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__ 9*4f1b2e48SStefano Zampini #define __FUNCT__ "MatDetectDisconnectedComponents" 10*4f1b2e48SStefano Zampini PetscErrorCode MatDetectDisconnectedComponents(Mat A, PetscBool filter, PetscInt *ncc, IS* cc[]) 11*4f1b2e48SStefano Zampini { 12*4f1b2e48SStefano Zampini Mat B; 13*4f1b2e48SStefano Zampini IS is_dummy,*cc_n; 14*4f1b2e48SStefano Zampini ISLocalToGlobalMapping l2gmap_dummy; 15*4f1b2e48SStefano Zampini PCBDDCGraph graph; 16*4f1b2e48SStefano Zampini PetscInt i,n; 17*4f1b2e48SStefano Zampini PetscInt *xadj,*adjncy; 18*4f1b2e48SStefano Zampini PetscInt *xadj_filtered,*adjncy_filtered; 19*4f1b2e48SStefano Zampini PetscBool flg_row,isseqaij; 20*4f1b2e48SStefano Zampini PetscErrorCode ierr; 21*4f1b2e48SStefano Zampini 22*4f1b2e48SStefano Zampini PetscFunctionBegin; 23*4f1b2e48SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)A,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 24*4f1b2e48SStefano Zampini if (!isseqaij && filter) { 25*4f1b2e48SStefano Zampini ierr = MatConvert(A,MATSEQAIJ,MAT_INITIAL_MATRIX,&B);CHKERRQ(ierr); 26*4f1b2e48SStefano Zampini } else { 27*4f1b2e48SStefano Zampini B = A; 28*4f1b2e48SStefano Zampini } 29*4f1b2e48SStefano Zampini ierr = MatGetRowIJ(B,0,PETSC_TRUE,PETSC_FALSE,&n,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 30*4f1b2e48SStefano Zampini 31*4f1b2e48SStefano Zampini /* if filter is true, then removes entries lower than PETSC_SMALL in magnitude */ 32*4f1b2e48SStefano Zampini if (filter) { 33*4f1b2e48SStefano Zampini PetscScalar *data; 34*4f1b2e48SStefano Zampini PetscInt j,cum; 35*4f1b2e48SStefano Zampini 36*4f1b2e48SStefano Zampini ierr = PetscCalloc2(n+1,&xadj_filtered,xadj[n],&adjncy_filtered);CHKERRQ(ierr); 37*4f1b2e48SStefano Zampini ierr = MatSeqAIJGetArray(B,&data);CHKERRQ(ierr); 38*4f1b2e48SStefano Zampini cum = 0; 39*4f1b2e48SStefano Zampini for (i=0;i<n;i++) { 40*4f1b2e48SStefano Zampini PetscInt t; 41*4f1b2e48SStefano Zampini 42*4f1b2e48SStefano Zampini for (j=xadj[i];j<xadj[i+1];j++) { 43*4f1b2e48SStefano Zampini if (PetscUnlikely(PetscAbsScalar(data[j]) < PETSC_SMALL)) { 44*4f1b2e48SStefano Zampini continue; 45*4f1b2e48SStefano Zampini } 46*4f1b2e48SStefano Zampini adjncy_filtered[cum+xadj_filtered[i]++] = adjncy[j]; 47*4f1b2e48SStefano Zampini } 48*4f1b2e48SStefano Zampini t = xadj_filtered[i]; 49*4f1b2e48SStefano Zampini xadj_filtered[i] = cum; 50*4f1b2e48SStefano Zampini cum += t; 51*4f1b2e48SStefano Zampini } 52*4f1b2e48SStefano Zampini ierr = MatSeqAIJRestoreArray(B,&data);CHKERRQ(ierr); 53*4f1b2e48SStefano Zampini } else { 54*4f1b2e48SStefano Zampini xadj_filtered = NULL; 55*4f1b2e48SStefano Zampini adjncy_filtered = NULL; 56*4f1b2e48SStefano Zampini } 57*4f1b2e48SStefano Zampini 58*4f1b2e48SStefano Zampini /* compute local connected components using PCBDDCGraph */ 59*4f1b2e48SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,n,0,1,&is_dummy);CHKERRQ(ierr); 60*4f1b2e48SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(is_dummy,&l2gmap_dummy);CHKERRQ(ierr); 61*4f1b2e48SStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 62*4f1b2e48SStefano Zampini ierr = PCBDDCGraphCreate(&graph);CHKERRQ(ierr); 63*4f1b2e48SStefano Zampini ierr = PCBDDCGraphInit(graph,l2gmap_dummy,n);CHKERRQ(ierr); 64*4f1b2e48SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&l2gmap_dummy);CHKERRQ(ierr); 65*4f1b2e48SStefano Zampini if (xadj_filtered) { 66*4f1b2e48SStefano Zampini graph->xadj = xadj_filtered; 67*4f1b2e48SStefano Zampini graph->adjncy = adjncy_filtered; 68*4f1b2e48SStefano Zampini } else { 69*4f1b2e48SStefano Zampini graph->xadj = xadj; 70*4f1b2e48SStefano Zampini graph->adjncy = adjncy; 71*4f1b2e48SStefano Zampini } 72*4f1b2e48SStefano Zampini ierr = PCBDDCGraphSetUp(graph,1,NULL,NULL,0,NULL,NULL);CHKERRQ(ierr); 73*4f1b2e48SStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(graph);CHKERRQ(ierr); 74*4f1b2e48SStefano Zampini 75*4f1b2e48SStefano Zampini /* partial clean up */ 76*4f1b2e48SStefano Zampini ierr = PetscFree2(xadj_filtered,adjncy_filtered);CHKERRQ(ierr); 77*4f1b2e48SStefano Zampini ierr = MatRestoreRowIJ(B,0,PETSC_TRUE,PETSC_FALSE,&n,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 78*4f1b2e48SStefano Zampini if (!isseqaij && filter) { 79*4f1b2e48SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 80*4f1b2e48SStefano Zampini } 81*4f1b2e48SStefano Zampini 82*4f1b2e48SStefano Zampini /* get back data */ 83*4f1b2e48SStefano Zampini *ncc = graph->ncc; 84*4f1b2e48SStefano Zampini ierr = PetscMalloc1(graph->ncc,&cc_n);CHKERRQ(ierr); 85*4f1b2e48SStefano Zampini for (i=0;i<graph->ncc;i++) { 86*4f1b2e48SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,graph->cptr[i+1]-graph->cptr[i],graph->queue+graph->cptr[i],PETSC_COPY_VALUES,&cc_n[i]);CHKERRQ(ierr); 87*4f1b2e48SStefano Zampini } 88*4f1b2e48SStefano Zampini *cc = cc_n; 89*4f1b2e48SStefano Zampini 90*4f1b2e48SStefano Zampini /* clean up graph */ 91*4f1b2e48SStefano Zampini graph->xadj = 0; 92*4f1b2e48SStefano Zampini graph->adjncy = 0; 93*4f1b2e48SStefano Zampini ierr = PCBDDCGraphDestroy(&graph);CHKERRQ(ierr); 94*4f1b2e48SStefano Zampini PetscFunctionReturn(0); 95*4f1b2e48SStefano Zampini } 96*4f1b2e48SStefano Zampini 97*4f1b2e48SStefano Zampini #undef __FUNCT__ 985408967cSStefano Zampini #define __FUNCT__ "PCBDDCBenignCheck" 995408967cSStefano Zampini PetscErrorCode PCBDDCBenignCheck(PC pc, IS zerodiag) 1005408967cSStefano Zampini { 1015408967cSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 1025408967cSStefano Zampini PC_IS* pcis = (PC_IS*)(pc->data); 103dee84bffSStefano Zampini IS dirIS = NULL; 104*4f1b2e48SStefano Zampini PetscInt i; 1055408967cSStefano Zampini PetscErrorCode ierr; 1065408967cSStefano Zampini 1075408967cSStefano Zampini PetscFunctionBegin; 108dee84bffSStefano Zampini ierr = PCBDDCGraphGetDirichletDofs(pcbddc->mat_graph,&dirIS);CHKERRQ(ierr); 1095408967cSStefano Zampini if (zerodiag) { 1105408967cSStefano Zampini Mat A; 1115408967cSStefano Zampini Vec vec3_N; 1125408967cSStefano Zampini PetscScalar *vals; 1135408967cSStefano Zampini const PetscInt *idxs; 114*4f1b2e48SStefano Zampini PetscInt nz; 1155408967cSStefano Zampini 1165408967cSStefano Zampini /* p0 */ 1175408967cSStefano Zampini ierr = VecSet(pcis->vec1_N,0.);CHKERRQ(ierr); 1185408967cSStefano Zampini ierr = PetscMalloc1(pcis->n,&vals);CHKERRQ(ierr); 1195408967cSStefano Zampini ierr = ISGetLocalSize(zerodiag,&nz);CHKERRQ(ierr); 1205408967cSStefano Zampini ierr = ISGetIndices(zerodiag,&idxs);CHKERRQ(ierr); 121*4f1b2e48SStefano Zampini for (i=0;i<nz;i++) vals[i] = 1.; 1225408967cSStefano Zampini ierr = VecSetValues(pcis->vec1_N,nz,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 1235408967cSStefano Zampini ierr = VecAssemblyBegin(pcis->vec1_N);CHKERRQ(ierr); 1245408967cSStefano Zampini ierr = VecAssemblyEnd(pcis->vec1_N);CHKERRQ(ierr); 1255408967cSStefano Zampini /* v_I */ 1265408967cSStefano Zampini ierr = VecSetRandom(pcis->vec2_N,NULL);CHKERRQ(ierr); 1275408967cSStefano Zampini for (i=0;i<nz;i++) vals[i] = 0.; 1285408967cSStefano Zampini ierr = VecSetValues(pcis->vec2_N,nz,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 1295408967cSStefano Zampini ierr = ISRestoreIndices(zerodiag,&idxs);CHKERRQ(ierr); 1305408967cSStefano Zampini ierr = ISGetIndices(pcis->is_B_local,&idxs);CHKERRQ(ierr); 1315408967cSStefano Zampini for (i=0;i<pcis->n_B;i++) vals[i] = 0.; 1325408967cSStefano Zampini ierr = VecSetValues(pcis->vec2_N,pcis->n_B,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 1335408967cSStefano Zampini ierr = ISRestoreIndices(pcis->is_B_local,&idxs);CHKERRQ(ierr); 1345408967cSStefano Zampini if (dirIS) { 1355408967cSStefano Zampini PetscInt n; 1365408967cSStefano Zampini 1375408967cSStefano Zampini ierr = ISGetLocalSize(dirIS,&n);CHKERRQ(ierr); 1385408967cSStefano Zampini ierr = ISGetIndices(dirIS,&idxs);CHKERRQ(ierr); 1395408967cSStefano Zampini for (i=0;i<n;i++) vals[i] = 0.; 1405408967cSStefano Zampini ierr = VecSetValues(pcis->vec2_N,n,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 1415408967cSStefano Zampini ierr = ISRestoreIndices(dirIS,&idxs);CHKERRQ(ierr); 1425408967cSStefano Zampini } 1435408967cSStefano Zampini ierr = VecAssemblyBegin(pcis->vec2_N);CHKERRQ(ierr); 1445408967cSStefano Zampini ierr = VecAssemblyEnd(pcis->vec2_N);CHKERRQ(ierr); 1455408967cSStefano Zampini ierr = VecDuplicate(pcis->vec1_N,&vec3_N);CHKERRQ(ierr); 1465408967cSStefano Zampini ierr = VecSet(vec3_N,0.);CHKERRQ(ierr); 1475408967cSStefano Zampini ierr = MatISGetLocalMat(pc->mat,&A);CHKERRQ(ierr); 1485408967cSStefano Zampini ierr = MatMult(A,pcis->vec1_N,vec3_N);CHKERRQ(ierr); 1495408967cSStefano Zampini ierr = VecDot(vec3_N,pcis->vec2_N,&vals[0]);CHKERRQ(ierr); 150*4f1b2e48SStefano Zampini if (PetscAbsScalar(vals[0]) > PETSC_SMALL) { 1515408967cSStefano 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])); 1525408967cSStefano Zampini } 1535408967cSStefano Zampini ierr = PetscFree(vals);CHKERRQ(ierr); 1545408967cSStefano Zampini ierr = VecDestroy(&vec3_N);CHKERRQ(ierr); 1555408967cSStefano Zampini } 156dee84bffSStefano Zampini ierr = ISDestroy(&dirIS);CHKERRQ(ierr); 1575408967cSStefano Zampini 1585408967cSStefano Zampini /* check PCBDDCBenignGetOrSetP0 */ 1595408967cSStefano Zampini ierr = VecSetRandom(pcis->vec1_global,NULL);CHKERRQ(ierr); 160*4f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) pcbddc->benign_p0[i] = -PetscGlobalRank-i; 1615408967cSStefano Zampini ierr = PCBDDCBenignGetOrSetP0(pc,pcis->vec1_global,PETSC_FALSE);CHKERRQ(ierr); 162*4f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) pcbddc->benign_p0[i] = 1; 1635408967cSStefano Zampini ierr = PCBDDCBenignGetOrSetP0(pc,pcis->vec1_global,PETSC_TRUE);CHKERRQ(ierr); 164*4f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 165*4f1b2e48SStefano Zampini if ((PetscInt)PetscRealPart(pcbddc->benign_p0[i]) != -PetscGlobalRank-i) { 166*4f1b2e48SStefano Zampini SETERRQ3(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Error testing PCBDDCBenignGetOrSetP0! Found %1.4e at %d instead of %1.4e\n",pcbddc->benign_p0[i],i,-PetscGlobalRank-i);CHKERRQ(ierr); 167*4f1b2e48SStefano Zampini } 1685408967cSStefano Zampini } 1695408967cSStefano Zampini PetscFunctionReturn(0); 1705408967cSStefano Zampini } 1715408967cSStefano Zampini 1725408967cSStefano Zampini #undef __FUNCT__ 173339f8db1SStefano Zampini #define __FUNCT__ "PCBDDCBenignDetectSaddlePoint" 174339f8db1SStefano Zampini PetscErrorCode PCBDDCBenignDetectSaddlePoint(PC pc, IS *zerodiaglocal) 175339f8db1SStefano Zampini { 176339f8db1SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 177*4f1b2e48SStefano Zampini IS pressures,zerodiag,*zerodiag_subs; 178b0f5fe93SStefano Zampini PetscInt nz,n; 179*4f1b2e48SStefano Zampini PetscBool sorted,have_null,has_null_pressures,recompute_zerodiag; 180339f8db1SStefano Zampini PetscErrorCode ierr; 181339f8db1SStefano Zampini 182339f8db1SStefano Zampini PetscFunctionBegin; 183339f8db1SStefano Zampini ierr = MatDestroy(&pcbddc->benign_original_mat);CHKERRQ(ierr); 184339f8db1SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->local_mat);CHKERRQ(ierr); 185339f8db1SStefano Zampini pcbddc->benign_original_mat = pcbddc->local_mat; 186*4f1b2e48SStefano Zampini /* if a local info on dofs is present, assumes the last field is represented by "pressures" 187*4f1b2e48SStefano Zampini otherwise, it uses only zerodiagonal dofs (ok if the pressure block is all zero; it could fail if it is not) 188*4f1b2e48SStefano Zampini Checks if all the pressure dofs in each subdomain have a zero diagonal 189*4f1b2e48SStefano Zampini If not, a change of basis on pressures is not needed 190*4f1b2e48SStefano Zampini since the local Schur complements are SPD 191*4f1b2e48SStefano Zampini */ 192*4f1b2e48SStefano Zampini has_null_pressures = PETSC_TRUE; 193*4f1b2e48SStefano Zampini have_null = PETSC_TRUE; 19440fa8d13SStefano Zampini if (pcbddc->n_ISForDofsLocal) { 195*4f1b2e48SStefano Zampini PetscInt npl,*idxs,p = pcbddc->n_ISForDofsLocal-1; 196*4f1b2e48SStefano Zampini 197*4f1b2e48SStefano Zampini /* Dofs splitting for BDDC cannot have PETSC_COMM_SELF, so create a sequential IS */ 198*4f1b2e48SStefano Zampini ierr = ISGetLocalSize(pcbddc->ISForDofsLocal[p],&npl);CHKERRQ(ierr); 199*4f1b2e48SStefano Zampini ierr = ISGetIndices(pcbddc->ISForDofsLocal[p],(const PetscInt**)&idxs);CHKERRQ(ierr); 200*4f1b2e48SStefano Zampini ierr = ISRestoreIndices(pcbddc->ISForDofsLocal[p],(const PetscInt**)&idxs);CHKERRQ(ierr); 201*4f1b2e48SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,npl,idxs,PETSC_COPY_VALUES,&pressures);CHKERRQ(ierr); 20240fa8d13SStefano Zampini ierr = ISSorted(pressures,&sorted);CHKERRQ(ierr); 20340fa8d13SStefano Zampini if (!sorted) { 20440fa8d13SStefano Zampini ierr = ISSort(pressures);CHKERRQ(ierr); 20540fa8d13SStefano Zampini } 20640fa8d13SStefano Zampini } else { 20740fa8d13SStefano Zampini pressures = NULL; 20840fa8d13SStefano Zampini } 209*4f1b2e48SStefano Zampini ierr = MatGetLocalSize(pcbddc->benign_original_mat,&n,NULL);CHKERRQ(ierr); 210*4f1b2e48SStefano Zampini /* TODO: add check for shared dofs and raise error */ 211339f8db1SStefano Zampini ierr = MatFindZeroDiagonals(pcbddc->benign_original_mat,&zerodiag);CHKERRQ(ierr); 212339f8db1SStefano Zampini ierr = ISSorted(zerodiag,&sorted);CHKERRQ(ierr); 213339f8db1SStefano Zampini if (!sorted) { 214339f8db1SStefano Zampini ierr = ISSort(zerodiag);CHKERRQ(ierr); 215339f8db1SStefano Zampini } 216339f8db1SStefano Zampini ierr = ISGetLocalSize(zerodiag,&nz);CHKERRQ(ierr); 217*4f1b2e48SStefano Zampini if (!nz) { 218*4f1b2e48SStefano Zampini if (n) have_null = PETSC_FALSE; 219*4f1b2e48SStefano Zampini has_null_pressures = PETSC_FALSE; 22040fa8d13SStefano Zampini ierr = ISDestroy(&zerodiag);CHKERRQ(ierr); 22140fa8d13SStefano Zampini } 222339f8db1SStefano Zampini 223*4f1b2e48SStefano Zampini recompute_zerodiag = PETSC_FALSE; 224*4f1b2e48SStefano Zampini /* in case disconnected subdomains info is present, split the pressures accordingly (otherwise the benign trick could fail) */ 225*4f1b2e48SStefano Zampini zerodiag_subs = NULL; 226*4f1b2e48SStefano Zampini pcbddc->benign_n = 0; 227*4f1b2e48SStefano Zampini if (has_null_pressures) { 228*4f1b2e48SStefano Zampini IS *subs; 229*4f1b2e48SStefano Zampini PetscInt nsubs,i; 230*4f1b2e48SStefano Zampini 231*4f1b2e48SStefano Zampini subs = pcbddc->local_subs; 232*4f1b2e48SStefano Zampini nsubs = pcbddc->n_local_subs; 233*4f1b2e48SStefano Zampini if (nsubs > 1) { 234*4f1b2e48SStefano Zampini ierr = PetscCalloc1(nsubs,&zerodiag_subs);CHKERRQ(ierr); 235*4f1b2e48SStefano Zampini for (i=0;i<nsubs;i++) { 236*4f1b2e48SStefano Zampini ISLocalToGlobalMapping l2g; 237*4f1b2e48SStefano Zampini IS t_zerodiag_subs; 238*4f1b2e48SStefano Zampini PetscInt nl; 239*4f1b2e48SStefano Zampini 240*4f1b2e48SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(subs[i],&l2g);CHKERRQ(ierr); 241*4f1b2e48SStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(l2g,IS_GTOLM_DROP,zerodiag,&t_zerodiag_subs);CHKERRQ(ierr); 242*4f1b2e48SStefano Zampini ierr = ISGetLocalSize(t_zerodiag_subs,&nl);CHKERRQ(ierr); 243*4f1b2e48SStefano Zampini if (nl) { 244*4f1b2e48SStefano Zampini PetscBool valid = PETSC_TRUE; 245*4f1b2e48SStefano Zampini 246*4f1b2e48SStefano Zampini if (pressures) { 247*4f1b2e48SStefano Zampini IS t_pressure_subs; 248*4f1b2e48SStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(l2g,IS_GTOLM_DROP,pressures,&t_pressure_subs);CHKERRQ(ierr); 249*4f1b2e48SStefano Zampini ierr = ISEqual(t_pressure_subs,t_zerodiag_subs,&valid);CHKERRQ(ierr); 250*4f1b2e48SStefano Zampini ierr = ISDestroy(&t_pressure_subs);CHKERRQ(ierr); 251*4f1b2e48SStefano Zampini } 252*4f1b2e48SStefano Zampini if (valid) { 253*4f1b2e48SStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(l2g,t_zerodiag_subs,&zerodiag_subs[pcbddc->benign_n]);CHKERRQ(ierr); 254*4f1b2e48SStefano Zampini pcbddc->benign_n++; 255*4f1b2e48SStefano Zampini } else { 256*4f1b2e48SStefano Zampini recompute_zerodiag = PETSC_TRUE; 257*4f1b2e48SStefano Zampini } 258*4f1b2e48SStefano Zampini } 259*4f1b2e48SStefano Zampini ierr = ISDestroy(&t_zerodiag_subs);CHKERRQ(ierr); 260*4f1b2e48SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&l2g);CHKERRQ(ierr); 261*4f1b2e48SStefano Zampini } 262*4f1b2e48SStefano Zampini } else { /* there's just one subdomain (or zero if they have not been detected */ 263*4f1b2e48SStefano Zampini PetscBool valid = PETSC_TRUE; 264*4f1b2e48SStefano Zampini if (pressures) { 265*4f1b2e48SStefano Zampini ierr = ISEqual(pressures,zerodiag,&valid);CHKERRQ(ierr); 266*4f1b2e48SStefano Zampini } 267*4f1b2e48SStefano Zampini if (valid) { 268*4f1b2e48SStefano Zampini pcbddc->benign_n = 1; 269*4f1b2e48SStefano Zampini ierr = PetscCalloc1(1,&zerodiag_subs);CHKERRQ(ierr); 270*4f1b2e48SStefano Zampini ierr = PetscObjectReference((PetscObject)zerodiag);CHKERRQ(ierr); 271*4f1b2e48SStefano Zampini zerodiag_subs[0] = zerodiag; 272*4f1b2e48SStefano Zampini } 273*4f1b2e48SStefano Zampini } 274*4f1b2e48SStefano Zampini } 275*4f1b2e48SStefano Zampini 276*4f1b2e48SStefano Zampini if (!pcbddc->benign_n) { 277*4f1b2e48SStefano Zampini ierr = ISDestroy(&zerodiag);CHKERRQ(ierr); 278*4f1b2e48SStefano Zampini recompute_zerodiag = PETSC_FALSE; 279*4f1b2e48SStefano Zampini has_null_pressures = PETSC_FALSE; 280*4f1b2e48SStefano Zampini have_null = PETSC_FALSE; 281*4f1b2e48SStefano Zampini } 282*4f1b2e48SStefano Zampini 283*4f1b2e48SStefano Zampini /* final check for null pressures */ 284*4f1b2e48SStefano Zampini if (zerodiag && pressures) { 285*4f1b2e48SStefano Zampini PetscInt nz,np; 286*4f1b2e48SStefano Zampini ierr = ISGetLocalSize(zerodiag,&nz);CHKERRQ(ierr); 287*4f1b2e48SStefano Zampini ierr = ISGetLocalSize(pressures,&np);CHKERRQ(ierr); 288*4f1b2e48SStefano Zampini if (nz != np) have_null = PETSC_FALSE; 289*4f1b2e48SStefano Zampini } 290*4f1b2e48SStefano Zampini 291*4f1b2e48SStefano Zampini if (recompute_zerodiag) { 292*4f1b2e48SStefano Zampini ierr = ISDestroy(&zerodiag);CHKERRQ(ierr); 293*4f1b2e48SStefano Zampini if (pcbddc->benign_n == 1) { 294*4f1b2e48SStefano Zampini ierr = PetscObjectReference((PetscObject)zerodiag_subs[0]);CHKERRQ(ierr); 295*4f1b2e48SStefano Zampini zerodiag = zerodiag_subs[0]; 296*4f1b2e48SStefano Zampini } else { 297*4f1b2e48SStefano Zampini PetscInt i,nzn,*new_idxs; 298*4f1b2e48SStefano Zampini 299*4f1b2e48SStefano Zampini nzn = 0; 300*4f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 301*4f1b2e48SStefano Zampini PetscInt ns; 302*4f1b2e48SStefano Zampini ierr = ISGetLocalSize(zerodiag_subs[i],&ns);CHKERRQ(ierr); 303*4f1b2e48SStefano Zampini nzn += ns; 304*4f1b2e48SStefano Zampini } 305*4f1b2e48SStefano Zampini ierr = PetscMalloc1(nzn,&new_idxs);CHKERRQ(ierr); 306*4f1b2e48SStefano Zampini nzn = 0; 307*4f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 308*4f1b2e48SStefano Zampini PetscInt ns,*idxs; 309*4f1b2e48SStefano Zampini ierr = ISGetLocalSize(zerodiag_subs[i],&ns);CHKERRQ(ierr); 310*4f1b2e48SStefano Zampini ierr = ISGetIndices(zerodiag_subs[i],(const PetscInt**)&idxs);CHKERRQ(ierr); 311*4f1b2e48SStefano Zampini ierr = PetscMemcpy(new_idxs+nzn,idxs,ns*sizeof(PetscInt));CHKERRQ(ierr); 312*4f1b2e48SStefano Zampini ierr = ISRestoreIndices(zerodiag_subs[i],(const PetscInt**)&idxs);CHKERRQ(ierr); 313*4f1b2e48SStefano Zampini nzn += ns; 314*4f1b2e48SStefano Zampini } 315*4f1b2e48SStefano Zampini ierr = PetscSortInt(nzn,new_idxs);CHKERRQ(ierr); 316*4f1b2e48SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,nzn,new_idxs,PETSC_OWN_POINTER,&zerodiag);CHKERRQ(ierr); 317*4f1b2e48SStefano Zampini } 318*4f1b2e48SStefano Zampini have_null = PETSC_FALSE; 319*4f1b2e48SStefano Zampini } 320*4f1b2e48SStefano Zampini 321*4f1b2e48SStefano Zampini if (has_null_pressures) { 322*4f1b2e48SStefano Zampini IS zerodiagc; 323*4f1b2e48SStefano Zampini const PetscInt *idxs,*idxsc; 324*4f1b2e48SStefano Zampini PetscInt i,s,*nnz; 325*4f1b2e48SStefano Zampini 326*4f1b2e48SStefano Zampini ierr = ISGetLocalSize(zerodiag,&nz);CHKERRQ(ierr); 327339f8db1SStefano Zampini ierr = ISComplement(zerodiag,0,n,&zerodiagc);CHKERRQ(ierr); 328339f8db1SStefano Zampini ierr = ISGetIndices(zerodiagc,&idxsc);CHKERRQ(ierr); 329339f8db1SStefano Zampini /* local change of basis for pressures */ 330339f8db1SStefano Zampini ierr = MatDestroy(&pcbddc->benign_change);CHKERRQ(ierr); 331339f8db1SStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)pcbddc->benign_original_mat),&pcbddc->benign_change);CHKERRQ(ierr); 332339f8db1SStefano Zampini ierr = MatSetType(pcbddc->benign_change,MATAIJ);CHKERRQ(ierr); 333339f8db1SStefano Zampini ierr = MatSetSizes(pcbddc->benign_change,n,n,PETSC_DECIDE,PETSC_DECIDE);CHKERRQ(ierr); 334339f8db1SStefano Zampini ierr = PetscMalloc1(n,&nnz);CHKERRQ(ierr); 335*4f1b2e48SStefano Zampini for (i=0;i<n-nz;i++) nnz[idxsc[i]] = 1; /* identity on velocities plus pressure dofs for non-singular subdomains */ 336*4f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 337*4f1b2e48SStefano Zampini PetscInt nzs,j; 338*4f1b2e48SStefano Zampini 339*4f1b2e48SStefano Zampini ierr = ISGetLocalSize(zerodiag_subs[i],&nzs);CHKERRQ(ierr); 340*4f1b2e48SStefano Zampini ierr = ISGetIndices(zerodiag_subs[i],&idxs);CHKERRQ(ierr); 341*4f1b2e48SStefano Zampini for (j=0;j<nzs-1;j++) nnz[idxs[j]] = 2; /* change on pressures */ 342*4f1b2e48SStefano Zampini nnz[idxs[nzs-1]] = nzs; /* last local pressure dof in subdomain */ 343*4f1b2e48SStefano Zampini ierr = ISRestoreIndices(zerodiag_subs[i],&idxs);CHKERRQ(ierr); 344*4f1b2e48SStefano Zampini } 345339f8db1SStefano Zampini ierr = MatSeqAIJSetPreallocation(pcbddc->benign_change,0,nnz);CHKERRQ(ierr); 346339f8db1SStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 347339f8db1SStefano Zampini /* set identity on velocities */ 348339f8db1SStefano Zampini for (i=0;i<n-nz;i++) { 349339f8db1SStefano Zampini ierr = MatSetValue(pcbddc->benign_change,idxsc[i],idxsc[i],1.,INSERT_VALUES);CHKERRQ(ierr); 350339f8db1SStefano Zampini } 351*4f1b2e48SStefano Zampini ierr = ISRestoreIndices(zerodiagc,&idxsc);CHKERRQ(ierr); 352*4f1b2e48SStefano Zampini ierr = ISDestroy(&zerodiagc);CHKERRQ(ierr); 353*4f1b2e48SStefano Zampini ierr = PetscMalloc3(pcbddc->benign_n,&pcbddc->benign_p0_lidx,pcbddc->benign_n,&pcbddc->benign_p0_gidx,pcbddc->benign_n,&pcbddc->benign_p0);CHKERRQ(ierr); 354339f8db1SStefano Zampini /* set change on pressures */ 355*4f1b2e48SStefano Zampini for (s=0;s<pcbddc->benign_n;s++) { 356*4f1b2e48SStefano Zampini PetscScalar *array; 357*4f1b2e48SStefano Zampini PetscInt nzs; 358*4f1b2e48SStefano Zampini 359*4f1b2e48SStefano Zampini ierr = ISGetLocalSize(zerodiag_subs[s],&nzs);CHKERRQ(ierr); 360*4f1b2e48SStefano Zampini ierr = ISGetIndices(zerodiag_subs[s],&idxs);CHKERRQ(ierr); 361*4f1b2e48SStefano Zampini for (i=0;i<nzs-1;i++) { 362339f8db1SStefano Zampini PetscScalar vals[2]; 363339f8db1SStefano Zampini PetscInt cols[2]; 364339f8db1SStefano Zampini 365339f8db1SStefano Zampini cols[0] = idxs[i]; 366*4f1b2e48SStefano Zampini cols[1] = idxs[nzs-1]; 367339f8db1SStefano Zampini vals[0] = 1.; 368b0f5fe93SStefano Zampini vals[1] = 1.; 369*4f1b2e48SStefano Zampini ierr = MatSetValues(pcbddc->benign_change,1,cols,2,cols,vals,INSERT_VALUES);CHKERRQ(ierr); 370339f8db1SStefano Zampini } 371*4f1b2e48SStefano Zampini ierr = PetscMalloc1(nzs,&array);CHKERRQ(ierr); 372*4f1b2e48SStefano Zampini for (i=0;i<nzs-1;i++) array[i] = -1.; 373*4f1b2e48SStefano Zampini array[nzs-1] = 1.; 374*4f1b2e48SStefano Zampini ierr = MatSetValues(pcbddc->benign_change,1,idxs+nzs-1,nzs,idxs,array,INSERT_VALUES);CHKERRQ(ierr); 375*4f1b2e48SStefano Zampini /* store local idxs for p0 */ 376*4f1b2e48SStefano Zampini pcbddc->benign_p0_lidx[s] = idxs[nzs-1]; 377*4f1b2e48SStefano Zampini ierr = ISRestoreIndices(zerodiag_subs[s],&idxs);CHKERRQ(ierr); 378*4f1b2e48SStefano Zampini ierr = ISDestroy(&zerodiag_subs[s]);CHKERRQ(ierr); 379339f8db1SStefano Zampini ierr = PetscFree(array);CHKERRQ(ierr); 380*4f1b2e48SStefano Zampini } 381339f8db1SStefano Zampini ierr = MatAssemblyBegin(pcbddc->benign_change,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 382339f8db1SStefano Zampini ierr = MatAssemblyEnd(pcbddc->benign_change,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 383339f8db1SStefano Zampini /* TODO: need optimization? */ 384339f8db1SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 385339f8db1SStefano Zampini ierr = MatPtAP(pcbddc->benign_original_mat,pcbddc->benign_change,MAT_INITIAL_MATRIX,2.0,&pcbddc->local_mat);CHKERRQ(ierr); 386*4f1b2e48SStefano Zampini /* store global idxs for p0 */ 387*4f1b2e48SStefano Zampini ierr = ISLocalToGlobalMappingApply(pc->pmat->rmap->mapping,pcbddc->benign_n,pcbddc->benign_p0_lidx,pcbddc->benign_p0_gidx);CHKERRQ(ierr); 388339f8db1SStefano Zampini } 389*4f1b2e48SStefano Zampini ierr = PetscFree(zerodiag_subs);CHKERRQ(ierr); 390*4f1b2e48SStefano Zampini ierr = ISDestroy(&pressures);CHKERRQ(ierr); 391b0f5fe93SStefano Zampini 392b0f5fe93SStefano Zampini /* determines if the coarse solver will be singular or not */ 393b0f5fe93SStefano Zampini ierr = MPI_Allreduce(&have_null,&pcbddc->benign_null,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 394339f8db1SStefano Zampini *zerodiaglocal = zerodiag; 395339f8db1SStefano Zampini PetscFunctionReturn(0); 396339f8db1SStefano Zampini } 397339f8db1SStefano Zampini 398339f8db1SStefano Zampini #undef __FUNCT__ 399015636ebSStefano Zampini #define __FUNCT__ "PCBDDCBenignGetOrSetP0" 400015636ebSStefano Zampini PetscErrorCode PCBDDCBenignGetOrSetP0(PC pc, Vec v, PetscBool get) 401efc2fbd9SStefano Zampini { 402efc2fbd9SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 403efc2fbd9SStefano Zampini PetscErrorCode ierr; 404efc2fbd9SStefano Zampini 405efc2fbd9SStefano Zampini PetscFunctionBegin; 406efc2fbd9SStefano Zampini if (!pcbddc->benign_sf) { 407efc2fbd9SStefano Zampini ierr = PetscSFCreate(PetscObjectComm((PetscObject)pc),&pcbddc->benign_sf);CHKERRQ(ierr); 408*4f1b2e48SStefano Zampini ierr = PetscSFSetGraphLayout(pcbddc->benign_sf,pc->pmat->rmap,pcbddc->benign_n,NULL,PETSC_OWN_POINTER,pcbddc->benign_p0_gidx);CHKERRQ(ierr); 409efc2fbd9SStefano Zampini } 410015636ebSStefano Zampini if (get) { /* use SF to get values */ 411efc2fbd9SStefano Zampini PetscScalar *array; 412efc2fbd9SStefano Zampini 413efc2fbd9SStefano Zampini ierr = VecGetArrayRead(v,(const PetscScalar**)&array);CHKERRQ(ierr); 414*4f1b2e48SStefano Zampini ierr = PetscSFBcastBegin(pcbddc->benign_sf,MPIU_SCALAR,array,pcbddc->benign_p0);CHKERRQ(ierr); 415*4f1b2e48SStefano Zampini ierr = PetscSFBcastEnd(pcbddc->benign_sf,MPIU_SCALAR,array,pcbddc->benign_p0);CHKERRQ(ierr); 416efc2fbd9SStefano Zampini ierr = VecRestoreArrayRead(v,(const PetscScalar**)&array);CHKERRQ(ierr); 417*4f1b2e48SStefano Zampini } else { /* use VecSetValues (not scalable, I should try to find a better solution (defining a new MPI_OP for reduction) */ 418*4f1b2e48SStefano Zampini ierr = VecSetValues(v,pcbddc->benign_n,pcbddc->benign_p0_gidx,pcbddc->benign_p0,INSERT_VALUES);CHKERRQ(ierr); 419efc2fbd9SStefano Zampini ierr = VecAssemblyBegin(v);CHKERRQ(ierr); 420efc2fbd9SStefano Zampini ierr = VecAssemblyEnd(v);CHKERRQ(ierr); 421efc2fbd9SStefano Zampini } 422efc2fbd9SStefano Zampini PetscFunctionReturn(0); 423efc2fbd9SStefano Zampini } 424efc2fbd9SStefano Zampini 425efc2fbd9SStefano Zampini #undef __FUNCT__ 426c263805aSStefano Zampini #define __FUNCT__ "PCBDDCBenignPopOrPushB0" 427c263805aSStefano Zampini PetscErrorCode PCBDDCBenignPopOrPushB0(PC pc, PetscBool pop) 428c263805aSStefano Zampini { 429c263805aSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 430c263805aSStefano Zampini PetscErrorCode ierr; 431c263805aSStefano Zampini 432c263805aSStefano Zampini PetscFunctionBegin; 433c263805aSStefano Zampini /* TODO: add error checking 434c263805aSStefano Zampini - avoid nested pop (or push) calls. 435c263805aSStefano Zampini - cannot push before pop. 4361c604dc7SStefano Zampini - cannot call this if pcbddc->local_mat is NULL 437c263805aSStefano Zampini */ 438*4f1b2e48SStefano Zampini if (!pcbddc->benign_n) { 439efc2fbd9SStefano Zampini PetscFunctionReturn(0); 440efc2fbd9SStefano Zampini } 441c263805aSStefano Zampini if (pop) { 442*4f1b2e48SStefano Zampini IS is_p0; 443*4f1b2e48SStefano Zampini MatReuse reuse; 444c263805aSStefano Zampini 445c263805aSStefano Zampini /* extract B_0 */ 446*4f1b2e48SStefano Zampini reuse = MAT_INITIAL_MATRIX; 447*4f1b2e48SStefano Zampini if (pcbddc->benign_B0) { 448*4f1b2e48SStefano Zampini reuse = MAT_REUSE_MATRIX; 449*4f1b2e48SStefano Zampini } 450*4f1b2e48SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,pcbddc->benign_n,pcbddc->benign_p0_lidx,PETSC_COPY_VALUES,&is_p0);CHKERRQ(ierr); 451*4f1b2e48SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,is_p0,NULL,reuse,&pcbddc->benign_B0);CHKERRQ(ierr); 452c263805aSStefano Zampini /* remove rows and cols from local problem */ 453c263805aSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_KEEP_NONZERO_PATTERN,PETSC_TRUE);CHKERRQ(ierr); 454*4f1b2e48SStefano Zampini ierr = MatZeroRowsColumnsIS(pcbddc->local_mat,is_p0,1.0,NULL,NULL);CHKERRQ(ierr); 455*4f1b2e48SStefano Zampini ierr = ISDestroy(&is_p0);CHKERRQ(ierr); 456c263805aSStefano Zampini } else { /* push */ 457*4f1b2e48SStefano Zampini PetscInt i; 458*4f1b2e48SStefano Zampini 459*4f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 460*4f1b2e48SStefano Zampini PetscScalar *B0_vals; 461*4f1b2e48SStefano Zampini PetscInt *B0_cols,B0_ncol; 462*4f1b2e48SStefano Zampini 463*4f1b2e48SStefano Zampini ierr = MatGetRow(pcbddc->benign_B0,i,&B0_ncol,(const PetscInt**)&B0_cols,(const PetscScalar**)&B0_vals);CHKERRQ(ierr); 464*4f1b2e48SStefano Zampini ierr = MatSetValues(pcbddc->local_mat,1,pcbddc->benign_p0_lidx+i,B0_ncol,B0_cols,B0_vals,INSERT_VALUES);CHKERRQ(ierr); 465*4f1b2e48SStefano Zampini ierr = MatSetValues(pcbddc->local_mat,B0_ncol,B0_cols,1,pcbddc->benign_p0_lidx+1,B0_vals,INSERT_VALUES);CHKERRQ(ierr); 466*4f1b2e48SStefano Zampini ierr = MatSetValue(pcbddc->local_mat,pcbddc->benign_p0_lidx[i],pcbddc->benign_p0_lidx[i],0.0,INSERT_VALUES);CHKERRQ(ierr); 467*4f1b2e48SStefano Zampini ierr = MatRestoreRow(pcbddc->benign_B0,i,&B0_ncol,(const PetscInt**)&B0_cols,(const PetscScalar**)&B0_vals);CHKERRQ(ierr); 468*4f1b2e48SStefano Zampini } 469c263805aSStefano Zampini ierr = MatAssemblyBegin(pcbddc->local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 470c263805aSStefano Zampini ierr = MatAssemblyEnd(pcbddc->local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 471c263805aSStefano Zampini } 472c263805aSStefano Zampini PetscFunctionReturn(0); 473c263805aSStefano Zampini } 474c263805aSStefano Zampini 475c263805aSStefano Zampini #undef __FUNCT__ 476b1b3d7a2SStefano Zampini #define __FUNCT__ "PCBDDCAdaptiveSelection" 47708122e43SStefano Zampini PetscErrorCode PCBDDCAdaptiveSelection(PC pc) 478b1b3d7a2SStefano Zampini { 479b1b3d7a2SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 48008122e43SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 48108122e43SStefano Zampini PetscBLASInt B_dummyint,B_neigs,B_ierr,B_lwork; 48208122e43SStefano Zampini PetscBLASInt *B_iwork,*B_ifail; 48308122e43SStefano Zampini PetscScalar *work,lwork; 48408122e43SStefano Zampini PetscScalar *St,*S,*eigv; 48508122e43SStefano Zampini PetscScalar *Sarray,*Starray; 48608122e43SStefano Zampini PetscReal *eigs,thresh; 4871b968477SStefano Zampini PetscInt i,nmax,nmin,nv,cum,mss,cum2,cumarray,maxneigs; 488f6f667cfSStefano Zampini PetscBool allocated_S_St; 48908122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 49008122e43SStefano Zampini PetscReal *rwork; 49108122e43SStefano Zampini #endif 492b1b3d7a2SStefano Zampini PetscErrorCode ierr; 493b1b3d7a2SStefano Zampini 494b1b3d7a2SStefano Zampini PetscFunctionBegin; 49508122e43SStefano Zampini if (!sub_schurs->use_mumps) { 49608122e43SStefano Zampini SETERRQ(PetscObjectComm((PetscObject)pc),PETSC_ERR_SUP,"Adaptive selection of constraints requires MUMPS"); 49708122e43SStefano Zampini } 49808122e43SStefano Zampini 49906a4e24aSStefano Zampini if (sub_schurs->n_subs && (!sub_schurs->is_hermitian || !sub_schurs->is_posdef)) { 50006a4e24aSStefano 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); 50106a4e24aSStefano Zampini } 50206a4e24aSStefano Zampini 503fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 504fd14bc51SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 505fd14bc51SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 506fd14bc51SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Check adaptive selection of constraints\n");CHKERRQ(ierr); 507fd14bc51SStefano Zampini ierr = PetscViewerASCIISynchronizedAllow(pcbddc->dbg_viewer,PETSC_TRUE);CHKERRQ(ierr); 508fd14bc51SStefano Zampini } 509fd14bc51SStefano Zampini 510e496cd5dSStefano Zampini if (pcbddc->dbg_flag) { 511e496cd5dSStefano 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); 512e496cd5dSStefano Zampini } 513e496cd5dSStefano Zampini 51408122e43SStefano Zampini /* max size of subsets */ 51508122e43SStefano Zampini mss = 0; 51608122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 51708122e43SStefano Zampini PetscInt subset_size; 518862806e4SStefano Zampini 51908122e43SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 52008122e43SStefano Zampini mss = PetscMax(mss,subset_size); 52108122e43SStefano Zampini } 52208122e43SStefano Zampini 52308122e43SStefano Zampini /* min/max and threshold */ 52408122e43SStefano Zampini nmax = pcbddc->adaptive_nmax > 0 ? pcbddc->adaptive_nmax : mss; 525f6f667cfSStefano Zampini nmin = pcbddc->adaptive_nmin > 0 ? pcbddc->adaptive_nmin : 0; 52608122e43SStefano Zampini nmax = PetscMax(nmin,nmax); 527f6f667cfSStefano Zampini allocated_S_St = PETSC_FALSE; 528f6f667cfSStefano Zampini if (nmin) { 529f6f667cfSStefano Zampini allocated_S_St = PETSC_TRUE; 530f6f667cfSStefano Zampini } 53108122e43SStefano Zampini 53208122e43SStefano Zampini /* allocate lapack workspace */ 53308122e43SStefano Zampini cum = cum2 = 0; 53408122e43SStefano Zampini maxneigs = 0; 53508122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 53608122e43SStefano Zampini PetscInt n,subset_size; 537f6f667cfSStefano Zampini 53808122e43SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 53908122e43SStefano Zampini n = PetscMin(subset_size,nmax); 5409162d606SStefano Zampini cum += subset_size; 5419162d606SStefano Zampini cum2 += subset_size*n; 54208122e43SStefano Zampini maxneigs = PetscMax(maxneigs,n); 54308122e43SStefano Zampini } 54408122e43SStefano Zampini if (mss) { 5459ab7bb16SStefano Zampini if (sub_schurs->is_hermitian && sub_schurs->is_posdef) { 54608122e43SStefano Zampini PetscBLASInt B_itype = 1; 54708122e43SStefano Zampini PetscBLASInt B_N = mss; 5484c6709b3SStefano Zampini PetscReal zero = 0.0; 5494c6709b3SStefano Zampini PetscReal eps = 0.0; /* dlamch? */ 55008122e43SStefano Zampini 55108122e43SStefano Zampini B_lwork = -1; 55208122e43SStefano Zampini S = NULL; 55308122e43SStefano Zampini St = NULL; 554a58a30b4SStefano Zampini eigs = NULL; 555a58a30b4SStefano Zampini eigv = NULL; 556a58a30b4SStefano Zampini B_iwork = NULL; 557a58a30b4SStefano Zampini B_ifail = NULL; 558d1710679SStefano Zampini #if defined(PETSC_USE_COMPLEX) 559d1710679SStefano Zampini rwork = NULL; 560d1710679SStefano Zampini #endif 5618bec7fa6SStefano Zampini thresh = 1.0; 56208122e43SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 56308122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 56408122e43SStefano 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)); 56508122e43SStefano Zampini #else 56608122e43SStefano 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)); 56708122e43SStefano Zampini #endif 56808122e43SStefano Zampini if (B_ierr != 0) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to SYGVX Lapack routine %d",(int)B_ierr); 56908122e43SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 57008122e43SStefano Zampini } else { 57108122e43SStefano Zampini /* TODO */ 57208122e43SStefano Zampini } 57308122e43SStefano Zampini } else { 57408122e43SStefano Zampini lwork = 0; 57508122e43SStefano Zampini } 57608122e43SStefano Zampini 57708122e43SStefano Zampini nv = 0; 578d62866d3SStefano 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) */ 579d62866d3SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_vertices,&nv);CHKERRQ(ierr); 58008122e43SStefano Zampini } 5814c6709b3SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lwork),&B_lwork);CHKERRQ(ierr); 582f6f667cfSStefano Zampini if (allocated_S_St) { 583f6f667cfSStefano Zampini ierr = PetscMalloc2(mss*mss,&S,mss*mss,&St);CHKERRQ(ierr); 584f6f667cfSStefano Zampini } 585f6f667cfSStefano Zampini ierr = PetscMalloc5(mss*mss,&eigv,mss,&eigs,B_lwork,&work,5*mss,&B_iwork,mss,&B_ifail);CHKERRQ(ierr); 58608122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 58708122e43SStefano Zampini ierr = PetscMalloc1(7*mss,&rwork);CHKERRQ(ierr); 58808122e43SStefano Zampini #endif 5899162d606SStefano Zampini ierr = PetscMalloc5(nv+sub_schurs->n_subs,&pcbddc->adaptive_constraints_n, 5909162d606SStefano Zampini nv+sub_schurs->n_subs+1,&pcbddc->adaptive_constraints_idxs_ptr, 5919162d606SStefano Zampini nv+sub_schurs->n_subs+1,&pcbddc->adaptive_constraints_data_ptr, 59208122e43SStefano Zampini nv+cum,&pcbddc->adaptive_constraints_idxs, 5939162d606SStefano Zampini nv+cum2,&pcbddc->adaptive_constraints_data);CHKERRQ(ierr); 59408122e43SStefano Zampini ierr = PetscMemzero(pcbddc->adaptive_constraints_n,(nv+sub_schurs->n_subs)*sizeof(PetscInt));CHKERRQ(ierr); 59508122e43SStefano Zampini 59608122e43SStefano Zampini maxneigs = 0; 59708122e43SStefano Zampini cum = cum2 = cumarray = 0; 5989162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr[0] = 0; 5999162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr[0] = 0; 600d62866d3SStefano Zampini if (sub_schurs->is_vertices && pcbddc->use_vertices) { 60108122e43SStefano Zampini const PetscInt *idxs; 60208122e43SStefano Zampini 603d62866d3SStefano Zampini ierr = ISGetIndices(sub_schurs->is_vertices,&idxs);CHKERRQ(ierr); 60408122e43SStefano Zampini for (cum=0;cum<nv;cum++) { 60508122e43SStefano Zampini pcbddc->adaptive_constraints_n[cum] = 1; 60608122e43SStefano Zampini pcbddc->adaptive_constraints_idxs[cum] = idxs[cum]; 60708122e43SStefano Zampini pcbddc->adaptive_constraints_data[cum] = 1.0; 6089162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr[cum+1] = pcbddc->adaptive_constraints_idxs_ptr[cum]+1; 6099162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr[cum+1] = pcbddc->adaptive_constraints_data_ptr[cum]+1; 61008122e43SStefano Zampini } 61108122e43SStefano Zampini cum2 = cum; 612d62866d3SStefano Zampini ierr = ISRestoreIndices(sub_schurs->is_vertices,&idxs);CHKERRQ(ierr); 61308122e43SStefano Zampini } 61408122e43SStefano Zampini 61508122e43SStefano Zampini if (mss) { /* multilevel */ 61608122e43SStefano Zampini ierr = MatSeqAIJGetArray(sub_schurs->sum_S_Ej_inv_all,&Sarray);CHKERRQ(ierr); 61708122e43SStefano Zampini ierr = MatSeqAIJGetArray(sub_schurs->sum_S_Ej_tilda_all,&Starray);CHKERRQ(ierr); 61808122e43SStefano Zampini } 61908122e43SStefano Zampini 62008122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 62108122e43SStefano Zampini 62208122e43SStefano Zampini const PetscInt *idxs; 623f6f667cfSStefano Zampini PetscReal infty = PETSC_MAX_REAL; 624862806e4SStefano Zampini PetscInt j,subset_size,eigs_start = 0; 62508122e43SStefano Zampini PetscBLASInt B_N; 626aff50787SStefano Zampini PetscBool same_data = PETSC_FALSE; 62708122e43SStefano Zampini 628862806e4SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 629f6f667cfSStefano Zampini ierr = PetscBLASIntCast(subset_size,&B_N);CHKERRQ(ierr); 630f6f667cfSStefano Zampini if (allocated_S_St) { /* S and S_t should be copied since we could need them later */ 6319ab7bb16SStefano Zampini if (sub_schurs->is_hermitian) { 632aff50787SStefano Zampini PetscInt j,k; 633aff50787SStefano Zampini if (sub_schurs->n_subs == 1) { /* zeroing memory to use PetscMemcmp later */ 634aff50787SStefano Zampini ierr = PetscMemzero(S,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 635aff50787SStefano Zampini ierr = PetscMemzero(St,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 63608122e43SStefano Zampini } 63708122e43SStefano Zampini for (j=0;j<subset_size;j++) { 638aff50787SStefano Zampini for (k=j;k<subset_size;k++) { 639aff50787SStefano Zampini S [j*subset_size+k] = Sarray [cumarray+j*subset_size+k]; 640aff50787SStefano Zampini St[j*subset_size+k] = Starray[cumarray+j*subset_size+k]; 641aff50787SStefano Zampini } 64208122e43SStefano Zampini } 64308122e43SStefano Zampini } else { 64408122e43SStefano Zampini ierr = PetscMemcpy(S,Sarray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 64508122e43SStefano Zampini ierr = PetscMemcpy(St,Starray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 64608122e43SStefano Zampini } 6478bec7fa6SStefano Zampini } else { 648f6f667cfSStefano Zampini S = Sarray + cumarray; 649f6f667cfSStefano Zampini St = Starray + cumarray; 6508bec7fa6SStefano Zampini } 65108122e43SStefano Zampini 652f6f667cfSStefano Zampini ierr = ISGetIndices(sub_schurs->is_subs[i],&idxs);CHKERRQ(ierr); 653aff50787SStefano Zampini /* see if we can save some work */ 654aff50787SStefano Zampini if (sub_schurs->n_subs == 1) { 655aff50787SStefano Zampini ierr = PetscMemcmp(S,St,subset_size*subset_size*sizeof(PetscScalar),&same_data);CHKERRQ(ierr); 656aff50787SStefano Zampini } 657aff50787SStefano Zampini 658aff50787SStefano Zampini if (same_data) { /* there's no need of constraints here, deluxe scaling is enough */ 659aff50787SStefano Zampini B_neigs = 0; 660aff50787SStefano Zampini } else { 661aff50787SStefano Zampini /* Threshold: this is an heuristic for edges */ 662f6f667cfSStefano Zampini thresh = pcbddc->mat_graph->count[idxs[0]]*pcbddc->adaptive_threshold; 663f6f667cfSStefano Zampini 6649ab7bb16SStefano Zampini if (sub_schurs->is_hermitian && sub_schurs->is_posdef) { 66508122e43SStefano Zampini PetscBLASInt B_itype = 1; 666f6f667cfSStefano Zampini PetscBLASInt B_IL, B_IU; 6674c6709b3SStefano Zampini PetscReal eps = -1.0; /* dlamch? */ 6689552c7c7SStefano Zampini PetscInt nmin_s; 66908122e43SStefano Zampini 670fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 6718bec7fa6SStefano 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]]); 672fd14bc51SStefano Zampini } 673d16cbb6bSStefano Zampini 67408122e43SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 675d16cbb6bSStefano Zampini if (thresh > 1.