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcis->N_to_B,pcis->vec1_N,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 5436fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcis->N_to_B,pcis->vec1_N,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 5437fa34dd3eSStefano Zampini /* global: \widetilde{B0}_B w_\Gamma */ 5438fa34dd3eSStefano Zampini if (zerodiag) { 5439fa34dd3eSStefano Zampini ierr = MatMult(B0_B,pcis->vec2_B,dummy_vec);CHKERRQ(ierr); 5440fa34dd3eSStefano Zampini ierr = VecGetArray(dummy_vec,&array);CHKERRQ(ierr); 5441fa34dd3eSStefano Zampini pcbddc->benign_p0 = array[0]; 5442fa34dd3eSStefano Zampini ierr = VecRestoreArray(dummy_vec,&array);CHKERRQ(ierr); 5443fa34dd3eSStefano Zampini } else { 5444fa34dd3eSStefano Zampini pcbddc->benign_p0 = 0.; 5445fa34dd3eSStefano Zampini } 5446fa34dd3eSStefano Zampini /* BDDC */ 5447fa34dd3eSStefano Zampini ierr = VecSet(pcis->vec1_D,0.);CHKERRQ(ierr); 5448fa34dd3eSStefano Zampini ierr = PCBDDCApplyInterfacePreconditioner(pc,PETSC_FALSE);CHKERRQ(ierr); 5449fa34dd3eSStefano Zampini 5450fa34dd3eSStefano Zampini ierr = VecCopy(pcis->vec1_B,pcis->vec2_B);CHKERRQ(ierr); 5451fa34dd3eSStefano Zampini ierr = VecAXPY(pcis->vec1_B,-1.0,vec_check_B);CHKERRQ(ierr); 5452fa34dd3eSStefano Zampini ierr = VecNorm(pcis->vec1_B,NORM_INFINITY,&norm);CHKERRQ(ierr); 5453fa34dd3eSStefano Zampini PetscPrintf(PETSC_COMM_SELF,"[%d] BDDC local error is %1.4e\n",PetscGlobalRank,norm); 5454fa34dd3eSStefano Zampini if (pcbddc->benign_p0_lidx >= 0) { 5455fa34dd3eSStefano Zampini PetscPrintf(PETSC_COMM_SELF,"[%d] BDDC p0 error is %1.4e\n",PetscGlobalRank,PetscAbsScalar(pcbddc->benign_p0-p0_check)); 5456fa34dd3eSStefano Zampini } 5457fa34dd3eSStefano Zampini 5458fa34dd3eSStefano Zampini ierr = VecDestroy(&vec_scale_P);CHKERRQ(ierr); 5459fa34dd3eSStefano Zampini ierr = VecDestroy(&vec_check_B);CHKERRQ(ierr); 5460fa34dd3eSStefano Zampini ierr = VecDestroy(&dummy_vec);CHKERRQ(ierr); 5461fa34dd3eSStefano Zampini ierr = MatDestroy(&S_j);CHKERRQ(ierr); 5462fa34dd3eSStefano Zampini ierr = MatDestroy(&B0);CHKERRQ(ierr); 5463fa34dd3eSStefano Zampini ierr = MatDestroy(&B0_B);CHKERRQ(ierr); 5464fa34dd3eSStefano Zampini } 5465fa34dd3eSStefano Zampini PetscFunctionReturn(0); 5466fa34dd3eSStefano Zampini } 5467