+PETSC_SMALL) { 676d16cbb6bSStefano Zampini 677d16cbb6bSStefano Zampini /* ask for eigenvalues larger than thresh */ 67808122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 679f6f667cfSStefano 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)); 68008122e43SStefano Zampini #else 681f6f667cfSStefano 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)); 68208122e43SStefano Zampini #endif 683d16cbb6bSStefano Zampini } else { 684d16cbb6bSStefano Zampini B_IU = PetscMax(1,PetscMin(B_N,nmax)); 685d16cbb6bSStefano Zampini B_IL = 1; 686d16cbb6bSStefano Zampini #if defined(PETSC_USE_COMPLEX) 687d16cbb6bSStefano 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)); 688d16cbb6bSStefano Zampini #else 689d16cbb6bSStefano 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)); 690d16cbb6bSStefano Zampini #endif 691d16cbb6bSStefano Zampini } 69208122e43SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 69308122e43SStefano Zampini if (B_ierr) { 69408122e43SStefano Zampini if (B_ierr < 0 ) { 69508122e43SStefano Zampini SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYGVX Lapack routine: illegal value for argument %d",-(int)B_ierr); 69608122e43SStefano Zampini } else if (B_ierr <= B_N) { 69708122e43SStefano Zampini SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYGVX Lapack routine: %d eigenvalues failed to converge",(int)B_ierr); 69808122e43SStefano Zampini } else { 6999552c7c7SStefano 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); 70008122e43SStefano Zampini } 70108122e43SStefano Zampini } 70208122e43SStefano Zampini 70308122e43SStefano Zampini if (B_neigs > nmax) { 704fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 705fd14bc51SStefano Zampini PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," found %d eigs, more than maximum required %d.\n",B_neigs,nmax); 706fd14bc51SStefano Zampini } 707f6f667cfSStefano Zampini eigs_start = B_neigs -nmax; 70808122e43SStefano Zampini B_neigs = nmax; 70908122e43SStefano Zampini } 71008122e43SStefano Zampini 7119552c7c7SStefano Zampini nmin_s = PetscMin(nmin,B_N); 7129552c7c7SStefano Zampini if (B_neigs < nmin_s) { 71308122e43SStefano Zampini PetscBLASInt B_neigs2; 71408122e43SStefano Zampini 715f6f667cfSStefano Zampini B_IU = B_N - B_neigs; 716f6f667cfSStefano Zampini B_IL = B_N - nmin_s + 1; 717fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 718fd14bc51SStefano 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); 719fd14bc51SStefano Zampini } 7209ab7bb16SStefano Zampini if (sub_schurs->is_hermitian) { 72108122e43SStefano Zampini PetscInt j; 72208122e43SStefano Zampini for (j=0;j<subset_size;j++) { 72308122e43SStefano Zampini ierr = PetscMemcpy(S+j*(subset_size+1),Sarray+cumarray+j*(subset_size+1),(subset_size-j)*sizeof(PetscScalar));CHKERRQ(ierr); 72408122e43SStefano Zampini } 72508122e43SStefano Zampini for (j=0;j<subset_size;j++) { 72608122e43SStefano Zampini ierr = PetscMemcpy(St+j*(subset_size+1),Starray+cumarray+j*(subset_size+1),(subset_size-j)*sizeof(PetscScalar));CHKERRQ(ierr); 72708122e43SStefano Zampini } 72808122e43SStefano Zampini } else { 72908122e43SStefano Zampini ierr = PetscMemcpy(S,Sarray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 73008122e43SStefano Zampini ierr = PetscMemcpy(St,Starray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 73108122e43SStefano Zampini } 73208122e43SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 73308122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 734f6f667cfSStefano 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)); 73508122e43SStefano Zampini #else 736f6f667cfSStefano 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)); 73708122e43SStefano Zampini #endif 73808122e43SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 73908122e43SStefano Zampini B_neigs += B_neigs2; 74008122e43SStefano Zampini } 74108122e43SStefano Zampini if (B_ierr) { 74208122e43SStefano Zampini if (B_ierr < 0 ) { 74308122e43SStefano Zampini SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYGVX Lapack routine: illegal value for argument %d",-(int)B_ierr); 74408122e43SStefano Zampini } else if (B_ierr <= B_N) { 74508122e43SStefano Zampini SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYGVX Lapack routine: %d eigenvalues failed to converge",(int)B_ierr); 74608122e43SStefano Zampini } else { 7479552c7c7SStefano 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); 74808122e43SStefano Zampini } 74908122e43SStefano Zampini } 750fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 751ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," -> Got %d eigs\n",B_neigs);CHKERRQ(ierr); 75208122e43SStefano Zampini for (j=0;j<B_neigs;j++) { 75308122e43SStefano Zampini if (eigs[j] == 0.0) { 754ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," Inf\n");CHKERRQ(ierr); 75508122e43SStefano Zampini } else { 756ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," %1.6e\n",eigs[j+eigs_start]);CHKERRQ(ierr); 757fd14bc51SStefano Zampini } 75808122e43SStefano Zampini } 75908122e43SStefano Zampini } 76008122e43SStefano Zampini } else { 76108122e43SStefano Zampini /* TODO */ 76208122e43SStefano Zampini } 763aff50787SStefano Zampini } 7648bec7fa6SStefano Zampini maxneigs = PetscMax(B_neigs,maxneigs); 7658bec7fa6SStefano Zampini pcbddc->adaptive_constraints_n[i+nv] = B_neigs; 7669162d606SStefano Zampini if (B_neigs) { 7679162d606SStefano 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); 768fd14bc51SStefano Zampini 769fd14bc51SStefano Zampini if (pcbddc->dbg_flag > 1) { 7709552c7c7SStefano Zampini PetscInt ii; 7719552c7c7SStefano Zampini for (ii=0;ii<B_neigs;ii++) { 772ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," -> Eigenvector %d/%d (%d)\n",ii,B_neigs,B_N);CHKERRQ(ierr); 7739552c7c7SStefano Zampini for (j=0;j<B_N;j++) { 774ac47001eSStefano Zampini #if defined(PETSC_USE_COMPLEX) 775ac47001eSStefano Zampini PetscReal r = PetscRealPart(pcbddc->adaptive_constraints_data[ii*subset_size+j+pcbddc->adaptive_constraints_data_ptr[cum]]); 776ac47001eSStefano Zampini PetscReal c = PetscImaginaryPart(pcbddc->adaptive_constraints_data[ii*subset_size+j+pcbddc->adaptive_constraints_data_ptr[cum]]); 777ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," %1.4e + %1.4e i\n",r,c);CHKERRQ(ierr); 778ac47001eSStefano Zampini #else 779ac47001eSStefano 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); 780ac47001eSStefano Zampini #endif 7819552c7c7SStefano Zampini } 7829552c7c7SStefano Zampini } 783fd14bc51SStefano Zampini } 78408122e43SStefano Zampini #if 0 7859162d606SStefano Zampini for (j=0;j<B_neigs;j++) { 78608122e43SStefano Zampini PetscBLASInt Blas_N,Blas_one = 1.0; 78708122e43SStefano Zampini PetscScalar norm; 78808122e43SStefano Zampini ierr = PetscBLASIntCast(subset_size,&Blas_N);CHKERRQ(ierr); 7899162d606SStefano Zampini PetscStackCallBLAS("BLASdot",norm = BLASdot_(&Blas_N,pcbddc->adaptive_constraints_data+pcbddc->adaptive_constraints_data_ptr[cum]+j*subset_size, 7909162d606SStefano Zampini &Blas_one,pcbddc->adaptive_constraints_data+pcbddc->adaptive_constraints_data_ptr[cum]+j*subset_size,&Blas_one)); 79108122e43SStefano Zampini if (pcbddc->adaptive_constraints_data[cum2] > 0.0) { 79208122e43SStefano Zampini norm = 1.0/PetscSqrtReal(PetscRealPart(norm)); 79308122e43SStefano Zampini } else { 79408122e43SStefano Zampini norm = -1.0/PetscSqrtReal(PetscRealPart(norm)); 79508122e43SStefano Zampini } 7969162d606SStefano Zampini PetscStackCallBLAS("BLASscal",BLASscal_(&Blas_N,&norm,pcbddc->adaptive_constraints_data+pcbddc->adaptive_constraints_data_ptr[cum]+j*subset_size,&Blas_one)); 797b1b3d7a2SStefano Zampini } 798b1b3d7a2SStefano Zampini #endif 7999162d606SStefano Zampini ierr = PetscMemcpy(pcbddc->adaptive_constraints_idxs+pcbddc->adaptive_constraints_idxs_ptr[cum],idxs,subset_size*sizeof(PetscInt));CHKERRQ(ierr); 8009162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr[cum+1] = pcbddc->adaptive_constraints_idxs_ptr[cum] + subset_size; 8019162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr[cum+1] = pcbddc->adaptive_constraints_data_ptr[cum] + subset_size*B_neigs; 8029162d606SStefano Zampini cum++; 80308122e43SStefano Zampini } 80408122e43SStefano Zampini ierr = ISRestoreIndices(sub_schurs->is_subs[i],&idxs);CHKERRQ(ierr); 80508122e43SStefano Zampini /* shift for next computation */ 80608122e43SStefano Zampini cumarray += subset_size*subset_size; 80708122e43SStefano Zampini } 808fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 809fd14bc51SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 810fd14bc51SStefano Zampini } 81108122e43SStefano Zampini 81208122e43SStefano Zampini if (mss) { 81308122e43SStefano Zampini ierr = MatSeqAIJRestoreArray(sub_schurs->sum_S_Ej_inv_all,&Sarray);CHKERRQ(ierr); 81408122e43SStefano Zampini ierr = MatSeqAIJRestoreArray(sub_schurs->sum_S_Ej_tilda_all,&Starray);CHKERRQ(ierr); 815f6f667cfSStefano Zampini /* destroy matrices (junk) */ 816f6f667cfSStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_inv_all);CHKERRQ(ierr); 817f6f667cfSStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_tilda_all);CHKERRQ(ierr); 81808122e43SStefano Zampini } 819f6f667cfSStefano Zampini if (allocated_S_St) { 820f6f667cfSStefano Zampini ierr = PetscFree2(S,St);CHKERRQ(ierr); 821f6f667cfSStefano Zampini } 822f6f667cfSStefano Zampini ierr = PetscFree5(eigv,eigs,work,B_iwork,B_ifail);CHKERRQ(ierr); 82308122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 82408122e43SStefano Zampini ierr = PetscFree(rwork);CHKERRQ(ierr); 82508122e43SStefano Zampini #endif 82608122e43SStefano Zampini if (pcbddc->dbg_flag) { 8271b968477SStefano Zampini PetscInt maxneigs_r; 82808122e43SStefano Zampini ierr = MPI_Allreduce(&maxneigs,&maxneigs_r,1,MPIU_INT,MPI_MAX,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 8299b28b941SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Maximum number of constraints per cc %d\n",maxneigs_r);CHKERRQ(ierr); 83008122e43SStefano Zampini } 83108122e43SStefano Zampini PetscFunctionReturn(0); 83208122e43SStefano Zampini } 833b1b3d7a2SStefano Zampini 834674ae819SStefano Zampini #undef __FUNCT__ 835c8587f34SStefano Zampini #define __FUNCT__ "PCBDDCSetUpSolvers" 836c8587f34SStefano Zampini PetscErrorCode PCBDDCSetUpSolvers(PC pc) 837c8587f34SStefano Zampini { 838c8587f34SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 8398629588bSStefano Zampini PetscScalar *coarse_submat_vals; 840c8587f34SStefano Zampini PetscErrorCode ierr; 841c8587f34SStefano Zampini 842c8587f34SStefano Zampini PetscFunctionBegin; 843f4ddd8eeSStefano Zampini /* Setup local scatters R_to_B and (optionally) R_to_D */ 8445e8657edSStefano Zampini /* PCBDDCSetUpLocalWorkVectors should be called first! */ 845c8587f34SStefano Zampini ierr = PCBDDCSetUpLocalScatters(pc);CHKERRQ(ierr); 846c8587f34SStefano Zampini 847684f6988SStefano Zampini /* Setup local neumann solver ksp_R */ 8480fccc4e9SStefano Zampini /* PCBDDCSetUpLocalScatters should be called first! */ 849684f6988SStefano Zampini ierr = PCBDDCSetUpLocalSolvers(pc,PETSC_FALSE,PETSC_TRUE);CHKERRQ(ierr); 850c8587f34SStefano Zampini 851c8587f34SStefano Zampini /* Change global null space passed in by the user if change of basis has been requested */ 852b9b85e73SStefano Zampini if (pcbddc->NullSpace && pcbddc->ChangeOfBasisMatrix) { 853c8587f34SStefano Zampini ierr = PCBDDCNullSpaceAdaptGlobal(pc);CHKERRQ(ierr); 854c8587f34SStefano Zampini } 855c8587f34SStefano Zampini 8568629588bSStefano Zampini /* 8578629588bSStefano Zampini Setup local correction and local part of coarse basis. 8588629588bSStefano Zampini Gives back the dense local part of the coarse matrix in column major ordering 8598629588bSStefano Zampini */ 86047f4ddc3SStefano Zampini ierr = PCBDDCSetUpCorrection(pc,&coarse_submat_vals);CHKERRQ(ierr); 8618629588bSStefano Zampini 8628629588bSStefano Zampini /* Compute total number of coarse nodes and setup coarse solver */ 8638629588bSStefano Zampini ierr = PCBDDCSetUpCoarseSolver(pc,coarse_submat_vals);CHKERRQ(ierr); 8648629588bSStefano Zampini 8658629588bSStefano Zampini /* free */ 8668629588bSStefano Zampini ierr = PetscFree(coarse_submat_vals);CHKERRQ(ierr); 867c8587f34SStefano Zampini PetscFunctionReturn(0); 868c8587f34SStefano Zampini } 869c8587f34SStefano Zampini 870c8587f34SStefano Zampini #undef __FUNCT__ 871674ae819SStefano Zampini #define __FUNCT__ "PCBDDCResetCustomization" 872674ae819SStefano Zampini PetscErrorCode PCBDDCResetCustomization(PC pc) 873674ae819SStefano Zampini { 874674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 875674ae819SStefano Zampini PetscErrorCode ierr; 876674ae819SStefano Zampini 877674ae819SStefano Zampini PetscFunctionBegin; 878674ae819SStefano Zampini ierr = PCBDDCGraphResetCSR(pcbddc->mat_graph);CHKERRQ(ierr); 879674ae819SStefano Zampini ierr = ISDestroy(&pcbddc->user_primal_vertices);CHKERRQ(ierr); 880674ae819SStefano Zampini ierr = MatNullSpaceDestroy(&pcbddc->NullSpace);CHKERRQ(ierr); 881674ae819SStefano Zampini ierr = ISDestroy(&pcbddc->NeumannBoundaries);CHKERRQ(ierr); 882785d1243SStefano Zampini ierr = ISDestroy(&pcbddc->NeumannBoundariesLocal);CHKERRQ(ierr); 883674ae819SStefano Zampini ierr = ISDestroy(&pcbddc->DirichletBoundaries);CHKERRQ(ierr); 884f4ddd8eeSStefano Zampini ierr = MatNullSpaceDestroy(&pcbddc->onearnullspace);CHKERRQ(ierr); 885f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->onearnullvecs_state);CHKERRQ(ierr); 886785d1243SStefano Zampini ierr = ISDestroy(&pcbddc->DirichletBoundariesLocal);CHKERRQ(ierr); 88763602bcaSStefano Zampini ierr = PCBDDCSetDofsSplitting(pc,0,NULL);CHKERRQ(ierr); 88863602bcaSStefano Zampini ierr = PCBDDCSetDofsSplittingLocal(pc,0,NULL);CHKERRQ(ierr); 889674ae819SStefano Zampini PetscFunctionReturn(0); 890674ae819SStefano Zampini } 891674ae819SStefano Zampini 892674ae819SStefano Zampini #undef __FUNCT__ 893674ae819SStefano Zampini #define __FUNCT__ "PCBDDCResetTopography" 894674ae819SStefano Zampini PetscErrorCode PCBDDCResetTopography(PC pc) 895674ae819SStefano Zampini { 896674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 897*4f1b2e48SStefano Zampini PetscInt i; 898674ae819SStefano Zampini PetscErrorCode ierr; 899674ae819SStefano Zampini 900674ae819SStefano Zampini PetscFunctionBegin; 901b9b85e73SStefano Zampini ierr = MatDestroy(&pcbddc->user_ChangeOfBasisMatrix);CHKERRQ(ierr); 902674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 903674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 904674ae819SStefano Zampini ierr = PCBDDCGraphReset(pcbddc->mat_graph);CHKERRQ(ierr); 905*4f1b2e48SStefano Zampini for (i=0;i<pcbddc->n_local_subs;i++) { 906*4f1b2e48SStefano Zampini ierr = ISDestroy(&pcbddc->local_subs[i]);CHKERRQ(ierr); 907*4f1b2e48SStefano Zampini } 908*4f1b2e48SStefano Zampini ierr = PetscFree(pcbddc->local_subs);CHKERRQ(ierr); 909b96c3477SStefano Zampini ierr = PCBDDCSubSchursReset(pcbddc->sub_schurs);CHKERRQ(ierr); 910674ae819SStefano Zampini PetscFunctionReturn(0); 911674ae819SStefano Zampini } 912674ae819SStefano Zampini 913674ae819SStefano Zampini #undef __FUNCT__ 914674ae819SStefano Zampini #define __FUNCT__ "PCBDDCResetSolvers" 915674ae819SStefano Zampini PetscErrorCode PCBDDCResetSolvers(PC pc) 916674ae819SStefano Zampini { 917674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 91806656605SStefano Zampini PetscScalar *array; 919674ae819SStefano Zampini PetscErrorCode ierr; 920674ae819SStefano Zampini 921674ae819SStefano Zampini PetscFunctionBegin; 922674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->coarse_vec);CHKERRQ(ierr); 92358da7f69SStefano Zampini if (pcbddc->coarse_phi_B) { 92406656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&array);CHKERRQ(ierr); 92506656605SStefano Zampini ierr = PetscFree(array);CHKERRQ(ierr); 92658da7f69SStefano Zampini } 927674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_B);CHKERRQ(ierr); 928674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_D);CHKERRQ(ierr); 92915aaf578SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_B);CHKERRQ(ierr); 93015aaf578SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_D);CHKERRQ(ierr); 931674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec1_P);CHKERRQ(ierr); 932674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec1_C);CHKERRQ(ierr); 933674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat2);CHKERRQ(ierr); 93406656605SStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat1);CHKERRQ(ierr); 935674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec1_R);CHKERRQ(ierr); 936674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec2_R);CHKERRQ(ierr); 9378ce42a96SStefano Zampini ierr = ISDestroy(&pcbddc->is_R_local);CHKERRQ(ierr); 938674ae819SStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_B);CHKERRQ(ierr); 939674ae819SStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_D);CHKERRQ(ierr); 940674ae819SStefano Zampini ierr = VecScatterDestroy(&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 941f4ddd8eeSStefano Zampini ierr = KSPDestroy(&pcbddc->ksp_D);CHKERRQ(ierr); 942f4ddd8eeSStefano Zampini ierr = KSPDestroy(&pcbddc->ksp_R);CHKERRQ(ierr); 943f4ddd8eeSStefano Zampini ierr = KSPDestroy(&pcbddc->coarse_ksp);CHKERRQ(ierr); 944f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 945727cdba6SStefano Zampini ierr = PetscFree(pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 9460e6343abSStefano Zampini ierr = PetscFree2(pcbddc->local_primal_ref_node,pcbddc->local_primal_ref_mult);CHKERRQ(ierr); 947f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->global_primal_indices);CHKERRQ(ierr); 94870cf5478SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling);CHKERRQ(ierr); 9496e683305SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 95081d14e9dSStefano Zampini ierr = MatDestroy(&pcbddc->benign_change);CHKERRQ(ierr); 9510369aaf7SStefano Zampini ierr = VecDestroy(&pcbddc->benign_vec);CHKERRQ(ierr); 9528b9f24d4SStefano Zampini ierr = MatDestroy(&pcbddc->benign_original_mat);CHKERRQ(ierr); 953*4f1b2e48SStefano Zampini ierr = MatDestroy(&pcbddc->benign_original_mat);CHKERRQ(ierr); 954*4f1b2e48SStefano Zampini ierr = MatDestroy(&pcbddc->benign_B0);CHKERRQ(ierr); 9558b9f24d4SStefano Zampini ierr = PetscSFDestroy(&pcbddc->benign_sf);CHKERRQ(ierr); 956*4f1b2e48SStefano Zampini ierr = PetscFree3(pcbddc->benign_p0_lidx,pcbddc->benign_p0_gidx,pcbddc->benign_p0);CHKERRQ(ierr); 957674ae819SStefano Zampini PetscFunctionReturn(0); 958674ae819SStefano Zampini } 959674ae819SStefano Zampini 960674ae819SStefano Zampini #undef __FUNCT__ 961f4ddd8eeSStefano Zampini #define __FUNCT__ "PCBDDCSetUpLocalWorkVectors" 962f4ddd8eeSStefano Zampini PetscErrorCode PCBDDCSetUpLocalWorkVectors(PC pc) 9636bfb1811SStefano Zampini { 9646bfb1811SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 9656bfb1811SStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 9666bfb1811SStefano Zampini VecType impVecType; 967*4f1b2e48SStefano Zampini PetscInt n_constraints,n_R,old_size; 9686bfb1811SStefano Zampini PetscErrorCode ierr; 9696bfb1811SStefano Zampini 9706bfb1811SStefano Zampini PetscFunctionBegin; 971f4ddd8eeSStefano Zampini if (!pcbddc->ConstraintMatrix) { 972019a44ceSStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"BDDC Constraint matrix has not been created"); 973f4ddd8eeSStefano Zampini } 974e7b262bdSStefano Zampini /* get sizes */ 975*4f1b2e48SStefano Zampini n_constraints = pcbddc->local_primal_size - pcbddc->benign_n - pcbddc->n_vertices; 976b371cd4fSStefano Zampini n_R = pcis->n - pcbddc->n_vertices; 9776bfb1811SStefano Zampini ierr = VecGetType(pcis->vec1_N,&impVecType);CHKERRQ(ierr); 978e7b262bdSStefano Zampini /* local work vectors (try to avoid unneeded work)*/ 979e7b262bdSStefano Zampini /* R nodes */ 980e7b262bdSStefano Zampini old_size = -1; 981e7b262bdSStefano Zampini if (pcbddc->vec1_R) { 982e7b262bdSStefano Zampini ierr = VecGetSize(pcbddc->vec1_R,&old_size);CHKERRQ(ierr); 983e7b262bdSStefano Zampini } 984e7b262bdSStefano Zampini if (n_R != old_size) { 985e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec1_R);CHKERRQ(ierr); 986e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec2_R);CHKERRQ(ierr); 9876bfb1811SStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_R);CHKERRQ(ierr); 9886bfb1811SStefano Zampini ierr = VecSetSizes(pcbddc->vec1_R,PETSC_DECIDE,n_R);CHKERRQ(ierr); 9896bfb1811SStefano Zampini ierr = VecSetType(pcbddc->vec1_R,impVecType);CHKERRQ(ierr); 9906bfb1811SStefano Zampini ierr = VecDuplicate(pcbddc->vec1_R,&pcbddc->vec2_R);CHKERRQ(ierr); 991e7b262bdSStefano Zampini } 992e7b262bdSStefano Zampini /* local primal dofs */ 993e7b262bdSStefano Zampini old_size = -1; 994e7b262bdSStefano Zampini if (pcbddc->vec1_P) { 995e7b262bdSStefano Zampini ierr = VecGetSize(pcbddc->vec1_P,&old_size);CHKERRQ(ierr); 996e7b262bdSStefano Zampini } 997e9189074SStefano Zampini if (pcbddc->local_primal_size != old_size) { 998e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec1_P);CHKERRQ(ierr); 99983b7ccabSStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_P);CHKERRQ(ierr); 1000e9189074SStefano Zampini ierr = VecSetSizes(pcbddc->vec1_P,PETSC_DECIDE,pcbddc->local_primal_size);CHKERRQ(ierr); 10016bfb1811SStefano Zampini ierr = VecSetType(pcbddc->vec1_P,impVecType);CHKERRQ(ierr); 1002e7b262bdSStefano Zampini } 1003e7b262bdSStefano Zampini /* local explicit constraints */ 1004e7b262bdSStefano Zampini old_size = -1; 1005e7b262bdSStefano Zampini if (pcbddc->vec1_C) { 1006e7b262bdSStefano Zampini ierr = VecGetSize(pcbddc->vec1_C,&old_size);CHKERRQ(ierr); 1007e7b262bdSStefano Zampini } 1008e7b262bdSStefano Zampini if (n_constraints && n_constraints != old_size) { 1009e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec1_C);CHKERRQ(ierr); 101083b7ccabSStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_C);CHKERRQ(ierr); 101183b7ccabSStefano Zampini ierr = VecSetSizes(pcbddc->vec1_C,PETSC_DECIDE,n_constraints);CHKERRQ(ierr); 101283b7ccabSStefano Zampini ierr = VecSetType(pcbddc->vec1_C,impVecType);CHKERRQ(ierr); 101383b7ccabSStefano Zampini } 10146bfb1811SStefano Zampini PetscFunctionReturn(0); 10156bfb1811SStefano Zampini } 10166bfb1811SStefano Zampini 10176bfb1811SStefano Zampini #undef __FUNCT__ 101847f4ddc3SStefano Zampini #define __FUNCT__ "PCBDDCSetUpCorrection" 101947f4ddc3SStefano Zampini PetscErrorCode PCBDDCSetUpCorrection(PC pc, PetscScalar **coarse_submat_vals_n) 102088ebb749SStefano Zampini { 102125084f0cSStefano Zampini PetscErrorCode ierr; 102225084f0cSStefano Zampini /* pointers to pcis and pcbddc */ 102388ebb749SStefano Zampini PC_IS* pcis = (PC_IS*)pc->data; 102488ebb749SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 1025d62866d3SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 102625084f0cSStefano Zampini /* submatrices of local problem */ 102780677318SStefano Zampini Mat A_RV,A_VR,A_VV,local_auxmat2_R; 1028019a44ceSStefano Zampini /* submatrices of benign trick */ 1029d16cbb6bSStefano Zampini Mat B0_V = NULL; 103006656605SStefano Zampini /* submatrices of local coarse problem */ 103106656605SStefano Zampini Mat S_VV,S_CV,S_VC,S_CC; 103225084f0cSStefano Zampini /* working matrices */ 103306656605SStefano Zampini Mat C_CR; 103425084f0cSStefano Zampini /* additional working stuff */ 103506656605SStefano Zampini PC pc_R; 1036*4f1b2e48SStefano Zampini Mat F; 103706656605SStefano Zampini PetscBool isLU,isCHOL,isILU; 103806656605SStefano Zampini 103925084f0cSStefano Zampini PetscScalar *coarse_submat_vals; /* TODO: use a PETSc matrix */ 104006656605SStefano Zampini PetscScalar *work; 104106656605SStefano Zampini PetscInt *idx_V_B; 1042*4f1b2e48SStefano Zampini PetscInt n,n_vertices,n_constraints,*p0_lidx_I; 104306656605SStefano Zampini PetscInt i,n_R,n_D,n_B; 1044b9d89cd5SStefano Zampini PetscBool unsymmetric_check; 104545a1bb75SStefano Zampini /* matrix type (vector type propagated downstream from vec1_C and local matrix type) */ 104688ebb749SStefano Zampini MatType impMatType; 104725084f0cSStefano Zampini /* some shortcuts to scalars */ 104806656605SStefano Zampini PetscScalar one=1.0,m_one=-1.0; 104988ebb749SStefano Zampini 105088ebb749SStefano Zampini PetscFunctionBegin; 1051b371cd4fSStefano Zampini n_vertices = pcbddc->n_vertices; 1052*4f1b2e48SStefano Zampini n_constraints = pcbddc->local_primal_size - pcbddc->benign_n - n_vertices; 105388ebb749SStefano Zampini /* Set Non-overlapping dimensions */ 1054b371cd4fSStefano Zampini n_B = pcis->n_B; 1055b371cd4fSStefano Zampini n_D = pcis->n - n_B; 105688ebb749SStefano Zampini n_R = pcis->n - n_vertices; 105788ebb749SStefano Zampini 105888ebb749SStefano Zampini /* Set types for local objects needed by BDDC precondtioner */ 105988ebb749SStefano Zampini impMatType = MATSEQDENSE; 106088ebb749SStefano Zampini 106188ebb749SStefano Zampini /* vertices in boundary numbering */ 1062785e854fSJed Brown ierr = PetscMalloc1(n_vertices,&idx_V_B);CHKERRQ(ierr); 10630e6343abSStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,n_vertices,pcbddc->local_primal_ref_node,&i,idx_V_B);CHKERRQ(ierr); 106488ebb749SStefano Zampini if (i != n_vertices) { 106522d5777bSStefano Zampini SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Error in boundary numbering for BDDC vertices! %d != %d\n",n_vertices,i); 106688ebb749SStefano Zampini } 106788ebb749SStefano Zampini 106806656605SStefano Zampini /* Subdomain contribution (Non-overlapping) to coarse matrix */ 1069019a44ceSStefano Zampini ierr = PetscCalloc1(pcbddc->local_primal_size*pcbddc->local_primal_size,&coarse_submat_vals);CHKERRQ(ierr); 107006656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_vertices,n_vertices,coarse_submat_vals,&S_VV);CHKERRQ(ierr); 107106656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_VV,pcbddc->local_primal_size);CHKERRQ(ierr); 107206656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_constraints,n_vertices,coarse_submat_vals+n_vertices,&S_CV);CHKERRQ(ierr); 107306656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_CV,pcbddc->local_primal_size);CHKERRQ(ierr); 107406656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_vertices,n_constraints,coarse_submat_vals+pcbddc->local_primal_size*n_vertices,&S_VC);CHKERRQ(ierr); 107506656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_VC,pcbddc->local_primal_size);CHKERRQ(ierr); 107606656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_constraints,n_constraints,coarse_submat_vals+(pcbddc->local_primal_size+1)*n_vertices,&S_CC);CHKERRQ(ierr); 107706656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_CC,pcbddc->local_primal_size);CHKERRQ(ierr); 107806656605SStefano Zampini 107906656605SStefano Zampini unsymmetric_check = PETSC_FALSE; 108006656605SStefano Zampini /* allocate workspace */ 108106656605SStefano Zampini n = 0; 108206656605SStefano Zampini if (n_constraints) { 108306656605SStefano Zampini n += n_R*n_constraints; 108406656605SStefano Zampini } 108506656605SStefano Zampini if (n_vertices) { 108606656605SStefano Zampini n = PetscMax(2*n_R*n_vertices,n); 108780677318SStefano Zampini n = PetscMax((n_R+n_B)*n_vertices,n); 108806656605SStefano Zampini } 10893301b35fSStefano Zampini if (!pcbddc->symmetric_primal) { 109006656605SStefano Zampini n = PetscMax(2*n_R*pcbddc->local_primal_size,n); 109106656605SStefano Zampini unsymmetric_check = PETSC_TRUE; 109206656605SStefano Zampini } 109306656605SStefano Zampini ierr = PetscMalloc1(n,&work);CHKERRQ(ierr); 109406656605SStefano Zampini 109506656605SStefano Zampini /* determine if can use MatSolve routines instead of calling KSPSolve on ksp_R */ 109606656605SStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_R);CHKERRQ(ierr); 109706656605SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_R,PCLU,&isLU);CHKERRQ(ierr); 109806656605SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_R,PCILU,&isILU);CHKERRQ(ierr); 109906656605SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_R,PCCHOLESKY,&isCHOL);CHKERRQ(ierr); 110006656605SStefano Zampini if (isLU || isILU || isCHOL) { 110106656605SStefano Zampini ierr = PCFactorGetMatrix(pc_R,&F);CHKERRQ(ierr); 1102d62866d3SStefano Zampini } else if (sub_schurs->reuse_mumps) { 1103d62866d3SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 1104d62866d3SStefano Zampini MatFactorType type; 1105d62866d3SStefano Zampini 11066816873aSStefano Zampini F = reuse_mumps->F; 11076816873aSStefano Zampini ierr = MatGetFactorType(F,&type);CHKERRQ(ierr); 1108d62866d3SStefano Zampini if (type == MAT_FACTOR_CHOLESKY) isCHOL = PETSC_TRUE; 110906656605SStefano Zampini } else { 111006656605SStefano Zampini F = NULL; 111106656605SStefano Zampini } 111206656605SStefano Zampini 111388ebb749SStefano Zampini /* Precompute stuffs needed for preprocessing and application of BDDC*/ 111488ebb749SStefano Zampini if (n_constraints) { 111506656605SStefano Zampini Mat M1,M2,M3; 111680677318SStefano Zampini Mat auxmat; 111706656605SStefano Zampini IS is_aux; 111880677318SStefano Zampini PetscScalar *array,*array2; 111906656605SStefano Zampini 1120f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat1);CHKERRQ(ierr); 112180677318SStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat2);CHKERRQ(ierr); 112288ebb749SStefano Zampini 112325084f0cSStefano Zampini /* Extract constraints on R nodes: C_{CR} */ 112425084f0cSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,n_constraints,n_vertices,1,&is_aux);CHKERRQ(ierr); 11258ce42a96SStefano Zampini ierr = MatGetSubMatrix(pcbddc->ConstraintMatrix,is_aux,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&C_CR);CHKERRQ(ierr); 112680677318SStefano Zampini ierr = MatGetSubMatrix(pcbddc->ConstraintMatrix,is_aux,pcis->is_B_local,MAT_INITIAL_MATRIX,&auxmat);CHKERRQ(ierr); 112788ebb749SStefano Zampini 112880677318SStefano Zampini /* Assemble local_auxmat2_R = (- A_{RR}^{-1} C^T_{CR}) needed by BDDC setup */ 112980677318SStefano Zampini /* Assemble pcbddc->local_auxmat2 = R_to_B (- A_{RR}^{-1} C^T_{CR}) needed by BDDC application */ 113006656605SStefano Zampini ierr = PetscMemzero(work,n_R*n_constraints*sizeof(PetscScalar));CHKERRQ(ierr); 113188ebb749SStefano Zampini for (i=0;i<n_constraints;i++) { 113206656605SStefano Zampini const PetscScalar *row_cmat_values; 113306656605SStefano Zampini const PetscInt *row_cmat_indices; 113406656605SStefano Zampini PetscInt size_of_constraint,j; 113588ebb749SStefano Zampini 113606656605SStefano Zampini ierr = MatGetRow(C_CR,i,&size_of_constraint,&row_cmat_indices,&row_cmat_values);CHKERRQ(ierr); 113706656605SStefano Zampini for (j=0;j<size_of_constraint;j++) { 113806656605SStefano Zampini work[row_cmat_indices[j]+i*n_R] = -row_cmat_values[j]; 113906656605SStefano Zampini } 114006656605SStefano Zampini ierr = MatRestoreRow(C_CR,i,&size_of_constraint,&row_cmat_indices,&row_cmat_values);CHKERRQ(ierr); 114106656605SStefano Zampini } 114280677318SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_constraints,NULL,&local_auxmat2_R);CHKERRQ(ierr); 114306656605SStefano Zampini if (F) { 114406656605SStefano Zampini Mat B; 114506656605SStefano Zampini 114606656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_constraints,work,&B);CHKERRQ(ierr); 114780677318SStefano Zampini ierr = MatMatSolve(F,B,local_auxmat2_R);CHKERRQ(ierr); 114806656605SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 114906656605SStefano Zampini } else { 115080677318SStefano Zampini PetscScalar *marr; 115180677318SStefano Zampini 115280677318SStefano Zampini ierr = MatDenseGetArray(local_auxmat2_R,&marr);CHKERRQ(ierr); 115306656605SStefano Zampini for (i=0;i<n_constraints;i++) { 115406656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*n_R);CHKERRQ(ierr); 115580677318SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,marr+i*n_R);CHKERRQ(ierr); 115606656605SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 115706656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 115806656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 115906656605SStefano Zampini } 116080677318SStefano Zampini ierr = MatDenseRestoreArray(local_auxmat2_R,&marr);CHKERRQ(ierr); 116106656605SStefano Zampini } 116280677318SStefano Zampini if (!pcbddc->switch_static) { 116380677318SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,n_constraints,NULL,&pcbddc->local_auxmat2);CHKERRQ(ierr); 116480677318SStefano Zampini ierr = MatDenseGetArray(pcbddc->local_auxmat2,&array);CHKERRQ(ierr); 116580677318SStefano Zampini ierr = MatDenseGetArray(local_auxmat2_R,&array2);CHKERRQ(ierr); 116680677318SStefano Zampini for (i=0;i<n_constraints;i++) { 116780677318SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,array2+i*n_R);CHKERRQ(ierr); 116880677318SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,array+i*n_B);CHKERRQ(ierr); 116980677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 117080677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 117180677318SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 117280677318SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 117380677318SStefano Zampini } 117480677318SStefano Zampini ierr = MatDenseRestoreArray(local_auxmat2_R,&array2);CHKERRQ(ierr); 117580677318SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->local_auxmat2,&array);CHKERRQ(ierr); 117680677318SStefano Zampini ierr = MatMatMult(auxmat,pcbddc->local_auxmat2,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&M3);CHKERRQ(ierr); 117780677318SStefano Zampini } else { 117880677318SStefano Zampini ierr = PetscObjectReference((PetscObject)local_auxmat2_R);CHKERRQ(ierr); 117980677318SStefano Zampini pcbddc->local_auxmat2 = local_auxmat2_R; 118025084f0cSStefano Zampini ierr = MatMatMult(C_CR,pcbddc->local_auxmat2,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&M3);CHKERRQ(ierr); 118180677318SStefano Zampini } 118280677318SStefano Zampini ierr = ISDestroy(&is_aux);CHKERRQ(ierr); 118380677318SStefano Zampini /* Assemble explicitly S_CC = ( C_{CR} A_{RR}^{-1} C^T_{CR} )^{-1} */ 118480677318SStefano Zampini ierr = MatScale(M3,m_one);CHKERRQ(ierr); 118506656605SStefano Zampini ierr = MatDuplicate(M3,MAT_DO_NOT_COPY_VALUES,&M1);CHKERRQ(ierr); 118606656605SStefano Zampini ierr = MatDuplicate(M3,MAT_DO_NOT_COPY_VALUES,&M2);CHKERRQ(ierr); 118780677318SStefano Zampini if (isCHOL) { 118880677318SStefano Zampini ierr = MatCholeskyFactor(M3,NULL,NULL);CHKERRQ(ierr); 118980677318SStefano Zampini } else { 119025084f0cSStefano Zampini ierr = MatLUFactor(M3,NULL,NULL,NULL);CHKERRQ(ierr); 119180677318SStefano Zampini } 119280677318SStefano Zampini ierr = VecSet(pcbddc->vec1_C,one);CHKERRQ(ierr); 119306656605SStefano Zampini ierr = MatDiagonalSet(M2,pcbddc->vec1_C,INSERT_VALUES);CHKERRQ(ierr); 119425084f0cSStefano Zampini ierr = MatMatSolve(M3,M2,M1);CHKERRQ(ierr); 119525084f0cSStefano Zampini ierr = MatDestroy(&M2);CHKERRQ(ierr); 119625084f0cSStefano Zampini ierr = MatDestroy(&M3);CHKERRQ(ierr); 119780677318SStefano Zampini /* Assemble local_auxmat1 = S_CC*C_{CB} needed by BDDC application in KSP and in preproc */ 119880677318SStefano Zampini ierr = MatMatMult(M1,auxmat,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&pcbddc->local_auxmat1);CHKERRQ(ierr); 119980677318SStefano Zampini ierr = MatDestroy(&auxmat);CHKERRQ(ierr); 120006656605SStefano Zampini ierr = MatCopy(M1,S_CC,SAME_NONZERO_PATTERN);CHKERRQ(ierr); /* S_CC can have a different LDA, MatMatSolve doesn't support it */ 120106656605SStefano Zampini ierr = MatDestroy(&M1);CHKERRQ(ierr); 1202f4ddd8eeSStefano Zampini } 120388ebb749SStefano Zampini /* Get submatrices from subdomain matrix */ 1204*4f1b2e48SStefano Zampini if (pcbddc->benign_n) { 1205d16cbb6bSStefano Zampini IS dummy; 1206d16cbb6bSStefano Zampini Mat B0_R; 1207d16cbb6bSStefano Zampini PetscReal norm; 1208d16cbb6bSStefano Zampini 1209*4f1b2e48SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcbddc->benign_n,0,1,&dummy);CHKERRQ(ierr); 1210*4f1b2e48SStefano Zampini ierr = MatGetSubMatrix(pcbddc->benign_B0,dummy,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&B0_R);CHKERRQ(ierr); 1211d16cbb6bSStefano Zampini ierr = MatNorm(B0_R,NORM_INFINITY,&norm);CHKERRQ(ierr); 1212d16cbb6bSStefano Zampini if (norm > PETSC_SMALL) { 1213d16cbb6bSStefano Zampini SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_SUP,"Benign trick can not be applied! ||B0_R|| = %f (should be numerically 0.)",norm); 1214d16cbb6bSStefano Zampini } 1215d16cbb6bSStefano Zampini ierr = MatDestroy(&B0_R);CHKERRQ(ierr); 1216d16cbb6bSStefano Zampini ierr = ISDestroy(&dummy);CHKERRQ(ierr); 1217d16cbb6bSStefano Zampini } 1218d16cbb6bSStefano Zampini 121988ebb749SStefano Zampini if (n_vertices) { 122006656605SStefano Zampini IS is_aux; 12213a50541eSStefano Zampini 12226816873aSStefano Zampini if (sub_schurs->reuse_mumps) { /* is_R_local is not sorted, ISComplement doesn't like it */ 12236816873aSStefano Zampini IS tis; 12246816873aSStefano Zampini 12256816873aSStefano Zampini ierr = ISDuplicate(pcbddc->is_R_local,&tis);CHKERRQ(ierr); 12266816873aSStefano Zampini ierr = ISSort(tis);CHKERRQ(ierr); 12276816873aSStefano Zampini ierr = ISComplement(tis,0,pcis->n,&is_aux);CHKERRQ(ierr); 12286816873aSStefano Zampini ierr = ISDestroy(&tis);CHKERRQ(ierr); 12296816873aSStefano Zampini } else { 12303a50541eSStefano Zampini ierr = ISComplement(pcbddc->is_R_local,0,pcis->n,&is_aux);CHKERRQ(ierr); 12316816873aSStefano Zampini } 12329577ea80SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,pcbddc->is_R_local,is_aux,MAT_INITIAL_MATRIX,&A_RV);CHKERRQ(ierr); 12339577ea80SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,is_aux,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&A_VR);CHKERRQ(ierr); 123404708bb6SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,is_aux,is_aux,MAT_INITIAL_MATRIX,&A_VV);CHKERRQ(ierr); 1235*4f1b2e48SStefano Zampini if (pcbddc->benign_n) { 1236019a44ceSStefano Zampini IS dummy; 1237019a44ceSStefano Zampini 1238*4f1b2e48SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcbddc->benign_n,0,1,&dummy);CHKERRQ(ierr); 1239*4f1b2e48SStefano Zampini ierr = MatGetSubMatrix(pcbddc->benign_B0,dummy,is_aux,MAT_INITIAL_MATRIX,&B0_V);CHKERRQ(ierr); 1240019a44ceSStefano Zampini ierr = ISDestroy(&dummy);CHKERRQ(ierr); 1241019a44ceSStefano Zampini } 124225084f0cSStefano Zampini ierr = ISDestroy(&is_aux);CHKERRQ(ierr); 124388ebb749SStefano Zampini } 124488ebb749SStefano Zampini 124588ebb749SStefano Zampini /* Matrix of coarse basis functions (local) */ 1246f4ddd8eeSStefano Zampini if (pcbddc->coarse_phi_B) { 124706656605SStefano Zampini PetscInt on_B,on_primal,on_D=n_D; 124806656605SStefano Zampini if (pcbddc->coarse_phi_D) { 124906656605SStefano Zampini ierr = MatGetSize(pcbddc->coarse_phi_D,&on_D,NULL);CHKERRQ(ierr); 125006656605SStefano Zampini } 1251f4ddd8eeSStefano Zampini ierr = MatGetSize(pcbddc->coarse_phi_B,&on_B,&on_primal);CHKERRQ(ierr); 125206656605SStefano Zampini if (on_B != n_B || on_primal != pcbddc->local_primal_size || on_D != n_D) { 125306656605SStefano Zampini PetscScalar *marray; 125406656605SStefano Zampini 125506656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&marray);CHKERRQ(ierr); 125606656605SStefano Zampini ierr = PetscFree(marray);CHKERRQ(ierr); 1257f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_B);CHKERRQ(ierr); 1258f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_B);CHKERRQ(ierr); 1259f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_D);CHKERRQ(ierr); 1260f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_D);CHKERRQ(ierr); 1261f4ddd8eeSStefano Zampini } 1262f4ddd8eeSStefano Zampini } 126306656605SStefano Zampini 1264f4ddd8eeSStefano Zampini if (!pcbddc->coarse_phi_B) { 126506656605SStefano Zampini PetscScalar *marray; 126688ebb749SStefano Zampini 126706656605SStefano Zampini n = n_B*pcbddc->local_primal_size; 12688eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 126906656605SStefano Zampini n += n_D*pcbddc->local_primal_size; 127088ebb749SStefano Zampini } 12713301b35fSStefano Zampini if (!pcbddc->symmetric_primal) { 127206656605SStefano Zampini n *= 2; 127388ebb749SStefano Zampini } 127406656605SStefano Zampini ierr = PetscCalloc1(n,&marray);CHKERRQ(ierr); 127506656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,pcbddc->local_primal_size,marray,&pcbddc->coarse_phi_B);CHKERRQ(ierr); 127606656605SStefano Zampini n = n_B*pcbddc->local_primal_size; 12778eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 127806656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_D,pcbddc->local_primal_size,marray+n,&pcbddc->coarse_phi_D);CHKERRQ(ierr); 127906656605SStefano Zampini n += n_D*pcbddc->local_primal_size; 128088ebb749SStefano Zampini } 12813301b35fSStefano Zampini if (!pcbddc->symmetric_primal) { 128206656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,pcbddc->local_primal_size,marray+n,&pcbddc->coarse_psi_B);CHKERRQ(ierr); 12838eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 128406656605SStefano Zampini n = n_B*pcbddc->local_primal_size; 128506656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_D,pcbddc->local_primal_size,marray+n,&pcbddc->coarse_psi_D);CHKERRQ(ierr); 128688ebb749SStefano Zampini } 128788ebb749SStefano Zampini } else { 1288c0553b1fSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->coarse_phi_B);CHKERRQ(ierr); 1289c0553b1fSStefano Zampini pcbddc->coarse_psi_B = pcbddc->coarse_phi_B; 12901b968477SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 1291c0553b1fSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->coarse_phi_D);CHKERRQ(ierr); 1292c0553b1fSStefano Zampini pcbddc->coarse_psi_D = pcbddc->coarse_phi_D; 1293c0553b1fSStefano Zampini } 129488ebb749SStefano Zampini } 129506656605SStefano Zampini } 1296019a44ceSStefano Zampini 129706656605SStefano Zampini /* We are now ready to evaluate coarse basis functions and subdomain contribution to coarse problem */ 1298*4f1b2e48SStefano Zampini p0_lidx_I = NULL; 1299*4f1b2e48SStefano Zampini if (pcbddc->benign_n && (pcbddc->switch_static || pcbddc->dbg_flag)) { 1300d12edf2fSStefano Zampini const PetscInt *idxs; 1301d12edf2fSStefano Zampini 1302d12edf2fSStefano Zampini ierr = ISGetIndices(pcis->is_I_local,&idxs);CHKERRQ(ierr); 1303*4f1b2e48SStefano Zampini ierr = PetscMalloc1(pcbddc->benign_n,&p0_lidx_I);CHKERRQ(ierr); 1304*4f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 1305*4f1b2e48SStefano Zampini ierr = PetscFindInt(pcbddc->benign_p0_lidx[i],pcis->n-pcis->n_B,idxs,&p0_lidx_I[i]);CHKERRQ(ierr); 1306*4f1b2e48SStefano Zampini } 1307d12edf2fSStefano Zampini ierr = ISRestoreIndices(pcis->is_I_local,&idxs);CHKERRQ(ierr); 1308d12edf2fSStefano Zampini } 1309d16cbb6bSStefano Zampini 131006656605SStefano Zampini /* vertices */ 131106656605SStefano Zampini if (n_vertices) { 131216f15bc4SStefano Zampini 131304708bb6SStefano Zampini ierr = MatConvert(A_VV,impMatType,MAT_REUSE_MATRIX,&A_VV);CHKERRQ(ierr); 131404708bb6SStefano Zampini 131516f15bc4SStefano Zampini if (n_R) { 131606656605SStefano Zampini Mat A_RRmA_RV,S_VVt; /* S_VVt with LDA=N */ 131706656605SStefano Zampini PetscBLASInt B_N,B_one = 1; 131816f15bc4SStefano Zampini PetscScalar *x,*y; 131904708bb6SStefano Zampini PetscBool isseqaij; 132006656605SStefano Zampini 132121eccb56SStefano Zampini ierr = MatScale(A_RV,m_one);CHKERRQ(ierr); 132206656605SStefano Zampini ierr = MatConvert(A_RV,impMatType,MAT_REUSE_MATRIX,&A_RV);CHKERRQ(ierr); 132304708bb6SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_vertices,work,&A_RRmA_RV);CHKERRQ(ierr); 13246816873aSStefano Zampini if (F) { /* TODO could be optimized for symmetric problems */ 132506656605SStefano Zampini ierr = MatMatSolve(F,A_RV,A_RRmA_RV);CHKERRQ(ierr); 132606656605SStefano Zampini } else { 132706656605SStefano Zampini ierr = MatDenseGetArray(A_RV,&y);CHKERRQ(ierr); 132806656605SStefano Zampini for (i=0;i<n_vertices;i++) { 132906656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,y+i*n_R);CHKERRQ(ierr); 133006656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,work+i*n_R);CHKERRQ(ierr); 133106656605SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 133206656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 133306656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 133406656605SStefano Zampini } 133506656605SStefano Zampini ierr = MatDenseRestoreArray(A_RV,&y);CHKERRQ(ierr); 133606656605SStefano Zampini } 133780677318SStefano Zampini ierr = MatDestroy(&A_RV);CHKERRQ(ierr); 133806656605SStefano Zampini /* S_VV and S_CV are the subdomain contribution to coarse matrix. WARNING -> column major ordering */ 133906656605SStefano Zampini if (n_constraints) { 134006656605SStefano Zampini Mat B; 134180677318SStefano Zampini 1342b3d85658SStefano Zampini ierr = PetscMemzero(work+n_R*n_vertices,n_B*n_vertices*sizeof(PetscScalar));CHKERRQ(ierr); 134380677318SStefano Zampini for (i=0;i<n_vertices;i++) { 134480677318SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*n_R);CHKERRQ(ierr); 134580677318SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,work+n_R*n_vertices+i*n_B);CHKERRQ(ierr); 134680677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 134780677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 134880677318SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 134980677318SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 135080677318SStefano Zampini } 135180677318SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,n_vertices,work+n_R*n_vertices,&B);CHKERRQ(ierr); 135280677318SStefano Zampini ierr = MatMatMult(pcbddc->local_auxmat1,B,MAT_REUSE_MATRIX,PETSC_DEFAULT,&S_CV);CHKERRQ(ierr); 135380677318SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 135406656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_vertices,work+n_R*n_vertices,&B);CHKERRQ(ierr); 135580677318SStefano Zampini ierr = MatMatMult(local_auxmat2_R,S_CV,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B);CHKERRQ(ierr); 135606656605SStefano Zampini ierr = MatScale(S_CV,m_one);CHKERRQ(ierr); 135706656605SStefano Zampini ierr = PetscBLASIntCast(n_R*n_vertices,&B_N);CHKERRQ(ierr); 135806656605SStefano Zampini PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&B_N,&one,work+n_R*n_vertices,&B_one,work,&B_one)); 135906656605SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 136006656605SStefano Zampini } 136104708bb6SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)A_VR,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 136204708bb6SStefano Zampini if (!isseqaij) { /* MatMatMult with SEQ(S)BAIJ below will raise an error */ 136304708bb6SStefano Zampini ierr = MatConvert(A_VR,MATSEQAIJ,MAT_REUSE_MATRIX,&A_VR);CHKERRQ(ierr); 136404708bb6SStefano Zampini } 136506656605SStefano Zampini ierr = MatMatMult(A_VR,A_RRmA_RV,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&S_VVt);CHKERRQ(ierr); 136680677318SStefano Zampini ierr = MatDestroy(&A_RRmA_RV);CHKERRQ(ierr); 136706656605SStefano Zampini ierr = PetscBLASIntCast(n_vertices*n_vertices,&B_N);CHKERRQ(ierr); 136806656605SStefano Zampini ierr = MatDenseGetArray(A_VV,&x);CHKERRQ(ierr); 136906656605SStefano Zampini ierr = MatDenseGetArray(S_VVt,&y);CHKERRQ(ierr); 137006656605SStefano Zampini PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&B_N,&one,x,&B_one,y,&B_one)); 137106656605SStefano Zampini ierr = MatDenseRestoreArray(A_VV,&x);CHKERRQ(ierr); 137206656605SStefano Zampini ierr = MatDenseRestoreArray(S_VVt,&y);CHKERRQ(ierr); 137306656605SStefano Zampini ierr = MatCopy(S_VVt,S_VV,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 1374d16cbb6bSStefano Zampini ierr = MatDestroy(&S_VVt);CHKERRQ(ierr); 1375019a44ceSStefano Zampini } else { 1376d16cbb6bSStefano Zampini ierr = MatCopy(A_VV,S_VV,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 1377d16cbb6bSStefano Zampini } 1378*4f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 1379019a44ceSStefano Zampini const PetscScalar *vals; 1380019a44ceSStefano Zampini const PetscInt *idxs; 1381*4f1b2e48SStefano Zampini PetscInt n,j,primal_idx; 1382019a44ceSStefano Zampini 1383*4f1b2e48SStefano Zampini ierr = MatGetRow(B0_V,i,&n,&idxs,&vals);CHKERRQ(ierr); 1384*4f1b2e48SStefano Zampini primal_idx = pcbddc->local_primal_size - pcbddc->benign_n + i; 1385d16cbb6bSStefano Zampini for (j=0;j<n;j++) { 1386*4f1b2e48SStefano Zampini coarse_submat_vals[primal_idx*pcbddc->local_primal_size+idxs[j]] = vals[j]; 1387*4f1b2e48SStefano Zampini coarse_submat_vals[idxs[j]*pcbddc->local_primal_size+primal_idx] = vals[j]; 1388019a44ceSStefano Zampini } 1389*4f1b2e48SStefano Zampini ierr = MatRestoreRow(B0_V,i,&n,&idxs,&vals);CHKERRQ(ierr); 139016f15bc4SStefano Zampini } 139121eccb56SStefano Zampini ierr = MatDestroy(&A_VV);CHKERRQ(ierr); 1392d16cbb6bSStefano Zampini 139306656605SStefano Zampini /* coarse basis functions */ 139406656605SStefano Zampini for (i=0;i<n_vertices;i++) { 139516f15bc4SStefano Zampini PetscScalar *y; 139616f15bc4SStefano Zampini 139706656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+n_R*i);CHKERRQ(ierr); 139806656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 139906656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*i);CHKERRQ(ierr); 140006656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 140106656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 140206656605SStefano Zampini y[n_B*i+idx_V_B[i]] = 1.0; 140306656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 140406656605SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 140506656605SStefano Zampini 140606656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 1407*4f1b2e48SStefano Zampini PetscInt j; 1408*4f1b2e48SStefano Zampini 140906656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 141006656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*i);CHKERRQ(ierr); 141106656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 141206656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 141306656605SStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 1414*4f1b2e48SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) y[n_D*i+p0_lidx_I[j]] = 0.0; 141506656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 141606656605SStefano Zampini } 141706656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 141806656605SStefano Zampini } 141904708bb6SStefano Zampini /* if n_R == 0 the object is not destroyed */ 142004708bb6SStefano Zampini ierr = MatDestroy(&A_RV);CHKERRQ(ierr); 142106656605SStefano Zampini } 142206656605SStefano Zampini 142306656605SStefano Zampini if (n_constraints) { 142406656605SStefano Zampini Mat B; 142506656605SStefano Zampini 142606656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_constraints,work,&B);CHKERRQ(ierr); 142706656605SStefano Zampini ierr = MatScale(S_CC,m_one);CHKERRQ(ierr); 142880677318SStefano Zampini ierr = MatMatMult(local_auxmat2_R,S_CC,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B);CHKERRQ(ierr); 142906656605SStefano Zampini ierr = MatScale(S_CC,m_one);CHKERRQ(ierr); 143006656605SStefano Zampini if (n_vertices) { 143180677318SStefano Zampini if (isCHOL) { /* if we can solve the interior problem with cholesky, we should also be fine with transposing here */ 143280677318SStefano Zampini ierr = MatTranspose(S_CV,MAT_REUSE_MATRIX,&S_VC);CHKERRQ(ierr); 143380677318SStefano Zampini } else { 143480677318SStefano Zampini Mat S_VCt; 143580677318SStefano Zampini 143680677318SStefano Zampini ierr = MatMatMult(A_VR,B,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&S_VCt);CHKERRQ(ierr); 143780677318SStefano Zampini ierr = MatCopy(S_VCt,S_VC,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 143880677318SStefano Zampini ierr = MatDestroy(&S_VCt);CHKERRQ(ierr); 143980677318SStefano Zampini } 144006656605SStefano Zampini } 144106656605SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 144206656605SStefano Zampini /* coarse basis functions */ 144306656605SStefano Zampini for (i=0;i<n_constraints;i++) { 144406656605SStefano Zampini PetscScalar *y; 144506656605SStefano Zampini 144606656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+n_R*i);CHKERRQ(ierr); 144706656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 144806656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*(i+n_vertices));CHKERRQ(ierr); 144906656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 145006656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 145106656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 145206656605SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 145306656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 1454*4f1b2e48SStefano Zampini PetscInt j; 1455*4f1b2e48SStefano Zampini 145606656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 145706656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*(i+n_vertices));CHKERRQ(ierr); 145806656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 145906656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 146006656605SStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 1461*4f1b2e48SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) y[n_D*i+p0_lidx_I[j]] = 0.0; 146206656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 146306656605SStefano Zampini } 146406656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 146506656605SStefano Zampini } 146606656605SStefano Zampini } 146780677318SStefano Zampini if (n_constraints) { 146880677318SStefano Zampini ierr = MatDestroy(&local_auxmat2_R);CHKERRQ(ierr); 146980677318SStefano Zampini } 1470*4f1b2e48SStefano Zampini ierr = PetscFree(p0_lidx_I);CHKERRQ(ierr); 1471019a44ceSStefano Zampini ierr = MatDestroy(&B0_V);CHKERRQ(ierr); 1472019a44ceSStefano Zampini 147306656605SStefano Zampini /* compute other basis functions for non-symmetric problems */ 14743301b35fSStefano Zampini if (!pcbddc->symmetric_primal) { 147506656605SStefano Zampini 147606656605SStefano Zampini if (n_constraints) { 147716f15bc4SStefano Zampini Mat S_CCT,B_C; 147806656605SStefano Zampini 147980677318SStefano Zampini /* this is a lazy thing */ 148080677318SStefano Zampini ierr = MatConvert(C_CR,impMatType,MAT_REUSE_MATRIX,&C_CR);CHKERRQ(ierr); 148106656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_constraints,work+n_vertices*n_R,&B_C);CHKERRQ(ierr); 148206656605SStefano Zampini ierr = MatTranspose(S_CC,MAT_INITIAL_MATRIX,&S_CCT);CHKERRQ(ierr); 148306656605SStefano Zampini ierr = MatTransposeMatMult(C_CR,S_CCT,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B_C);CHKERRQ(ierr); 148416f15bc4SStefano Zampini ierr = MatDestroy(&S_CCT);CHKERRQ(ierr); 148506656605SStefano Zampini if (n_vertices) { 148616f15bc4SStefano Zampini Mat B_V,S_VCT; 148706656605SStefano Zampini 148806656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_vertices,work,&B_V);CHKERRQ(ierr); 148906656605SStefano Zampini ierr = MatTranspose(S_VC,MAT_INITIAL_MATRIX,&S_VCT);CHKERRQ(ierr); 149006656605SStefano Zampini ierr = MatTransposeMatMult(C_CR,S_VCT,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B_V);CHKERRQ(ierr); 149106656605SStefano Zampini ierr = MatDestroy(&B_V);CHKERRQ(ierr); 149216f15bc4SStefano Zampini ierr = MatDestroy(&S_VCT);CHKERRQ(ierr); 149306656605SStefano Zampini } 149406656605SStefano Zampini ierr = MatDestroy(&B_C);CHKERRQ(ierr); 149580677318SStefano Zampini } else { /* if there are no constraints, reset work */ 149680677318SStefano Zampini ierr = PetscMemzero(work,n_R*pcbddc->local_primal_size*sizeof(PetscScalar));CHKERRQ(ierr); 149706656605SStefano Zampini } 149816f15bc4SStefano Zampini if (n_vertices && n_R) { 149906656605SStefano Zampini Mat A_VRT; 150080677318SStefano Zampini PetscScalar *marray; 150106656605SStefano Zampini PetscBLASInt B_N,B_one = 1; 150206656605SStefano Zampini 150380677318SStefano Zampini ierr = MatTranspose(A_VR,MAT_INITIAL_MATRIX,&A_VRT);CHKERRQ(ierr); 150480677318SStefano Zampini ierr = MatConvert(A_VRT,impMatType,MAT_REUSE_MATRIX,&A_VRT);CHKERRQ(ierr); 150580677318SStefano Zampini ierr = MatDenseGetArray(A_VRT,&marray);CHKERRQ(ierr); 150606656605SStefano Zampini ierr = PetscBLASIntCast(n_vertices*n_R,&B_N);CHKERRQ(ierr); 150780677318SStefano Zampini PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&B_N,&m_one,marray,&B_one,work,&B_one)); 150880677318SStefano Zampini ierr = MatDenseRestoreArray(A_VRT,&marray);CHKERRQ(ierr); 150916f15bc4SStefano Zampini ierr = MatDestroy(&A_VRT);CHKERRQ(ierr); 151006656605SStefano Zampini } 151106656605SStefano Zampini 151206656605SStefano Zampini if (F) { /* currently there's no support for MatTransposeMatSolve(F,B,X) */ 151306656605SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 151406656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*n_R);CHKERRQ(ierr); 151506656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,work+(i+pcbddc->local_primal_size)*n_R);CHKERRQ(ierr); 151606656605SStefano Zampini ierr = MatSolveTranspose(F,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 151706656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 151806656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 151906656605SStefano Zampini } 152006656605SStefano Zampini } else { 152106656605SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 152206656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*n_R);CHKERRQ(ierr); 152306656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,work+(i+pcbddc->local_primal_size)*n_R);CHKERRQ(ierr); 152406656605SStefano Zampini ierr = KSPSolveTranspose(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 152506656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 152606656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 152706656605SStefano Zampini } 152806656605SStefano Zampini } 152906656605SStefano Zampini /* coarse basis functions */ 153006656605SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 153106656605SStefano Zampini PetscScalar *y; 153206656605SStefano Zampini 153306656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+n_R*(i+pcbddc->local_primal_size));CHKERRQ(ierr); 153406656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_psi_B,&y);CHKERRQ(ierr); 153506656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*i);CHKERRQ(ierr); 153606656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 153706656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 153806656605SStefano Zampini if (i<n_vertices) { 153906656605SStefano Zampini y[n_B*i+idx_V_B[i]] = 1.0; 154006656605SStefano Zampini } 154106656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_psi_B,&y);CHKERRQ(ierr); 154206656605SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 154306656605SStefano Zampini 154406656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 154506656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_psi_D,&y);CHKERRQ(ierr); 154606656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*i);CHKERRQ(ierr); 154706656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 154806656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 154906656605SStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 155006656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_psi_D,&y);CHKERRQ(ierr); 155106656605SStefano Zampini } 155206656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 155306656605SStefano Zampini } 155406656605SStefano Zampini } 1555d62866d3SStefano Zampini /* free memory */ 155688ebb749SStefano Zampini ierr = PetscFree(idx_V_B);CHKERRQ(ierr); 155706656605SStefano Zampini ierr = MatDestroy(&S_VV);CHKERRQ(ierr); 155806656605SStefano Zampini ierr = MatDestroy(&S_CV);CHKERRQ(ierr); 155906656605SStefano Zampini ierr = MatDestroy(&S_VC);CHKERRQ(ierr); 156006656605SStefano Zampini ierr = MatDestroy(&S_CC);CHKERRQ(ierr); 1561d62866d3SStefano Zampini ierr = PetscFree(work);CHKERRQ(ierr); 1562d62866d3SStefano Zampini if (n_vertices) { 1563d62866d3SStefano Zampini ierr = MatDestroy(&A_VR);CHKERRQ(ierr); 1564d62866d3SStefano Zampini } 1565d62866d3SStefano Zampini if (n_constraints) { 1566d62866d3SStefano Zampini ierr = MatDestroy(&C_CR);CHKERRQ(ierr); 1567d62866d3SStefano Zampini } 156888ebb749SStefano Zampini /* Checking coarse_sub_mat and coarse basis functios */ 156988ebb749SStefano Zampini /* Symmetric case : It should be \Phi^{(j)^T} A^{(j)} \Phi^{(j)}=coarse_sub_mat */ 157088ebb749SStefano Zampini /* Non-symmetric case : It should be \Psi^{(j)^T} A^{(j)} \Phi^{(j)}=coarse_sub_mat */ 1571d12edf2fSStefano Zampini if (pcbddc->dbg_flag) { 157288ebb749SStefano Zampini Mat coarse_sub_mat; 157325084f0cSStefano Zampini Mat AUXMAT,TM1,TM2,TM3,TM4; 157488ebb749SStefano Zampini Mat coarse_phi_D,coarse_phi_B; 157588ebb749SStefano Zampini Mat coarse_psi_D,coarse_psi_B; 157688ebb749SStefano Zampini Mat A_II,A_BB,A_IB,A_BI; 15778bec7fa6SStefano Zampini Mat C_B,CPHI; 15788bec7fa6SStefano Zampini IS is_dummy; 15798bec7fa6SStefano Zampini Vec mones; 158088ebb749SStefano Zampini MatType checkmattype=MATSEQAIJ; 158188ebb749SStefano Zampini PetscReal real_value; 158288ebb749SStefano Zampini 158388ebb749SStefano Zampini ierr = MatConvert(pcis->A_II,checkmattype,MAT_INITIAL_MATRIX,&A_II);CHKERRQ(ierr); 158488ebb749SStefano Zampini ierr = MatConvert(pcis->A_IB,checkmattype,MAT_INITIAL_MATRIX,&A_IB);CHKERRQ(ierr); 158588ebb749SStefano Zampini ierr = MatConvert(pcis->A_BI,checkmattype,MAT_INITIAL_MATRIX,&A_BI);CHKERRQ(ierr); 158688ebb749SStefano Zampini ierr = MatConvert(pcis->A_BB,checkmattype,MAT_INITIAL_MATRIX,&A_BB);CHKERRQ(ierr); 158788ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_phi_D,checkmattype,MAT_INITIAL_MATRIX,&coarse_phi_D);CHKERRQ(ierr); 158888ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_phi_B,checkmattype,MAT_INITIAL_MATRIX,&coarse_phi_B);CHKERRQ(ierr); 1589c0553b1fSStefano Zampini if (unsymmetric_check) { 159088ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_psi_D,checkmattype,MAT_INITIAL_MATRIX,&coarse_psi_D);CHKERRQ(ierr); 159188ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_psi_B,checkmattype,MAT_INITIAL_MATRIX,&coarse_psi_B);CHKERRQ(ierr); 159288ebb749SStefano Zampini } 159388ebb749SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,pcbddc->local_primal_size,pcbddc->local_primal_size,coarse_submat_vals,&coarse_sub_mat);CHKERRQ(ierr); 159488ebb749SStefano Zampini 159525084f0cSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 15963301b35fSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Check coarse sub mat computation (symmetric %d)\n",pcbddc->symmetric_primal);CHKERRQ(ierr); 159725084f0cSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1598c0553b1fSStefano Zampini if (unsymmetric_check) { 159988ebb749SStefano Zampini ierr = MatMatMult(A_II,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 160088ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM1);CHKERRQ(ierr); 160188ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 160288ebb749SStefano Zampini ierr = MatMatMult(A_BB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 160388ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM2);CHKERRQ(ierr); 160488ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 160588ebb749SStefano Zampini ierr = MatMatMult(A_IB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 160688ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM3);CHKERRQ(ierr); 160788ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 160888ebb749SStefano Zampini ierr = MatMatMult(A_BI,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 160988ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM4);CHKERRQ(ierr); 161088ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 161188ebb749SStefano Zampini } else { 161288ebb749SStefano Zampini ierr = MatPtAP(A_II,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&TM1);CHKERRQ(ierr); 161388ebb749SStefano Zampini ierr = MatPtAP(A_BB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&TM2);CHKERRQ(ierr); 161488ebb749SStefano Zampini ierr = MatMatMult(A_IB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 161588ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_phi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM3);CHKERRQ(ierr); 161688ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 161788ebb749SStefano Zampini ierr = MatMatMult(A_BI,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 161888ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_phi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM4);CHKERRQ(ierr); 161988ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 162088ebb749SStefano Zampini } 162188ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM2,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 162288ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM3,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 162388ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM4,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 162488ebb749SStefano Zampini ierr = MatConvert(TM1,MATSEQDENSE,MAT_REUSE_MATRIX,&TM1);CHKERRQ(ierr); 1625*4f1b2e48SStefano Zampini if (pcbddc->benign_n) { 1626d12edf2fSStefano Zampini Mat B0_I,B0_B,B0_BPHI,B0_IPHI; 1627d12edf2fSStefano Zampini PetscScalar *data,*data2; 1628*4f1b2e48SStefano Zampini PetscInt j; 1629d12edf2fSStefano Zampini 1630*4f1b2e48SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcbddc->benign_n,0,1,&is_dummy);CHKERRQ(ierr); 1631*4f1b2e48SStefano Zampini ierr = MatGetSubMatrix(pcbddc->benign_B0,is_dummy,pcis->is_B_local,MAT_INITIAL_MATRIX,&B0_B); 1632*4f1b2e48SStefano Zampini ierr = MatGetSubMatrix(pcbddc->benign_B0,is_dummy,pcis->is_I_local,MAT_INITIAL_MATRIX,&B0_I); 1633d12edf2fSStefano Zampini ierr = MatMatMult(B0_B,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&B0_BPHI);CHKERRQ(ierr); 1634d12edf2fSStefano Zampini ierr = MatConvert(B0_BPHI,MATSEQDENSE,MAT_REUSE_MATRIX,&B0_BPHI);CHKERRQ(ierr); 1635d12edf2fSStefano Zampini ierr = MatDenseGetArray(TM1,&data);CHKERRQ(ierr); 1636d12edf2fSStefano Zampini ierr = MatDenseGetArray(B0_BPHI,&data2);CHKERRQ(ierr); 1637*4f1b2e48SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) { 1638*4f1b2e48SStefano Zampini PetscInt primal_idx = pcbddc->local_primal_size - pcbddc->benign_n + j; 1639d12edf2fSStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 1640*4f1b2e48SStefano Zampini data[primal_idx*pcbddc->local_primal_size+i] += data2[i*pcbddc->benign_n+j]; 1641*4f1b2e48SStefano Zampini data[i*pcbddc->local_primal_size+primal_idx] += data2[i*pcbddc->benign_n+j]; 1642*4f1b2e48SStefano Zampini } 1643d12edf2fSStefano Zampini } 1644d12edf2fSStefano Zampini ierr = MatDenseRestoreArray(TM1,&data);CHKERRQ(ierr); 1645d12edf2fSStefano Zampini ierr = MatDenseRestoreArray(B0_BPHI,&data2);CHKERRQ(ierr); 1646d12edf2fSStefano Zampini ierr = MatDestroy(&B0_B);CHKERRQ(ierr); 1647d12edf2fSStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 1648d12edf2fSStefano Zampini ierr = MatDestroy(&B0_BPHI);CHKERRQ(ierr); 1649d12edf2fSStefano Zampini ierr = MatMatMult(B0_I,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&B0_IPHI);CHKERRQ(ierr); 1650d12edf2fSStefano Zampini ierr = MatDestroy(&B0_I);CHKERRQ(ierr); 1651d12edf2fSStefano Zampini ierr = MatNorm(B0_IPHI,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 1652d12edf2fSStefano Zampini ierr = MatDestroy(&B0_IPHI);CHKERRQ(ierr); 1653d12edf2fSStefano Zampini } 1654d12edf2fSStefano Zampini #if 0 1655d12edf2fSStefano Zampini { 1656d12edf2fSStefano Zampini PetscViewer viewer; 1657d12edf2fSStefano Zampini char filename[256]; 1658d12edf2fSStefano Zampini sprintf(filename,"proj_local_coarse_mat%d.m",PetscGlobalRank); 1659d12edf2fSStefano Zampini ierr = PetscViewerASCIIOpen(PETSC_COMM_SELF,filename,&viewer);CHKERRQ(ierr); 1660d12edf2fSStefano Zampini ierr = PetscViewerSetFormat(viewer,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 1661d12edf2fSStefano Zampini ierr = MatView(TM1,viewer);CHKERRQ(ierr); 1662d12edf2fSStefano Zampini ierr = PetscViewerDestroy(&viewer);CHKERRQ(ierr); 1663d12edf2fSStefano Zampini } 1664d12edf2fSStefano Zampini #endif 166581d9aea3SBarry Smith ierr = MatAXPY(TM1,m_one,coarse_sub_mat,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 16668bec7fa6SStefano Zampini ierr = MatNorm(TM1,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 16670fccc4e9SStefano Zampini ierr = PetscViewerASCIISynchronizedAllow(pcbddc->dbg_viewer,PETSC_TRUE);CHKERRQ(ierr); 166806656605SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d matrix error % 1.14e\n",PetscGlobalRank,real_value);CHKERRQ(ierr); 16698bec7fa6SStefano Zampini 16708bec7fa6SStefano Zampini /* check constraints */ 1671*4f1b2e48SStefano Zampini if (!pcbddc->benign_n) { /* TODO: add benign case */ 16728bec7fa6SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcbddc->local_primal_size,0,1,&is_dummy);CHKERRQ(ierr); 16738bec7fa6SStefano Zampini ierr = MatGetSubMatrix(pcbddc->ConstraintMatrix,is_dummy,pcis->is_B_local,MAT_INITIAL_MATRIX,&C_B); 16748bec7fa6SStefano Zampini ierr = MatMatMult(C_B,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&CPHI);CHKERRQ(ierr); 16758bec7fa6SStefano Zampini ierr = MatCreateVecs(CPHI,&mones,NULL);CHKERRQ(ierr); 16768bec7fa6SStefano Zampini ierr = VecSet(mones,-1.0);CHKERRQ(ierr); 16778bec7fa6SStefano Zampini ierr = MatDiagonalSet(CPHI,mones,ADD_VALUES);CHKERRQ(ierr); 16788bec7fa6SStefano Zampini ierr = MatNorm(CPHI,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 1679bdae7319SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d phi constraints error % 1.14e\n",PetscGlobalRank,real_value);CHKERRQ(ierr); 1680c0553b1fSStefano Zampini if (unsymmetric_check) { 1681bdae7319SStefano Zampini ierr = MatMatMult(C_B,coarse_psi_B,MAT_REUSE_MATRIX,1.0,&CPHI);CHKERRQ(ierr); 1682bdae7319SStefano Zampini ierr = VecSet(mones,-1.0);CHKERRQ(ierr); 1683bdae7319SStefano Zampini ierr = MatDiagonalSet(CPHI,mones,ADD_VALUES);CHKERRQ(ierr); 1684bdae7319SStefano Zampini ierr = MatNorm(CPHI,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 1685bdae7319SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d psi constraints error % 1.14e\n",PetscGlobalRank,real_value);CHKERRQ(ierr); 168688ebb749SStefano Zampini } 16878bec7fa6SStefano Zampini ierr = MatDestroy(&C_B);CHKERRQ(ierr); 16888bec7fa6SStefano Zampini ierr = MatDestroy(&CPHI);CHKERRQ(ierr); 16898bec7fa6SStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 16908bec7fa6SStefano Zampini ierr = VecDestroy(&mones);CHKERRQ(ierr); 1691d12edf2fSStefano Zampini } 169225084f0cSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 169388ebb749SStefano Zampini ierr = MatDestroy(&A_II);CHKERRQ(ierr); 169488ebb749SStefano Zampini ierr = MatDestroy(&A_BB);CHKERRQ(ierr); 169588ebb749SStefano Zampini ierr = MatDestroy(&A_IB);CHKERRQ(ierr); 169688ebb749SStefano Zampini ierr = MatDestroy(&A_BI);CHKERRQ(ierr); 169788ebb749SStefano Zampini ierr = MatDestroy(&TM1);CHKERRQ(ierr); 169888ebb749SStefano Zampini ierr = MatDestroy(&TM2);CHKERRQ(ierr); 169988ebb749SStefano Zampini ierr = MatDestroy(&TM3);CHKERRQ(ierr); 170088ebb749SStefano Zampini ierr = MatDestroy(&TM4);CHKERRQ(ierr); 170188ebb749SStefano Zampini ierr = MatDestroy(&coarse_phi_D);CHKERRQ(ierr); 170288ebb749SStefano Zampini ierr = MatDestroy(&coarse_phi_B);CHKERRQ(ierr); 1703c0553b1fSStefano Zampini if (unsymmetric_check) { 170488ebb749SStefano Zampini ierr = MatDestroy(&coarse_psi_D);CHKERRQ(ierr); 170588ebb749SStefano Zampini ierr = MatDestroy(&coarse_psi_B);CHKERRQ(ierr); 170688ebb749SStefano Zampini } 170788ebb749SStefano Zampini ierr = MatDestroy(&coarse_sub_mat);CHKERRQ(ierr); 170888ebb749SStefano Zampini } 17098629588bSStefano Zampini /* get back data */ 17108629588bSStefano Zampini *coarse_submat_vals_n = coarse_submat_vals; 171188ebb749SStefano Zampini PetscFunctionReturn(0); 171288ebb749SStefano Zampini } 171388ebb749SStefano Zampini 171488ebb749SStefano Zampini #undef __FUNCT__ 1715d65f70fdSStefano Zampini #define __FUNCT__ "MatGetSubMatrixUnsorted" 1716d648f858SStefano Zampini PetscErrorCode MatGetSubMatrixUnsorted(Mat A, IS isrow, IS iscol, Mat* B) 1717aa0d41d4SStefano Zampini { 1718d65f70fdSStefano Zampini Mat *work_mat; 1719d65f70fdSStefano Zampini IS isrow_s,iscol_s; 1720d65f70fdSStefano Zampini PetscBool rsorted,csorted; 1721d65f70fdSStefano Zampini PetscInt rsize,*idxs_perm_r,csize,*idxs_perm_c; 1722aa0d41d4SStefano Zampini PetscErrorCode ierr; 1723aa0d41d4SStefano Zampini 1724aa0d41d4SStefano Zampini PetscFunctionBegin; 1725d65f70fdSStefano Zampini ierr = ISSorted(isrow,&rsorted);CHKERRQ(ierr); 1726d65f70fdSStefano Zampini ierr = ISSorted(iscol,&csorted);CHKERRQ(ierr); 1727d65f70fdSStefano Zampini ierr = ISGetLocalSize(isrow,&rsize);CHKERRQ(ierr); 1728d65f70fdSStefano Zampini ierr = ISGetLocalSize(iscol,&csize);CHKERRQ(ierr); 1729aa0d41d4SStefano Zampini 1730d65f70fdSStefano Zampini if (!rsorted) { 1731906d46d4SStefano Zampini const PetscInt *idxs; 1732906d46d4SStefano Zampini PetscInt *idxs_sorted,i; 1733aa0d41d4SStefano Zampini 1734d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_perm_r);CHKERRQ(ierr); 1735d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_sorted);CHKERRQ(ierr); 1736d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 1737d65f70fdSStefano Zampini idxs_perm_r[i] = i; 1738aa0d41d4SStefano Zampini } 1739d65f70fdSStefano Zampini ierr = ISGetIndices(isrow,&idxs);CHKERRQ(ierr); 1740d65f70fdSStefano Zampini ierr = PetscSortIntWithPermutation(rsize,idxs,idxs_perm_r);CHKERRQ(ierr); 1741d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 1742d65f70fdSStefano Zampini idxs_sorted[i] = idxs[idxs_perm_r[i]]; 1743aa0d41d4SStefano Zampini } 1744d65f70fdSStefano Zampini ierr = ISRestoreIndices(isrow,&idxs);CHKERRQ(ierr); 1745d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,rsize,idxs_sorted,PETSC_OWN_POINTER,&isrow_s);CHKERRQ(ierr); 1746d65f70fdSStefano Zampini } else { 1747d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)isrow);CHKERRQ(ierr); 1748d65f70fdSStefano Zampini isrow_s = isrow; 1749aa0d41d4SStefano Zampini } 1750906d46d4SStefano Zampini 1751d65f70fdSStefano Zampini if (!csorted) { 1752d65f70fdSStefano Zampini if (isrow == iscol) { 1753d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)isrow_s);CHKERRQ(ierr); 1754d65f70fdSStefano Zampini iscol_s = isrow_s; 1755d65f70fdSStefano Zampini } else { 1756d65f70fdSStefano Zampini const PetscInt *idxs; 1757d65f70fdSStefano Zampini PetscInt *idxs_sorted,i; 1758906d46d4SStefano Zampini 1759d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_perm_c);CHKERRQ(ierr); 1760d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_sorted);CHKERRQ(ierr); 1761d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 1762d65f70fdSStefano Zampini idxs_perm_c[i] = i; 1763d65f70fdSStefano Zampini } 1764d65f70fdSStefano Zampini ierr = ISGetIndices(iscol,&idxs);CHKERRQ(ierr); 1765d65f70fdSStefano Zampini ierr = PetscSortIntWithPermutation(csize,idxs,idxs_perm_c);CHKERRQ(ierr); 1766d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 1767d65f70fdSStefano Zampini idxs_sorted[i] = idxs[idxs_perm_c[i]]; 1768d65f70fdSStefano Zampini } 1769d65f70fdSStefano Zampini ierr = ISRestoreIndices(iscol,&idxs);CHKERRQ(ierr); 1770d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,csize,idxs_sorted,PETSC_OWN_POINTER,&iscol_s);CHKERRQ(ierr); 1771d65f70fdSStefano Zampini } 1772d65f70fdSStefano Zampini } else { 1773d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)iscol);CHKERRQ(ierr); 1774d65f70fdSStefano Zampini iscol_s = iscol; 1775d65f70fdSStefano Zampini } 1776d65f70fdSStefano Zampini 1777d648f858SStefano Zampini ierr = MatGetSubMatrices(A,1,&isrow_s,&iscol_s,MAT_INITIAL_MATRIX,&work_mat);CHKERRQ(ierr); 1778d65f70fdSStefano Zampini 1779d65f70fdSStefano Zampini if (!rsorted || !csorted) { 1780906d46d4SStefano Zampini Mat new_mat; 1781d65f70fdSStefano Zampini IS is_perm_r,is_perm_c; 1782906d46d4SStefano Zampini 1783d65f70fdSStefano Zampini if (!rsorted) { 1784d65f70fdSStefano Zampini PetscInt *idxs_r,i; 1785d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_r);CHKERRQ(ierr); 1786d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 1787d65f70fdSStefano Zampini idxs_r[idxs_perm_r[i]] = i; 1788906d46d4SStefano Zampini } 1789d65f70fdSStefano Zampini ierr = PetscFree(idxs_perm_r);CHKERRQ(ierr); 1790d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,rsize,idxs_r,PETSC_OWN_POINTER,&is_perm_r);CHKERRQ(ierr); 1791d65f70fdSStefano Zampini } else { 1792d65f70fdSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,rsize,0,1,&is_perm_r);CHKERRQ(ierr); 1793906d46d4SStefano Zampini } 1794d65f70fdSStefano Zampini ierr = ISSetPermutation(is_perm_r);CHKERRQ(ierr); 1795d65f70fdSStefano Zampini 1796d65f70fdSStefano Zampini if (!csorted) { 1797d65f70fdSStefano Zampini if (isrow_s == iscol_s) { 1798d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)is_perm_r);CHKERRQ(ierr); 1799d65f70fdSStefano Zampini is_perm_c = is_perm_r; 1800d65f70fdSStefano Zampini } else { 1801d65f70fdSStefano Zampini PetscInt *idxs_c,i; 1802d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_c);CHKERRQ(ierr); 1803d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 1804d65f70fdSStefano Zampini idxs_c[idxs_perm_c[i]] = i; 1805d65f70fdSStefano Zampini } 1806d65f70fdSStefano Zampini ierr = PetscFree(idxs_perm_c);CHKERRQ(ierr); 1807d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,csize,idxs_c,PETSC_OWN_POINTER,&is_perm_c);CHKERRQ(ierr); 1808d65f70fdSStefano Zampini } 1809d65f70fdSStefano Zampini } else { 1810d65f70fdSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,csize,0,1,&is_perm_c);CHKERRQ(ierr); 1811d65f70fdSStefano Zampini } 1812d65f70fdSStefano Zampini ierr = ISSetPermutation(is_perm_c);CHKERRQ(ierr); 1813d65f70fdSStefano Zampini 1814d65f70fdSStefano Zampini ierr = MatPermute(work_mat[0],is_perm_r,is_perm_c,&new_mat);CHKERRQ(ierr); 1815d65f70fdSStefano Zampini ierr = MatDestroy(&work_mat[0]);CHKERRQ(ierr); 1816d65f70fdSStefano Zampini work_mat[0] = new_mat; 1817d65f70fdSStefano Zampini ierr = ISDestroy(&is_perm_r);CHKERRQ(ierr); 1818d65f70fdSStefano Zampini ierr = ISDestroy(&is_perm_c);CHKERRQ(ierr); 1819d65f70fdSStefano Zampini } 1820d65f70fdSStefano Zampini 1821d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)work_mat[0]);CHKERRQ(ierr); 1822d65f70fdSStefano Zampini *B = work_mat[0]; 1823d65f70fdSStefano Zampini ierr = MatDestroyMatrices(1,&work_mat);CHKERRQ(ierr); 1824d65f70fdSStefano Zampini ierr = ISDestroy(&isrow_s);CHKERRQ(ierr); 1825d65f70fdSStefano Zampini ierr = ISDestroy(&iscol_s);CHKERRQ(ierr); 1826d65f70fdSStefano Zampini PetscFunctionReturn(0); 1827d65f70fdSStefano Zampini } 1828d65f70fdSStefano Zampini 1829d65f70fdSStefano Zampini #undef __FUNCT__ 18305e8657edSStefano Zampini #define __FUNCT__ "PCBDDCComputeLocalMatrix" 18315e8657edSStefano Zampini PetscErrorCode PCBDDCComputeLocalMatrix(PC pc, Mat ChangeOfBasisMatrix) 1832aa0d41d4SStefano Zampini { 1833aa0d41d4SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 18345e8657edSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 1835d65f70fdSStefano Zampini Mat new_mat; 18365e8657edSStefano Zampini IS is_local,is_global; 1837d65f70fdSStefano Zampini PetscInt local_size; 1838d65f70fdSStefano Zampini PetscBool isseqaij; 1839aa0d41d4SStefano Zampini PetscErrorCode ierr; 1840aa0d41d4SStefano Zampini 1841aa0d41d4SStefano Zampini PetscFunctionBegin; 1842aa0d41d4SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 18435e8657edSStefano Zampini ierr = MatGetSize(matis->A,&local_size,NULL);CHKERRQ(ierr); 18445e8657edSStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)matis->A),local_size,0,1,&is_local);CHKERRQ(ierr); 1845b087196eSStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(pc->pmat->rmap->mapping,is_local,&is_global);CHKERRQ(ierr); 1846aa0d41d4SStefano Zampini ierr = ISDestroy(&is_local);CHKERRQ(ierr); 1847d648f858SStefano Zampini ierr = MatGetSubMatrixUnsorted(ChangeOfBasisMatrix,is_global,is_global,&new_mat);CHKERRQ(ierr); 1848aa0d41d4SStefano Zampini ierr = ISDestroy(&is_global);CHKERRQ(ierr); 1849906d46d4SStefano Zampini 1850906d46d4SStefano Zampini /* check */ 1851906d46d4SStefano Zampini if (pcbddc->dbg_flag) { 1852906d46d4SStefano Zampini Vec x,x_change; 1853906d46d4SStefano Zampini PetscReal error; 1854906d46d4SStefano Zampini 18555e8657edSStefano Zampini ierr = MatCreateVecs(ChangeOfBasisMatrix,&x,&x_change);CHKERRQ(ierr); 1856906d46d4SStefano Zampini ierr = VecSetRandom(x,NULL);CHKERRQ(ierr); 18575e8657edSStefano Zampini ierr = MatMult(ChangeOfBasisMatrix,x,x_change);CHKERRQ(ierr); 1858e176bc59SStefano Zampini ierr = VecScatterBegin(matis->cctx,x,matis->x,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1859e176bc59SStefano Zampini ierr = VecScatterEnd(matis->cctx,x,matis->x,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1860d65f70fdSStefano Zampini ierr = MatMult(new_mat,matis->x,matis->y);CHKERRQ(ierr); 1861e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,matis->y,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1862e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,matis->y,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1863906d46d4SStefano Zampini ierr = VecAXPY(x,-1.0,x_change);CHKERRQ(ierr); 1864906d46d4SStefano Zampini ierr = VecNorm(x,NORM_INFINITY,&error);CHKERRQ(ierr); 1865906d46d4SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1866906d46d4SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Error global vs local change on N: %1.6e\n",error);CHKERRQ(ierr); 1867906d46d4SStefano Zampini ierr = VecDestroy(&x);CHKERRQ(ierr); 1868906d46d4SStefano Zampini ierr = VecDestroy(&x_change);CHKERRQ(ierr); 1869906d46d4SStefano Zampini } 1870906d46d4SStefano Zampini 187122d5777bSStefano Zampini /* TODO: HOW TO WORK WITH BAIJ and SBAIJ and SEQDENSE? */ 18729b28b3ffSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)matis->A,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 187322d5777bSStefano Zampini if (isseqaij) { 1874d65f70fdSStefano Zampini ierr = MatPtAP(matis->A,new_mat,MAT_INITIAL_MATRIX,2.0,&pcbddc->local_mat);CHKERRQ(ierr); 1875aa0d41d4SStefano Zampini } else { 1876aa0d41d4SStefano Zampini Mat work_mat; 1877aa0d41d4SStefano Zampini ierr = MatConvert(matis->A,MATSEQAIJ,MAT_INITIAL_MATRIX,&work_mat);CHKERRQ(ierr); 1878d65f70fdSStefano Zampini ierr = MatPtAP(work_mat,new_mat,MAT_INITIAL_MATRIX,2.0,&pcbddc->local_mat);CHKERRQ(ierr); 1879aa0d41d4SStefano Zampini ierr = MatDestroy(&work_mat);CHKERRQ(ierr); 1880aa0d41d4SStefano Zampini } 18813301b35fSStefano Zampini if (matis->A->symmetric_set) { 18823301b35fSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_SYMMETRIC,matis->A->symmetric);CHKERRQ(ierr); 1883e496cd5dSStefano Zampini #if !defined(PETSC_USE_COMPLEX) 18843301b35fSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_HERMITIAN,matis->A->symmetric);CHKERRQ(ierr); 1885e496cd5dSStefano Zampini #endif 18863301b35fSStefano Zampini } 188745a1bb75SStefano Zampini /* 188845a1bb75SStefano Zampini ierr = PetscViewerSetFormat(PETSC_VIEWER_STDOUT_SELF,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 1889d65f70fdSStefano Zampini ierr = MatView(new_mat,(PetscViewer)0);CHKERRQ(ierr); 189045a1bb75SStefano Zampini */ 1891d65f70fdSStefano Zampini ierr = MatDestroy(&new_mat);CHKERRQ(ierr); 1892aa0d41d4SStefano Zampini PetscFunctionReturn(0); 1893aa0d41d4SStefano Zampini } 1894aa0d41d4SStefano Zampini 1895aa0d41d4SStefano Zampini #undef __FUNCT__ 1896a64d13efSStefano Zampini #define __FUNCT__ "PCBDDCSetUpLocalScatters" 18978ce42a96SStefano Zampini PetscErrorCode PCBDDCSetUpLocalScatters(PC pc) 1898a64d13efSStefano Zampini { 1899a64d13efSStefano Zampini PC_IS* pcis = (PC_IS*)(pc->data); 1900a64d13efSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 1901d62866d3SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 190253892102SStefano Zampini PetscInt *idx_R_local=NULL; 19033a50541eSStefano Zampini PetscInt n_vertices,i,j,n_R,n_D,n_B; 19043a50541eSStefano Zampini PetscInt vbs,bs; 19056816873aSStefano Zampini PetscBT bitmask=NULL; 1906a64d13efSStefano Zampini PetscErrorCode ierr; 1907a64d13efSStefano Zampini 1908a64d13efSStefano Zampini PetscFunctionBegin; 1909b23d619eSStefano Zampini /* 1910b23d619eSStefano Zampini No need to setup local scatters if 1911b23d619eSStefano Zampini - primal space is unchanged 1912b23d619eSStefano Zampini AND 1913b23d619eSStefano Zampini - we actually have locally some primal dofs (could not be true in multilevel or for isolated subdomains) 1914b23d619eSStefano Zampini AND 1915b23d619eSStefano Zampini - we are not in debugging mode (this is needed since there are Synchronized prints at the end of the subroutine 1916b23d619eSStefano Zampini */ 1917b23d619eSStefano Zampini if (!pcbddc->new_primal_space_local && pcbddc->local_primal_size && !pcbddc->dbg_flag) { 1918f4ddd8eeSStefano Zampini PetscFunctionReturn(0); 1919f4ddd8eeSStefano Zampini } 1920f4ddd8eeSStefano Zampini /* destroy old objects */ 1921f4ddd8eeSStefano Zampini ierr = ISDestroy(&pcbddc->is_R_local);CHKERRQ(ierr); 1922f4ddd8eeSStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_B);CHKERRQ(ierr); 1923f4ddd8eeSStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_D);CHKERRQ(ierr); 1924a64d13efSStefano Zampini /* Set Non-overlapping dimensions */ 1925b371cd4fSStefano Zampini n_B = pcis->n_B; 1926b371cd4fSStefano Zampini n_D = pcis->n - n_B; 1927b371cd4fSStefano Zampini n_vertices = pcbddc->n_vertices; 19283a50541eSStefano Zampini 1929a64d13efSStefano Zampini /* Dohrmann's notation: dofs splitted in R (Remaining: all dofs but the vertices) and V (Vertices) */ 19306816873aSStefano Zampini 193153892102SStefano Zampini /* create auxiliary bitmask and allocate workspace */ 19326816873aSStefano Zampini if (!sub_schurs->reuse_mumps) { 1933854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n-n_vertices,&idx_R_local);CHKERRQ(ierr); 1934a64d13efSStefano Zampini ierr = PetscBTCreate(pcis->n,&bitmask);CHKERRQ(ierr); 1935a64d13efSStefano Zampini for (i=0;i<n_vertices;i++) { 19360e6343abSStefano Zampini ierr = PetscBTSet(bitmask,pcbddc->local_primal_ref_node[i]);CHKERRQ(ierr); 1937a64d13efSStefano Zampini } 1938a64d13efSStefano Zampini 1939a64d13efSStefano Zampini for (i=0, n_R=0; i<pcis->n; i++) { 19404641a718SStefano Zampini if (!PetscBTLookup(bitmask,i)) { 19416816873aSStefano Zampini idx_R_local[n_R++] = i; 1942a64d13efSStefano Zampini } 1943a64d13efSStefano Zampini } 194453892102SStefano Zampini } else { /* A different ordering (already computed) is present if we are reusing MUMPS Schur solver */ 19456816873aSStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 19466816873aSStefano Zampini 194753892102SStefano Zampini ierr = ISGetIndices(reuse_mumps->is_R,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 194853892102SStefano Zampini ierr = ISGetLocalSize(reuse_mumps->is_R,&n_R);CHKERRQ(ierr); 19496816873aSStefano Zampini } 19503a50541eSStefano Zampini 19513a50541eSStefano Zampini /* Block code */ 19523a50541eSStefano Zampini vbs = 1; 19533a50541eSStefano Zampini ierr = MatGetBlockSize(pcbddc->local_mat,&bs);CHKERRQ(ierr); 19543a50541eSStefano Zampini if (bs>1 && !(n_vertices%bs)) { 19553a50541eSStefano Zampini PetscBool is_blocked = PETSC_TRUE; 19563a50541eSStefano Zampini PetscInt *vary; 1957d3df7717SStefano Zampini if (!sub_schurs->reuse_mumps) { 1958785e854fSJed Brown ierr = PetscMalloc1(pcis->n/bs,&vary);CHKERRQ(ierr); 19593a50541eSStefano Zampini ierr = PetscMemzero(vary,pcis->n/bs*sizeof(PetscInt));CHKERRQ(ierr); 1960d3df7717SStefano Zampini /* Verify that the vertex indices correspond to each element in a block (code taken from sbaij2.c) */ 1961d3df7717SStefano 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 */ 19620e6343abSStefano Zampini for (i=0; i<n_vertices; i++) vary[pcbddc->local_primal_ref_node[i]/bs]++; 1963d3df7717SStefano Zampini for (i=0; i<pcis->n/bs; i++) { 19643a50541eSStefano Zampini if (vary[i]!=0 && vary[i]!=bs) { 19653a50541eSStefano Zampini is_blocked = PETSC_FALSE; 19663a50541eSStefano Zampini break; 19673a50541eSStefano Zampini } 19683a50541eSStefano Zampini } 1969d3df7717SStefano Zampini ierr = PetscFree(vary);CHKERRQ(ierr); 1970d3df7717SStefano Zampini } else { 1971d3df7717SStefano Zampini /* Verify directly the R set */ 1972d3df7717SStefano Zampini for (i=0; i<n_R/bs; i++) { 1973d3df7717SStefano Zampini PetscInt j,node=idx_R_local[bs*i]; 1974d3df7717SStefano Zampini for (j=1; j<bs; j++) { 1975d3df7717SStefano Zampini if (node != idx_R_local[bs*i+j]-j) { 1976d3df7717SStefano Zampini is_blocked = PETSC_FALSE; 1977d3df7717SStefano Zampini break; 1978d3df7717SStefano Zampini } 1979d3df7717SStefano Zampini } 1980d3df7717SStefano Zampini } 1981d3df7717SStefano Zampini } 19823a50541eSStefano Zampini if (is_blocked) { /* build compressed IS for R nodes (complement of vertices) */ 19833a50541eSStefano Zampini vbs = bs; 19843a50541eSStefano Zampini for (i=0;i<n_R/vbs;i++) { 19853a50541eSStefano Zampini idx_R_local[i] = idx_R_local[vbs*i]/vbs; 19863a50541eSStefano Zampini } 19873a50541eSStefano Zampini } 19883a50541eSStefano Zampini } 19893a50541eSStefano Zampini ierr = ISCreateBlock(PETSC_COMM_SELF,vbs,n_R/vbs,idx_R_local,PETSC_COPY_VALUES,&pcbddc->is_R_local);CHKERRQ(ierr); 199053892102SStefano Zampini if (sub_schurs->reuse_mumps) { 199153892102SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 199253892102SStefano Zampini 199353892102SStefano Zampini ierr = ISRestoreIndices(reuse_mumps->is_R,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 199453892102SStefano Zampini ierr = ISDestroy(&reuse_mumps->is_R);CHKERRQ(ierr); 199553892102SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->is_R_local);CHKERRQ(ierr); 199653892102SStefano Zampini reuse_mumps->is_R = pcbddc->is_R_local; 199753892102SStefano Zampini } else { 19983a50541eSStefano Zampini ierr = PetscFree(idx_R_local);CHKERRQ(ierr); 199953892102SStefano Zampini } 2000a64d13efSStefano Zampini 2001a64d13efSStefano Zampini /* print some info if requested */ 2002a64d13efSStefano Zampini if (pcbddc->dbg_flag) { 2003a64d13efSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 2004a64d13efSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 20050fccc4e9SStefano Zampini ierr = PetscViewerASCIISynchronizedAllow(pcbddc->dbg_viewer,PETSC_TRUE);CHKERRQ(ierr); 2006a64d13efSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d local dimensions\n",PetscGlobalRank);CHKERRQ(ierr); 2007a64d13efSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"local_size = %d, dirichlet_size = %d, boundary_size = %d\n",pcis->n,n_D,n_B);CHKERRQ(ierr); 2008*4f1b2e48SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"r_size = %d, v_size = %d, constraints = %d, local_primal_size = %d\n",n_R,n_vertices,pcbddc->local_primal_size-n_vertices-pcbddc->benign_n,pcbddc->local_primal_size);CHKERRQ(ierr); 2009a64d13efSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 2010a64d13efSStefano Zampini } 2011a64d13efSStefano Zampini 2012a64d13efSStefano Zampini /* VecScatters pcbddc->R_to_B and (optionally) pcbddc->R_to_D */ 20136816873aSStefano Zampini if (!sub_schurs->reuse_mumps) { 20146816873aSStefano Zampini IS is_aux1,is_aux2; 20156816873aSStefano Zampini PetscInt *aux_array1,*aux_array2,*is_indices,*idx_R_local; 20166816873aSStefano Zampini 20173a50541eSStefano Zampini ierr = ISGetIndices(pcbddc->is_R_local,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 2018854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n_B-n_vertices,&aux_array1);CHKERRQ(ierr); 2019854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n_B-n_vertices,&aux_array2);CHKERRQ(ierr); 2020a64d13efSStefano Zampini ierr = ISGetIndices(pcis->is_I_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 20214641a718SStefano Zampini for (i=0; i<n_D; i++) { 20224641a718SStefano Zampini ierr = PetscBTSet(bitmask,is_indices[i]);CHKERRQ(ierr); 20234641a718SStefano Zampini } 2024a64d13efSStefano Zampini ierr = ISRestoreIndices(pcis->is_I_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 2025a64d13efSStefano Zampini for (i=0, j=0; i<n_R; i++) { 20264641a718SStefano Zampini if (!PetscBTLookup(bitmask,idx_R_local[i])) { 20274641a718SStefano Zampini aux_array1[j++] = i; 2028a64d13efSStefano Zampini } 2029a64d13efSStefano Zampini } 2030a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array1,PETSC_OWN_POINTER,&is_aux1);CHKERRQ(ierr); 2031a64d13efSStefano Zampini ierr = ISGetIndices(pcis->is_B_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 2032a64d13efSStefano Zampini for (i=0, j=0; i<n_B; i++) { 20334641a718SStefano Zampini if (!PetscBTLookup(bitmask,is_indices[i])) { 20344641a718SStefano Zampini aux_array2[j++] = i; 2035a64d13efSStefano Zampini } 2036a64d13efSStefano Zampini } 2037a64d13efSStefano Zampini ierr = ISRestoreIndices(pcis->is_B_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 2038a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array2,PETSC_OWN_POINTER,&is_aux2);CHKERRQ(ierr); 2039a64d13efSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,is_aux1,pcis->vec1_B,is_aux2,&pcbddc->R_to_B);CHKERRQ(ierr); 2040a64d13efSStefano Zampini ierr = ISDestroy(&is_aux1);CHKERRQ(ierr); 2041a64d13efSStefano Zampini ierr = ISDestroy(&is_aux2);CHKERRQ(ierr); 2042a64d13efSStefano Zampini 20438eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 2044785e854fSJed Brown ierr = PetscMalloc1(n_D,&aux_array1);CHKERRQ(ierr); 2045a64d13efSStefano Zampini for (i=0, j=0; i<n_R; i++) { 20464641a718SStefano Zampini if (PetscBTLookup(bitmask,idx_R_local[i])) { 20474641a718SStefano Zampini aux_array1[j++] = i; 2048a64d13efSStefano Zampini } 2049a64d13efSStefano Zampini } 2050a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array1,PETSC_OWN_POINTER,&is_aux1);CHKERRQ(ierr); 2051a64d13efSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,is_aux1,pcis->vec1_D,(IS)0,&pcbddc->R_to_D);CHKERRQ(ierr); 2052a64d13efSStefano Zampini ierr = ISDestroy(&is_aux1);CHKERRQ(ierr); 2053a64d13efSStefano Zampini } 20544641a718SStefano Zampini ierr = PetscBTDestroy(&bitmask);CHKERRQ(ierr); 20553a50541eSStefano Zampini ierr = ISRestoreIndices(pcbddc->is_R_local,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 2056d62866d3SStefano Zampini } else { 205753892102SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 20586816873aSStefano Zampini IS tis; 20596816873aSStefano Zampini PetscInt schur_size; 20606816873aSStefano Zampini 206153892102SStefano Zampini ierr = ISGetLocalSize(reuse_mumps->is_B,&schur_size);CHKERRQ(ierr); 20626816873aSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,schur_size,n_D,1,&tis);CHKERRQ(ierr); 206353892102SStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,tis,pcis->vec1_B,reuse_mumps->is_B,&pcbddc->R_to_B);CHKERRQ(ierr); 20646816873aSStefano Zampini ierr = ISDestroy(&tis);CHKERRQ(ierr); 20656816873aSStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 20666816873aSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,n_D,0,1,&tis);CHKERRQ(ierr); 20676816873aSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,tis,pcis->vec1_D,(IS)0,&pcbddc->R_to_D);CHKERRQ(ierr); 20686816873aSStefano Zampini ierr = ISDestroy(&tis);CHKERRQ(ierr); 2069d62866d3SStefano Zampini } 2070d62866d3SStefano Zampini } 2071a64d13efSStefano Zampini PetscFunctionReturn(0); 2072a64d13efSStefano Zampini } 2073a64d13efSStefano Zampini 2074304d26faSStefano Zampini 2075304d26faSStefano Zampini #undef __FUNCT__ 2076304d26faSStefano Zampini #define __FUNCT__ "PCBDDCSetUpLocalSolvers" 2077684f6988SStefano Zampini PetscErrorCode PCBDDCSetUpLocalSolvers(PC pc, PetscBool dirichlet, PetscBool neumann) 2078304d26faSStefano Zampini { 2079304d26faSStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 2080304d26faSStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 2081304d26faSStefano Zampini PC pc_temp; 2082304d26faSStefano Zampini Mat A_RR; 2083f4ddd8eeSStefano Zampini MatReuse reuse; 2084304d26faSStefano Zampini PetscScalar m_one = -1.0; 2085304d26faSStefano Zampini PetscReal value; 208604708bb6SStefano Zampini PetscInt n_D,n_R; 20879577ea80SStefano Zampini PetscBool use_exact,use_exact_reduced,issbaij; 2088304d26faSStefano Zampini PetscErrorCode ierr; 2089e604994aSStefano Zampini /* prefixes stuff */ 2090312be037SStefano Zampini char dir_prefix[256],neu_prefix[256],str_level[16]; 2091e604994aSStefano Zampini size_t len; 2092304d26faSStefano Zampini 2093304d26faSStefano Zampini PetscFunctionBegin; 2094304d26faSStefano Zampini 2095e604994aSStefano Zampini /* compute prefixes */ 2096e604994aSStefano Zampini ierr = PetscStrcpy(dir_prefix,"");CHKERRQ(ierr); 2097e604994aSStefano Zampini ierr = PetscStrcpy(neu_prefix,"");CHKERRQ(ierr); 2098e604994aSStefano Zampini if (!pcbddc->current_level) { 2099e604994aSStefano Zampini ierr = PetscStrcpy(dir_prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 2100e604994aSStefano Zampini ierr = PetscStrcpy(neu_prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 2101e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,"pc_bddc_dirichlet_");CHKERRQ(ierr); 2102e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,"pc_bddc_neumann_");CHKERRQ(ierr); 2103e604994aSStefano Zampini } else { 2104e604994aSStefano Zampini ierr = PetscStrcpy(str_level,"");CHKERRQ(ierr); 2105312be037SStefano Zampini sprintf(str_level,"l%d_",(int)(pcbddc->current_level)); 2106e604994aSStefano Zampini ierr = PetscStrlen(((PetscObject)pc)->prefix,&len);CHKERRQ(ierr); 2107e604994aSStefano Zampini len -= 15; /* remove "pc_bddc_coarse_" */ 2108312be037SStefano Zampini if (pcbddc->current_level>1) len -= 3; /* remove "lX_" with X level number */ 2109312be037SStefano Zampini if (pcbddc->current_level>10) len -= 1; /* remove another char from level number */ 211034d6797cSStefano Zampini ierr = PetscStrncpy(dir_prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 211134d6797cSStefano Zampini ierr = PetscStrncpy(neu_prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 2112e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,"pc_bddc_dirichlet_");CHKERRQ(ierr); 2113e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,"pc_bddc_neumann_");CHKERRQ(ierr); 2114e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,str_level);CHKERRQ(ierr); 2115e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,str_level);CHKERRQ(ierr); 2116e604994aSStefano Zampini } 2117e604994aSStefano Zampini 2118304d26faSStefano Zampini /* DIRICHLET PROBLEM */ 2119684f6988SStefano Zampini if (dirichlet) { 2120d5574798SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 21213301b35fSStefano Zampini if (pcbddc->local_mat->symmetric_set) { 21223301b35fSStefano Zampini ierr = MatSetOption(pcis->A_II,MAT_SYMMETRIC,pcbddc->local_mat->symmetric_set);CHKERRQ(ierr); 2123964fefecSStefano Zampini } 2124ac78edfcSStefano Zampini /* Matrix for Dirichlet problem is pcis->A_II */ 2125964fefecSStefano Zampini n_D = pcis->n - pcis->n_B; 2126304d26faSStefano Zampini if (!pcbddc->ksp_D) { /* create object if not yet build */ 2127304d26faSStefano Zampini ierr = KSPCreate(PETSC_COMM_SELF,&pcbddc->ksp_D);CHKERRQ(ierr); 2128304d26faSStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->ksp_D,(PetscObject)pc,1);CHKERRQ(ierr); 2129304d26faSStefano Zampini /* default */ 2130304d26faSStefano Zampini ierr = KSPSetType(pcbddc->ksp_D,KSPPREONLY);CHKERRQ(ierr); 2131e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->ksp_D,dir_prefix);CHKERRQ(ierr); 21329577ea80SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcis->A_II,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 2133304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_D,&pc_temp);CHKERRQ(ierr); 21349577ea80SStefano Zampini if (issbaij) { 21359577ea80SStefano Zampini ierr = PCSetType(pc_temp,PCCHOLESKY);CHKERRQ(ierr); 21369577ea80SStefano Zampini } else { 2137304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCLU);CHKERRQ(ierr); 21389577ea80SStefano Zampini } 2139304d26faSStefano Zampini /* Allow user's customization */ 2140304d26faSStefano Zampini ierr = KSPSetFromOptions(pcbddc->ksp_D);CHKERRQ(ierr); 2141304d26faSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 2142304d26faSStefano Zampini } 2143d1e9a80fSBarry Smith ierr = KSPSetOperators(pcbddc->ksp_D,pcis->A_II,pcis->A_II);CHKERRQ(ierr); 2144d62866d3SStefano Zampini if (sub_schurs->reuse_mumps) { 2145d62866d3SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 2146d62866d3SStefano Zampini 2147d62866d3SStefano Zampini ierr = KSPSetPC(pcbddc->ksp_D,reuse_mumps->interior_solver);CHKERRQ(ierr); 2148d5574798SStefano Zampini } 2149304d26faSStefano Zampini /* umfpack interface has a bug when matrix dimension is zero. TODO solve from umfpack interface */ 2150304d26faSStefano Zampini if (!n_D) { 2151304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_D,&pc_temp);CHKERRQ(ierr); 2152304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCNONE);CHKERRQ(ierr); 2153304d26faSStefano Zampini } 2154304d26faSStefano Zampini /* Set Up KSP for Dirichlet problem of BDDC */ 2155304d26faSStefano Zampini ierr = KSPSetUp(pcbddc->ksp_D);CHKERRQ(ierr); 2156304d26faSStefano Zampini /* set ksp_D into pcis data */ 2157304d26faSStefano Zampini ierr = KSPDestroy(&pcis->ksp_D);CHKERRQ(ierr); 2158304d26faSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->ksp_D);CHKERRQ(ierr); 2159304d26faSStefano Zampini pcis->ksp_D = pcbddc->ksp_D; 2160684f6988SStefano Zampini } 2161304d26faSStefano Zampini 2162304d26faSStefano Zampini /* NEUMANN PROBLEM */ 2163684f6988SStefano Zampini A_RR = 0; 2164684f6988SStefano Zampini if (neumann) { 2165d62866d3SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 216604708bb6SStefano Zampini PetscInt ibs,mbs; 216704708bb6SStefano Zampini PetscBool issbaij; 216804708bb6SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 2169f4ddd8eeSStefano Zampini /* Matrix for Neumann problem is A_RR -> we need to create/reuse it at this point */ 21708ce42a96SStefano Zampini ierr = ISGetSize(pcbddc->is_R_local,&n_R);CHKERRQ(ierr); 2171f4ddd8eeSStefano Zampini if (pcbddc->ksp_R) { /* already created ksp */ 2172f4ddd8eeSStefano Zampini PetscInt nn_R; 217381d9aea3SBarry Smith ierr = KSPGetOperators(pcbddc->ksp_R,NULL,&A_RR);CHKERRQ(ierr); 2174f4ddd8eeSStefano Zampini ierr = PetscObjectReference((PetscObject)A_RR);CHKERRQ(ierr); 2175f4ddd8eeSStefano Zampini ierr = MatGetSize(A_RR,&nn_R,NULL);CHKERRQ(ierr); 2176f4ddd8eeSStefano Zampini if (nn_R != n_R) { /* old ksp is not reusable, so reset it */ 2177f4ddd8eeSStefano Zampini ierr = KSPReset(pcbddc->ksp_R);CHKERRQ(ierr); 2178f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 2179f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 2180f4ddd8eeSStefano Zampini } else { /* same sizes, but nonzero pattern depend on primal vertices so it can be changed */ 2181727cdba6SStefano Zampini if (pcbddc->new_primal_space_local) { /* we are not sure the matrix will have the same nonzero pattern */ 2182f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 2183f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 2184f4ddd8eeSStefano Zampini } else { /* safe to reuse the matrix */ 2185f4ddd8eeSStefano Zampini reuse = MAT_REUSE_MATRIX; 2186f4ddd8eeSStefano Zampini } 2187f4ddd8eeSStefano Zampini } 2188f4ddd8eeSStefano Zampini /* last check */ 2189d1e9a80fSBarry Smith if (pc->flag == DIFFERENT_NONZERO_PATTERN) { 2190f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 2191f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 2192f4ddd8eeSStefano Zampini } 2193f4ddd8eeSStefano Zampini } else { /* first time, so we need to create the matrix */ 2194f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 2195f4ddd8eeSStefano Zampini } 2196f4ddd8eeSStefano Zampini /* extract A_RR */ 2197af732b37SStefano Zampini ierr = MatGetBlockSize(pcbddc->local_mat,&mbs);CHKERRQ(ierr); 2198af732b37SStefano Zampini ierr = ISGetBlockSize(pcbddc->is_R_local,&ibs);CHKERRQ(ierr); 219904708bb6SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcbddc->local_mat,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 220004708bb6SStefano Zampini if (ibs != mbs) { /* need to convert to SEQAIJ to extract any submatrix with is_R_local */ 220104708bb6SStefano Zampini if (matis->A == pcbddc->local_mat) { 220204708bb6SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 220304708bb6SStefano Zampini ierr = MatConvert(matis->A,MATSEQAIJ,MAT_INITIAL_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 2204af732b37SStefano Zampini } else { 220504708bb6SStefano Zampini ierr = MatConvert(pcbddc->local_mat,MATSEQAIJ,MAT_REUSE_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 22066816873aSStefano Zampini } 220704708bb6SStefano Zampini } else if (issbaij) { /* need to convert to BAIJ to get offdiagonal blocks */ 220804708bb6SStefano Zampini if (matis->A == pcbddc->local_mat) { 220904708bb6SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 221004708bb6SStefano Zampini ierr = MatConvert(matis->A,MATSEQBAIJ,MAT_INITIAL_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 221104708bb6SStefano Zampini } else { 221204708bb6SStefano Zampini ierr = MatConvert(pcbddc->local_mat,MATSEQBAIJ,MAT_REUSE_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 221304708bb6SStefano Zampini } 221404708bb6SStefano Zampini } 221504708bb6SStefano Zampini if (!sub_schurs->reuse_mumps) { 2216f4ddd8eeSStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,pcbddc->is_R_local,pcbddc->is_R_local,reuse,&A_RR);CHKERRQ(ierr); 22173301b35fSStefano Zampini if (pcbddc->local_mat->symmetric_set) { 22183301b35fSStefano Zampini ierr = MatSetOption(A_RR,MAT_SYMMETRIC,pcbddc->local_mat->symmetric_set);CHKERRQ(ierr); 22196816873aSStefano Zampini } 22206816873aSStefano Zampini } else { 22216816873aSStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 22226816873aSStefano Zampini 22236816873aSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 22246816873aSStefano Zampini ierr = PCGetOperators(reuse_mumps->correction_solver,&A_RR,NULL);CHKERRQ(ierr); 22256816873aSStefano Zampini ierr = PetscObjectReference((PetscObject)A_RR);CHKERRQ(ierr); 2226af732b37SStefano Zampini } 2227f4ddd8eeSStefano Zampini if (!pcbddc->ksp_R) { /* create object if not present */ 2228304d26faSStefano Zampini ierr = KSPCreate(PETSC_COMM_SELF,&pcbddc->ksp_R);CHKERRQ(ierr); 2229304d26faSStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->ksp_R,(PetscObject)pc,1);CHKERRQ(ierr); 2230304d26faSStefano Zampini /* default */ 2231304d26faSStefano Zampini ierr = KSPSetType(pcbddc->ksp_R,KSPPREONLY);CHKERRQ(ierr); 2232e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->ksp_R,neu_prefix);CHKERRQ(ierr); 2233304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_temp);CHKERRQ(ierr); 22349577ea80SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)A_RR,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 22359577ea80SStefano Zampini if (issbaij) { 22369577ea80SStefano Zampini ierr = PCSetType(pc_temp,PCCHOLESKY);CHKERRQ(ierr); 22379577ea80SStefano Zampini } else { 2238304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCLU);CHKERRQ(ierr); 22399577ea80SStefano Zampini } 2240304d26faSStefano Zampini /* Allow user's customization */ 2241304d26faSStefano Zampini ierr = KSPSetFromOptions(pcbddc->ksp_R);CHKERRQ(ierr); 2242304d26faSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 2243304d26faSStefano Zampini } 2244d1e9a80fSBarry Smith ierr = KSPSetOperators(pcbddc->ksp_R,A_RR,A_RR);CHKERRQ(ierr); 2245304d26faSStefano Zampini /* umfpack interface has a bug when matrix dimension is zero. TODO solve from umfpack interface */ 2246304d26faSStefano Zampini if (!n_R) { 2247304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_temp);CHKERRQ(ierr); 2248304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCNONE);CHKERRQ(ierr); 2249304d26faSStefano Zampini } 2250d62866d3SStefano Zampini /* Reuse MUMPS solver if it is present */ 2251d62866d3SStefano Zampini if (sub_schurs->reuse_mumps) { 2252d62866d3SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 2253d62866d3SStefano Zampini 2254d62866d3SStefano Zampini ierr = KSPSetPC(pcbddc->ksp_R,reuse_mumps->correction_solver);CHKERRQ(ierr); 2255d62866d3SStefano Zampini } 2256304d26faSStefano Zampini /* Set Up KSP for Neumann problem of BDDC */ 2257304d26faSStefano Zampini ierr = KSPSetUp(pcbddc->ksp_R);CHKERRQ(ierr); 2258684f6988SStefano Zampini } 22596816873aSStefano Zampini /* free Neumann problem's matrix */ 22606816873aSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 2261304d26faSStefano Zampini 2262304d26faSStefano Zampini /* check Dirichlet and Neumann solvers and adapt them if a nullspace correction is needed */ 22630fccc4e9SStefano Zampini if (pcbddc->NullSpace || pcbddc->dbg_flag) { 2264684f6988SStefano Zampini if (pcbddc->dbg_flag) { 2265684f6988SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 2266684f6988SStefano Zampini ierr = PetscViewerASCIISynchronizedAllow(pcbddc->dbg_viewer,PETSC_TRUE);CHKERRQ(ierr); 2267684f6988SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 2268684f6988SStefano Zampini } 2269684f6988SStefano Zampini if (dirichlet) { /* Dirichlet */ 22700fccc4e9SStefano Zampini ierr = VecSetRandom(pcis->vec1_D,NULL);CHKERRQ(ierr); 22710fccc4e9SStefano Zampini ierr = MatMult(pcis->A_II,pcis->vec1_D,pcis->vec2_D);CHKERRQ(ierr); 22720fccc4e9SStefano Zampini ierr = KSPSolve(pcbddc->ksp_D,pcis->vec2_D,pcis->vec2_D);CHKERRQ(ierr); 22730fccc4e9SStefano Zampini ierr = VecAXPY(pcis->vec1_D,m_one,pcis->vec2_D);CHKERRQ(ierr); 22740fccc4e9SStefano Zampini ierr = VecNorm(pcis->vec1_D,NORM_INFINITY,&value);CHKERRQ(ierr); 2275304d26faSStefano Zampini /* need to be adapted? */ 2276b8ffe317SStefano Zampini use_exact = (PetscAbsReal(value) > 1.e-4 ? PETSC_FALSE : PETSC_TRUE); 2277b8ffe317SStefano Zampini ierr = MPI_Allreduce(&use_exact,&use_exact_reduced,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 2278b8ffe317SStefano Zampini ierr = PCBDDCSetUseExactDirichlet(pc,use_exact_reduced);CHKERRQ(ierr); 2279304d26faSStefano Zampini /* print info */ 2280304d26faSStefano Zampini if (pcbddc->dbg_flag) { 2281e604994aSStefano 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); 2282304d26faSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 2283304d26faSStefano Zampini } 2284b8ffe317SStefano Zampini if (pcbddc->NullSpace && !use_exact_reduced && !pcbddc->switch_static) { 2285298c0119SStefano Zampini ierr = PCBDDCNullSpaceAssembleCorrection(pc,PETSC_TRUE,pcis->is_I_local);CHKERRQ(ierr); 2286304d26faSStefano Zampini } 2287684f6988SStefano Zampini } 2288684f6988SStefano Zampini if (neumann) { /* Neumann */ 22896816873aSStefano Zampini ierr = KSPGetOperators(pcbddc->ksp_R,&A_RR,NULL);CHKERRQ(ierr); 22900fccc4e9SStefano Zampini ierr = VecSetRandom(pcbddc->vec1_R,NULL);CHKERRQ(ierr); 22910fccc4e9SStefano Zampini ierr = MatMult(A_RR,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 22920fccc4e9SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec2_R,pcbddc->vec2_R);CHKERRQ(ierr); 22930fccc4e9SStefano Zampini ierr = VecAXPY(pcbddc->vec1_R,m_one,pcbddc->vec2_R);CHKERRQ(ierr); 22940fccc4e9SStefano Zampini ierr = VecNorm(pcbddc->vec1_R,NORM_INFINITY,&value);CHKERRQ(ierr); 2295304d26faSStefano Zampini /* need to be adapted? */ 2296b8ffe317SStefano Zampini use_exact = (PetscAbsReal(value) > 1.e-4 ? PETSC_FALSE : PETSC_TRUE); 2297b8ffe317SStefano Zampini ierr = MPI_Allreduce(&use_exact,&use_exact_reduced,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 2298304d26faSStefano Zampini /* print info */ 2299304d26faSStefano Zampini if (pcbddc->dbg_flag) { 2300e604994aSStefano 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); 2301304d26faSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 2302304d26faSStefano Zampini } 2303b8ffe317SStefano Zampini if (pcbddc->NullSpace && !use_exact_reduced) { /* is it the right logic? */ 2304298c0119SStefano Zampini ierr = PCBDDCNullSpaceAssembleCorrection(pc,PETSC_FALSE,pcbddc->is_R_local);CHKERRQ(ierr); 2305304d26faSStefano Zampini } 23060fccc4e9SStefano Zampini } 2307684f6988SStefano Zampini } 2308304d26faSStefano Zampini PetscFunctionReturn(0); 2309304d26faSStefano Zampini } 2310304d26faSStefano Zampini 2311304d26faSStefano Zampini #undef __FUNCT__ 2312ba15a52eSStefano Zampini #define __FUNCT__ "PCBDDCSolveSubstructureCorrection" 231380677318SStefano Zampini static PetscErrorCode PCBDDCSolveSubstructureCorrection(PC pc, Vec inout_B, Vec inout_D, PetscBool applytranspose) 2314674ae819SStefano Zampini { 2315674ae819SStefano Zampini PetscErrorCode ierr; 2316674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 2317be83ff47SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 2318674ae819SStefano Zampini 2319674ae819SStefano Zampini PetscFunctionBegin; 2320be83ff47SStefano Zampini if (!sub_schurs->reuse_mumps) { 232180677318SStefano Zampini ierr = VecSet(pcbddc->vec1_R,0.);CHKERRQ(ierr); 232220c7b377SStefano Zampini } 232380677318SStefano Zampini if (!pcbddc->switch_static) { 232480677318SStefano Zampini if (applytranspose && pcbddc->local_auxmat1) { 232580677318SStefano Zampini ierr = MatMultTranspose(pcbddc->local_auxmat2,inout_B,pcbddc->vec1_C);CHKERRQ(ierr); 232680677318SStefano Zampini ierr = MatMultTransposeAdd(pcbddc->local_auxmat1,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr); 232720c7b377SStefano Zampini } 2328be83ff47SStefano Zampini if (!sub_schurs->reuse_mumps) { 232980677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 233080677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 233120c7b377SStefano Zampini } else { 2332be83ff47SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 2333be83ff47SStefano Zampini 233453892102SStefano Zampini ierr = VecScatterBegin(reuse_mumps->correction_scatter_B,inout_B,reuse_mumps->rhs_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 233553892102SStefano Zampini ierr = VecScatterEnd(reuse_mumps->correction_scatter_B,inout_B,reuse_mumps->rhs_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 233620c7b377SStefano Zampini } 2337be83ff47SStefano Zampini } else { 233880677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 233980677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 234080677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,inout_D,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 234180677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,inout_D,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 234280677318SStefano Zampini if (applytranspose && pcbddc->local_auxmat1) { 234380677318SStefano Zampini ierr = MatMultTranspose(pcbddc->local_auxmat2,pcbddc->vec1_R,pcbddc->vec1_C);CHKERRQ(ierr); 234480677318SStefano Zampini ierr = MatMultTransposeAdd(pcbddc->local_auxmat1,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr); 234580677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 234680677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2347674ae819SStefano Zampini } 2348674ae819SStefano Zampini } 2349be83ff47SStefano Zampini if (!sub_schurs->reuse_mumps) { 235080677318SStefano Zampini if (applytranspose) { 235180677318SStefano Zampini ierr = KSPSolveTranspose(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 235280677318SStefano Zampini } else { 235380677318SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 235480677318SStefano Zampini } 235553892102SStefano Zampini #if defined(PETSC_HAVE_MUMPS) 2356be83ff47SStefano Zampini } else { 2357be83ff47SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 2358be83ff47SStefano Zampini 2359be83ff47SStefano Zampini if (applytranspose) { 236053892102SStefano Zampini ierr = MatMumpsSolveSchurComplementTranspose(reuse_mumps->F,reuse_mumps->rhs_B,reuse_mumps->sol_B);CHKERRQ(ierr); 2361be83ff47SStefano Zampini } else { 236253892102SStefano Zampini ierr = MatMumpsSolveSchurComplement(reuse_mumps->F,reuse_mumps->rhs_B,reuse_mumps->sol_B);CHKERRQ(ierr); 2363be83ff47SStefano Zampini } 236453892102SStefano Zampini #endif 2365be83ff47SStefano Zampini } 236680677318SStefano Zampini ierr = VecSet(inout_B,0.);CHKERRQ(ierr); 236780677318SStefano Zampini if (!pcbddc->switch_static) { 2368be83ff47SStefano Zampini if (!sub_schurs->reuse_mumps) { 236980677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 237080677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2371be83ff47SStefano Zampini } else { 2372be83ff47SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 2373be83ff47SStefano Zampini 237453892102SStefano Zampini ierr = VecScatterBegin(reuse_mumps->correction_scatter_B,reuse_mumps->sol_B,inout_B,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 237553892102SStefano Zampini ierr = VecScatterEnd(reuse_mumps->correction_scatter_B,reuse_mumps->sol_B,inout_B,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2376be83ff47SStefano Zampini } 237780677318SStefano Zampini if (!applytranspose && pcbddc->local_auxmat1) { 237880677318SStefano Zampini ierr = MatMult(pcbddc->local_auxmat1,inout_B,pcbddc->vec1_C);CHKERRQ(ierr); 237980677318SStefano Zampini ierr = MatMultAdd(pcbddc->local_auxmat2,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr); 238080677318SStefano Zampini } 238180677318SStefano Zampini } else { 238280677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 238380677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 238480677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 238580677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 238680677318SStefano Zampini if (!applytranspose && pcbddc->local_auxmat1) { 238780677318SStefano Zampini ierr = MatMult(pcbddc->local_auxmat1,inout_B,pcbddc->vec1_C);CHKERRQ(ierr); 238880677318SStefano Zampini ierr = MatMultAdd(pcbddc->local_auxmat2,pcbddc->vec1_C,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 238980677318SStefano Zampini } 239080677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 239180677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 239280677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 239380677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2394674ae819SStefano Zampini } 2395674ae819SStefano Zampini PetscFunctionReturn(0); 2396674ae819SStefano Zampini } 2397674ae819SStefano Zampini 2398dc359a40SStefano Zampini /* parameter apply transpose determines if the interface preconditioner should be applied transposed or not */ 2399674ae819SStefano Zampini #undef __FUNCT__ 2400674ae819SStefano Zampini #define __FUNCT__ "PCBDDCApplyInterfacePreconditioner" 2401dc359a40SStefano Zampini PetscErrorCode PCBDDCApplyInterfacePreconditioner(PC pc, PetscBool applytranspose) 2402674ae819SStefano Zampini { 2403674ae819SStefano Zampini PetscErrorCode ierr; 2404674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 2405674ae819SStefano Zampini PC_IS* pcis = (PC_IS*) (pc->data); 2406674ae819SStefano Zampini const PetscScalar zero = 0.0; 2407674ae819SStefano Zampini 2408674ae819SStefano Zampini PetscFunctionBegin; 2409dc359a40SStefano Zampini /* Application of PSI^T or PHI^T (depending on applytranspose, see comment above) */ 2410dc359a40SStefano Zampini if (applytranspose) { 2411674ae819SStefano Zampini ierr = MatMultTranspose(pcbddc->coarse_phi_B,pcis->vec1_B,pcbddc->vec1_P);CHKERRQ(ierr); 24128eeda7d8SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultTransposeAdd(pcbddc->coarse_phi_D,pcis->vec1_D,pcbddc->vec1_P,pcbddc->vec1_P);CHKERRQ(ierr); } 2413dc359a40SStefano Zampini } else { 2414674ae819SStefano Zampini ierr = MatMultTranspose(pcbddc->coarse_psi_B,pcis->vec1_B,pcbddc->vec1_P);CHKERRQ(ierr); 2415674ae819SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultTransposeAdd(pcbddc->coarse_psi_D,pcis->vec1_D,pcbddc->vec1_P,pcbddc->vec1_P);CHKERRQ(ierr); } 241615aaf578SStefano Zampini } 2417efc2fbd9SStefano Zampini 2418efc2fbd9SStefano Zampini /* add p0 to the last value of vec1_P holding the coarse dof relative to p0 */ 2419*4f1b2e48SStefano Zampini if (pcbddc->benign_n) { 2420efc2fbd9SStefano Zampini PetscScalar *array; 2421*4f1b2e48SStefano Zampini PetscInt j; 2422efc2fbd9SStefano Zampini 2423efc2fbd9SStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 2424*4f1b2e48SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) array[pcbddc->local_primal_size-pcbddc->benign_n+j] += pcbddc->benign_p0[j]; 2425efc2fbd9SStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 2426efc2fbd9SStefano Zampini } 2427efc2fbd9SStefano Zampini 242812edc857SStefano Zampini /* start communications from local primal nodes to rhs of coarse solver */ 242912edc857SStefano Zampini ierr = VecSet(pcbddc->coarse_vec,zero);CHKERRQ(ierr); 243012edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataBegin(pc,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 243112edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataEnd(pc,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 243212edc857SStefano Zampini 24339f00e9b4SStefano Zampini /* Coarse solution -> rhs and sol updated inside PCBDDCScattarCoarseDataBegin/End */ 243412edc857SStefano Zampini /* TODO remove null space when doing multilevel */ 243512edc857SStefano Zampini if (pcbddc->coarse_ksp) { 243651694757SStefano Zampini Mat coarse_mat; 2437964fefecSStefano Zampini Vec rhs,sol; 243851694757SStefano Zampini MatNullSpace nullsp; 2439964fefecSStefano Zampini 2440964fefecSStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&rhs);CHKERRQ(ierr); 2441964fefecSStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&sol);CHKERRQ(ierr); 244251694757SStefano Zampini ierr = KSPGetOperators(pcbddc->coarse_ksp,&coarse_mat,NULL);CHKERRQ(ierr); 244351694757SStefano Zampini ierr = MatGetNullSpace(coarse_mat,&nullsp);CHKERRQ(ierr); 244451694757SStefano Zampini if (nullsp) { 244551694757SStefano Zampini ierr = MatNullSpaceRemove(nullsp,rhs);CHKERRQ(ierr); 244651694757SStefano Zampini } 244712edc857SStefano Zampini if (applytranspose) { 2448964fefecSStefano Zampini ierr = KSPSolveTranspose(pcbddc->coarse_ksp,rhs,sol);CHKERRQ(ierr); 244912edc857SStefano Zampini } else { 2450964fefecSStefano Zampini ierr = KSPSolve(pcbddc->coarse_ksp,rhs,sol);CHKERRQ(ierr); 245112edc857SStefano Zampini } 245251694757SStefano Zampini if (nullsp) { 245351694757SStefano Zampini ierr = MatNullSpaceRemove(nullsp,sol);CHKERRQ(ierr); 245451694757SStefano Zampini } 245512edc857SStefano Zampini } 2456674ae819SStefano Zampini 2457674ae819SStefano Zampini /* Local solution on R nodes */ 245880677318SStefano Zampini if (pcis->n) { /* in/out pcbddc->vec1_B,pcbddc->vec1_D */ 245980677318SStefano Zampini ierr = PCBDDCSolveSubstructureCorrection(pc,pcis->vec1_B,pcis->vec1_D,applytranspose);CHKERRQ(ierr); 24609f00e9b4SStefano Zampini } 2461674ae819SStefano Zampini 24629f00e9b4SStefano Zampini /* communications from coarse sol to local primal nodes */ 24639f00e9b4SStefano Zampini ierr = PCBDDCScatterCoarseDataBegin(pc,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 246412edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataEnd(pc,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2465674ae819SStefano Zampini 2466674ae819SStefano Zampini /* Sum contributions from two levels */ 2467dc359a40SStefano Zampini if (applytranspose) { 2468dc359a40SStefano Zampini ierr = MatMultAdd(pcbddc->coarse_psi_B,pcbddc->vec1_P,pcis->vec1_B,pcis->vec1_B);CHKERRQ(ierr); 2469dc359a40SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultAdd(pcbddc->coarse_psi_D,pcbddc->vec1_P,pcis->vec1_D,pcis->vec1_D);CHKERRQ(ierr); } 2470dc359a40SStefano Zampini } else { 2471674ae819SStefano Zampini ierr = MatMultAdd(pcbddc->coarse_phi_B,pcbddc->vec1_P,pcis->vec1_B,pcis->vec1_B);CHKERRQ(ierr); 24728eeda7d8SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultAdd(pcbddc->coarse_phi_D,pcbddc->vec1_P,pcis->vec1_D,pcis->vec1_D);CHKERRQ(ierr); } 2473dc359a40SStefano Zampini } 2474efc2fbd9SStefano Zampini /* store p0 */ 2475*4f1b2e48SStefano Zampini if (pcbddc->benign_n) { 2476efc2fbd9SStefano Zampini PetscScalar *array; 2477*4f1b2e48SStefano Zampini PetscInt j; 2478efc2fbd9SStefano Zampini 2479efc2fbd9SStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 2480*4f1b2e48SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) pcbddc->benign_p0[j] = array[pcbddc->local_primal_size-pcbddc->benign_n+j]; 2481efc2fbd9SStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 2482efc2fbd9SStefano Zampini } 2483674ae819SStefano Zampini PetscFunctionReturn(0); 2484674ae819SStefano Zampini } 2485674ae819SStefano Zampini 2486674ae819SStefano Zampini #undef __FUNCT__ 2487674ae819SStefano Zampini #define __FUNCT__ "PCBDDCScatterCoarseDataBegin" 248812edc857SStefano Zampini PetscErrorCode PCBDDCScatterCoarseDataBegin(PC pc,InsertMode imode, ScatterMode smode) 2489674ae819SStefano Zampini { 2490674ae819SStefano Zampini PetscErrorCode ierr; 2491674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 249258da7f69SStefano Zampini PetscScalar *array; 249312edc857SStefano Zampini Vec from,to; 2494674ae819SStefano Zampini 2495674ae819SStefano Zampini PetscFunctionBegin; 249612edc857SStefano Zampini if (smode == SCATTER_REVERSE) { /* from global to local -> get data from coarse solution */ 249712edc857SStefano Zampini from = pcbddc->coarse_vec; 249812edc857SStefano Zampini to = pcbddc->vec1_P; 249912edc857SStefano Zampini if (pcbddc->coarse_ksp) { /* get array from coarse processes */ 250012edc857SStefano Zampini Vec tvec; 250158da7f69SStefano Zampini 250258da7f69SStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 250358da7f69SStefano Zampini ierr = VecResetArray(tvec);CHKERRQ(ierr); 250412edc857SStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 250558da7f69SStefano Zampini ierr = VecGetArray(tvec,&array);CHKERRQ(ierr); 250658da7f69SStefano Zampini ierr = VecPlaceArray(from,array);CHKERRQ(ierr); 250758da7f69SStefano Zampini ierr = VecRestoreArray(tvec,&array);CHKERRQ(ierr); 250812edc857SStefano Zampini } 250912edc857SStefano Zampini } else { /* from local to global -> put data in coarse right hand side */ 251012edc857SStefano Zampini from = pcbddc->vec1_P; 251112edc857SStefano Zampini to = pcbddc->coarse_vec; 251212edc857SStefano Zampini } 251312edc857SStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,from,to,imode,smode);CHKERRQ(ierr); 2514674ae819SStefano Zampini PetscFunctionReturn(0); 2515674ae819SStefano Zampini } 2516674ae819SStefano Zampini 2517674ae819SStefano Zampini #undef __FUNCT__ 2518674ae819SStefano Zampini #define __FUNCT__ "PCBDDCScatterCoarseDataEnd" 251912edc857SStefano Zampini PetscErrorCode PCBDDCScatterCoarseDataEnd(PC pc, InsertMode imode, ScatterMode smode) 2520674ae819SStefano Zampini { 2521674ae819SStefano Zampini PetscErrorCode ierr; 2522674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 252358da7f69SStefano Zampini PetscScalar *array; 252412edc857SStefano Zampini Vec from,to; 2525674ae819SStefano Zampini 2526674ae819SStefano Zampini PetscFunctionBegin; 252712edc857SStefano Zampini if (smode == SCATTER_REVERSE) { /* from global to local -> get data from coarse solution */ 252812edc857SStefano Zampini from = pcbddc->coarse_vec; 252912edc857SStefano Zampini to = pcbddc->vec1_P; 253012edc857SStefano Zampini } else { /* from local to global -> put data in coarse right hand side */ 253112edc857SStefano Zampini from = pcbddc->vec1_P; 253212edc857SStefano Zampini to = pcbddc->coarse_vec; 253312edc857SStefano Zampini } 253412edc857SStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,from,to,imode,smode);CHKERRQ(ierr); 253512edc857SStefano Zampini if (smode == SCATTER_FORWARD) { 253612edc857SStefano Zampini if (pcbddc->coarse_ksp) { /* get array from coarse processes */ 253712edc857SStefano Zampini Vec tvec; 253858da7f69SStefano Zampini 253912edc857SStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 254058da7f69SStefano Zampini ierr = VecGetArray(to,&array);CHKERRQ(ierr); 254158da7f69SStefano Zampini ierr = VecPlaceArray(tvec,array);CHKERRQ(ierr); 254258da7f69SStefano Zampini ierr = VecRestoreArray(to,&array);CHKERRQ(ierr); 254358da7f69SStefano Zampini } 254458da7f69SStefano Zampini } else { 254558da7f69SStefano Zampini if (pcbddc->coarse_ksp) { /* restore array of pcbddc->coarse_vec */ 254658da7f69SStefano Zampini ierr = VecResetArray(from);CHKERRQ(ierr); 254712edc857SStefano Zampini } 254812edc857SStefano Zampini } 2549674ae819SStefano Zampini PetscFunctionReturn(0); 2550674ae819SStefano Zampini } 2551674ae819SStefano Zampini 2552984c4197SStefano Zampini /* uncomment for testing purposes */ 2553984c4197SStefano Zampini /* #define PETSC_MISSING_LAPACK_GESVD 1 */ 2554674ae819SStefano Zampini #undef __FUNCT__ 2555674ae819SStefano Zampini #define __FUNCT__ "PCBDDCConstraintsSetUp" 2556674ae819SStefano Zampini PetscErrorCode PCBDDCConstraintsSetUp(PC pc) 2557674ae819SStefano Zampini { 2558674ae819SStefano Zampini PetscErrorCode ierr; 2559674ae819SStefano Zampini PC_IS* pcis = (PC_IS*)(pc->data); 2560674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 2561674ae819SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 2562984c4197SStefano Zampini /* one and zero */ 2563984c4197SStefano Zampini PetscScalar one=1.0,zero=0.0; 2564984c4197SStefano Zampini /* space to store constraints and their local indices */ 25659162d606SStefano Zampini PetscScalar *constraints_data; 25669162d606SStefano Zampini PetscInt *constraints_idxs,*constraints_idxs_B; 25679162d606SStefano Zampini PetscInt *constraints_idxs_ptr,*constraints_data_ptr; 25689162d606SStefano Zampini PetscInt *constraints_n; 2569984c4197SStefano Zampini /* iterators */ 2570b3d85658SStefano Zampini PetscInt i,j,k,total_counts,total_counts_cc,cum; 2571984c4197SStefano Zampini /* BLAS integers */ 2572e310c8b4SStefano Zampini PetscBLASInt lwork,lierr; 2573e310c8b4SStefano Zampini PetscBLASInt Blas_N,Blas_M,Blas_K,Blas_one=1; 2574c4303822SStefano Zampini PetscBLASInt Blas_LDA,Blas_LDB,Blas_LDC; 2575727cdba6SStefano Zampini /* reuse */ 25760e6343abSStefano Zampini PetscInt olocal_primal_size,olocal_primal_size_cc; 25770e6343abSStefano Zampini PetscInt *olocal_primal_ref_node,*olocal_primal_ref_mult; 2578984c4197SStefano Zampini /* change of basis */ 2579b3d85658SStefano Zampini PetscBool qr_needed; 25809162d606SStefano Zampini PetscBT change_basis,qr_needed_idx; 2581984c4197SStefano Zampini /* auxiliary stuff */ 258264efe560SStefano Zampini PetscInt *nnz,*is_indices; 25838a0068c3SStefano Zampini PetscInt ncc; 2584984c4197SStefano Zampini /* some quantities */ 258545a1bb75SStefano Zampini PetscInt n_vertices,total_primal_vertices,valid_constraints; 2586a58a30b4SStefano Zampini PetscInt size_of_constraint,max_size_of_constraint=0,max_constraints,temp_constraints; 2587984c4197SStefano Zampini 2588674ae819SStefano Zampini PetscFunctionBegin; 25898e61c736SStefano Zampini /* Destroy Mat objects computed previously */ 25908e61c736SStefano Zampini ierr = MatDestroy(&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 25918e61c736SStefano Zampini ierr = MatDestroy(&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 2592088faed8SStefano Zampini /* save info on constraints from previous setup (if any) */ 2593088faed8SStefano Zampini olocal_primal_size = pcbddc->local_primal_size; 25940e6343abSStefano Zampini olocal_primal_size_cc = pcbddc->local_primal_size_cc; 25950e6343abSStefano Zampini ierr = PetscMalloc2(olocal_primal_size_cc,&olocal_primal_ref_node,olocal_primal_size_cc,&olocal_primal_ref_mult);CHKERRQ(ierr); 25960e6343abSStefano Zampini ierr = PetscMemcpy(olocal_primal_ref_node,pcbddc->local_primal_ref_node,olocal_primal_size_cc*sizeof(PetscInt));CHKERRQ(ierr); 25970e6343abSStefano Zampini ierr = PetscMemcpy(olocal_primal_ref_mult,pcbddc->local_primal_ref_mult,olocal_primal_size_cc*sizeof(PetscInt));CHKERRQ(ierr); 25980e6343abSStefano Zampini ierr = PetscFree2(pcbddc->local_primal_ref_node,pcbddc->local_primal_ref_mult);CHKERRQ(ierr); 2599088faed8SStefano Zampini ierr = PetscFree(pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 2600cf5a6209SStefano Zampini 2601cf5a6209SStefano Zampini /* print some info */ 2602cf5a6209SStefano Zampini if (pcbddc->dbg_flag) { 2603cf5a6209SStefano Zampini IS vertices; 2604cf5a6209SStefano Zampini PetscInt nv,nedges,nfaces; 2605cf5a6209SStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,&nfaces,NULL,&nedges,NULL,&vertices);CHKERRQ(ierr); 2606cf5a6209SStefano Zampini ierr = ISGetSize(vertices,&nv);CHKERRQ(ierr); 2607cf5a6209SStefano Zampini ierr = ISDestroy(&vertices);CHKERRQ(ierr); 2608cf5a6209SStefano Zampini ierr = PetscViewerASCIISynchronizedAllow(pcbddc->dbg_viewer,PETSC_TRUE);CHKERRQ(ierr); 2609cf5a6209SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"--------------------------------------------------------------\n");CHKERRQ(ierr); 2610cf5a6209SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate vertices (%d)\n",PetscGlobalRank,nv,pcbddc->use_vertices);CHKERRQ(ierr); 2611fd14bc51SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate edges (%d)\n",PetscGlobalRank,nedges,pcbddc->use_edges);CHKERRQ(ierr); 2612fd14bc51SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate faces (%d)\n",PetscGlobalRank,nfaces,pcbddc->use_faces);CHKERRQ(ierr); 2613cf5a6209SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 2614cf5a6209SStefano Zampini } 2615cf5a6209SStefano Zampini 2616cf5a6209SStefano Zampini if (!pcbddc->adaptive_selection) { 26179162d606SStefano Zampini IS ISForVertices,*ISForFaces,*ISForEdges; 2618cf5a6209SStefano Zampini MatNullSpace nearnullsp; 2619cf5a6209SStefano Zampini const Vec *nearnullvecs; 2620cf5a6209SStefano Zampini Vec *localnearnullsp; 2621cf5a6209SStefano Zampini PetscScalar *array; 2622cf5a6209SStefano Zampini PetscInt n_ISForFaces,n_ISForEdges,nnsp_size; 2623cf5a6209SStefano Zampini PetscBool nnsp_has_cnst; 2624674ae819SStefano Zampini /* LAPACK working arrays for SVD or POD */ 2625b3d85658SStefano Zampini PetscBool skip_lapack,boolforchange; 2626674ae819SStefano Zampini PetscScalar *work; 2627674ae819SStefano Zampini PetscReal *singular_vals; 2628674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX) 2629674ae819SStefano Zampini PetscReal *rwork; 2630674ae819SStefano Zampini #endif 2631674ae819SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 2632674ae819SStefano Zampini PetscScalar *temp_basis,*correlation_mat; 2633674ae819SStefano Zampini #else 2634964fefecSStefano Zampini PetscBLASInt dummy_int=1; 2635964fefecSStefano Zampini PetscScalar dummy_scalar=1.; 2636674ae819SStefano Zampini #endif 2637674ae819SStefano Zampini 2638674ae819SStefano Zampini /* Get index sets for faces, edges and vertices from graph */ 2639d06fc5fdSStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,&n_ISForFaces,&ISForFaces,&n_ISForEdges,&ISForEdges,&ISForVertices);CHKERRQ(ierr); 2640d06fc5fdSStefano Zampini /* free unneeded index sets */ 2641d06fc5fdSStefano Zampini if (!pcbddc->use_vertices) { 2642d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForVertices);CHKERRQ(ierr); 2643674ae819SStefano Zampini } 2644d06fc5fdSStefano Zampini if (!pcbddc->use_edges) { 2645d06fc5fdSStefano Zampini for (i=0;i<n_ISForEdges;i++) { 2646d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForEdges[i]);CHKERRQ(ierr); 2647d06fc5fdSStefano Zampini } 2648d06fc5fdSStefano Zampini ierr = PetscFree(ISForEdges);CHKERRQ(ierr); 2649d06fc5fdSStefano Zampini n_ISForEdges = 0; 2650d06fc5fdSStefano Zampini } 2651d06fc5fdSStefano Zampini if (!pcbddc->use_faces) { 2652d06fc5fdSStefano Zampini for (i=0;i<n_ISForFaces;i++) { 2653d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForFaces[i]);CHKERRQ(ierr); 2654d06fc5fdSStefano Zampini } 2655d06fc5fdSStefano Zampini ierr = PetscFree(ISForFaces);CHKERRQ(ierr); 2656d06fc5fdSStefano Zampini n_ISForFaces = 0; 2657d06fc5fdSStefano Zampini } 265870022509SStefano Zampini 265970022509SStefano Zampini #if defined(PETSC_USE_DEBUG) 266070022509SStefano Zampini /* HACK: when solving singular problems not using vertices, a change of basis is mandatory. 266170022509SStefano Zampini Also use_change_of_basis should be consistent among processors */ 266270022509SStefano Zampini if (pcbddc->NullSpace) { 266370022509SStefano Zampini PetscBool tbool[2],gbool[2]; 266470022509SStefano Zampini 266570022509SStefano Zampini if (!ISForVertices && !pcbddc->user_ChangeOfBasisMatrix) { 2666b8ffe317SStefano Zampini pcbddc->use_change_of_basis = PETSC_TRUE; 2667d06fc5fdSStefano Zampini if (!ISForEdges) { 2668d06fc5fdSStefano Zampini pcbddc->use_change_on_faces = PETSC_TRUE; 2669d06fc5fdSStefano Zampini } 2670b8ffe317SStefano Zampini } 2671d06fc5fdSStefano Zampini tbool[0] = pcbddc->use_change_of_basis; 2672d06fc5fdSStefano Zampini tbool[1] = pcbddc->use_change_on_faces; 2673d06fc5fdSStefano Zampini ierr = MPI_Allreduce(tbool,gbool,2,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 2674d06fc5fdSStefano Zampini pcbddc->use_change_of_basis = gbool[0]; 2675d06fc5fdSStefano Zampini pcbddc->use_change_on_faces = gbool[1]; 267698a51de6SStefano Zampini } 267770022509SStefano Zampini #endif 267808122e43SStefano Zampini 2679674ae819SStefano Zampini /* check if near null space is attached to global mat */ 2680674ae819SStefano Zampini ierr = MatGetNearNullSpace(pc->pmat,&nearnullsp);CHKERRQ(ierr); 2681674ae819SStefano Zampini if (nearnullsp) { 2682674ae819SStefano Zampini ierr = MatNullSpaceGetVecs(nearnullsp,&nnsp_has_cnst,&nnsp_size,&nearnullvecs);CHKERRQ(ierr); 2683f4ddd8eeSStefano Zampini /* remove any stored info */ 2684f4ddd8eeSStefano Zampini ierr = MatNullSpaceDestroy(&pcbddc->onearnullspace);CHKERRQ(ierr); 2685f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->onearnullvecs_state);CHKERRQ(ierr); 2686f4ddd8eeSStefano Zampini /* store information for BDDC solver reuse */ 2687f4ddd8eeSStefano Zampini ierr = PetscObjectReference((PetscObject)nearnullsp);CHKERRQ(ierr); 2688f4ddd8eeSStefano Zampini pcbddc->onearnullspace = nearnullsp; 2689473ba861SJed Brown ierr = PetscMalloc1(nnsp_size,&pcbddc->onearnullvecs_state);CHKERRQ(ierr); 2690f4ddd8eeSStefano Zampini for (i=0;i<nnsp_size;i++) { 2691f4ddd8eeSStefano Zampini ierr = PetscObjectStateGet((PetscObject)nearnullvecs[i],&pcbddc->onearnullvecs_state[i]);CHKERRQ(ierr); 2692f4ddd8eeSStefano Zampini } 2693984c4197SStefano Zampini } else { /* if near null space is not provided BDDC uses constants by default */ 2694984c4197SStefano Zampini nnsp_size = 0; 2695674ae819SStefano Zampini nnsp_has_cnst = PETSC_TRUE; 2696674ae819SStefano Zampini } 2697984c4197SStefano Zampini /* get max number of constraints on a single cc */ 2698984c4197SStefano Zampini max_constraints = nnsp_size; 2699984c4197SStefano Zampini if (nnsp_has_cnst) max_constraints++; 2700984c4197SStefano Zampini 2701674ae819SStefano Zampini /* 2702674ae819SStefano Zampini Evaluate maximum storage size needed by the procedure 27039162d606SStefano Zampini - Indices for connected component i stored at "constraints_idxs + constraints_idxs_ptr[i]" 27049162d606SStefano Zampini - Values for constraints on connected component i stored at "constraints_data + constraints_data_ptr[i]" 27059162d606SStefano Zampini There can be multiple constraints per connected component 2706674ae819SStefano Zampini */ 2707674ae819SStefano Zampini n_vertices = 0; 2708674ae819SStefano Zampini if (ISForVertices) { 2709674ae819SStefano Zampini ierr = ISGetSize(ISForVertices,&n_vertices);CHKERRQ(ierr); 2710674ae819SStefano Zampini } 27119162d606SStefano Zampini ncc = n_vertices+n_ISForFaces+n_ISForEdges; 27129162d606SStefano Zampini ierr = PetscMalloc3(ncc+1,&constraints_idxs_ptr,ncc+1,&constraints_data_ptr,ncc,&constraints_n);CHKERRQ(ierr); 27139162d606SStefano Zampini 27149162d606SStefano Zampini total_counts = n_ISForFaces+n_ISForEdges; 27159162d606SStefano Zampini total_counts *= max_constraints; 2716674ae819SStefano Zampini total_counts += n_vertices; 27174641a718SStefano Zampini ierr = PetscBTCreate(total_counts,&change_basis);CHKERRQ(ierr); 27189162d606SStefano Zampini 2719674ae819SStefano Zampini total_counts = 0; 2720674ae819SStefano Zampini max_size_of_constraint = 0; 2721674ae819SStefano Zampini for (i=0;i<n_ISForEdges+n_ISForFaces;i++) { 27229162d606SStefano Zampini IS used_is; 2723674ae819SStefano Zampini if (i<n_ISForEdges) { 27249162d606SStefano Zampini used_is = ISForEdges[i]; 2725674ae819SStefano Zampini } else { 27269162d606SStefano Zampini used_is = ISForFaces[i-n_ISForEdges]; 2727674ae819SStefano Zampini } 27289162d606SStefano Zampini ierr = ISGetSize(used_is,&j);CHKERRQ(ierr); 2729674ae819SStefano Zampini total_counts += j; 2730674ae819SStefano Zampini max_size_of_constraint = PetscMax(j,max_size_of_constraint); 2731674ae819SStefano Zampini } 27329162d606SStefano 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); 27339162d606SStefano Zampini 2734984c4197SStefano Zampini /* get local part of global near null space vectors */ 2735785e854fSJed Brown ierr = PetscMalloc1(nnsp_size,&localnearnullsp);CHKERRQ(ierr); 2736984c4197SStefano Zampini for (k=0;k<nnsp_size;k++) { 2737984c4197SStefano Zampini ierr = VecDuplicate(pcis->vec1_N,&localnearnullsp[k]);CHKERRQ(ierr); 2738e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,nearnullvecs[k],localnearnullsp[k],INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2739e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,nearnullvecs[k],localnearnullsp[k],INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2740984c4197SStefano Zampini } 2741674ae819SStefano Zampini 2742242a89d7SStefano Zampini /* whether or not to skip lapack calls */ 2743242a89d7SStefano Zampini skip_lapack = PETSC_TRUE; 2744a773dcb8SStefano Zampini if (n_ISForFaces+n_ISForEdges && max_constraints > 1 && !pcbddc->use_nnsp_true) skip_lapack = PETSC_FALSE; 2745242a89d7SStefano Zampini 2746984c4197SStefano Zampini /* First we issue queries to allocate optimal workspace for LAPACKgesvd (or LAPACKsyev if SVD is missing) */ 2747a773dcb8SStefano Zampini if (!skip_lapack) { 2748674ae819SStefano Zampini PetscScalar temp_work; 2749911cabfeSStefano Zampini 2750674ae819SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 2751984c4197SStefano Zampini /* Proper Orthogonal Decomposition (POD) using the snapshot method */ 2752785e854fSJed Brown ierr = PetscMalloc1(max_constraints*max_constraints,&correlation_mat);CHKERRQ(ierr); 2753785e854fSJed Brown ierr = PetscMalloc1(max_constraints,&singular_vals);CHKERRQ(ierr); 2754785e854fSJed Brown ierr = PetscMalloc1(max_size_of_constraint*max_constraints,&temp_basis);CHKERRQ(ierr); 2755674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX) 2756785e854fSJed Brown ierr = PetscMalloc1(3*max_constraints,&rwork);CHKERRQ(ierr); 2757674ae819SStefano Zampini #endif 2758674ae819SStefano Zampini /* now we evaluate the optimal workspace using query with lwork=-1 */ 2759c8244a33SStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_N);CHKERRQ(ierr); 2760c8244a33SStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_LDA);CHKERRQ(ierr); 2761674ae819SStefano Zampini lwork = -1; 2762674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 2763674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 2764c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,&temp_work,&lwork,&lierr)); 2765674ae819SStefano Zampini #else 2766c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,&temp_work,&lwork,rwork,&lierr)); 2767674ae819SStefano Zampini #endif 2768674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2769984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to SYEV Lapack routine %d",(int)lierr); 2770674ae819SStefano Zampini #else /* on missing GESVD */ 2771674ae819SStefano Zampini /* SVD */ 2772674ae819SStefano Zampini PetscInt max_n,min_n; 2773674ae819SStefano Zampini max_n = max_size_of_constraint; 2774984c4197SStefano Zampini min_n = max_constraints; 2775984c4197SStefano Zampini if (max_size_of_constraint < max_constraints) { 2776674ae819SStefano Zampini min_n = max_size_of_constraint; 2777984c4197SStefano Zampini max_n = max_constraints; 2778674ae819SStefano Zampini } 2779785e854fSJed Brown ierr = PetscMalloc1(min_n,&singular_vals);CHKERRQ(ierr); 2780674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX) 2781785e854fSJed Brown ierr = PetscMalloc1(5*min_n,&rwork);CHKERRQ(ierr); 2782674ae819SStefano Zampini #endif 2783674ae819SStefano Zampini /* now we evaluate the optimal workspace using query with lwork=-1 */ 2784674ae819SStefano Zampini lwork = -1; 2785e310c8b4SStefano Zampini ierr = PetscBLASIntCast(max_n,&Blas_M);CHKERRQ(ierr); 2786e310c8b4SStefano Zampini ierr = PetscBLASIntCast(min_n,&Blas_N);CHKERRQ(ierr); 2787b7d8b9f8SStefano Zampini ierr = PetscBLASIntCast(max_n,&Blas_LDA);CHKERRQ(ierr); 2788674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 2789674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 27909162d606SStefano 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)); 2791674ae819SStefano Zampini #else 27929162d606SStefano 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)); 2793674ae819SStefano Zampini #endif 2794674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2795984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to GESVD Lapack routine %d",(int)lierr); 2796984c4197SStefano Zampini #endif /* on missing GESVD */ 2797674ae819SStefano Zampini /* Allocate optimal workspace */ 2798674ae819SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(temp_work),&lwork);CHKERRQ(ierr); 2799854ce69bSBarry Smith ierr = PetscMalloc1(lwork,&work);CHKERRQ(ierr); 2800674ae819SStefano Zampini } 2801674ae819SStefano Zampini /* Now we can loop on constraining sets */ 2802674ae819SStefano Zampini total_counts = 0; 28039162d606SStefano Zampini constraints_idxs_ptr[0] = 0; 28049162d606SStefano Zampini constraints_data_ptr[0] = 0; 2805674ae819SStefano Zampini /* vertices */ 28069162d606SStefano Zampini if (n_vertices) { 2807674ae819SStefano Zampini ierr = ISGetIndices(ISForVertices,(const PetscInt**)&is_indices);CHKERRQ(ierr); 28089162d606SStefano Zampini ierr = PetscMemcpy(constraints_idxs,is_indices,n_vertices*sizeof(PetscInt));CHKERRQ(ierr); 2809674ae819SStefano Zampini for (i=0;i<n_vertices;i++) { 28109162d606SStefano Zampini constraints_n[total_counts] = 1; 28119162d606SStefano Zampini constraints_data[total_counts] = 1.0; 28129162d606SStefano Zampini constraints_idxs_ptr[total_counts+1] = constraints_idxs_ptr[total_counts]+1; 28139162d606SStefano Zampini constraints_data_ptr[total_counts+1] = constraints_data_ptr[total_counts]+1; 2814674ae819SStefano Zampini total_counts++; 2815674ae819SStefano Zampini } 2816674ae819SStefano Zampini ierr = ISRestoreIndices(ISForVertices,(const PetscInt**)&is_indices);CHKERRQ(ierr); 2817674ae819SStefano Zampini n_vertices = total_counts; 2818674ae819SStefano Zampini } 2819984c4197SStefano Zampini 2820674ae819SStefano Zampini /* edges and faces */ 28219162d606SStefano Zampini total_counts_cc = total_counts; 2822911cabfeSStefano Zampini for (ncc=0;ncc<n_ISForEdges+n_ISForFaces;ncc++) { 28239162d606SStefano Zampini IS used_is; 28249162d606SStefano Zampini PetscBool idxs_copied = PETSC_FALSE; 28259162d606SStefano Zampini 2826911cabfeSStefano Zampini if (ncc<n_ISForEdges) { 28279162d606SStefano Zampini used_is = ISForEdges[ncc]; 2828984c4197SStefano Zampini boolforchange = pcbddc->use_change_of_basis; /* change or not the basis on the edge */ 2829674ae819SStefano Zampini } else { 28309162d606SStefano Zampini used_is = ISForFaces[ncc-n_ISForEdges]; 2831984c4197SStefano Zampini boolforchange = (PetscBool)(pcbddc->use_change_of_basis && pcbddc->use_change_on_faces); /* change or not the basis on the face */ 2832674ae819SStefano Zampini } 2833674ae819SStefano Zampini temp_constraints = 0; /* zero the number of constraints I have on this conn comp */ 28349162d606SStefano Zampini 28359162d606SStefano Zampini ierr = ISGetSize(used_is,&size_of_constraint);CHKERRQ(ierr); 28369162d606SStefano Zampini ierr = ISGetIndices(used_is,(const PetscInt**)&is_indices);CHKERRQ(ierr); 2837984c4197SStefano Zampini /* change of basis should not be performed on local periodic nodes */ 2838984c4197SStefano Zampini if (pcbddc->mat_graph->mirrors && pcbddc->mat_graph->mirrors[is_indices[0]]) boolforchange = PETSC_FALSE; 2839674ae819SStefano Zampini if (nnsp_has_cnst) { 28405b08dc53SStefano Zampini PetscScalar quad_value; 28419162d606SStefano Zampini 28429162d606SStefano Zampini ierr = PetscMemcpy(constraints_idxs + constraints_idxs_ptr[total_counts_cc],is_indices,size_of_constraint*sizeof(PetscInt));CHKERRQ(ierr); 28439162d606SStefano Zampini idxs_copied = PETSC_TRUE; 28449162d606SStefano Zampini 2845a773dcb8SStefano Zampini if (!pcbddc->use_nnsp_true) { 2846674ae819SStefano Zampini quad_value = (PetscScalar)(1.0/PetscSqrtReal((PetscReal)size_of_constraint)); 2847a773dcb8SStefano Zampini } else { 2848a773dcb8SStefano Zampini quad_value = 1.0; 2849a773dcb8SStefano Zampini } 2850674ae819SStefano Zampini for (j=0;j<size_of_constraint;j++) { 28519162d606SStefano Zampini constraints_data[constraints_data_ptr[total_counts_cc]+j] = quad_value; 2852674ae819SStefano Zampini } 28539162d606SStefano Zampini temp_constraints++; 2854674ae819SStefano Zampini total_counts++; 2855674ae819SStefano Zampini } 2856674ae819SStefano Zampini for (k=0;k<nnsp_size;k++) { 2857984c4197SStefano Zampini PetscReal real_value; 28589162d606SStefano Zampini PetscScalar *ptr_to_data; 28599162d606SStefano Zampini 2860984c4197SStefano Zampini ierr = VecGetArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 28619162d606SStefano Zampini ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]+temp_constraints*size_of_constraint]; 2862674ae819SStefano Zampini for (j=0;j<size_of_constraint;j++) { 28639162d606SStefano Zampini ptr_to_data[j] = array[is_indices[j]]; 2864674ae819SStefano Zampini } 2865984c4197SStefano Zampini ierr = VecRestoreArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 2866984c4197SStefano Zampini /* check if array is null on the connected component */ 2867e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 28689162d606SStefano Zampini PetscStackCallBLAS("BLASasum",real_value = BLASasum_(&Blas_N,ptr_to_data,&Blas_one)); 28695b08dc53SStefano Zampini if (real_value > 0.0) { /* keep indices and values */ 2870674ae819SStefano Zampini temp_constraints++; 2871674ae819SStefano Zampini total_counts++; 28729162d606SStefano Zampini if (!idxs_copied) { 28739162d606SStefano Zampini ierr = PetscMemcpy(constraints_idxs + constraints_idxs_ptr[total_counts_cc],is_indices,size_of_constraint*sizeof(PetscInt));CHKERRQ(ierr); 28749162d606SStefano Zampini idxs_copied = PETSC_TRUE; 2875674ae819SStefano Zampini } 2876674ae819SStefano Zampini } 28779162d606SStefano Zampini } 28789162d606SStefano Zampini ierr = ISRestoreIndices(used_is,(const PetscInt**)&is_indices);CHKERRQ(ierr); 287945a1bb75SStefano Zampini valid_constraints = temp_constraints; 2880eb97c9d2SStefano Zampini if (!pcbddc->use_nnsp_true && temp_constraints) { 2881a773dcb8SStefano Zampini if (temp_constraints == 1) { /* just normalize the constraint */ 28829162d606SStefano Zampini PetscScalar norm,*ptr_to_data; 28839162d606SStefano Zampini 28849162d606SStefano Zampini ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]]; 2885a773dcb8SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 28869162d606SStefano Zampini PetscStackCallBLAS("BLASdot",norm = BLASdot_(&Blas_N,ptr_to_data,&Blas_one,ptr_to_data,&Blas_one)); 2887a773dcb8SStefano Zampini norm = 1.0/PetscSqrtReal(PetscRealPart(norm)); 28889162d606SStefano Zampini PetscStackCallBLAS("BLASscal",BLASscal_(&Blas_N,&norm,ptr_to_data,&Blas_one)); 2889a773dcb8SStefano Zampini } else { /* perform SVD */ 2890984c4197SStefano Zampini PetscReal tol = 1.0e-8; /* tolerance for retaining eigenmodes */ 28919162d606SStefano Zampini PetscScalar *ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]]; 2892674ae819SStefano Zampini 2893674ae819SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 2894984c4197SStefano Zampini /* SVD: Y = U*S*V^H -> U (eigenvectors of Y*Y^H) = Y*V*(S)^\dag 2895984c4197SStefano Zampini POD: Y^H*Y = V*D*V^H, D = S^H*S -> U = Y*V*D^(-1/2) 2896984c4197SStefano Zampini -> When PETSC_USE_COMPLEX and PETSC_MISSING_LAPACK_GESVD are defined 2897984c4197SStefano Zampini the constraints basis will differ (by a complex factor with absolute value equal to 1) 2898984c4197SStefano Zampini from that computed using LAPACKgesvd 2899984c4197SStefano Zampini -> This is due to a different computation of eigenvectors in LAPACKheev 2900984c4197SStefano Zampini -> The quality of the POD-computed basis will be the same */ 2901984c4197SStefano Zampini ierr = PetscMemzero(correlation_mat,temp_constraints*temp_constraints*sizeof(PetscScalar));CHKERRQ(ierr); 2902674ae819SStefano Zampini /* Store upper triangular part of correlation matrix */ 2903e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 2904984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 2905674ae819SStefano Zampini for (j=0;j<temp_constraints;j++) { 2906674ae819SStefano Zampini for (k=0;k<j+1;k++) { 29079162d606SStefano 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)); 2908674ae819SStefano Zampini } 2909674ae819SStefano Zampini } 2910e310c8b4SStefano Zampini /* compute eigenvalues and eigenvectors of correlation matrix */ 2911e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 2912e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_LDA);CHKERRQ(ierr); 2913674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 2914c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,work,&lwork,&lierr)); 2915674ae819SStefano Zampini #else 2916c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,work,&lwork,rwork,&lierr)); 2917674ae819SStefano Zampini #endif 2918674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2919984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYEV Lapack routine %d",(int)lierr); 2920984c4197SStefano Zampini /* retain eigenvalues greater than tol: note that LAPACKsyev gives eigs in ascending order */ 2921674ae819SStefano Zampini j = 0; 2922984c4197SStefano Zampini while (j < temp_constraints && singular_vals[j] < tol) j++; 2923674ae819SStefano Zampini total_counts = total_counts-j; 292445a1bb75SStefano Zampini valid_constraints = temp_constraints-j; 2925e310c8b4SStefano Zampini /* scale and copy POD basis into used quadrature memory */ 2926c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 2927c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 2928c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_K);CHKERRQ(ierr); 2929c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 2930c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_LDB);CHKERRQ(ierr); 2931c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDC);CHKERRQ(ierr); 2932674ae819SStefano Zampini if (j<temp_constraints) { 2933984c4197SStefano Zampini PetscInt ii; 2934984c4197SStefano Zampini for (k=j;k<temp_constraints;k++) singular_vals[k] = 1.0/PetscSqrtReal(singular_vals[k]); 2935674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 29369162d606SStefano 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)); 2937674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2938984c4197SStefano Zampini for (k=0;k<temp_constraints-j;k++) { 2939674ae819SStefano Zampini for (ii=0;ii<size_of_constraint;ii++) { 29409162d606SStefano 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]; 2941674ae819SStefano Zampini } 2942674ae819SStefano Zampini } 2943674ae819SStefano Zampini } 2944674ae819SStefano Zampini #else /* on missing GESVD */ 2945e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 2946e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 2947b7d8b9f8SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 2948674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 2949674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 29509162d606SStefano 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)); 2951674ae819SStefano Zampini #else 29529162d606SStefano 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)); 2953674ae819SStefano Zampini #endif 2954984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GESVD Lapack routine %d",(int)lierr); 2955674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2956984c4197SStefano Zampini /* retain eigenvalues greater than tol: note that LAPACKgesvd gives eigs in descending order */ 2957e310c8b4SStefano Zampini k = temp_constraints; 2958e310c8b4SStefano Zampini if (k > size_of_constraint) k = size_of_constraint; 2959674ae819SStefano Zampini j = 0; 2960e310c8b4SStefano Zampini while (j < k && singular_vals[k-j-1] < tol) j++; 296145a1bb75SStefano Zampini valid_constraints = k-j; 2962911cabfeSStefano Zampini total_counts = total_counts-temp_constraints+valid_constraints; 2963984c4197SStefano Zampini #endif /* on missing GESVD */ 2964674ae819SStefano Zampini } 2965a773dcb8SStefano Zampini } 29669162d606SStefano Zampini /* update pointers information */ 29679162d606SStefano Zampini if (valid_constraints) { 29689162d606SStefano Zampini constraints_n[total_counts_cc] = valid_constraints; 29699162d606SStefano Zampini constraints_idxs_ptr[total_counts_cc+1] = constraints_idxs_ptr[total_counts_cc]+size_of_constraint; 29709162d606SStefano Zampini constraints_data_ptr[total_counts_cc+1] = constraints_data_ptr[total_counts_cc]+size_of_constraint*valid_constraints; 29719162d606SStefano Zampini /* set change_of_basis flag */ 297245a1bb75SStefano Zampini if (boolforchange) { 2973b3d85658SStefano Zampini PetscBTSet(change_basis,total_counts_cc); 29749162d606SStefano Zampini } 2975b3d85658SStefano Zampini total_counts_cc++; 297645a1bb75SStefano Zampini } 297745a1bb75SStefano Zampini } 2978984c4197SStefano Zampini /* free workspace */ 29798f1c130eSStefano Zampini if (!skip_lapack) { 2980984c4197SStefano Zampini ierr = PetscFree(work);CHKERRQ(ierr); 2981984c4197SStefano Zampini #if defined(PETSC_USE_COMPLEX) 2982984c4197SStefano Zampini ierr = PetscFree(rwork);CHKERRQ(ierr); 2983984c4197SStefano Zampini #endif 2984984c4197SStefano Zampini ierr = PetscFree(singular_vals);CHKERRQ(ierr); 2985984c4197SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 2986984c4197SStefano Zampini ierr = PetscFree(correlation_mat);CHKERRQ(ierr); 2987984c4197SStefano Zampini ierr = PetscFree(temp_basis);CHKERRQ(ierr); 2988984c4197SStefano Zampini #endif 2989984c4197SStefano Zampini } 2990984c4197SStefano Zampini for (k=0;k<nnsp_size;k++) { 2991984c4197SStefano Zampini ierr = VecDestroy(&localnearnullsp[k]);CHKERRQ(ierr); 2992984c4197SStefano Zampini } 2993984c4197SStefano Zampini ierr = PetscFree(localnearnullsp);CHKERRQ(ierr); 2994cf5a6209SStefano Zampini /* free index sets of faces, edges and vertices */ 2995cf5a6209SStefano Zampini for (i=0;i<n_ISForFaces;i++) { 2996cf5a6209SStefano Zampini ierr = ISDestroy(&ISForFaces[i]);CHKERRQ(ierr); 2997cf5a6209SStefano Zampini } 2998cf5a6209SStefano Zampini if (n_ISForFaces) { 2999cf5a6209SStefano Zampini ierr = PetscFree(ISForFaces);CHKERRQ(ierr); 3000cf5a6209SStefano Zampini } 3001cf5a6209SStefano Zampini for (i=0;i<n_ISForEdges;i++) { 3002cf5a6209SStefano Zampini ierr = ISDestroy(&ISForEdges[i]);CHKERRQ(ierr); 3003cf5a6209SStefano Zampini } 3004cf5a6209SStefano Zampini if (n_ISForEdges) { 3005cf5a6209SStefano Zampini ierr = PetscFree(ISForEdges);CHKERRQ(ierr); 3006cf5a6209SStefano Zampini } 3007cf5a6209SStefano Zampini ierr = ISDestroy(&ISForVertices);CHKERRQ(ierr); 300808122e43SStefano Zampini } else { 300908122e43SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 3010984c4197SStefano Zampini 301108122e43SStefano Zampini total_counts = 0; 301208122e43SStefano Zampini n_vertices = 0; 3013d62866d3SStefano Zampini if (sub_schurs->is_vertices && pcbddc->use_vertices) { 3014d62866d3SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_vertices,&n_vertices);CHKERRQ(ierr); 301508122e43SStefano Zampini } 301608122e43SStefano Zampini max_constraints = 0; 30179162d606SStefano Zampini total_counts_cc = 0; 301808122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs+n_vertices;i++) { 301908122e43SStefano Zampini total_counts += pcbddc->adaptive_constraints_n[i]; 30209162d606SStefano Zampini if (pcbddc->adaptive_constraints_n[i]) total_counts_cc++; 302108122e43SStefano Zampini max_constraints = PetscMax(max_constraints,pcbddc->adaptive_constraints_n[i]); 302208122e43SStefano Zampini } 30239162d606SStefano Zampini constraints_idxs_ptr = pcbddc->adaptive_constraints_idxs_ptr; 30249162d606SStefano Zampini constraints_data_ptr = pcbddc->adaptive_constraints_data_ptr; 30259162d606SStefano Zampini constraints_idxs = pcbddc->adaptive_constraints_idxs; 30269162d606SStefano Zampini constraints_data = pcbddc->adaptive_constraints_data; 302774d5cdf7SStefano Zampini /* constraints_n differs from pcbddc->adaptive_constraints_n */ 30289162d606SStefano Zampini ierr = PetscMalloc1(total_counts_cc,&constraints_n);CHKERRQ(ierr); 30299162d606SStefano Zampini total_counts_cc = 0; 30309162d606SStefano Zampini for (i=0;i<sub_schurs->n_subs+n_vertices;i++) { 30319162d606SStefano Zampini if (pcbddc->adaptive_constraints_n[i]) { 30329162d606SStefano Zampini constraints_n[total_counts_cc++] = pcbddc->adaptive_constraints_n[i]; 303308122e43SStefano Zampini } 303408122e43SStefano Zampini } 30359162d606SStefano Zampini #if 0 30369162d606SStefano Zampini printf("Found %d totals (%d)\n",total_counts_cc,total_counts); 30379162d606SStefano Zampini for (i=0;i<total_counts_cc;i++) { 30389162d606SStefano Zampini printf("const %d, start %d",i,constraints_idxs_ptr[i]); 30399162d606SStefano Zampini printf(" end %d:\n",constraints_idxs_ptr[i+1]); 30409162d606SStefano Zampini for (j=constraints_idxs_ptr[i];j<constraints_idxs_ptr[i+1];j++) { 30419162d606SStefano Zampini printf(" %d",constraints_idxs[j]); 30429162d606SStefano Zampini } 30439162d606SStefano Zampini printf("\n"); 30449162d606SStefano Zampini printf("number of cc: %d\n",constraints_n[i]); 30459162d606SStefano Zampini } 30461b968477SStefano Zampini for (i=0;i<n_vertices;i++) { 30478bec7fa6SStefano Zampini PetscPrintf(PETSC_COMM_SELF,"[%d] vertex %d, n %d\n",PetscGlobalRank,i,pcbddc->adaptive_constraints_n[i]); 30481b968477SStefano Zampini } 30491b968477SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 30508bec7fa6SStefano 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]); 30511b968477SStefano Zampini } 305208122e43SStefano Zampini #endif 305308122e43SStefano Zampini 30548bec7fa6SStefano Zampini max_size_of_constraint = 0; 30559162d606SStefano 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]); 30569162d606SStefano Zampini ierr = PetscMalloc1(constraints_idxs_ptr[total_counts_cc],&constraints_idxs_B);CHKERRQ(ierr); 305708122e43SStefano Zampini /* Change of basis */ 3058b3d85658SStefano Zampini ierr = PetscBTCreate(total_counts_cc,&change_basis);CHKERRQ(ierr); 305908122e43SStefano Zampini if (pcbddc->use_change_of_basis) { 306008122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 306108122e43SStefano Zampini if (PetscBTLookup(sub_schurs->is_edge,i) || pcbddc->use_change_on_faces) { 3062b3d85658SStefano Zampini ierr = PetscBTSet(change_basis,i+n_vertices);CHKERRQ(ierr); 306308122e43SStefano Zampini } 306408122e43SStefano Zampini } 306508122e43SStefano Zampini } 306608122e43SStefano Zampini } 3067984c4197SStefano Zampini pcbddc->local_primal_size = total_counts; 3068*4f1b2e48SStefano Zampini ierr = PetscMalloc1(pcbddc->local_primal_size+pcbddc->benign_n,&pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 306908122e43SStefano Zampini 30709162d606SStefano Zampini /* map constraints_idxs in boundary numbering */ 30719162d606SStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,constraints_idxs_ptr[total_counts_cc],constraints_idxs,&i,constraints_idxs_B);CHKERRQ(ierr); 30729162d606SStefano Zampini if (i != constraints_idxs_ptr[total_counts_cc]) { 30739162d606SStefano 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); 307408122e43SStefano Zampini } 3075674ae819SStefano Zampini 3076674ae819SStefano Zampini /* Create constraint matrix */ 3077674ae819SStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 307816f15bc4SStefano Zampini ierr = MatSetType(pcbddc->ConstraintMatrix,MATAIJ);CHKERRQ(ierr); 3079984c4197SStefano Zampini ierr = MatSetSizes(pcbddc->ConstraintMatrix,pcbddc->local_primal_size,pcis->n,pcbddc->local_primal_size,pcis->n);CHKERRQ(ierr); 3080984c4197SStefano Zampini 3081984c4197SStefano Zampini /* find primal_dofs: subdomain corners plus dofs selected as primal after change of basis */ 3082a717540cSStefano Zampini /* determine if a QR strategy is needed for change of basis */ 3083a717540cSStefano Zampini qr_needed = PETSC_FALSE; 308474d5cdf7SStefano Zampini ierr = PetscBTCreate(total_counts_cc,&qr_needed_idx);CHKERRQ(ierr); 3085984c4197SStefano Zampini total_primal_vertices=0; 3086b3d85658SStefano Zampini pcbddc->local_primal_size_cc = 0; 30879162d606SStefano Zampini for (i=0;i<total_counts_cc;i++) { 30889162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 3089984c4197SStefano Zampini if (size_of_constraint == 1) { 30909162d606SStefano Zampini pcbddc->primal_indices_local_idxs[total_primal_vertices++] = constraints_idxs[constraints_idxs_ptr[i]]; 3091b3d85658SStefano Zampini pcbddc->local_primal_size_cc += 1; 309264efe560SStefano Zampini } else if (PetscBTLookup(change_basis,i)) { 30939162d606SStefano Zampini for (k=0;k<constraints_n[i];k++) { 30949162d606SStefano Zampini pcbddc->primal_indices_local_idxs[total_primal_vertices++] = constraints_idxs[constraints_idxs_ptr[i]+k]; 3095a717540cSStefano Zampini } 3096b3d85658SStefano Zampini pcbddc->local_primal_size_cc += constraints_n[i]; 309774d5cdf7SStefano Zampini if (constraints_n[i] > 1 || pcbddc->use_qr_single || pcbddc->faster_deluxe) { 3098a717540cSStefano Zampini PetscBTSet(qr_needed_idx,i); 3099a717540cSStefano Zampini qr_needed = PETSC_TRUE; 3100a717540cSStefano Zampini } 3101fa434743SStefano Zampini } else { 3102b3d85658SStefano Zampini pcbddc->local_primal_size_cc += 1; 3103fa434743SStefano Zampini } 3104a717540cSStefano Zampini } 3105b371cd4fSStefano Zampini /* note that the local variable n_vertices used below stores the number of pointwise constraints */ 3106b371cd4fSStefano Zampini pcbddc->n_vertices = total_primal_vertices; 3107674ae819SStefano Zampini /* permute indices in order to have a sorted set of vertices */ 310870022509SStefano Zampini ierr = PetscSortInt(total_primal_vertices,pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 3109b3d85658SStefano Zampini 3110*4f1b2e48SStefano Zampini ierr = PetscMalloc2(pcbddc->local_primal_size_cc+pcbddc->benign_n,&pcbddc->local_primal_ref_node,pcbddc->local_primal_size_cc+pcbddc->benign_n,&pcbddc->local_primal_ref_mult);CHKERRQ(ierr); 31110e6343abSStefano Zampini ierr = PetscMemcpy(pcbddc->local_primal_ref_node,pcbddc->primal_indices_local_idxs,total_primal_vertices*sizeof(PetscInt));CHKERRQ(ierr); 31120e6343abSStefano Zampini for (i=0;i<total_primal_vertices;i++) pcbddc->local_primal_ref_mult[i] = 1; 3113984c4197SStefano Zampini 3114984c4197SStefano Zampini /* nonzero structure of constraint matrix */ 311574d5cdf7SStefano Zampini /* and get reference dof for local constraints */ 3116785e854fSJed Brown ierr = PetscMalloc1(pcbddc->local_primal_size,&nnz);CHKERRQ(ierr); 3117984c4197SStefano Zampini for (i=0;i<total_primal_vertices;i++) nnz[i] = 1; 311874d5cdf7SStefano Zampini 3119984c4197SStefano Zampini j = total_primal_vertices; 312074d5cdf7SStefano Zampini total_counts = total_primal_vertices; 3121b3d85658SStefano Zampini cum = total_primal_vertices; 31229162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 31234641a718SStefano Zampini if (!PetscBTLookup(change_basis,i)) { 3124b3d85658SStefano Zampini pcbddc->local_primal_ref_node[cum] = constraints_idxs[constraints_idxs_ptr[i]]; 3125b3d85658SStefano Zampini pcbddc->local_primal_ref_mult[cum] = constraints_n[i]; 3126b3d85658SStefano Zampini cum++; 31279162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 312874d5cdf7SStefano Zampini for (k=0;k<constraints_n[i];k++) { 312974d5cdf7SStefano Zampini pcbddc->primal_indices_local_idxs[total_counts++] = constraints_idxs[constraints_idxs_ptr[i]+k]; 313074d5cdf7SStefano Zampini nnz[j+k] = size_of_constraint; 313174d5cdf7SStefano Zampini } 31329162d606SStefano Zampini j += constraints_n[i]; 3133674ae819SStefano Zampini } 3134674ae819SStefano Zampini } 3135674ae819SStefano Zampini ierr = MatSeqAIJSetPreallocation(pcbddc->ConstraintMatrix,0,nnz);CHKERRQ(ierr); 3136674ae819SStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 3137088faed8SStefano Zampini 3138674ae819SStefano Zampini /* set values in constraint matrix */ 3139984c4197SStefano Zampini for (i=0;i<total_primal_vertices;i++) { 31400e6343abSStefano Zampini ierr = MatSetValue(pcbddc->ConstraintMatrix,i,pcbddc->local_primal_ref_node[i],1.0,INSERT_VALUES);CHKERRQ(ierr); 3141674ae819SStefano Zampini } 3142984c4197SStefano Zampini total_counts = total_primal_vertices; 31439162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 31444641a718SStefano Zampini if (!PetscBTLookup(change_basis,i)) { 31459162d606SStefano Zampini PetscInt *cols; 31469162d606SStefano Zampini 31479162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 31489162d606SStefano Zampini cols = constraints_idxs+constraints_idxs_ptr[i]; 31499162d606SStefano Zampini for (k=0;k<constraints_n[i];k++) { 31509162d606SStefano Zampini PetscInt row = total_counts+k; 31519162d606SStefano Zampini PetscScalar *vals; 31529162d606SStefano Zampini 31539162d606SStefano Zampini vals = constraints_data+constraints_data_ptr[i]+k*size_of_constraint; 31549162d606SStefano Zampini ierr = MatSetValues(pcbddc->ConstraintMatrix,1,&row,size_of_constraint,cols,vals,INSERT_VALUES);CHKERRQ(ierr); 31559162d606SStefano Zampini } 31569162d606SStefano Zampini total_counts += constraints_n[i]; 3157674ae819SStefano Zampini } 3158674ae819SStefano Zampini } 3159674ae819SStefano Zampini /* assembling */ 3160674ae819SStefano Zampini ierr = MatAssemblyBegin(pcbddc->ConstraintMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3161674ae819SStefano Zampini ierr = MatAssemblyEnd(pcbddc->ConstraintMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3162088faed8SStefano Zampini 3163984c4197SStefano Zampini /* 316445a1bb75SStefano Zampini ierr = PetscViewerSetFormat(PETSC_VIEWER_STDOUT_SELF,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 3165984c4197SStefano Zampini ierr = MatView(pcbddc->ConstraintMatrix,(PetscViewer)0);CHKERRQ(ierr); 3166984c4197SStefano Zampini */ 3167674ae819SStefano Zampini /* Create matrix for change of basis. We don't need it in case pcbddc->use_change_of_basis is FALSE */ 3168674ae819SStefano Zampini if (pcbddc->use_change_of_basis) { 3169026de310SStefano Zampini /* dual and primal dofs on a single cc */ 3170984c4197SStefano Zampini PetscInt dual_dofs,primal_dofs; 3171984c4197SStefano Zampini /* working stuff for GEQRF */ 317281d9aea3SBarry Smith PetscScalar *qr_basis,*qr_tau = NULL,*qr_work,lqr_work_t; 3173984c4197SStefano Zampini PetscBLASInt lqr_work; 3174984c4197SStefano Zampini /* working stuff for UNGQR */ 3175984c4197SStefano Zampini PetscScalar *gqr_work,lgqr_work_t; 3176984c4197SStefano Zampini PetscBLASInt lgqr_work; 3177984c4197SStefano Zampini /* working stuff for TRTRS */ 3178984c4197SStefano Zampini PetscScalar *trs_rhs; 31793f08241aSStefano Zampini PetscBLASInt Blas_NRHS; 3180984c4197SStefano Zampini /* pointers for values insertion into change of basis matrix */ 3181984c4197SStefano Zampini PetscInt *start_rows,*start_cols; 3182984c4197SStefano Zampini PetscScalar *start_vals; 3183984c4197SStefano Zampini /* working stuff for values insertion */ 31844641a718SStefano Zampini PetscBT is_primal; 318564efe560SStefano Zampini PetscInt *aux_primal_numbering_B; 3186906d46d4SStefano Zampini /* matrix sizes */ 3187906d46d4SStefano Zampini PetscInt global_size,local_size; 3188906d46d4SStefano Zampini /* temporary change of basis */ 3189906d46d4SStefano Zampini Mat localChangeOfBasisMatrix; 3190cf5a6209SStefano Zampini /* extra space for debugging */ 3191cf5a6209SStefano Zampini PetscScalar *dbg_work; 3192984c4197SStefano Zampini 3193906d46d4SStefano Zampini /* local temporary change of basis acts on local interfaces -> dimension is n_B x n_B */ 3194906d46d4SStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&localChangeOfBasisMatrix);CHKERRQ(ierr); 319516f15bc4SStefano Zampini ierr = MatSetType(localChangeOfBasisMatrix,MATAIJ);CHKERRQ(ierr); 3196bbb9e6c6SStefano Zampini ierr = MatSetSizes(localChangeOfBasisMatrix,pcis->n,pcis->n,pcis->n,pcis->n);CHKERRQ(ierr); 3197906d46d4SStefano Zampini /* nonzeros for local mat */ 3198bbb9e6c6SStefano Zampini ierr = PetscMalloc1(pcis->n,&nnz);CHKERRQ(ierr); 3199bbb9e6c6SStefano Zampini for (i=0;i<pcis->n;i++) nnz[i]=1; 32009162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 3201a717540cSStefano Zampini if (PetscBTLookup(change_basis,i)) { 32029162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 3203a717540cSStefano Zampini if (PetscBTLookup(qr_needed_idx,i)) { 32049162d606SStefano Zampini for (j=0;j<size_of_constraint;j++) nnz[constraints_idxs[constraints_idxs_ptr[i]+j]] = size_of_constraint; 3205a717540cSStefano Zampini } else { 32069162d606SStefano Zampini nnz[constraints_idxs[constraints_idxs_ptr[i]]] = size_of_constraint; 32079162d606SStefano Zampini for (j=1;j<size_of_constraint;j++) nnz[constraints_idxs[constraints_idxs_ptr[i]+j]] = 2; 3208a717540cSStefano Zampini } 3209a717540cSStefano Zampini } 3210a717540cSStefano Zampini } 3211906d46d4SStefano Zampini ierr = MatSeqAIJSetPreallocation(localChangeOfBasisMatrix,0,nnz);CHKERRQ(ierr); 3212bbb9e6c6SStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 3213a717540cSStefano Zampini /* Set initial identity in the matrix */ 3214bbb9e6c6SStefano Zampini for (i=0;i<pcis->n;i++) { 3215906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,i,i,1.0,INSERT_VALUES);CHKERRQ(ierr); 3216a717540cSStefano Zampini } 3217a717540cSStefano Zampini 3218a717540cSStefano Zampini if (pcbddc->dbg_flag) { 3219a717540cSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"--------------------------------------------------------------\n");CHKERRQ(ierr); 3220a717540cSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Checking change of basis computation for subdomain %04d\n",PetscGlobalRank);CHKERRQ(ierr); 3221a717540cSStefano Zampini } 3222a717540cSStefano Zampini 3223a717540cSStefano Zampini 3224a717540cSStefano Zampini /* Now we loop on the constraints which need a change of basis */ 3225a717540cSStefano Zampini /* 3226a717540cSStefano Zampini Change of basis matrix is evaluated similarly to the FIRST APPROACH in 3227a717540cSStefano Zampini Klawonn and Widlund, Dual-primal FETI-DP methods for linear elasticity, (see Sect 6.2.1) 3228a717540cSStefano Zampini 3229a6b551f4SStefano Zampini Basic blocks of change of basis matrix T computed by 3230a717540cSStefano Zampini 3231a6b551f4SStefano 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) 3232a6b551f4SStefano Zampini 3233a6b551f4SStefano Zampini | 1 0 ... 0 s_1/S | 3234a6b551f4SStefano Zampini | 0 1 ... 0 s_2/S | 3235a717540cSStefano Zampini | ... | 3236a6b551f4SStefano Zampini | 0 ... 1 s_{n-1}/S | 3237a6b551f4SStefano Zampini | -s_1/s_n ... -s_{n-1}/s_n s_n/S | 3238a717540cSStefano Zampini 3239a6b551f4SStefano Zampini with S = \sum_{i=1}^n s_i^2 3240a6b551f4SStefano Zampini NOTE: in the above example, the primal dof is the last one of the edge in LOCAL ordering 3241a6b551f4SStefano Zampini in the current implementation, the primal dof is the first one of the edge in GLOBAL ordering 3242a6b551f4SStefano Zampini 3243a6b551f4SStefano Zampini - QR decomposition of constraints otherwise 3244a717540cSStefano Zampini */ 3245a717540cSStefano Zampini if (qr_needed) { 3246984c4197SStefano Zampini /* space to store Q */ 3247854ce69bSBarry Smith ierr = PetscMalloc1(max_size_of_constraint*max_size_of_constraint,&qr_basis);CHKERRQ(ierr); 3248984c4197SStefano Zampini /* first we issue queries for optimal work */ 32493f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_M);CHKERRQ(ierr); 32503f08241aSStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_N);CHKERRQ(ierr); 32513f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 3252984c4197SStefano Zampini lqr_work = -1; 32533f08241aSStefano Zampini PetscStackCallBLAS("LAPACKgeqrf",LAPACKgeqrf_(&Blas_M,&Blas_N,qr_basis,&Blas_LDA,qr_tau,&lqr_work_t,&lqr_work,&lierr)); 3254984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to GEQRF Lapack routine %d",(int)lierr); 3255984c4197SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lqr_work_t),&lqr_work);CHKERRQ(ierr); 3256785e854fSJed Brown ierr = PetscMalloc1((PetscInt)PetscRealPart(lqr_work_t),&qr_work);CHKERRQ(ierr); 3257984c4197SStefano Zampini lgqr_work = -1; 32583f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_M);CHKERRQ(ierr); 32593f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_N);CHKERRQ(ierr); 32603f08241aSStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_K);CHKERRQ(ierr); 32613f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 32623f08241aSStefano Zampini if (Blas_K>Blas_M) Blas_K=Blas_M; /* adjust just for computing optimal work */ 32633f08241aSStefano Zampini PetscStackCallBLAS("LAPACKungqr",LAPACKungqr_(&Blas_M,&Blas_N,&Blas_K,qr_basis,&Blas_LDA,qr_tau,&lgqr_work_t,&lgqr_work,&lierr)); 3264984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to UNGQR Lapack routine %d",(int)lierr); 3265984c4197SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lgqr_work_t),&lgqr_work);CHKERRQ(ierr); 3266785e854fSJed Brown ierr = PetscMalloc1((PetscInt)PetscRealPart(lgqr_work_t),&gqr_work);CHKERRQ(ierr); 3267984c4197SStefano Zampini /* array to store scaling factors for reflectors */ 3268785e854fSJed Brown ierr = PetscMalloc1(max_constraints,&qr_tau);CHKERRQ(ierr); 3269984c4197SStefano Zampini /* array to store rhs and solution of triangular solver */ 3270785e854fSJed Brown ierr = PetscMalloc1(max_constraints*max_constraints,&trs_rhs);CHKERRQ(ierr); 3271a717540cSStefano Zampini /* allocating workspace for check */ 3272a717540cSStefano Zampini if (pcbddc->dbg_flag) { 3273cf5a6209SStefano Zampini ierr = PetscMalloc1(max_size_of_constraint*(max_constraints+max_size_of_constraint),&dbg_work);CHKERRQ(ierr); 3274a717540cSStefano Zampini } 3275a717540cSStefano Zampini } 3276984c4197SStefano Zampini /* array to store whether a node is primal or not */ 32774641a718SStefano Zampini ierr = PetscBTCreate(pcis->n_B,&is_primal);CHKERRQ(ierr); 3278473ba861SJed Brown ierr = PetscMalloc1(total_primal_vertices,&aux_primal_numbering_B);CHKERRQ(ierr); 32790e6343abSStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,total_primal_vertices,pcbddc->local_primal_ref_node,&i,aux_primal_numbering_B);CHKERRQ(ierr); 328039e2fb2aSStefano Zampini if (i != total_primal_vertices) { 328139e2fb2aSStefano Zampini SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_SUP,"Error in boundary numbering for BDDC vertices! %d != %d\n",total_primal_vertices,i); 32824641a718SStefano Zampini } 328339e2fb2aSStefano Zampini for (i=0;i<total_primal_vertices;i++) { 328439e2fb2aSStefano Zampini ierr = PetscBTSet(is_primal,aux_primal_numbering_B[i]);CHKERRQ(ierr); 328539e2fb2aSStefano Zampini } 328639e2fb2aSStefano Zampini ierr = PetscFree(aux_primal_numbering_B);CHKERRQ(ierr); 3287984c4197SStefano Zampini 3288a717540cSStefano Zampini /* loop on constraints and see whether or not they need a change of basis and compute it */ 32899162d606SStefano Zampini for (total_counts=n_vertices;total_counts<total_counts_cc;total_counts++) { 32909162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[total_counts+1]-constraints_idxs_ptr[total_counts]; 32914641a718SStefano Zampini if (PetscBTLookup(change_basis,total_counts)) { 3292984c4197SStefano Zampini /* get constraint info */ 32939162d606SStefano Zampini primal_dofs = constraints_n[total_counts]; 3294984c4197SStefano Zampini dual_dofs = size_of_constraint-primal_dofs; 3295984c4197SStefano Zampini 3296984c4197SStefano Zampini if (pcbddc->dbg_flag) { 32979162d606SStefano 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); 3298674ae819SStefano Zampini } 3299984c4197SStefano Zampini 3300fa434743SStefano Zampini if (PetscBTLookup(qr_needed_idx,total_counts)) { /* QR */ 3301a717540cSStefano Zampini 3302a717540cSStefano Zampini /* copy quadrature constraints for change of basis check */ 3303a717540cSStefano Zampini if (pcbddc->dbg_flag) { 33049162d606SStefano Zampini ierr = PetscMemcpy(dbg_work,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 3305a717540cSStefano Zampini } 3306984c4197SStefano Zampini /* copy temporary constraints into larger work vector (in order to store all columns of Q) */ 33079162d606SStefano Zampini ierr = PetscMemcpy(qr_basis,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 3308984c4197SStefano Zampini 3309984c4197SStefano Zampini /* compute QR decomposition of constraints */ 33103f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 33113f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 33123f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 3313674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 33143f08241aSStefano Zampini PetscStackCallBLAS("LAPACKgeqrf",LAPACKgeqrf_(&Blas_M,&Blas_N,qr_basis,&Blas_LDA,qr_tau,qr_work,&lqr_work,&lierr)); 3315984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GEQRF Lapack routine %d",(int)lierr); 3316674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 3317984c4197SStefano Zampini 3318984c4197SStefano Zampini /* explictly compute R^-T */ 3319984c4197SStefano Zampini ierr = PetscMemzero(trs_rhs,primal_dofs*primal_dofs*sizeof(*trs_rhs));CHKERRQ(ierr); 3320984c4197SStefano Zampini for (j=0;j<primal_dofs;j++) trs_rhs[j*(primal_dofs+1)] = 1.0; 33213f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 33223f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_NRHS);CHKERRQ(ierr); 33233f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 33243f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDB);CHKERRQ(ierr); 3325984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 33263f08241aSStefano Zampini PetscStackCallBLAS("LAPACKtrtrs",LAPACKtrtrs_("U","T","N",&Blas_N,&Blas_NRHS,qr_basis,&Blas_LDA,trs_rhs,&Blas_LDB,&lierr)); 3327984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in TRTRS Lapack routine %d",(int)lierr); 3328984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 3329984c4197SStefano Zampini 3330a717540cSStefano Zampini /* explicitly compute all columns of Q (Q = [Q1 | Q2] ) overwriting QR factorization in qr_basis */ 33313f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 33323f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 33333f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_K);CHKERRQ(ierr); 33343f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 3335984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 33363f08241aSStefano Zampini PetscStackCallBLAS("LAPACKungqr",LAPACKungqr_(&Blas_M,&Blas_N,&Blas_K,qr_basis,&Blas_LDA,qr_tau,gqr_work,&lgqr_work,&lierr)); 3337984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in UNGQR Lapack routine %d",(int)lierr); 3338984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 3339984c4197SStefano Zampini 3340984c4197SStefano Zampini /* first primal_dofs columns of Q need to be re-scaled in order to be unitary w.r.t constraints 3341984c4197SStefano Zampini i.e. C_{pxn}*Q_{nxn} should be equal to [I_pxp | 0_pxd] (see check below) 3342984c4197SStefano Zampini where n=size_of_constraint, p=primal_dofs, d=dual_dofs (n=p+d), I and 0 identity and null matrix resp. */ 33433f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 33443f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 33453f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_K);CHKERRQ(ierr); 33463f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 33473f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDB);CHKERRQ(ierr); 33483f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDC);CHKERRQ(ierr); 3349984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 33509162d606SStefano 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)); 3351984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 33529162d606SStefano Zampini ierr = PetscMemcpy(qr_basis,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 3353984c4197SStefano Zampini 3354984c4197SStefano Zampini /* insert values in change of basis matrix respecting global ordering of new primal dofs */ 33559162d606SStefano Zampini start_rows = &constraints_idxs[constraints_idxs_ptr[total_counts]]; 3356984c4197SStefano Zampini /* insert cols for primal dofs */ 3357984c4197SStefano Zampini for (j=0;j<primal_dofs;j++) { 3358984c4197SStefano Zampini start_vals = &qr_basis[j*size_of_constraint]; 33599162d606SStefano Zampini start_cols = &constraints_idxs[constraints_idxs_ptr[total_counts]+j]; 3360906d46d4SStefano Zampini ierr = MatSetValues(localChangeOfBasisMatrix,size_of_constraint,start_rows,1,start_cols,start_vals,INSERT_VALUES);CHKERRQ(ierr); 3361984c4197SStefano Zampini } 3362984c4197SStefano Zampini /* insert cols for dual dofs */ 3363984c4197SStefano Zampini for (j=0,k=0;j<dual_dofs;k++) { 33649162d606SStefano Zampini if (!PetscBTLookup(is_primal,constraints_idxs_B[constraints_idxs_ptr[total_counts]+k])) { 3365984c4197SStefano Zampini start_vals = &qr_basis[(primal_dofs+j)*size_of_constraint]; 33669162d606SStefano Zampini start_cols = &constraints_idxs[constraints_idxs_ptr[total_counts]+k]; 3367906d46d4SStefano Zampini ierr = MatSetValues(localChangeOfBasisMatrix,size_of_constraint,start_rows,1,start_cols,start_vals,INSERT_VALUES);CHKERRQ(ierr); 3368984c4197SStefano Zampini j++; 3369674ae819SStefano Zampini } 3370674ae819SStefano Zampini } 3371984c4197SStefano Zampini 3372984c4197SStefano Zampini /* check change of basis */ 3373984c4197SStefano Zampini if (pcbddc->dbg_flag) { 3374984c4197SStefano Zampini PetscInt ii,jj; 3375984c4197SStefano Zampini PetscBool valid_qr=PETSC_TRUE; 3376c4303822SStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_M);CHKERRQ(ierr); 3377c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 3378c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_K);CHKERRQ(ierr); 3379c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 3380c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDB);CHKERRQ(ierr); 3381c4303822SStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDC);CHKERRQ(ierr); 3382984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 3383cf5a6209SStefano 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)); 3384984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 3385984c4197SStefano Zampini for (jj=0;jj<size_of_constraint;jj++) { 3386984c4197SStefano Zampini for (ii=0;ii<primal_dofs;ii++) { 3387cf5a6209SStefano Zampini if (ii != jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]) > 1.e-12) valid_qr = PETSC_FALSE; 3388cf5a6209SStefano 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; 3389674ae819SStefano Zampini } 3390674ae819SStefano Zampini } 3391984c4197SStefano Zampini if (!valid_qr) { 339222d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> wrong change of basis!\n");CHKERRQ(ierr); 3393984c4197SStefano Zampini for (jj=0;jj<size_of_constraint;jj++) { 3394984c4197SStefano Zampini for (ii=0;ii<primal_dofs;ii++) { 3395cf5a6209SStefano Zampini if (ii != jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]) > 1.e-12) { 3396cf5a6209SStefano 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])); 3397674ae819SStefano Zampini } 3398cf5a6209SStefano Zampini if (ii == jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]-1.0) > 1.e-12) { 3399cf5a6209SStefano 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])); 3400984c4197SStefano Zampini } 3401984c4197SStefano Zampini } 3402984c4197SStefano Zampini } 3403674ae819SStefano Zampini } else { 340422d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> right change of basis!\n");CHKERRQ(ierr); 3405674ae819SStefano Zampini } 3406674ae819SStefano Zampini } 3407a717540cSStefano Zampini } else { /* simple transformation block */ 3408a717540cSStefano Zampini PetscInt row,col; 3409a6b551f4SStefano Zampini PetscScalar val,norm; 3410a6b551f4SStefano Zampini 3411a6b551f4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 34129162d606SStefano 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)); 3413a717540cSStefano Zampini for (j=0;j<size_of_constraint;j++) { 34149162d606SStefano Zampini PetscInt row_B = constraints_idxs_B[constraints_idxs_ptr[total_counts]+j]; 34159162d606SStefano Zampini row = constraints_idxs[constraints_idxs_ptr[total_counts]+j]; 3416bbb9e6c6SStefano Zampini if (!PetscBTLookup(is_primal,row_B)) { 34179162d606SStefano Zampini col = constraints_idxs[constraints_idxs_ptr[total_counts]]; 3418906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,row,1.0,INSERT_VALUES);CHKERRQ(ierr); 34199162d606SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,col,constraints_data[constraints_data_ptr[total_counts]+j]/norm,INSERT_VALUES);CHKERRQ(ierr); 3420a717540cSStefano Zampini } else { 3421a717540cSStefano Zampini for (k=0;k<size_of_constraint;k++) { 34229162d606SStefano Zampini col = constraints_idxs[constraints_idxs_ptr[total_counts]+k]; 3423a717540cSStefano Zampini if (row != col) { 34249162d606SStefano Zampini val = -constraints_data[constraints_data_ptr[total_counts]+k]/constraints_data[constraints_data_ptr[total_counts]]; 3425a717540cSStefano Zampini } else { 34269162d606SStefano Zampini val = constraints_data[constraints_data_ptr[total_counts]]/norm; 3427a717540cSStefano Zampini } 3428906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,col,val,INSERT_VALUES);CHKERRQ(ierr); 3429a717540cSStefano Zampini } 3430a717540cSStefano Zampini } 3431a717540cSStefano Zampini } 343298a51de6SStefano Zampini if (pcbddc->dbg_flag) { 343322d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> using standard change of basis\n");CHKERRQ(ierr); 3434a717540cSStefano Zampini } 3435674ae819SStefano Zampini } 3436984c4197SStefano Zampini } else { 3437984c4197SStefano Zampini if (pcbddc->dbg_flag) { 34389162d606SStefano 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); 3439674ae819SStefano Zampini } 3440674ae819SStefano Zampini } 3441674ae819SStefano Zampini } 3442a717540cSStefano Zampini 3443a717540cSStefano Zampini /* free workspace */ 3444a717540cSStefano Zampini if (qr_needed) { 3445984c4197SStefano Zampini if (pcbddc->dbg_flag) { 3446cf5a6209SStefano Zampini ierr = PetscFree(dbg_work);CHKERRQ(ierr); 3447984c4197SStefano Zampini } 3448984c4197SStefano Zampini ierr = PetscFree(trs_rhs);CHKERRQ(ierr); 3449984c4197SStefano Zampini ierr = PetscFree(qr_tau);CHKERRQ(ierr); 3450984c4197SStefano Zampini ierr = PetscFree(qr_work);CHKERRQ(ierr); 3451984c4197SStefano Zampini ierr = PetscFree(gqr_work);CHKERRQ(ierr); 3452984c4197SStefano Zampini ierr = PetscFree(qr_basis);CHKERRQ(ierr); 3453674ae819SStefano Zampini } 3454a717540cSStefano Zampini ierr = PetscBTDestroy(&is_primal);CHKERRQ(ierr); 3455906d46d4SStefano Zampini ierr = MatAssemblyBegin(localChangeOfBasisMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3456906d46d4SStefano Zampini ierr = MatAssemblyEnd(localChangeOfBasisMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3457906d46d4SStefano Zampini 3458906d46d4SStefano Zampini /* assembling of global change of variable */ 3459bbb9e6c6SStefano Zampini { 3460bbb9e6c6SStefano Zampini Mat tmat; 346116f15bc4SStefano Zampini PetscInt bs; 346216f15bc4SStefano Zampini 3463906d46d4SStefano Zampini ierr = VecGetSize(pcis->vec1_global,&global_size);CHKERRQ(ierr); 3464906d46d4SStefano Zampini ierr = VecGetLocalSize(pcis->vec1_global,&local_size);CHKERRQ(ierr); 3465bbb9e6c6SStefano Zampini ierr = MatDuplicate(pc->pmat,MAT_DO_NOT_COPY_VALUES,&tmat);CHKERRQ(ierr); 3466bbb9e6c6SStefano Zampini ierr = MatISSetLocalMat(tmat,localChangeOfBasisMatrix);CHKERRQ(ierr); 3467bbb9e6c6SStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)pc),&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 3468bbb9e6c6SStefano Zampini ierr = MatSetType(pcbddc->ChangeOfBasisMatrix,MATAIJ);CHKERRQ(ierr); 346916f15bc4SStefano Zampini ierr = MatGetBlockSize(pc->pmat,&bs);CHKERRQ(ierr); 347016f15bc4SStefano Zampini ierr = MatSetBlockSize(pcbddc->ChangeOfBasisMatrix,bs);CHKERRQ(ierr); 3471906d46d4SStefano Zampini ierr = MatSetSizes(pcbddc->ChangeOfBasisMatrix,local_size,local_size,global_size,global_size);CHKERRQ(ierr); 3472bbb9e6c6SStefano Zampini ierr = MatISSetMPIXAIJPreallocation_Private(tmat,pcbddc->ChangeOfBasisMatrix,PETSC_TRUE);CHKERRQ(ierr); 3473bbb9e6c6SStefano Zampini ierr = MatISGetMPIXAIJ(tmat,MAT_REUSE_MATRIX,&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 3474bbb9e6c6SStefano Zampini ierr = MatDestroy(&tmat);CHKERRQ(ierr); 3475bbb9e6c6SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 3476bbb9e6c6SStefano Zampini ierr = VecSet(pcis->vec1_N,1.0);CHKERRQ(ierr); 3477e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 3478e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 3479bbb9e6c6SStefano Zampini ierr = VecReciprocal(pcis->vec1_global);CHKERRQ(ierr); 3480bbb9e6c6SStefano Zampini ierr = MatDiagonalScale(pcbddc->ChangeOfBasisMatrix,pcis->vec1_global,NULL);CHKERRQ(ierr); 3481906d46d4SStefano Zampini } 3482906d46d4SStefano Zampini /* check */ 3483906d46d4SStefano Zampini if (pcbddc->dbg_flag) { 3484906d46d4SStefano Zampini PetscReal error; 3485906d46d4SStefano Zampini Vec x,x_change; 3486906d46d4SStefano Zampini 3487906d46d4SStefano Zampini ierr = VecDuplicate(pcis->vec1_global,&x);CHKERRQ(ierr); 3488906d46d4SStefano Zampini ierr = VecDuplicate(pcis->vec1_global,&x_change);CHKERRQ(ierr); 3489906d46d4SStefano Zampini ierr = VecSetRandom(x,NULL);CHKERRQ(ierr); 3490906d46d4SStefano Zampini ierr = VecCopy(x,pcis->vec1_global);CHKERRQ(ierr); 3491e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,x,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 3492e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,x,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 3493bbb9e6c6SStefano Zampini ierr = MatMult(localChangeOfBasisMatrix,pcis->vec1_N,pcis->vec2_N);CHKERRQ(ierr); 3494e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec2_N,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 3495e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec2_N,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 3496906d46d4SStefano Zampini ierr = MatMult(pcbddc->ChangeOfBasisMatrix,pcis->vec1_global,x_change);CHKERRQ(ierr); 3497906d46d4SStefano Zampini ierr = VecAXPY(x,-1.0,x_change);CHKERRQ(ierr); 3498906d46d4SStefano Zampini ierr = VecNorm(x,NORM_INFINITY,&error);CHKERRQ(ierr); 3499906d46d4SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 3500bbb9e6c6SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Error global vs local change: %1.6e\n",error);CHKERRQ(ierr); 3501906d46d4SStefano Zampini ierr = VecDestroy(&x);CHKERRQ(ierr); 3502906d46d4SStefano Zampini ierr = VecDestroy(&x_change);CHKERRQ(ierr); 3503906d46d4SStefano Zampini } 3504b96c3477SStefano Zampini 3505b96c3477SStefano Zampini /* adapt sub_schurs computed (if any) */ 3506b96c3477SStefano Zampini if (pcbddc->use_deluxe_scaling) { 3507b96c3477SStefano Zampini PCBDDCSubSchurs sub_schurs=pcbddc->sub_schurs; 3508b96c3477SStefano Zampini if (sub_schurs->S_Ej_all) { 3509ac632422SStefano Zampini Mat S_new,tmat; 3510b087196eSStefano Zampini ISLocalToGlobalMapping NtoSall; 3511b087196eSStefano Zampini IS is_all_N,is_V,is_V_Sall; 3512b087196eSStefano Zampini const PetscScalar *array; 3513b087196eSStefano Zampini const PetscInt *idxs_V,*idxs_all; 3514b087196eSStefano Zampini PetscInt i,n_V; 3515bbb9e6c6SStefano Zampini 3516bbb9e6c6SStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(pcis->BtoNmap,sub_schurs->is_Ej_all,&is_all_N);CHKERRQ(ierr); 35176816873aSStefano Zampini ierr = MatGetSubMatrix(localChangeOfBasisMatrix,is_all_N,is_all_N,MAT_INITIAL_MATRIX,&tmat);CHKERRQ(ierr); 3518b087196eSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,pcbddc->n_vertices,pcbddc->local_primal_ref_node,PETSC_COPY_VALUES,&is_V);CHKERRQ(ierr); 3519b087196eSStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(is_all_N,&NtoSall);CHKERRQ(ierr); 3520b087196eSStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(NtoSall,IS_GTOLM_DROP,is_V,&is_V_Sall);CHKERRQ(ierr); 3521b087196eSStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&NtoSall);CHKERRQ(ierr); 3522bbb9e6c6SStefano Zampini ierr = ISDestroy(&is_all_N);CHKERRQ(ierr); 3523b087196eSStefano Zampini ierr = ISDestroy(&is_V);CHKERRQ(ierr); 3524ac632422SStefano Zampini ierr = MatPtAP(sub_schurs->S_Ej_all,tmat,MAT_INITIAL_MATRIX,1.0,&S_new);CHKERRQ(ierr); 3525b96c3477SStefano Zampini ierr = MatDestroy(&sub_schurs->S_Ej_all);CHKERRQ(ierr); 3526ac632422SStefano Zampini ierr = PetscObjectReference((PetscObject)S_new);CHKERRQ(ierr); 3527b087196eSStefano Zampini ierr = MatZeroRowsColumnsIS(S_new,is_V_Sall,1.,NULL,NULL);CHKERRQ(ierr); 3528b087196eSStefano Zampini ierr = ISGetLocalSize(is_V_Sall,&n_V);CHKERRQ(ierr); 3529b087196eSStefano Zampini ierr = ISGetIndices(is_V_Sall,&idxs_V);CHKERRQ(ierr); 3530b087196eSStefano Zampini ierr = ISGetIndices(sub_schurs->is_Ej_all,&idxs_all);CHKERRQ(ierr); 3531b087196eSStefano Zampini ierr = VecGetArrayRead(pcis->D,&array);CHKERRQ(ierr); 3532b087196eSStefano Zampini for (i=0;i<n_V;i++) { 3533b087196eSStefano Zampini PetscScalar val; 3534b087196eSStefano Zampini PetscInt idx; 3535b087196eSStefano Zampini 3536b087196eSStefano Zampini idx = idxs_V[i]; 3537b087196eSStefano Zampini val = array[idxs_all[idxs_V[i]]]; 3538b087196eSStefano Zampini ierr = MatSetValue(S_new,idx,idx,val,INSERT_VALUES);CHKERRQ(ierr); 3539b087196eSStefano Zampini } 3540b087196eSStefano Zampini ierr = MatAssemblyBegin(S_new,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3541b087196eSStefano Zampini ierr = MatAssemblyEnd(S_new,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3542ac632422SStefano Zampini sub_schurs->S_Ej_all = S_new; 3543ac632422SStefano Zampini ierr = MatDestroy(&S_new);CHKERRQ(ierr); 3544ac632422SStefano Zampini if (sub_schurs->sum_S_Ej_all) { 3545ac632422SStefano Zampini ierr = MatPtAP(sub_schurs->sum_S_Ej_all,tmat,MAT_INITIAL_MATRIX,1.0,&S_new);CHKERRQ(ierr); 3546b96c3477SStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_all);CHKERRQ(ierr); 3547ac632422SStefano Zampini ierr = PetscObjectReference((PetscObject)S_new);CHKERRQ(ierr); 3548b087196eSStefano Zampini ierr = MatZeroRowsColumnsIS(S_new,is_V_Sall,1.,NULL,NULL);CHKERRQ(ierr); 3549ac632422SStefano Zampini sub_schurs->sum_S_Ej_all = S_new; 3550ac632422SStefano Zampini ierr = MatDestroy(&S_new);CHKERRQ(ierr); 3551ac632422SStefano Zampini } 3552b087196eSStefano Zampini ierr = VecRestoreArrayRead(pcis->D,&array);CHKERRQ(ierr); 3553b087196eSStefano Zampini ierr = ISRestoreIndices(sub_schurs->is_Ej_all,&idxs_all);CHKERRQ(ierr); 3554b087196eSStefano Zampini ierr = ISRestoreIndices(is_V_Sall,&idxs_V);CHKERRQ(ierr); 3555b96c3477SStefano Zampini ierr = MatDestroy(&tmat);CHKERRQ(ierr); 3556b087196eSStefano Zampini ierr = ISDestroy(&is_V_Sall);CHKERRQ(ierr); 3557b96c3477SStefano Zampini } 3558b96c3477SStefano Zampini } 3559906d46d4SStefano Zampini ierr = MatDestroy(&localChangeOfBasisMatrix);CHKERRQ(ierr); 3560906d46d4SStefano Zampini } else if (pcbddc->user_ChangeOfBasisMatrix) { 3561b9b85e73SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->user_ChangeOfBasisMatrix);CHKERRQ(ierr); 3562b9b85e73SStefano Zampini pcbddc->ChangeOfBasisMatrix = pcbddc->user_ChangeOfBasisMatrix; 3563b9b85e73SStefano Zampini } 3564906d46d4SStefano Zampini 3565906d46d4SStefano Zampini /* set up change of basis context */ 3566906d46d4SStefano Zampini if (pcbddc->ChangeOfBasisMatrix) { 3567906d46d4SStefano Zampini PCBDDCChange_ctx change_ctx; 3568906d46d4SStefano Zampini 3569906d46d4SStefano Zampini if (!pcbddc->new_global_mat) { 3570906d46d4SStefano Zampini PetscInt global_size,local_size; 3571906d46d4SStefano Zampini 3572906d46d4SStefano Zampini ierr = VecGetSize(pcis->vec1_global,&global_size);CHKERRQ(ierr); 3573906d46d4SStefano Zampini ierr = VecGetLocalSize(pcis->vec1_global,&local_size);CHKERRQ(ierr); 3574906d46d4SStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)pc),&pcbddc->new_global_mat);CHKERRQ(ierr); 3575906d46d4SStefano Zampini ierr = MatSetSizes(pcbddc->new_global_mat,local_size,local_size,global_size,global_size);CHKERRQ(ierr); 3576906d46d4SStefano Zampini ierr = MatSetType(pcbddc->new_global_mat,MATSHELL);CHKERRQ(ierr); 3577906d46d4SStefano Zampini ierr = MatShellSetOperation(pcbddc->new_global_mat,MATOP_MULT,(void (*)(void))PCBDDCMatMult_Private);CHKERRQ(ierr); 3578906d46d4SStefano Zampini ierr = MatShellSetOperation(pcbddc->new_global_mat,MATOP_MULT_TRANSPOSE,(void (*)(void))PCBDDCMatMultTranspose_Private);CHKERRQ(ierr); 3579906d46d4SStefano Zampini ierr = PetscNew(&change_ctx);CHKERRQ(ierr); 3580906d46d4SStefano Zampini ierr = MatShellSetContext(pcbddc->new_global_mat,change_ctx);CHKERRQ(ierr); 3581906d46d4SStefano Zampini } else { 3582906d46d4SStefano Zampini ierr = MatShellGetContext(pcbddc->new_global_mat,&change_ctx);CHKERRQ(ierr); 3583906d46d4SStefano Zampini ierr = MatDestroy(&change_ctx->global_change);CHKERRQ(ierr); 3584906d46d4SStefano Zampini ierr = VecDestroyVecs(2,&change_ctx->work);CHKERRQ(ierr); 3585906d46d4SStefano Zampini } 3586906d46d4SStefano Zampini if (!pcbddc->user_ChangeOfBasisMatrix) { 3587906d46d4SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 3588906d46d4SStefano Zampini change_ctx->global_change = pcbddc->ChangeOfBasisMatrix; 3589906d46d4SStefano Zampini } else { 3590906d46d4SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->user_ChangeOfBasisMatrix);CHKERRQ(ierr); 3591906d46d4SStefano Zampini change_ctx->global_change = pcbddc->user_ChangeOfBasisMatrix; 3592906d46d4SStefano Zampini } 3593906d46d4SStefano Zampini ierr = VecDuplicateVecs(pcis->vec1_global,2,&change_ctx->work);CHKERRQ(ierr); 3594906d46d4SStefano Zampini ierr = MatSetUp(pcbddc->new_global_mat);CHKERRQ(ierr); 3595906d46d4SStefano Zampini ierr = MatAssemblyBegin(pcbddc->new_global_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3596906d46d4SStefano Zampini ierr = MatAssemblyEnd(pcbddc->new_global_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3597b9b85e73SStefano Zampini } 3598a717540cSStefano Zampini 3599*4f1b2e48SStefano Zampini /* add pressure dofs to set of primal nodes for numbering purposes */ 3600*4f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 3601*4f1b2e48SStefano Zampini pcbddc->local_primal_ref_node[pcbddc->local_primal_size_cc] = pcbddc->benign_p0_lidx[i]; 3602*4f1b2e48SStefano Zampini pcbddc->primal_indices_local_idxs[pcbddc->local_primal_size] = pcbddc->benign_p0_lidx[i]; 3603019a44ceSStefano Zampini pcbddc->local_primal_ref_mult[pcbddc->local_primal_size_cc] = 1; 3604019a44ceSStefano Zampini pcbddc->local_primal_size_cc++; 3605019a44ceSStefano Zampini pcbddc->local_primal_size++; 3606019a44ceSStefano Zampini } 3607019a44ceSStefano Zampini 3608019a44ceSStefano Zampini /* check if a new primal space has been introduced (also take into account benign trick) */ 3609727cdba6SStefano Zampini pcbddc->new_primal_space_local = PETSC_TRUE; 3610727cdba6SStefano Zampini if (olocal_primal_size == pcbddc->local_primal_size) { 3611aff50787SStefano 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); 3612c1c8e736SStefano Zampini pcbddc->new_primal_space_local = (PetscBool)(!pcbddc->new_primal_space_local); 36130e6343abSStefano Zampini if (!pcbddc->new_primal_space_local) { 3614aff50787SStefano 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); 3615727cdba6SStefano Zampini pcbddc->new_primal_space_local = (PetscBool)(!pcbddc->new_primal_space_local); 3616727cdba6SStefano Zampini } 36170e6343abSStefano Zampini } 36180e6343abSStefano Zampini ierr = PetscFree2(olocal_primal_ref_node,olocal_primal_ref_mult);CHKERRQ(ierr); 3619727cdba6SStefano Zampini /* new_primal_space will be used for numbering of coarse dofs, so it should be the same across all subdomains */ 3620727cdba6SStefano Zampini ierr = MPI_Allreduce(&pcbddc->new_primal_space_local,&pcbddc->new_primal_space,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 3621727cdba6SStefano Zampini 3622a717540cSStefano Zampini /* flush dbg viewer */ 3623b8ffe317SStefano Zampini if (pcbddc->dbg_flag) { 3624b8ffe317SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 3625b8ffe317SStefano Zampini } 3626a717540cSStefano Zampini 3627e310c8b4SStefano Zampini /* free workspace */ 3628a717540cSStefano Zampini ierr = PetscBTDestroy(&qr_needed_idx);CHKERRQ(ierr); 36294641a718SStefano Zampini ierr = PetscBTDestroy(&change_basis);CHKERRQ(ierr); 363008122e43SStefano Zampini if (!pcbddc->adaptive_selection) { 36319162d606SStefano Zampini ierr = PetscFree3(constraints_idxs_ptr,constraints_data_ptr,constraints_n);CHKERRQ(ierr); 36329162d606SStefano Zampini ierr = PetscFree3(constraints_data,constraints_idxs,constraints_idxs_B);CHKERRQ(ierr); 363308122e43SStefano Zampini } else { 36349162d606SStefano Zampini ierr = PetscFree5(pcbddc->adaptive_constraints_n, 36359162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr, 36369162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr, 363708122e43SStefano Zampini pcbddc->adaptive_constraints_idxs, 363808122e43SStefano Zampini pcbddc->adaptive_constraints_data);CHKERRQ(ierr); 36399162d606SStefano Zampini ierr = PetscFree(constraints_n);CHKERRQ(ierr); 36409162d606SStefano Zampini ierr = PetscFree(constraints_idxs_B);CHKERRQ(ierr); 364108122e43SStefano Zampini } 3642674ae819SStefano Zampini PetscFunctionReturn(0); 3643674ae819SStefano Zampini } 3644674ae819SStefano Zampini 3645674ae819SStefano Zampini #undef __FUNCT__ 3646674ae819SStefano Zampini #define __FUNCT__ "PCBDDCAnalyzeInterface" 3647674ae819SStefano Zampini PetscErrorCode PCBDDCAnalyzeInterface(PC pc) 3648674ae819SStefano Zampini { 3649674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 3650674ae819SStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 3651674ae819SStefano Zampini Mat_IS *matis = (Mat_IS*)pc->pmat->data; 36527fb0e2dbSStefano Zampini PetscInt ierr,i,vertex_size,N; 3653674ae819SStefano Zampini PetscViewer viewer=pcbddc->dbg_viewer; 3654674ae819SStefano Zampini 3655674ae819SStefano Zampini PetscFunctionBegin; 36568e61c736SStefano Zampini /* Reset previously computed graph */ 36578e61c736SStefano Zampini ierr = PCBDDCGraphReset(pcbddc->mat_graph);CHKERRQ(ierr); 3658674ae819SStefano Zampini /* Init local Graph struct */ 36597fb0e2dbSStefano Zampini ierr = MatGetSize(pc->pmat,&N,NULL);CHKERRQ(ierr); 36603bbff08aSStefano Zampini ierr = PCBDDCGraphInit(pcbddc->mat_graph,pcis->mapping,N);CHKERRQ(ierr); 3661674ae819SStefano Zampini 3662575ad6abSStefano Zampini /* Check validity of the csr graph passed in by the user */ 36635099eff2SStefano Zampini if (pcbddc->mat_graph->nvtxs_csr && pcbddc->mat_graph->nvtxs_csr != pcbddc->mat_graph->nvtxs) { 36645099eff2SStefano 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); 3665575ad6abSStefano Zampini } 36669577ea80SStefano Zampini 3667674ae819SStefano Zampini /* Set default CSR adjacency of local dofs if not provided by the user with PCBDDCSetLocalAdjacencyGraph */ 36684d379d7bSStefano Zampini if (!pcbddc->mat_graph->xadj || !pcbddc->mat_graph->adjncy) { 36694d379d7bSStefano Zampini PetscInt *xadj,*adjncy; 36704d379d7bSStefano Zampini PetscInt nvtxs; 3671e496cd5dSStefano Zampini PetscBool flg_row=PETSC_FALSE; 3672674ae819SStefano Zampini 36734d379d7bSStefano Zampini if (pcbddc->use_local_adj) { 36742fffb893SStefano Zampini 36752fffb893SStefano Zampini ierr = MatGetRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 36762fffb893SStefano Zampini if (flg_row) { 36774d379d7bSStefano Zampini ierr = PCBDDCSetLocalAdjacencyGraph(pc,nvtxs,xadj,adjncy,PETSC_COPY_VALUES);CHKERRQ(ierr); 3678b96c3477SStefano Zampini pcbddc->computed_rowadj = PETSC_TRUE; 36792fffb893SStefano Zampini } 36802fffb893SStefano Zampini ierr = MatRestoreRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 36819b28b941SStefano 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 */ 36824d379d7bSStefano Zampini IS is_dummy; 36834d379d7bSStefano Zampini ISLocalToGlobalMapping l2gmap_dummy; 36844d379d7bSStefano Zampini PetscInt j,sum; 36854d379d7bSStefano Zampini PetscInt *cxadj,*cadjncy; 36864d379d7bSStefano Zampini const PetscInt *idxs; 36874d379d7bSStefano Zampini PCBDDCGraph graph; 36884d379d7bSStefano Zampini PetscBT is_on_boundary; 36894d379d7bSStefano Zampini 36904d379d7bSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcis->n,0,1,&is_dummy);CHKERRQ(ierr); 36914d379d7bSStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(is_dummy,&l2gmap_dummy);CHKERRQ(ierr); 36924d379d7bSStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 36934d379d7bSStefano Zampini ierr = PCBDDCGraphCreate(&graph);CHKERRQ(ierr); 36947fb0e2dbSStefano Zampini ierr = PCBDDCGraphInit(graph,l2gmap_dummy,pcis->n);CHKERRQ(ierr); 36954d379d7bSStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&l2gmap_dummy);CHKERRQ(ierr); 3696e496cd5dSStefano Zampini ierr = MatGetRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 3697e496cd5dSStefano Zampini if (flg_row) { 36984d379d7bSStefano Zampini graph->xadj = xadj; 36994d379d7bSStefano Zampini graph->adjncy = adjncy; 3700e496cd5dSStefano Zampini } 37014d379d7bSStefano Zampini ierr = PCBDDCGraphSetUp(graph,1,NULL,NULL,0,NULL,NULL);CHKERRQ(ierr); 37024d379d7bSStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(graph);CHKERRQ(ierr); 3703e496cd5dSStefano Zampini ierr = MatRestoreRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 37044d379d7bSStefano Zampini 37054d379d7bSStefano Zampini if (pcbddc->dbg_flag) { 37069b28b941SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"[%d] Found %d subdomains (local size %d)\n",PetscGlobalRank,graph->ncc,pcis->n);CHKERRQ(ierr); 37074d379d7bSStefano Zampini for (i=0;i<graph->ncc;i++) { 37084d379d7bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"[%d] %d cc size %d\n",PetscGlobalRank,i,graph->cptr[i+1]-graph->cptr[i]);CHKERRQ(ierr); 37094d379d7bSStefano Zampini } 37104d379d7bSStefano Zampini } 37114d379d7bSStefano Zampini 3712e496cd5dSStefano Zampini ierr = PetscBTCreate(pcis->n,&is_on_boundary);CHKERRQ(ierr); 37134d379d7bSStefano Zampini ierr = ISGetIndices(pcis->is_B_local,&idxs);CHKERRQ(ierr); 37144d379d7bSStefano Zampini for (i=0;i<pcis->n_B;i++) { 37154d379d7bSStefano Zampini ierr = PetscBTSet(is_on_boundary,idxs[i]);CHKERRQ(ierr); 37164d379d7bSStefano Zampini } 37174d379d7bSStefano Zampini ierr = ISRestoreIndices(pcis->is_B_local,&idxs);CHKERRQ(ierr); 37184d379d7bSStefano Zampini 3719e496cd5dSStefano Zampini ierr = PetscCalloc1(pcis->n+1,&cxadj);CHKERRQ(ierr); 37204d379d7bSStefano Zampini sum = 0; 37214d379d7bSStefano Zampini for (i=0;i<graph->ncc;i++) { 37224d379d7bSStefano Zampini PetscInt sizecc = 0; 37234d379d7bSStefano Zampini for (j=graph->cptr[i];j<graph->cptr[i+1];j++) { 37244d379d7bSStefano Zampini if (PetscBTLookup(is_on_boundary,graph->queue[j])) { 37254d379d7bSStefano Zampini sizecc++; 37264d379d7bSStefano Zampini } 37274d379d7bSStefano Zampini } 37284d379d7bSStefano Zampini for (j=graph->cptr[i];j<graph->cptr[i+1];j++) { 37294d379d7bSStefano Zampini if (PetscBTLookup(is_on_boundary,graph->queue[j])) { 37304d379d7bSStefano Zampini cxadj[graph->queue[j]] = sizecc; 37314d379d7bSStefano Zampini } 37324d379d7bSStefano Zampini } 37334d379d7bSStefano Zampini sum += sizecc*sizecc; 37344d379d7bSStefano Zampini } 37354d379d7bSStefano Zampini ierr = PetscMalloc1(sum,&cadjncy);CHKERRQ(ierr); 37364d379d7bSStefano Zampini sum = 0; 3737e496cd5dSStefano Zampini for (i=0;i<pcis->n;i++) { 37384d379d7bSStefano Zampini PetscInt temp = cxadj[i]; 37394d379d7bSStefano Zampini cxadj[i] = sum; 37404d379d7bSStefano Zampini sum += temp; 37414d379d7bSStefano Zampini } 3742e496cd5dSStefano Zampini cxadj[pcis->n] = sum; 37434d379d7bSStefano Zampini for (i=0;i<graph->ncc;i++) { 37444d379d7bSStefano Zampini for (j=graph->cptr[i];j<graph->cptr[i+1];j++) { 37454d379d7bSStefano Zampini if (PetscBTLookup(is_on_boundary,graph->queue[j])) { 37464d379d7bSStefano Zampini PetscInt k,sizecc = 0; 37474d379d7bSStefano Zampini for (k=graph->cptr[i];k<graph->cptr[i+1];k++) { 37484d379d7bSStefano Zampini if (PetscBTLookup(is_on_boundary,graph->queue[k])) { 37494d379d7bSStefano Zampini cadjncy[cxadj[graph->queue[j]]+sizecc]=graph->queue[k]; 37504d379d7bSStefano Zampini sizecc++; 37514d379d7bSStefano Zampini } 37524d379d7bSStefano Zampini } 37534d379d7bSStefano Zampini } 37544d379d7bSStefano Zampini } 37554d379d7bSStefano Zampini } 37569b28b941SStefano Zampini if (sum) { 3757e496cd5dSStefano Zampini ierr = PCBDDCSetLocalAdjacencyGraph(pc,pcis->n,cxadj,cadjncy,PETSC_OWN_POINTER);CHKERRQ(ierr); 37584d379d7bSStefano Zampini } else { 37594d379d7bSStefano Zampini ierr = PetscFree(cxadj);CHKERRQ(ierr); 37604d379d7bSStefano Zampini ierr = PetscFree(cadjncy);CHKERRQ(ierr); 37614d379d7bSStefano Zampini } 37624d379d7bSStefano Zampini graph->xadj = 0; 37634d379d7bSStefano Zampini graph->adjncy = 0; 37644d379d7bSStefano Zampini ierr = PCBDDCGraphDestroy(&graph);CHKERRQ(ierr); 37654d379d7bSStefano Zampini ierr = PetscBTDestroy(&is_on_boundary);CHKERRQ(ierr); 37664d379d7bSStefano Zampini } 3767674ae819SStefano Zampini } 37689b28b941SStefano Zampini if (pcbddc->dbg_flag) { 37699b28b941SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 3770674ae819SStefano Zampini } 3771674ae819SStefano Zampini 377263602bcaSStefano Zampini /* Set default dofs' splitting if no information has been provided by the user with PCBDDCSetDofsSplitting or PCBDDCSetDofsSplittingLocal */ 3773674ae819SStefano Zampini vertex_size = 1; 377463602bcaSStefano Zampini if (pcbddc->user_provided_isfordofs) { 377563602bcaSStefano Zampini if (pcbddc->n_ISForDofs) { /* need to convert from global to local and remove references to global dofs splitting */ 377695ecbf38SStefano Zampini ierr = PetscMalloc1(pcbddc->n_ISForDofs,&pcbddc->ISForDofsLocal);CHKERRQ(ierr); 377763602bcaSStefano Zampini for (i=0;i<pcbddc->n_ISForDofs;i++) { 3778e176bc59SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->rctx,pcis->vec1_global,pcis->vec1_N,pcbddc->ISForDofs[i],&pcbddc->ISForDofsLocal[i]);CHKERRQ(ierr); 377963602bcaSStefano Zampini ierr = ISDestroy(&pcbddc->ISForDofs[i]);CHKERRQ(ierr); 3780674ae819SStefano Zampini } 378163602bcaSStefano Zampini pcbddc->n_ISForDofsLocal = pcbddc->n_ISForDofs; 378263602bcaSStefano Zampini pcbddc->n_ISForDofs = 0; 378363602bcaSStefano Zampini ierr = PetscFree(pcbddc->ISForDofs);CHKERRQ(ierr); 3784674ae819SStefano Zampini } 378563602bcaSStefano Zampini /* mat block size as vertex size (used for elasticity with rigid body modes as nearnullspace) */ 3786674ae819SStefano Zampini ierr = MatGetBlockSize(matis->A,&vertex_size);CHKERRQ(ierr); 378763602bcaSStefano Zampini } else { 378863602bcaSStefano Zampini if (!pcbddc->n_ISForDofsLocal) { /* field split not present, create it in local ordering */ 378963602bcaSStefano Zampini ierr = MatGetBlockSize(pc->pmat,&pcbddc->n_ISForDofsLocal);CHKERRQ(ierr); 3790854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->n_ISForDofsLocal,&pcbddc->ISForDofsLocal);CHKERRQ(ierr); 379163602bcaSStefano Zampini for (i=0;i<pcbddc->n_ISForDofsLocal;i++) { 379263602bcaSStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)pc),pcis->n/pcbddc->n_ISForDofsLocal,i,pcbddc->n_ISForDofsLocal,&pcbddc->ISForDofsLocal[i]);CHKERRQ(ierr); 379363602bcaSStefano Zampini } 379463602bcaSStefano Zampini } 3795674ae819SStefano Zampini } 3796674ae819SStefano Zampini 3797674ae819SStefano Zampini /* Setup of Graph */ 3798785d1243SStefano Zampini if (!pcbddc->DirichletBoundariesLocal && pcbddc->DirichletBoundaries) { /* need to convert from global to local */ 3799e176bc59SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->rctx,pcis->vec1_global,pcis->vec1_N,pcbddc->DirichletBoundaries,&pcbddc->DirichletBoundariesLocal);CHKERRQ(ierr); 3800785d1243SStefano Zampini } 3801785d1243SStefano Zampini if (!pcbddc->NeumannBoundariesLocal && pcbddc->NeumannBoundaries) { /* need to convert from global to local */ 3802e176bc59SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->rctx,pcis->vec1_global,pcis->vec1_N,pcbddc->NeumannBoundaries,&pcbddc->NeumannBoundariesLocal);CHKERRQ(ierr); 3803785d1243SStefano Zampini } 3804302440fdSBarry Smith ierr = PCBDDCGraphSetUp(pcbddc->mat_graph,vertex_size,pcbddc->NeumannBoundariesLocal,pcbddc->DirichletBoundariesLocal,pcbddc->n_ISForDofsLocal,pcbddc->ISForDofsLocal,pcbddc->user_primal_vertices);CHKERRQ(ierr); 3805674ae819SStefano Zampini 3806*4f1b2e48SStefano Zampini /* attach info on disconnected subdomains if present */ 3807*4f1b2e48SStefano Zampini if (pcbddc->n_local_subs) { 3808*4f1b2e48SStefano Zampini PetscInt *local_subs; 3809*4f1b2e48SStefano Zampini 3810*4f1b2e48SStefano Zampini ierr = PetscMalloc1(N,&local_subs);CHKERRQ(ierr); 3811*4f1b2e48SStefano Zampini for (i=0;i<pcbddc->n_local_subs;i++) { 3812*4f1b2e48SStefano Zampini const PetscInt *idxs; 3813*4f1b2e48SStefano Zampini PetscInt nl,j; 3814*4f1b2e48SStefano Zampini 3815*4f1b2e48SStefano Zampini ierr = ISGetLocalSize(pcbddc->local_subs[i],&nl);CHKERRQ(ierr); 3816*4f1b2e48SStefano Zampini ierr = ISGetIndices(pcbddc->local_subs[i],&idxs);CHKERRQ(ierr); 3817*4f1b2e48SStefano Zampini for (j=0;j<nl;j++) { 3818*4f1b2e48SStefano Zampini local_subs[idxs[j]] = i; 3819*4f1b2e48SStefano Zampini } 3820*4f1b2e48SStefano Zampini ierr = ISRestoreIndices(pcbddc->local_subs[i],&idxs);CHKERRQ(ierr); 3821*4f1b2e48SStefano Zampini } 3822*4f1b2e48SStefano Zampini pcbddc->mat_graph->n_local_subs = pcbddc->n_local_subs; 3823*4f1b2e48SStefano Zampini pcbddc->mat_graph->local_subs = local_subs; 3824*4f1b2e48SStefano Zampini } 3825*4f1b2e48SStefano Zampini 3826674ae819SStefano Zampini /* Graph's connected components analysis */ 3827674ae819SStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(pcbddc->mat_graph);CHKERRQ(ierr); 3828674ae819SStefano Zampini 3829674ae819SStefano Zampini /* print some info to stdout */ 3830674ae819SStefano Zampini if (pcbddc->dbg_flag) { 3831302440fdSBarry Smith ierr = PCBDDCGraphASCIIView(pcbddc->mat_graph,pcbddc->dbg_flag,viewer);CHKERRQ(ierr); 3832674ae819SStefano Zampini } 3833fb180af4SStefano Zampini 3834fb180af4SStefano Zampini /* mark topography has done */ 3835fb180af4SStefano Zampini pcbddc->recompute_topography = PETSC_FALSE; 3836674ae819SStefano Zampini PetscFunctionReturn(0); 3837674ae819SStefano Zampini } 3838674ae819SStefano Zampini 3839dc456d91SStefano Zampini /* given an index sets possibly with holes, renumbers the indexes removing the holes */ 3840674ae819SStefano Zampini #undef __FUNCT__ 3841674ae819SStefano Zampini #define __FUNCT__ "PCBDDCSubsetNumbering" 3842dc456d91SStefano Zampini PetscErrorCode PCBDDCSubsetNumbering(IS subset, IS subset_mult, PetscInt *N_n, IS *subset_n) 3843674ae819SStefano Zampini { 3844dc456d91SStefano Zampini PetscSF sf; 3845dc456d91SStefano Zampini PetscLayout map; 3846dc456d91SStefano Zampini const PetscInt *idxs; 3847dc456d91SStefano Zampini PetscInt *leaf_data,*root_data,*gidxs; 3848dc456d91SStefano Zampini PetscInt N,n,i,lbounds[2],gbounds[2],Nl; 3849dc456d91SStefano Zampini PetscInt n_n,nlocals,start,first_index; 3850dc456d91SStefano Zampini PetscMPIInt commsize; 3851674ae819SStefano Zampini PetscBool first_found; 3852674ae819SStefano Zampini PetscErrorCode ierr; 3853674ae819SStefano Zampini 3854674ae819SStefano Zampini PetscFunctionBegin; 3855dc456d91SStefano Zampini ierr = ISGetLocalSize(subset,&n);CHKERRQ(ierr); 3856dc456d91SStefano Zampini if (subset_mult) { 3857dc456d91SStefano Zampini PetscCheckSameComm(subset,1,subset_mult,2); 3858dc456d91SStefano Zampini ierr = ISGetLocalSize(subset,&i);CHKERRQ(ierr); 3859dc456d91SStefano Zampini if (i != n) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Local subset and multiplicity sizes don't match! %d != %d",n,i); 3860674ae819SStefano Zampini } 3861dc456d91SStefano Zampini /* create workspace layout for computing global indices of subset */ 3862dc456d91SStefano Zampini ierr = ISGetIndices(subset,&idxs);CHKERRQ(ierr); 3863dc456d91SStefano Zampini lbounds[0] = lbounds[1] = 0; 3864dc456d91SStefano Zampini for (i=0;i<n;i++) { 3865dc456d91SStefano Zampini if (idxs[i] < lbounds[0]) lbounds[0] = idxs[i]; 3866dc456d91SStefano Zampini else if (idxs[i] > lbounds[1]) lbounds[1] = idxs[i]; 3867674ae819SStefano Zampini } 3868dc456d91SStefano Zampini lbounds[0] = -lbounds[0]; 3869dc456d91SStefano Zampini ierr = MPI_Allreduce(lbounds,gbounds,2,MPIU_INT,MPI_MAX,PetscObjectComm((PetscObject)subset));CHKERRQ(ierr); 3870dc456d91SStefano Zampini gbounds[0] = -gbounds[0]; 3871dc456d91SStefano Zampini N = gbounds[1] - gbounds[0] + 1; 3872dc456d91SStefano Zampini ierr = PetscLayoutCreate(PetscObjectComm((PetscObject)subset),&map);CHKERRQ(ierr); 3873dc456d91SStefano Zampini ierr = PetscLayoutSetBlockSize(map,1);CHKERRQ(ierr); 3874dc456d91SStefano Zampini ierr = PetscLayoutSetSize(map,N);CHKERRQ(ierr); 3875dc456d91SStefano Zampini ierr = PetscLayoutSetUp(map);CHKERRQ(ierr); 3876dc456d91SStefano Zampini ierr = PetscLayoutGetLocalSize(map,&Nl);CHKERRQ(ierr); 3877dc456d91SStefano Zampini 3878dc456d91SStefano Zampini /* create sf : leaf_data == multiplicity of indexes, root data == global index in layout */ 3879dc456d91SStefano Zampini ierr = PetscMalloc2(n,&leaf_data,Nl,&root_data);CHKERRQ(ierr); 3880dc456d91SStefano Zampini if (subset_mult) { 3881dc456d91SStefano Zampini const PetscInt* idxs_mult; 3882dc456d91SStefano Zampini 3883dc456d91SStefano Zampini ierr = ISGetIndices(subset_mult,&idxs_mult);CHKERRQ(ierr); 3884dc456d91SStefano Zampini ierr = PetscMemcpy(leaf_data,idxs_mult,n*sizeof(PetscInt));CHKERRQ(ierr); 3885dc456d91SStefano Zampini ierr = ISRestoreIndices(subset_mult,&idxs_mult);CHKERRQ(ierr); 3886674ae819SStefano Zampini } else { 3887dc456d91SStefano Zampini for (i=0;i<n;i++) leaf_data[i] = 1; 3888674ae819SStefano Zampini } 3889dc456d91SStefano Zampini /* local size of new subset */ 3890dc456d91SStefano Zampini n_n = 0; 3891dc456d91SStefano Zampini for (i=0;i<n;i++) n_n += leaf_data[i]; 3892dc456d91SStefano Zampini 3893dc456d91SStefano Zampini /* global indexes in layout */ 3894dc456d91SStefano Zampini ierr = PetscMalloc1(n_n,&gidxs);CHKERRQ(ierr); /* allocating possibly extra space in gidxs which will be used later */ 3895dc456d91SStefano Zampini for (i=0;i<n;i++) gidxs[i] = idxs[i] - gbounds[0]; 3896dc456d91SStefano Zampini ierr = ISRestoreIndices(subset,&idxs);CHKERRQ(ierr); 3897dc456d91SStefano Zampini ierr = PetscSFCreate(PetscObjectComm((PetscObject)subset),&sf);CHKERRQ(ierr); 3898dc456d91SStefano Zampini ierr = PetscSFSetGraphLayout(sf,map,n,NULL,PETSC_COPY_VALUES,gidxs);CHKERRQ(ierr); 3899dc456d91SStefano Zampini ierr = PetscLayoutDestroy(&map);CHKERRQ(ierr); 3900dc456d91SStefano Zampini 3901dc456d91SStefano Zampini /* reduce from leaves to roots */ 3902dc456d91SStefano Zampini ierr = PetscMemzero(root_data,Nl*sizeof(PetscInt));CHKERRQ(ierr); 390364a8e5bcSStefano Zampini ierr = PetscSFReduceBegin(sf,MPIU_INT,leaf_data,root_data,MPI_MAX);CHKERRQ(ierr); 390464a8e5bcSStefano Zampini ierr = PetscSFReduceEnd(sf,MPIU_INT,leaf_data,root_data,MPI_MAX);CHKERRQ(ierr); 3905dc456d91SStefano Zampini 3906dc456d91SStefano Zampini /* count indexes in local part of layout */ 3907674ae819SStefano Zampini nlocals = 0; 3908674ae819SStefano Zampini first_index = -1; 3909674ae819SStefano Zampini first_found = PETSC_FALSE; 3910dc456d91SStefano Zampini for (i=0;i<Nl;i++) { 3911dc456d91SStefano Zampini if (!first_found && root_data[i]) { 3912674ae819SStefano Zampini first_found = PETSC_TRUE; 3913674ae819SStefano Zampini first_index = i; 3914674ae819SStefano Zampini } 3915dc456d91SStefano Zampini nlocals += root_data[i]; 3916674ae819SStefano Zampini } 3917dc456d91SStefano Zampini 3918dc456d91SStefano Zampini /* cumulative of number of indexes and size of subset without holes */ 39195fa240b1SStefano Zampini #if defined(PETSC_HAVE_MPI_EXSCAN) 3920dc456d91SStefano Zampini start = 0; 392164a8e5bcSStefano Zampini ierr = MPI_Exscan(&nlocals,&start,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)subset));CHKERRQ(ierr); 39225fa240b1SStefano Zampini #else 392364a8e5bcSStefano Zampini ierr = MPI_Scan(&nlocals,&start,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)subset));CHKERRQ(ierr); 39245fa240b1SStefano Zampini start = start-nlocals; 39255fa240b1SStefano Zampini #endif 39265fa240b1SStefano Zampini 3927dc456d91SStefano Zampini if (N_n) { /* compute total size of new subset if requested */ 3928dc456d91SStefano Zampini *N_n = start + nlocals; 3929dc456d91SStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)subset),&commsize);CHKERRQ(ierr); 3930dc456d91SStefano Zampini ierr = MPI_Bcast(N_n,1,MPIU_INT,commsize-1,PetscObjectComm((PetscObject)subset));CHKERRQ(ierr); 3931674ae819SStefano Zampini } 39325fa240b1SStefano Zampini 39335fa240b1SStefano Zampini /* adapt root data with cumulative */ 3934674ae819SStefano Zampini if (first_found) { 3935dc456d91SStefano Zampini PetscInt old_index; 3936dc456d91SStefano Zampini 3937dc456d91SStefano Zampini root_data[first_index] += start; 3938674ae819SStefano Zampini old_index = first_index; 3939dc456d91SStefano Zampini for (i=first_index+1;i<Nl;i++) { 3940dc456d91SStefano Zampini if (root_data[i]) { 3941dc456d91SStefano Zampini root_data[i] += root_data[old_index]; 3942674ae819SStefano Zampini old_index = i; 3943674ae819SStefano Zampini } 3944674ae819SStefano Zampini } 3945674ae819SStefano Zampini } 3946dc456d91SStefano Zampini 3947dc456d91SStefano Zampini /* from roots to leaves */ 3948dc456d91SStefano Zampini ierr = PetscSFBcastBegin(sf,MPIU_INT,root_data,leaf_data);CHKERRQ(ierr); 3949dc456d91SStefano Zampini ierr = PetscSFBcastEnd(sf,MPIU_INT,root_data,leaf_data);CHKERRQ(ierr); 3950dc456d91SStefano Zampini ierr = PetscSFDestroy(&sf);CHKERRQ(ierr); 3951dc456d91SStefano Zampini 3952dc456d91SStefano Zampini /* create new IS with global indexes without holes */ 3953dc456d91SStefano Zampini if (subset_mult) { 3954dc456d91SStefano Zampini const PetscInt* idxs_mult; 3955dc456d91SStefano Zampini PetscInt cum; 3956dc456d91SStefano Zampini 3957dc456d91SStefano Zampini cum = 0; 3958dc456d91SStefano Zampini ierr = ISGetIndices(subset_mult,&idxs_mult);CHKERRQ(ierr); 3959dc456d91SStefano Zampini for (i=0;i<n;i++) { 3960dc456d91SStefano Zampini PetscInt j; 3961dc456d91SStefano Zampini for (j=0;j<idxs_mult[i];j++) gidxs[cum++] = leaf_data[i] - idxs_mult[i] + j; 3962674ae819SStefano Zampini } 3963dc456d91SStefano Zampini ierr = ISRestoreIndices(subset_mult,&idxs_mult);CHKERRQ(ierr); 3964674ae819SStefano Zampini } else { 3965dc456d91SStefano Zampini for (i=0;i<n;i++) { 3966dc456d91SStefano Zampini gidxs[i] = leaf_data[i]-1; 3967674ae819SStefano Zampini } 3968674ae819SStefano Zampini } 3969dc456d91SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)subset),n_n,gidxs,PETSC_OWN_POINTER,subset_n);CHKERRQ(ierr); 3970dc456d91SStefano Zampini ierr = PetscFree2(leaf_data,root_data);CHKERRQ(ierr); 3971674ae819SStefano Zampini PetscFunctionReturn(0); 3972674ae819SStefano Zampini } 39739a7d3425SStefano Zampini 39749a7d3425SStefano Zampini #undef __FUNCT__ 39759a7d3425SStefano Zampini #define __FUNCT__ "PCBDDCOrthonormalizeVecs" 39769a7d3425SStefano Zampini PetscErrorCode PCBDDCOrthonormalizeVecs(PetscInt n, Vec vecs[]) 39779a7d3425SStefano Zampini { 39789a7d3425SStefano Zampini PetscInt i,j; 39799a7d3425SStefano Zampini PetscScalar *alphas; 39809a7d3425SStefano Zampini PetscErrorCode ierr; 39819a7d3425SStefano Zampini 39829a7d3425SStefano Zampini PetscFunctionBegin; 39839a7d3425SStefano Zampini /* this implements stabilized Gram-Schmidt */ 3984785e854fSJed Brown ierr = PetscMalloc1(n,&alphas);CHKERRQ(ierr); 39859a7d3425SStefano Zampini for (i=0;i<n;i++) { 39869a7d3425SStefano Zampini ierr = VecNormalize(vecs[i],NULL);CHKERRQ(ierr); 39879a7d3425SStefano Zampini if (i<n) { ierr = VecMDot(vecs[i],n-i-1,&vecs[i+1],&alphas[i+1]);CHKERRQ(ierr); } 39889a7d3425SStefano Zampini for (j=i+1;j<n;j++) { ierr = VecAXPY(vecs[j],PetscConj(-alphas[j]),vecs[i]);CHKERRQ(ierr); } 39899a7d3425SStefano Zampini } 39909a7d3425SStefano Zampini ierr = PetscFree(alphas);CHKERRQ(ierr); 39919a7d3425SStefano Zampini PetscFunctionReturn(0); 39929a7d3425SStefano Zampini } 39939a7d3425SStefano Zampini 3994e7931f94SStefano Zampini #undef __FUNCT__ 399570cf5478SStefano Zampini #define __FUNCT__ "MatISGetSubassemblingPattern" 3996b0c7d250SStefano Zampini PetscErrorCode MatISGetSubassemblingPattern(Mat mat, PetscInt n_subdomains, PetscInt redprocs, IS* is_sends) 3997e7931f94SStefano Zampini { 399852e5ac9dSStefano Zampini IS ranks_send_to; 3999e7931f94SStefano Zampini PetscInt n_neighs,*neighs,*n_shared,**shared; 4000e7931f94SStefano Zampini PetscMPIInt size,rank,color; 400152e5ac9dSStefano Zampini PetscInt *xadj,*adjncy; 400252e5ac9dSStefano Zampini PetscInt *adjncy_wgt,*v_wgt,*ranks_send_to_idx; 40033837a79fSStefano Zampini PetscInt i,local_size,threshold=0; 40042b510759SStefano Zampini PetscBool use_vwgt=PETSC_FALSE,use_square=PETSC_FALSE; 4005e7931f94SStefano Zampini PetscSubcomm subcomm; 400652e5ac9dSStefano Zampini PetscErrorCode ierr; 4007a57a6d2fSStefano Zampini 4008e7931f94SStefano Zampini PetscFunctionBegin; 40092b510759SStefano Zampini ierr = PetscOptionsGetBool(NULL,"-matis_partitioning_use_square",&use_square,NULL);CHKERRQ(ierr); 40102b510759SStefano Zampini ierr = PetscOptionsGetBool(NULL,"-matis_partitioning_use_vwgt",&use_vwgt,NULL);CHKERRQ(ierr); 40112b510759SStefano Zampini ierr = PetscOptionsGetInt(NULL,"-matis_partitioning_threshold",&threshold,NULL);CHKERRQ(ierr); 4012e7931f94SStefano Zampini 4013e7931f94SStefano Zampini /* Get info on mapping */ 40143bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetSize(mat->rmap->mapping,&local_size);CHKERRQ(ierr); 40153bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetInfo(mat->rmap->mapping,&n_neighs,&neighs,&n_shared,&shared);CHKERRQ(ierr); 4016e7931f94SStefano Zampini 4017e7931f94SStefano Zampini /* build local CSR graph of subdomains' connectivity */ 4018785e854fSJed Brown ierr = PetscMalloc1(2,&xadj);CHKERRQ(ierr); 4019e7931f94SStefano Zampini xadj[0] = 0; 4020e7931f94SStefano Zampini xadj[1] = PetscMax(n_neighs-1,0); 4021785e854fSJed Brown ierr = PetscMalloc1(xadj[1],&adjncy);CHKERRQ(ierr); 4022785e854fSJed Brown ierr = PetscMalloc1(xadj[1],&adjncy_wgt);CHKERRQ(ierr); 4023e7931f94SStefano Zampini 40242b510759SStefano Zampini if (threshold) { 4025d023bfaeSStefano Zampini PetscInt xadj_count = 0; 40262b510759SStefano Zampini for (i=1;i<n_neighs;i++) { 4027d023bfaeSStefano Zampini if (n_shared[i] > threshold) { 4028d023bfaeSStefano Zampini adjncy[xadj_count] = neighs[i]; 4029d023bfaeSStefano Zampini adjncy_wgt[xadj_count] = n_shared[i]; 4030d023bfaeSStefano Zampini xadj_count++; 4031e7931f94SStefano Zampini } 4032e7931f94SStefano Zampini } 4033d023bfaeSStefano Zampini xadj[1] = xadj_count; 4034c8587f34SStefano Zampini } else { 4035e7931f94SStefano Zampini if (xadj[1]) { 4036e7931f94SStefano Zampini ierr = PetscMemcpy(adjncy,&neighs[1],xadj[1]*sizeof(*adjncy));CHKERRQ(ierr); 4037e7931f94SStefano Zampini ierr = PetscMemcpy(adjncy_wgt,&n_shared[1],xadj[1]*sizeof(*adjncy_wgt));CHKERRQ(ierr); 4038c8587f34SStefano Zampini } 4039e7931f94SStefano Zampini } 40403bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingRestoreInfo(mat->rmap->mapping,&n_neighs,&neighs,&n_shared,&shared);CHKERRQ(ierr); 4041e7931f94SStefano Zampini if (use_square) { 4042e7931f94SStefano Zampini for (i=0;i<xadj[1];i++) { 4043e7931f94SStefano Zampini adjncy_wgt[i] = adjncy_wgt[i]*adjncy_wgt[i]; 4044e7931f94SStefano Zampini } 4045e7931f94SStefano Zampini } 4046e7931f94SStefano Zampini ierr = PetscSortIntWithArray(xadj[1],adjncy,adjncy_wgt);CHKERRQ(ierr); 4047e7931f94SStefano Zampini 40483837a79fSStefano Zampini ierr = PetscMalloc1(1,&ranks_send_to_idx);CHKERRQ(ierr); 4049e7931f94SStefano Zampini 4050e7931f94SStefano Zampini /* 4051e7931f94SStefano Zampini Restrict work on active processes only. 4052e7931f94SStefano Zampini */ 4053e7931f94SStefano Zampini ierr = PetscSubcommCreate(PetscObjectComm((PetscObject)mat),&subcomm);CHKERRQ(ierr); 4054e7931f94SStefano Zampini ierr = PetscSubcommSetNumber(subcomm,2);CHKERRQ(ierr); /* 2 groups, active process and not active processes */ 4055e7931f94SStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)mat),&rank);CHKERRQ(ierr); 40562b510759SStefano Zampini ierr = PetscMPIIntCast(!local_size,&color);CHKERRQ(ierr); 4057d3531aaaSJed Brown ierr = PetscSubcommSetTypeGeneral(subcomm,color,rank);CHKERRQ(ierr); 4058e7931f94SStefano Zampini if (color) { 4059e7931f94SStefano Zampini ierr = PetscFree(xadj);CHKERRQ(ierr); 4060e7931f94SStefano Zampini ierr = PetscFree(adjncy);CHKERRQ(ierr); 4061e7931f94SStefano Zampini ierr = PetscFree(adjncy_wgt);CHKERRQ(ierr); 4062c8587f34SStefano Zampini } else { 406352e5ac9dSStefano Zampini Mat subdomain_adj; 406452e5ac9dSStefano Zampini IS new_ranks,new_ranks_contig; 406552e5ac9dSStefano Zampini MatPartitioning partitioner; 406652e5ac9dSStefano Zampini PetscInt prank,rstart=0,rend=0; 406752e5ac9dSStefano Zampini PetscInt *is_indices,*oldranks; 4068b0c7d250SStefano Zampini PetscBool aggregate; 4069b0c7d250SStefano Zampini 4070306c2d5bSBarry Smith ierr = MPI_Comm_size(PetscSubcommChild(subcomm),&size);CHKERRQ(ierr); 4071785e854fSJed Brown ierr = PetscMalloc1(size,&oldranks);CHKERRQ(ierr); 4072e7931f94SStefano Zampini prank = rank; 4073306c2d5bSBarry Smith ierr = MPI_Allgather(&prank,1,MPIU_INT,oldranks,1,MPIU_INT,PetscSubcommChild(subcomm));CHKERRQ(ierr); 40748002ef2cSStefano Zampini /* 4075e7931f94SStefano Zampini for (i=0;i<size;i++) { 4076e7931f94SStefano Zampini PetscPrintf(subcomm->comm,"oldranks[%d] = %d\n",i,oldranks[i]); 4077c8587f34SStefano Zampini } 40788002ef2cSStefano Zampini */ 4079e7931f94SStefano Zampini for (i=0;i<xadj[1];i++) { 4080e7931f94SStefano Zampini ierr = PetscFindInt(adjncy[i],size,oldranks,&adjncy[i]);CHKERRQ(ierr); 4081c8587f34SStefano Zampini } 4082e7931f94SStefano Zampini ierr = PetscSortIntWithArray(xadj[1],adjncy,adjncy_wgt);CHKERRQ(ierr); 4083b0c7d250SStefano Zampini aggregate = ((redprocs > 0 && redprocs < size) ? PETSC_TRUE : PETSC_FALSE); 4084b0c7d250SStefano Zampini if (aggregate) { 4085b0c7d250SStefano Zampini PetscInt lrows,row,ncols,*cols; 4086b0c7d250SStefano Zampini PetscMPIInt nrank; 4087b0c7d250SStefano Zampini PetscScalar *vals; 4088b0c7d250SStefano Zampini 4089b0c7d250SStefano Zampini ierr = MPI_Comm_rank(PetscSubcommChild(subcomm),&nrank);CHKERRQ(ierr); 4090b0c7d250SStefano Zampini lrows = 0; 4091b0c7d250SStefano Zampini if (nrank<redprocs) { 4092b0c7d250SStefano Zampini lrows = size/redprocs; 4093b0c7d250SStefano Zampini if (nrank<size%redprocs) lrows++; 4094b0c7d250SStefano Zampini } 40955fa240b1SStefano Zampini ierr = MatCreateAIJ(PetscSubcommChild(subcomm),lrows,lrows,size,size,50,NULL,50,NULL,&subdomain_adj);CHKERRQ(ierr); 4096b0c7d250SStefano Zampini ierr = MatGetOwnershipRange(subdomain_adj,&rstart,&rend);CHKERRQ(ierr); 4097b0c7d250SStefano Zampini ierr = MatSetOption(subdomain_adj,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_FALSE);CHKERRQ(ierr); 4098b0c7d250SStefano Zampini ierr = MatSetOption(subdomain_adj,MAT_NEW_NONZERO_ALLOCATION_ERR,PETSC_FALSE);CHKERRQ(ierr); 4099b0c7d250SStefano Zampini row = nrank; 4100b0c7d250SStefano Zampini ncols = xadj[1]-xadj[0]; 4101b0c7d250SStefano Zampini cols = adjncy; 4102b0c7d250SStefano Zampini ierr = PetscMalloc1(ncols,&vals);CHKERRQ(ierr); 4103b0c7d250SStefano Zampini for (i=0;i<ncols;i++) vals[i] = adjncy_wgt[i]; 4104b0c7d250SStefano Zampini ierr = MatSetValues(subdomain_adj,1,&row,ncols,cols,vals,INSERT_VALUES);CHKERRQ(ierr); 4105b0c7d250SStefano Zampini ierr = MatAssemblyBegin(subdomain_adj,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4106b0c7d250SStefano Zampini ierr = MatAssemblyEnd(subdomain_adj,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 410752e5ac9dSStefano Zampini ierr = PetscFree(xadj);CHKERRQ(ierr); 410852e5ac9dSStefano Zampini ierr = PetscFree(adjncy);CHKERRQ(ierr); 410952e5ac9dSStefano Zampini ierr = PetscFree(adjncy_wgt);CHKERRQ(ierr); 4110b0c7d250SStefano Zampini ierr = PetscFree(vals);CHKERRQ(ierr); 4111b0c7d250SStefano Zampini } else { 4112306c2d5bSBarry Smith ierr = MatCreateMPIAdj(PetscSubcommChild(subcomm),1,(PetscInt)size,xadj,adjncy,adjncy_wgt,&subdomain_adj);CHKERRQ(ierr); 4113b0c7d250SStefano Zampini } 411422b6e8a2SStefano Zampini /* ierr = MatView(subdomain_adj,0);CHKERRQ(ierr); */ 4115e7931f94SStefano Zampini 4116e7931f94SStefano Zampini /* Partition */ 4117306c2d5bSBarry Smith ierr = MatPartitioningCreate(PetscSubcommChild(subcomm),&partitioner);CHKERRQ(ierr); 4118e7931f94SStefano Zampini ierr = MatPartitioningSetAdjacency(partitioner,subdomain_adj);CHKERRQ(ierr); 4119e7931f94SStefano Zampini if (use_vwgt) { 41203837a79fSStefano Zampini ierr = PetscMalloc1(1,&v_wgt);CHKERRQ(ierr); 4121e7931f94SStefano Zampini v_wgt[0] = local_size; 4122e7931f94SStefano Zampini ierr = MatPartitioningSetVertexWeights(partitioner,v_wgt);CHKERRQ(ierr); 4123c8587f34SStefano Zampini } 412428143c3dSStefano Zampini n_subdomains = PetscMin((PetscInt)size,n_subdomains); 412528143c3dSStefano Zampini ierr = MatPartitioningSetNParts(partitioner,n_subdomains);CHKERRQ(ierr); 4126e7931f94SStefano Zampini ierr = MatPartitioningSetFromOptions(partitioner);CHKERRQ(ierr); 4127e7931f94SStefano Zampini ierr = MatPartitioningApply(partitioner,&new_ranks);CHKERRQ(ierr); 412822b6e8a2SStefano Zampini /* ierr = MatPartitioningView(partitioner,0);CHKERRQ(ierr); */ 4129e7931f94SStefano Zampini 413052e5ac9dSStefano Zampini /* renumber new_ranks to avoid "holes" in new set of processors */ 413152e5ac9dSStefano Zampini ierr = PCBDDCSubsetNumbering(new_ranks,NULL,NULL,&new_ranks_contig);CHKERRQ(ierr); 413252e5ac9dSStefano Zampini ierr = ISDestroy(&new_ranks);CHKERRQ(ierr); 413352e5ac9dSStefano Zampini ierr = ISGetIndices(new_ranks_contig,(const PetscInt**)&is_indices);CHKERRQ(ierr); 4134b0c7d250SStefano Zampini if (!redprocs) { 4135b0c7d250SStefano Zampini ranks_send_to_idx[0] = oldranks[is_indices[0]]; 413628143c3dSStefano Zampini } else { 4137b0c7d250SStefano Zampini PetscInt idxs[1]; 4138b0c7d250SStefano Zampini PetscMPIInt tag; 4139b0c7d250SStefano Zampini MPI_Request *reqs; 4140b0c7d250SStefano Zampini 4141b0c7d250SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)subdomain_adj,&tag);CHKERRQ(ierr); 4142b0c7d250SStefano Zampini ierr = PetscMalloc1(rend-rstart,&reqs);CHKERRQ(ierr); 4143b0c7d250SStefano Zampini for (i=rstart;i<rend;i++) { 4144b0c7d250SStefano Zampini ierr = MPI_Isend(is_indices+i-rstart,1,MPIU_INT,i,tag,PetscSubcommChild(subcomm),&reqs[i-rstart]);CHKERRQ(ierr); 414528143c3dSStefano Zampini } 4146b0c7d250SStefano Zampini ierr = MPI_Recv(idxs,1,MPIU_INT,MPI_ANY_SOURCE,tag,PetscSubcommChild(subcomm),MPI_STATUS_IGNORE);CHKERRQ(ierr); 4147b0c7d250SStefano Zampini ierr = MPI_Waitall(rend-rstart,reqs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 4148b0c7d250SStefano Zampini ierr = PetscFree(reqs);CHKERRQ(ierr); 4149b0c7d250SStefano Zampini ranks_send_to_idx[0] = oldranks[idxs[0]]; 4150e7931f94SStefano Zampini } 415152e5ac9dSStefano Zampini ierr = ISRestoreIndices(new_ranks_contig,(const PetscInt**)&is_indices);CHKERRQ(ierr); 4152e7931f94SStefano Zampini /* clean up */ 4153e7931f94SStefano Zampini ierr = PetscFree(oldranks);CHKERRQ(ierr); 415452e5ac9dSStefano Zampini ierr = ISDestroy(&new_ranks_contig);CHKERRQ(ierr); 4155e7931f94SStefano Zampini ierr = MatDestroy(&subdomain_adj);CHKERRQ(ierr); 4156e7931f94SStefano Zampini ierr = MatPartitioningDestroy(&partitioner);CHKERRQ(ierr); 4157e7931f94SStefano Zampini } 4158e7931f94SStefano Zampini ierr = PetscSubcommDestroy(&subcomm);CHKERRQ(ierr); 4159e7931f94SStefano Zampini 4160e7931f94SStefano Zampini /* assemble parallel IS for sends */ 4161e7931f94SStefano Zampini i = 1; 4162e7931f94SStefano Zampini if (color) i=0; 4163e7931f94SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)mat),i,ranks_send_to_idx,PETSC_OWN_POINTER,&ranks_send_to);CHKERRQ(ierr); 4164e7931f94SStefano Zampini /* get back IS */ 4165e7931f94SStefano Zampini *is_sends = ranks_send_to; 4166e7931f94SStefano Zampini PetscFunctionReturn(0); 4167e7931f94SStefano Zampini } 4168e7931f94SStefano Zampini 4169e7931f94SStefano Zampini typedef enum {MATDENSE_PRIVATE=0,MATAIJ_PRIVATE,MATBAIJ_PRIVATE,MATSBAIJ_PRIVATE}MatTypePrivate; 4170e7931f94SStefano Zampini 4171e7931f94SStefano Zampini #undef __FUNCT__ 4172e7931f94SStefano Zampini #define __FUNCT__ "MatISSubassemble" 417353a05cb3SStefano 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[]) 4174e7931f94SStefano Zampini { 417570cf5478SStefano Zampini Mat local_mat; 4176e7931f94SStefano Zampini IS is_sends_internal; 41779d30be91SStefano Zampini PetscInt rows,cols,new_local_rows; 417828143c3dSStefano Zampini PetscInt i,bs,buf_size_idxs,buf_size_idxs_is,buf_size_vals; 41799d30be91SStefano Zampini PetscBool ismatis,isdense,newisdense,destroy_mat; 4180e7931f94SStefano Zampini ISLocalToGlobalMapping l2gmap; 4181e7931f94SStefano Zampini PetscInt* l2gmap_indices; 4182e7931f94SStefano Zampini const PetscInt* is_indices; 4183e7931f94SStefano Zampini MatType new_local_type; 4184e7931f94SStefano Zampini /* buffers */ 4185e7931f94SStefano Zampini PetscInt *ptr_idxs,*send_buffer_idxs,*recv_buffer_idxs; 418628143c3dSStefano Zampini PetscInt *ptr_idxs_is,*send_buffer_idxs_is,*recv_buffer_idxs_is; 41879d30be91SStefano Zampini PetscInt *recv_buffer_idxs_local; 4188e7931f94SStefano Zampini PetscScalar *ptr_vals,*send_buffer_vals,*recv_buffer_vals; 4189e7931f94SStefano Zampini /* MPI */ 419028143c3dSStefano Zampini MPI_Comm comm,comm_n; 419128143c3dSStefano Zampini PetscSubcomm subcomm; 4192e7931f94SStefano Zampini PetscMPIInt n_sends,n_recvs,commsize; 419328143c3dSStefano Zampini PetscMPIInt *iflags,*ilengths_idxs,*ilengths_vals,*ilengths_idxs_is; 419428143c3dSStefano Zampini PetscMPIInt *onodes,*onodes_is,*olengths_idxs,*olengths_idxs_is,*olengths_vals; 419528143c3dSStefano Zampini PetscMPIInt len,tag_idxs,tag_idxs_is,tag_vals,source_dest; 419628143c3dSStefano Zampini MPI_Request *send_req_idxs,*send_req_idxs_is,*send_req_vals; 419728143c3dSStefano Zampini MPI_Request *recv_req_idxs,*recv_req_idxs_is,*recv_req_vals; 4198e7931f94SStefano Zampini PetscErrorCode ierr; 4199e7931f94SStefano Zampini 4200e7931f94SStefano Zampini PetscFunctionBegin; 420128143c3dSStefano Zampini /* TODO: add missing checks */ 420228143c3dSStefano Zampini PetscValidLogicalCollectiveInt(mat,n_subdomains,3); 420328143c3dSStefano Zampini PetscValidLogicalCollectiveBool(mat,restrict_comm,4); 420428143c3dSStefano Zampini PetscValidLogicalCollectiveEnum(mat,reuse,5); 420528143c3dSStefano Zampini PetscValidLogicalCollectiveInt(mat,nis,7); 4206e7931f94SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)mat,MATIS,&ismatis);CHKERRQ(ierr); 420728143c3dSStefano Zampini if (!ismatis) SETERRQ1(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot use %s on a matrix object which is not of type MATIS",__FUNCT__); 4208e7931f94SStefano Zampini ierr = MatISGetLocalMat(mat,&local_mat);CHKERRQ(ierr); 4209e7931f94SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)local_mat,MATSEQDENSE,&isdense);CHKERRQ(ierr); 4210e7931f94SStefano Zampini if (!isdense) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Currently cannot subassemble MATIS when local matrix type is not of type SEQDENSE"); 4211e7931f94SStefano Zampini ierr = MatGetSize(local_mat,&rows,&cols);CHKERRQ(ierr); 4212e7931f94SStefano Zampini if (rows != cols) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Local MATIS matrices should be square"); 421328143c3dSStefano Zampini if (reuse == MAT_REUSE_MATRIX && *mat_n) { 421470cf5478SStefano Zampini PetscInt mrows,mcols,mnrows,mncols; 421570cf5478SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)*mat_n,MATIS,&ismatis);CHKERRQ(ierr); 421628143c3dSStefano Zampini if (!ismatis) SETERRQ(PetscObjectComm((PetscObject)*mat_n),PETSC_ERR_SUP,"Cannot reuse a matrix which is not of type MATIS"); 421770cf5478SStefano Zampini ierr = MatGetSize(mat,&mrows,&mcols);CHKERRQ(ierr); 421870cf5478SStefano Zampini ierr = MatGetSize(*mat_n,&mnrows,&mncols);CHKERRQ(ierr); 421970cf5478SStefano Zampini if (mrows != mnrows) SETERRQ2(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot reuse matrix! Wrong number of rows %D != %D",mrows,mnrows); 422070cf5478SStefano Zampini if (mcols != mncols) SETERRQ2(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot reuse matrix! Wrong number of cols %D != %D",mcols,mncols); 422170cf5478SStefano Zampini } 4222e7931f94SStefano Zampini ierr = MatGetBlockSize(local_mat,&bs);CHKERRQ(ierr); 4223e7931f94SStefano Zampini PetscValidLogicalCollectiveInt(mat,bs,0); 4224e7931f94SStefano Zampini /* prepare IS for sending if not provided */ 4225e7931f94SStefano Zampini if (!is_sends) { 422628143c3dSStefano Zampini if (!n_subdomains) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"You should specify either an IS or a target number of subdomains"); 4227b0c7d250SStefano Zampini ierr = MatISGetSubassemblingPattern(mat,n_subdomains,0,&is_sends_internal);CHKERRQ(ierr); 4228c8587f34SStefano Zampini } else { 4229e7931f94SStefano Zampini ierr = PetscObjectReference((PetscObject)is_sends);CHKERRQ(ierr); 4230e7931f94SStefano Zampini is_sends_internal = is_sends; 4231c8587f34SStefano Zampini } 4232e7931f94SStefano Zampini 4233e7931f94SStefano Zampini /* get comm */ 4234a316fed8SStefano Zampini ierr = PetscObjectGetComm((PetscObject)mat,&comm);CHKERRQ(ierr); 4235e7931f94SStefano Zampini 4236e7931f94SStefano Zampini /* compute number of sends */ 4237e7931f94SStefano Zampini ierr = ISGetLocalSize(is_sends_internal,&i);CHKERRQ(ierr); 4238e7931f94SStefano Zampini ierr = PetscMPIIntCast(i,&n_sends);CHKERRQ(ierr); 4239e7931f94SStefano Zampini 4240e7931f94SStefano Zampini /* compute number of receives */ 4241e7931f94SStefano Zampini ierr = MPI_Comm_size(comm,&commsize);CHKERRQ(ierr); 4242785e854fSJed Brown ierr = PetscMalloc1(commsize,&iflags);CHKERRQ(ierr); 4243e7931f94SStefano Zampini ierr = PetscMemzero(iflags,commsize*sizeof(*iflags));CHKERRQ(ierr); 4244e7931f94SStefano Zampini ierr = ISGetIndices(is_sends_internal,&is_indices);CHKERRQ(ierr); 4245e7931f94SStefano Zampini for (i=0;i<n_sends;i++) iflags[is_indices[i]] = 1; 4246e7931f94SStefano Zampini ierr = PetscGatherNumberOfMessages(comm,iflags,NULL,&n_recvs);CHKERRQ(ierr); 4247e7931f94SStefano Zampini ierr = PetscFree(iflags);CHKERRQ(ierr); 4248e7931f94SStefano Zampini 424928143c3dSStefano Zampini /* restrict comm if requested */ 425028143c3dSStefano Zampini subcomm = 0; 425128143c3dSStefano Zampini destroy_mat = PETSC_FALSE; 425228143c3dSStefano Zampini if (restrict_comm) { 4253779c1cceSStefano Zampini PetscMPIInt color,subcommsize; 4254779c1cceSStefano Zampini 425528143c3dSStefano Zampini color = 0; 425653a05cb3SStefano Zampini if (restrict_full) { 425753a05cb3SStefano Zampini if (!n_recvs) color = 1; /* processes not receiving anything will not partecipate in new comm (full restriction) */ 425853a05cb3SStefano Zampini } else { 425953a05cb3SStefano Zampini if (!n_recvs && n_sends) color = 1; /* just those processes that are sending but not receiving anything will not partecipate in new comm */ 426053a05cb3SStefano Zampini } 426128143c3dSStefano Zampini ierr = MPI_Allreduce(&color,&subcommsize,1,MPI_INT,MPI_SUM,comm);CHKERRQ(ierr); 426228143c3dSStefano Zampini subcommsize = commsize - subcommsize; 426328143c3dSStefano Zampini /* check if reuse has been requested */ 426428143c3dSStefano Zampini if (reuse == MAT_REUSE_MATRIX) { 426528143c3dSStefano Zampini if (*mat_n) { 426628143c3dSStefano Zampini PetscMPIInt subcommsize2; 426728143c3dSStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)*mat_n),&subcommsize2);CHKERRQ(ierr); 426828143c3dSStefano Zampini if (subcommsize != subcommsize2) SETERRQ2(PetscObjectComm((PetscObject)*mat_n),PETSC_ERR_PLIB,"Cannot reuse matrix! wrong subcomm size %d != %d",subcommsize,subcommsize2); 426928143c3dSStefano Zampini comm_n = PetscObjectComm((PetscObject)*mat_n); 427028143c3dSStefano Zampini } else { 427128143c3dSStefano Zampini comm_n = PETSC_COMM_SELF; 427228143c3dSStefano Zampini } 427328143c3dSStefano Zampini } else { /* MAT_INITIAL_MATRIX */ 4274779c1cceSStefano Zampini PetscMPIInt rank; 4275779c1cceSStefano Zampini 4276779c1cceSStefano Zampini ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 427728143c3dSStefano Zampini ierr = PetscSubcommCreate(comm,&subcomm);CHKERRQ(ierr); 427828143c3dSStefano Zampini ierr = PetscSubcommSetNumber(subcomm,2);CHKERRQ(ierr); 427928143c3dSStefano Zampini ierr = PetscSubcommSetTypeGeneral(subcomm,color,rank);CHKERRQ(ierr); 4280306c2d5bSBarry Smith comm_n = PetscSubcommChild(subcomm); 428128143c3dSStefano Zampini } 428228143c3dSStefano Zampini /* flag to destroy *mat_n if not significative */ 428328143c3dSStefano Zampini if (color) destroy_mat = PETSC_TRUE; 428428143c3dSStefano Zampini } else { 428528143c3dSStefano Zampini comm_n = comm; 428628143c3dSStefano Zampini } 428728143c3dSStefano Zampini 4288e7931f94SStefano Zampini /* prepare send/receive buffers */ 4289785e854fSJed Brown ierr = PetscMalloc1(commsize,&ilengths_idxs);CHKERRQ(ierr); 4290e7931f94SStefano Zampini ierr = PetscMemzero(ilengths_idxs,commsize*sizeof(*ilengths_idxs));CHKERRQ(ierr); 4291785e854fSJed Brown ierr = PetscMalloc1(commsize,&ilengths_vals);CHKERRQ(ierr); 4292e7931f94SStefano Zampini ierr = PetscMemzero(ilengths_vals,commsize*sizeof(*ilengths_vals));CHKERRQ(ierr); 429328143c3dSStefano Zampini if (nis) { 4294854ce69bSBarry Smith ierr = PetscCalloc1(commsize,&ilengths_idxs_is);CHKERRQ(ierr); 429528143c3dSStefano Zampini } 4296e7931f94SStefano Zampini 429728143c3dSStefano Zampini /* Get data from local matrices */ 4298e7931f94SStefano Zampini if (!isdense) { 4299a26c9d0eSStefano Zampini SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Subassembling of AIJ local matrices not yet implemented"); 4300e7931f94SStefano Zampini /* TODO: See below some guidelines on how to prepare the local buffers */ 4301e7931f94SStefano Zampini /* 4302e7931f94SStefano Zampini send_buffer_vals should contain the raw values of the local matrix 4303e7931f94SStefano Zampini send_buffer_idxs should contain: 4304e7931f94SStefano Zampini - MatType_PRIVATE type 4305e7931f94SStefano Zampini - PetscInt size_of_l2gmap 4306e7931f94SStefano Zampini - PetscInt global_row_indices[size_of_l2gmap] 4307e7931f94SStefano Zampini - PetscInt all_other_info_which_is_needed_to_compute_preallocation_and_set_values 4308e7931f94SStefano Zampini */ 4309e7931f94SStefano Zampini } else { 4310e7931f94SStefano Zampini ierr = MatDenseGetArray(local_mat,&send_buffer_vals);CHKERRQ(ierr); 43113bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetSize(mat->rmap->mapping,&i);CHKERRQ(ierr); 4312854ce69bSBarry Smith ierr = PetscMalloc1(i+2,&send_buffer_idxs);CHKERRQ(ierr); 4313e7931f94SStefano Zampini send_buffer_idxs[0] = (PetscInt)MATDENSE_PRIVATE; 4314e7931f94SStefano Zampini send_buffer_idxs[1] = i; 43153bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetIndices(mat->rmap->mapping,(const PetscInt**)&ptr_idxs);CHKERRQ(ierr); 4316e7931f94SStefano Zampini ierr = PetscMemcpy(&send_buffer_idxs[2],ptr_idxs,i*sizeof(PetscInt));CHKERRQ(ierr); 43173bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingRestoreIndices(mat->rmap->mapping,(const PetscInt**)&ptr_idxs);CHKERRQ(ierr); 4318e7931f94SStefano Zampini ierr = PetscMPIIntCast(i,&len);CHKERRQ(ierr); 4319e7931f94SStefano Zampini for (i=0;i<n_sends;i++) { 4320e7931f94SStefano Zampini ilengths_vals[is_indices[i]] = len*len; 4321e7931f94SStefano Zampini ilengths_idxs[is_indices[i]] = len+2; 4322c8587f34SStefano Zampini } 4323c8587f34SStefano Zampini } 4324e7931f94SStefano Zampini ierr = PetscGatherMessageLengths2(comm,n_sends,n_recvs,ilengths_idxs,ilengths_vals,&onodes,&olengths_idxs,&olengths_vals);CHKERRQ(ierr); 432528143c3dSStefano Zampini /* additional is (if any) */ 432628143c3dSStefano Zampini if (nis) { 432728143c3dSStefano Zampini PetscMPIInt psum; 432828143c3dSStefano Zampini PetscInt j; 432928143c3dSStefano Zampini for (j=0,psum=0;j<nis;j++) { 433028143c3dSStefano Zampini PetscInt plen; 433128143c3dSStefano Zampini ierr = ISGetLocalSize(isarray[j],&plen);CHKERRQ(ierr); 433228143c3dSStefano Zampini ierr = PetscMPIIntCast(plen,&len);CHKERRQ(ierr); 433328143c3dSStefano Zampini psum += len+1; /* indices + lenght */ 433428143c3dSStefano Zampini } 4335854ce69bSBarry Smith ierr = PetscMalloc1(psum,&send_buffer_idxs_is);CHKERRQ(ierr); 433628143c3dSStefano Zampini for (j=0,psum=0;j<nis;j++) { 433728143c3dSStefano Zampini PetscInt plen; 433828143c3dSStefano Zampini const PetscInt *is_array_idxs; 433928143c3dSStefano Zampini ierr = ISGetLocalSize(isarray[j],&plen);CHKERRQ(ierr); 434028143c3dSStefano Zampini send_buffer_idxs_is[psum] = plen; 434128143c3dSStefano Zampini ierr = ISGetIndices(isarray[j],&is_array_idxs);CHKERRQ(ierr); 434228143c3dSStefano Zampini ierr = PetscMemcpy(&send_buffer_idxs_is[psum+1],is_array_idxs,plen*sizeof(PetscInt));CHKERRQ(ierr); 434328143c3dSStefano Zampini ierr = ISRestoreIndices(isarray[j],&is_array_idxs);CHKERRQ(ierr); 434428143c3dSStefano Zampini psum += plen+1; /* indices + lenght */ 434528143c3dSStefano Zampini } 434628143c3dSStefano Zampini for (i=0;i<n_sends;i++) { 434728143c3dSStefano Zampini ilengths_idxs_is[is_indices[i]] = psum; 434828143c3dSStefano Zampini } 434928143c3dSStefano Zampini ierr = PetscGatherMessageLengths(comm,n_sends,n_recvs,ilengths_idxs_is,&onodes_is,&olengths_idxs_is);CHKERRQ(ierr); 435028143c3dSStefano Zampini } 435128143c3dSStefano Zampini 4352e7931f94SStefano Zampini buf_size_idxs = 0; 4353e7931f94SStefano Zampini buf_size_vals = 0; 435428143c3dSStefano Zampini buf_size_idxs_is = 0; 4355e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 4356e7931f94SStefano Zampini buf_size_idxs += (PetscInt)olengths_idxs[i]; 4357e7931f94SStefano Zampini buf_size_vals += (PetscInt)olengths_vals[i]; 435828143c3dSStefano Zampini if (nis) buf_size_idxs_is += (PetscInt)olengths_idxs_is[i]; 4359e7931f94SStefano Zampini } 4360785e854fSJed Brown ierr = PetscMalloc1(buf_size_idxs,&recv_buffer_idxs);CHKERRQ(ierr); 4361785e854fSJed Brown ierr = PetscMalloc1(buf_size_vals,&recv_buffer_vals);CHKERRQ(ierr); 436295ecbf38SStefano Zampini ierr = PetscMalloc1(buf_size_idxs_is,&recv_buffer_idxs_is);CHKERRQ(ierr); 4363e7931f94SStefano Zampini 4364e7931f94SStefano Zampini /* get new tags for clean communications */ 4365e7931f94SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_idxs);CHKERRQ(ierr); 4366e7931f94SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_vals);CHKERRQ(ierr); 436728143c3dSStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_idxs_is);CHKERRQ(ierr); 4368e7931f94SStefano Zampini 4369e7931f94SStefano Zampini /* allocate for requests */ 4370785e854fSJed Brown ierr = PetscMalloc1(n_sends,&send_req_idxs);CHKERRQ(ierr); 4371785e854fSJed Brown ierr = PetscMalloc1(n_sends,&send_req_vals);CHKERRQ(ierr); 437295ecbf38SStefano Zampini ierr = PetscMalloc1(n_sends,&send_req_idxs_is);CHKERRQ(ierr); 4373785e854fSJed Brown ierr = PetscMalloc1(n_recvs,&recv_req_idxs);CHKERRQ(ierr); 4374785e854fSJed Brown ierr = PetscMalloc1(n_recvs,&recv_req_vals);CHKERRQ(ierr); 437595ecbf38SStefano Zampini ierr = PetscMalloc1(n_recvs,&recv_req_idxs_is);CHKERRQ(ierr); 4376e7931f94SStefano Zampini 4377e7931f94SStefano Zampini /* communications */ 4378e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 4379e7931f94SStefano Zampini ptr_vals = recv_buffer_vals; 438028143c3dSStefano Zampini ptr_idxs_is = recv_buffer_idxs_is; 4381e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 4382e7931f94SStefano Zampini source_dest = onodes[i]; 4383e7931f94SStefano Zampini ierr = MPI_Irecv(ptr_idxs,olengths_idxs[i],MPIU_INT,source_dest,tag_idxs,comm,&recv_req_idxs[i]);CHKERRQ(ierr); 4384e7931f94SStefano Zampini ierr = MPI_Irecv(ptr_vals,olengths_vals[i],MPIU_SCALAR,source_dest,tag_vals,comm,&recv_req_vals[i]);CHKERRQ(ierr); 4385e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 4386e7931f94SStefano Zampini ptr_vals += olengths_vals[i]; 438728143c3dSStefano Zampini if (nis) { 438828143c3dSStefano 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); 438928143c3dSStefano Zampini ptr_idxs_is += olengths_idxs_is[i]; 439028143c3dSStefano Zampini } 4391e7931f94SStefano Zampini } 4392e7931f94SStefano Zampini for (i=0;i<n_sends;i++) { 4393e7931f94SStefano Zampini ierr = PetscMPIIntCast(is_indices[i],&source_dest);CHKERRQ(ierr); 4394e7931f94SStefano Zampini ierr = MPI_Isend(send_buffer_idxs,ilengths_idxs[source_dest],MPIU_INT,source_dest,tag_idxs,comm,&send_req_idxs[i]);CHKERRQ(ierr); 4395e7931f94SStefano Zampini ierr = MPI_Isend(send_buffer_vals,ilengths_vals[source_dest],MPIU_SCALAR,source_dest,tag_vals,comm,&send_req_vals[i]);CHKERRQ(ierr); 439628143c3dSStefano Zampini if (nis) { 439728143c3dSStefano 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); 439828143c3dSStefano Zampini } 4399e7931f94SStefano Zampini } 4400e7931f94SStefano Zampini ierr = ISRestoreIndices(is_sends_internal,&is_indices);CHKERRQ(ierr); 4401e7931f94SStefano Zampini ierr = ISDestroy(&is_sends_internal);CHKERRQ(ierr); 4402e7931f94SStefano Zampini 4403e7931f94SStefano Zampini /* assemble new l2g map */ 4404e7931f94SStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_idxs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 4405e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 44069d30be91SStefano Zampini new_local_rows = 0; 4407e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 44089d30be91SStefano Zampini new_local_rows += *(ptr_idxs+1); /* second element is the local size of the l2gmap */ 4409e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 4410e7931f94SStefano Zampini } 44119d30be91SStefano Zampini ierr = PetscMalloc1(new_local_rows,&l2gmap_indices);CHKERRQ(ierr); 4412e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 44139d30be91SStefano Zampini new_local_rows = 0; 4414e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 44159d30be91SStefano Zampini ierr = PetscMemcpy(&l2gmap_indices[new_local_rows],ptr_idxs+2,(*(ptr_idxs+1))*sizeof(PetscInt));CHKERRQ(ierr); 44169d30be91SStefano Zampini new_local_rows += *(ptr_idxs+1); /* second element is the local size of the l2gmap */ 4417e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 4418e7931f94SStefano Zampini } 44199d30be91SStefano Zampini ierr = PetscSortRemoveDupsInt(&new_local_rows,l2gmap_indices);CHKERRQ(ierr); 44209d30be91SStefano Zampini ierr = ISLocalToGlobalMappingCreate(comm_n,1,new_local_rows,l2gmap_indices,PETSC_COPY_VALUES,&l2gmap);CHKERRQ(ierr); 4421e7931f94SStefano Zampini ierr = PetscFree(l2gmap_indices);CHKERRQ(ierr); 4422e7931f94SStefano Zampini 4423e7931f94SStefano Zampini /* infer new local matrix type from received local matrices type */ 4424e7931f94SStefano Zampini /* currently if all local matrices are of type X, then the resulting matrix will be of type X, except for the dense case */ 4425e7931f94SStefano 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) */ 4426e7931f94SStefano Zampini if (n_recvs) { 442728143c3dSStefano Zampini MatTypePrivate new_local_type_private = (MatTypePrivate)send_buffer_idxs[0]; 4428e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 4429e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 4430e7931f94SStefano Zampini if ((PetscInt)new_local_type_private != *ptr_idxs) { 4431e7931f94SStefano Zampini new_local_type_private = MATAIJ_PRIVATE; 4432e7931f94SStefano Zampini break; 4433e7931f94SStefano Zampini } 4434e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 4435e7931f94SStefano Zampini } 4436e7931f94SStefano Zampini switch (new_local_type_private) { 443728143c3dSStefano Zampini case MATDENSE_PRIVATE: 443828143c3dSStefano Zampini if (n_recvs>1) { /* subassembling of dense matrices does not give a dense matrix! */ 4439e7931f94SStefano Zampini new_local_type = MATSEQAIJ; 4440e7931f94SStefano Zampini bs = 1; 444128143c3dSStefano Zampini } else { /* if I receive only 1 dense matrix */ 444228143c3dSStefano Zampini new_local_type = MATSEQDENSE; 444328143c3dSStefano Zampini bs = 1; 444428143c3dSStefano Zampini } 4445e7931f94SStefano Zampini break; 4446e7931f94SStefano Zampini case MATAIJ_PRIVATE: 4447e7931f94SStefano Zampini new_local_type = MATSEQAIJ; 4448e7931f94SStefano Zampini bs = 1; 4449e7931f94SStefano Zampini break; 4450e7931f94SStefano Zampini case MATBAIJ_PRIVATE: 4451e7931f94SStefano Zampini new_local_type = MATSEQBAIJ; 4452e7931f94SStefano Zampini break; 4453e7931f94SStefano Zampini case MATSBAIJ_PRIVATE: 4454e7931f94SStefano Zampini new_local_type = MATSEQSBAIJ; 4455e7931f94SStefano Zampini break; 4456e7931f94SStefano Zampini default: 44579d30be91SStefano Zampini SETERRQ2(comm,PETSC_ERR_SUP,"Unsupported private type %d in %s",new_local_type_private,__FUNCT__); 4458e7931f94SStefano Zampini break; 4459e7931f94SStefano Zampini } 446028143c3dSStefano Zampini } else { /* by default, new_local_type is seqdense */ 446128143c3dSStefano Zampini new_local_type = MATSEQDENSE; 446228143c3dSStefano Zampini bs = 1; 4463e7931f94SStefano Zampini } 4464e7931f94SStefano Zampini 446570cf5478SStefano Zampini /* create MATIS object if needed */ 446670cf5478SStefano Zampini if (reuse == MAT_INITIAL_MATRIX) { 4467e7931f94SStefano Zampini ierr = MatGetSize(mat,&rows,&cols);CHKERRQ(ierr); 4468e176bc59SStefano Zampini ierr = MatCreateIS(comm_n,bs,PETSC_DECIDE,PETSC_DECIDE,rows,cols,l2gmap,NULL,mat_n);CHKERRQ(ierr); 446970cf5478SStefano Zampini } else { 447070cf5478SStefano Zampini /* it also destroys the local matrices */ 447170cf5478SStefano Zampini ierr = MatSetLocalToGlobalMapping(*mat_n,l2gmap,l2gmap);CHKERRQ(ierr); 447270cf5478SStefano Zampini } 447370cf5478SStefano Zampini ierr = MatISGetLocalMat(*mat_n,&local_mat);CHKERRQ(ierr); 4474e7931f94SStefano Zampini ierr = MatSetType(local_mat,new_local_type);CHKERRQ(ierr); 44759d30be91SStefano Zampini 44769d30be91SStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_vals,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 44779d30be91SStefano Zampini 44789d30be91SStefano Zampini /* Global to local map of received indices */ 44799d30be91SStefano Zampini ierr = PetscMalloc1(buf_size_idxs,&recv_buffer_idxs_local);CHKERRQ(ierr); /* needed for values insertion */ 44809d30be91SStefano Zampini ierr = ISGlobalToLocalMappingApply(l2gmap,IS_GTOLM_MASK,buf_size_idxs,recv_buffer_idxs,&i,recv_buffer_idxs_local);CHKERRQ(ierr); 44819d30be91SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&l2gmap);CHKERRQ(ierr); 44829d30be91SStefano Zampini 44839d30be91SStefano Zampini /* restore attributes -> type of incoming data and its size */ 44849d30be91SStefano Zampini buf_size_idxs = 0; 44859d30be91SStefano Zampini for (i=0;i<n_recvs;i++) { 44869d30be91SStefano Zampini recv_buffer_idxs_local[buf_size_idxs] = recv_buffer_idxs[buf_size_idxs]; 44879d30be91SStefano Zampini recv_buffer_idxs_local[buf_size_idxs+1] = recv_buffer_idxs[buf_size_idxs+1]; 44889d30be91SStefano Zampini buf_size_idxs += (PetscInt)olengths_idxs[i]; 44899d30be91SStefano Zampini } 44909d30be91SStefano Zampini ierr = PetscFree(recv_buffer_idxs);CHKERRQ(ierr); 44919d30be91SStefano Zampini 44929d30be91SStefano Zampini /* set preallocation */ 44939d30be91SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)local_mat,MATSEQDENSE,&newisdense);CHKERRQ(ierr); 44949d30be91SStefano Zampini if (!newisdense) { 44959d30be91SStefano Zampini PetscInt *new_local_nnz=0; 44969d30be91SStefano Zampini 44979d30be91SStefano Zampini ptr_vals = recv_buffer_vals; 44989d30be91SStefano Zampini ptr_idxs = recv_buffer_idxs_local; 44999d30be91SStefano Zampini if (n_recvs) { 45009d30be91SStefano Zampini ierr = PetscCalloc1(new_local_rows,&new_local_nnz);CHKERRQ(ierr); 45019d30be91SStefano Zampini } 45029d30be91SStefano Zampini for (i=0;i<n_recvs;i++) { 45039d30be91SStefano Zampini PetscInt j; 45049d30be91SStefano Zampini if (*ptr_idxs == (PetscInt)MATDENSE_PRIVATE) { /* preallocation provided for dense case only */ 45059d30be91SStefano Zampini for (j=0;j<*(ptr_idxs+1);j++) { 45069d30be91SStefano Zampini new_local_nnz[*(ptr_idxs+2+j)] += *(ptr_idxs+1); 45079d30be91SStefano Zampini } 45089d30be91SStefano Zampini } else { 45099d30be91SStefano Zampini /* TODO */ 45109d30be91SStefano Zampini } 45119d30be91SStefano Zampini ptr_idxs += olengths_idxs[i]; 45129d30be91SStefano Zampini } 45139d30be91SStefano Zampini if (new_local_nnz) { 45149d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] = PetscMin(new_local_nnz[i],new_local_rows); 45159d30be91SStefano Zampini ierr = MatSeqAIJSetPreallocation(local_mat,0,new_local_nnz);CHKERRQ(ierr); 45169d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] /= bs; 45179d30be91SStefano Zampini ierr = MatSeqBAIJSetPreallocation(local_mat,bs,0,new_local_nnz);CHKERRQ(ierr); 45189d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] = PetscMax(new_local_nnz[i]-i,0); 45199d30be91SStefano Zampini ierr = MatSeqSBAIJSetPreallocation(local_mat,bs,0,new_local_nnz);CHKERRQ(ierr); 45209d30be91SStefano Zampini } else { 45219d30be91SStefano Zampini ierr = MatSetUp(local_mat);CHKERRQ(ierr); 45229d30be91SStefano Zampini } 45239d30be91SStefano Zampini ierr = PetscFree(new_local_nnz);CHKERRQ(ierr); 45249d30be91SStefano Zampini } else { 45259d30be91SStefano Zampini ierr = MatSetUp(local_mat);CHKERRQ(ierr); 45269d30be91SStefano Zampini } 4527e7931f94SStefano Zampini 4528e7931f94SStefano Zampini /* set values */ 4529e7931f94SStefano Zampini ptr_vals = recv_buffer_vals; 45309d30be91SStefano Zampini ptr_idxs = recv_buffer_idxs_local; 4531e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 4532e7931f94SStefano Zampini if (*ptr_idxs == (PetscInt)MATDENSE_PRIVATE) { /* values insertion provided for dense case only */ 4533e7931f94SStefano Zampini ierr = MatSetOption(local_mat,MAT_ROW_ORIENTED,PETSC_FALSE);CHKERRQ(ierr); 45349d30be91SStefano Zampini ierr = MatSetValues(local_mat,*(ptr_idxs+1),ptr_idxs+2,*(ptr_idxs+1),ptr_idxs+2,ptr_vals,ADD_VALUES);CHKERRQ(ierr); 4535e7931f94SStefano Zampini ierr = MatAssemblyBegin(local_mat,MAT_FLUSH_ASSEMBLY);CHKERRQ(ierr); 4536e7931f94SStefano Zampini ierr = MatAssemblyEnd(local_mat,MAT_FLUSH_ASSEMBLY);CHKERRQ(ierr); 4537e7931f94SStefano Zampini ierr = MatSetOption(local_mat,MAT_ROW_ORIENTED,PETSC_TRUE);CHKERRQ(ierr); 453828143c3dSStefano Zampini } else { 453928143c3dSStefano Zampini /* TODO */ 4540e7931f94SStefano Zampini } 4541e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 4542e7931f94SStefano Zampini ptr_vals += olengths_vals[i]; 4543e7931f94SStefano Zampini } 4544e7931f94SStefano Zampini ierr = MatAssemblyBegin(local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4545e7931f94SStefano Zampini ierr = MatAssemblyEnd(local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 454670cf5478SStefano Zampini ierr = MatAssemblyBegin(*mat_n,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 454770cf5478SStefano Zampini ierr = MatAssemblyEnd(*mat_n,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 45489d30be91SStefano Zampini ierr = PetscFree(recv_buffer_vals);CHKERRQ(ierr); 45499d30be91SStefano Zampini ierr = PetscFree(recv_buffer_idxs_local);CHKERRQ(ierr); 4550e7931f94SStefano Zampini 4551dfd14d43SStefano Zampini #if 0 455228143c3dSStefano Zampini if (!restrict_comm) { /* check */ 4553e7931f94SStefano Zampini Vec lvec,rvec; 4554e7931f94SStefano Zampini PetscReal infty_error; 4555e7931f94SStefano Zampini 45562a7a6963SBarry Smith ierr = MatCreateVecs(mat,&rvec,&lvec);CHKERRQ(ierr); 4557e7931f94SStefano Zampini ierr = VecSetRandom(rvec,NULL);CHKERRQ(ierr); 4558e7931f94SStefano Zampini ierr = MatMult(mat,rvec,lvec);CHKERRQ(ierr); 4559e7931f94SStefano Zampini ierr = VecScale(lvec,-1.0);CHKERRQ(ierr); 456070cf5478SStefano Zampini ierr = MatMultAdd(*mat_n,rvec,lvec,lvec);CHKERRQ(ierr); 4561e7931f94SStefano Zampini ierr = VecNorm(lvec,NORM_INFINITY,&infty_error);CHKERRQ(ierr); 4562e7931f94SStefano Zampini ierr = PetscPrintf(PetscObjectComm((PetscObject)mat),"Infinity error subassembling %1.6e\n",infty_error); 4563e7931f94SStefano Zampini ierr = VecDestroy(&rvec);CHKERRQ(ierr); 4564e7931f94SStefano Zampini ierr = VecDestroy(&lvec);CHKERRQ(ierr); 4565e7931f94SStefano Zampini } 456628143c3dSStefano Zampini #endif 4567e7931f94SStefano Zampini 456828143c3dSStefano Zampini /* assemble new additional is (if any) */ 456928143c3dSStefano Zampini if (nis) { 457028143c3dSStefano Zampini PetscInt **temp_idxs,*count_is,j,psum; 457128143c3dSStefano Zampini 457228143c3dSStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_idxs_is,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 4573854ce69bSBarry Smith ierr = PetscCalloc1(nis,&count_is);CHKERRQ(ierr); 457428143c3dSStefano Zampini ptr_idxs = recv_buffer_idxs_is; 457528143c3dSStefano Zampini psum = 0; 457628143c3dSStefano Zampini for (i=0;i<n_recvs;i++) { 457728143c3dSStefano Zampini for (j=0;j<nis;j++) { 457828143c3dSStefano Zampini PetscInt plen = *(ptr_idxs); /* first element is the local size of IS's indices */ 457928143c3dSStefano Zampini count_is[j] += plen; /* increment counting of buffer for j-th IS */ 458028143c3dSStefano Zampini psum += plen; 458128143c3dSStefano Zampini ptr_idxs += plen+1; /* shift pointer to received data */ 458228143c3dSStefano Zampini } 458328143c3dSStefano Zampini } 4584854ce69bSBarry Smith ierr = PetscMalloc1(nis,&temp_idxs);CHKERRQ(ierr); 4585854ce69bSBarry Smith ierr = PetscMalloc1(psum,&temp_idxs[0]);CHKERRQ(ierr); 458628143c3dSStefano Zampini for (i=1;i<nis;i++) { 458728143c3dSStefano Zampini temp_idxs[i] = temp_idxs[i-1]+count_is[i-1]; 458828143c3dSStefano Zampini } 458928143c3dSStefano Zampini ierr = PetscMemzero(count_is,nis*sizeof(PetscInt));CHKERRQ(ierr); 459028143c3dSStefano Zampini ptr_idxs = recv_buffer_idxs_is; 459128143c3dSStefano Zampini for (i=0;i<n_recvs;i++) { 459228143c3dSStefano Zampini for (j=0;j<nis;j++) { 459328143c3dSStefano Zampini PetscInt plen = *(ptr_idxs); /* first element is the local size of IS's indices */ 459428143c3dSStefano Zampini ierr = PetscMemcpy(&temp_idxs[j][count_is[j]],ptr_idxs+1,plen*sizeof(PetscInt));CHKERRQ(ierr); 459528143c3dSStefano Zampini count_is[j] += plen; /* increment starting point of buffer for j-th IS */ 459628143c3dSStefano Zampini ptr_idxs += plen+1; /* shift pointer to received data */ 459728143c3dSStefano Zampini } 459828143c3dSStefano Zampini } 459928143c3dSStefano Zampini for (i=0;i<nis;i++) { 460028143c3dSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 460128143c3dSStefano Zampini ierr = PetscSortRemoveDupsInt(&count_is[i],temp_idxs[i]);CHKERRQ(ierr);CHKERRQ(ierr); 460228143c3dSStefano Zampini ierr = ISCreateGeneral(comm_n,count_is[i],temp_idxs[i],PETSC_COPY_VALUES,&isarray[i]);CHKERRQ(ierr); 460328143c3dSStefano Zampini } 460428143c3dSStefano Zampini ierr = PetscFree(count_is);CHKERRQ(ierr); 460528143c3dSStefano Zampini ierr = PetscFree(temp_idxs[0]);CHKERRQ(ierr); 460628143c3dSStefano Zampini ierr = PetscFree(temp_idxs);CHKERRQ(ierr); 460728143c3dSStefano Zampini } 4608e7931f94SStefano Zampini /* free workspace */ 460928143c3dSStefano Zampini ierr = PetscFree(recv_buffer_idxs_is);CHKERRQ(ierr); 4610e7931f94SStefano Zampini ierr = MPI_Waitall(n_sends,send_req_idxs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 4611e7931f94SStefano Zampini ierr = PetscFree(send_buffer_idxs);CHKERRQ(ierr); 4612e7931f94SStefano Zampini ierr = MPI_Waitall(n_sends,send_req_vals,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 4613e7931f94SStefano Zampini if (isdense) { 4614e7931f94SStefano Zampini ierr = MatISGetLocalMat(mat,&local_mat);CHKERRQ(ierr); 4615e7931f94SStefano Zampini ierr = MatDenseRestoreArray(local_mat,&send_buffer_vals);CHKERRQ(ierr); 4616e7931f94SStefano Zampini } else { 4617e7931f94SStefano Zampini /* ierr = PetscFree(send_buffer_vals);CHKERRQ(ierr); */ 4618e7931f94SStefano Zampini } 461928143c3dSStefano Zampini if (nis) { 462028143c3dSStefano Zampini ierr = MPI_Waitall(n_sends,send_req_idxs_is,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 462128143c3dSStefano Zampini ierr = PetscFree(send_buffer_idxs_is);CHKERRQ(ierr); 462228143c3dSStefano Zampini } 4623e7931f94SStefano Zampini ierr = PetscFree(recv_req_idxs);CHKERRQ(ierr); 4624e7931f94SStefano Zampini ierr = PetscFree(recv_req_vals);CHKERRQ(ierr); 462528143c3dSStefano Zampini ierr = PetscFree(recv_req_idxs_is);CHKERRQ(ierr); 4626e7931f94SStefano Zampini ierr = PetscFree(send_req_idxs);CHKERRQ(ierr); 4627e7931f94SStefano Zampini ierr = PetscFree(send_req_vals);CHKERRQ(ierr); 462828143c3dSStefano Zampini ierr = PetscFree(send_req_idxs_is);CHKERRQ(ierr); 4629e7931f94SStefano Zampini ierr = PetscFree(ilengths_vals);CHKERRQ(ierr); 4630e7931f94SStefano Zampini ierr = PetscFree(ilengths_idxs);CHKERRQ(ierr); 4631e7931f94SStefano Zampini ierr = PetscFree(olengths_vals);CHKERRQ(ierr); 4632e7931f94SStefano Zampini ierr = PetscFree(olengths_idxs);CHKERRQ(ierr); 4633e7931f94SStefano Zampini ierr = PetscFree(onodes);CHKERRQ(ierr); 463428143c3dSStefano Zampini if (nis) { 463528143c3dSStefano Zampini ierr = PetscFree(ilengths_idxs_is);CHKERRQ(ierr); 463628143c3dSStefano Zampini ierr = PetscFree(olengths_idxs_is);CHKERRQ(ierr); 463728143c3dSStefano Zampini ierr = PetscFree(onodes_is);CHKERRQ(ierr); 463828143c3dSStefano Zampini } 463928143c3dSStefano Zampini ierr = PetscSubcommDestroy(&subcomm);CHKERRQ(ierr); 464028143c3dSStefano Zampini if (destroy_mat) { /* destroy mat is true only if restrict comm is true and process will not partecipate */ 464128143c3dSStefano Zampini ierr = MatDestroy(mat_n);CHKERRQ(ierr); 464228143c3dSStefano Zampini for (i=0;i<nis;i++) { 464328143c3dSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 464428143c3dSStefano Zampini } 464553a05cb3SStefano Zampini *mat_n = NULL; 464628143c3dSStefano Zampini } 4647e7931f94SStefano Zampini PetscFunctionReturn(0); 4648e7931f94SStefano Zampini } 4649a57a6d2fSStefano Zampini 465012edc857SStefano Zampini /* temporary hack into ksp private data structure */ 4651af0996ceSBarry Smith #include <petsc/private/kspimpl.h> 465212edc857SStefano Zampini 4653c8587f34SStefano Zampini #undef __FUNCT__ 4654c8587f34SStefano Zampini #define __FUNCT__ "PCBDDCSetUpCoarseSolver" 4655c8587f34SStefano Zampini PetscErrorCode PCBDDCSetUpCoarseSolver(PC pc,PetscScalar* coarse_submat_vals) 4656c8587f34SStefano Zampini { 4657c8587f34SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 4658c8587f34SStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 465920a2ab83SStefano Zampini Mat coarse_mat,coarse_mat_is,coarse_submat_dense; 46609881197aSStefano Zampini MatNullSpace CoarseNullSpace=NULL; 466120a2ab83SStefano Zampini ISLocalToGlobalMapping coarse_islg; 46626e683305SStefano Zampini IS coarse_is,*isarray; 46636e683305SStefano Zampini PetscInt i,im_active=-1,active_procs=-1; 46646e683305SStefano Zampini PetscInt nis,nisdofs,nisneu; 4665f9eb5b7dSStefano Zampini PC pc_temp; 4666c8587f34SStefano Zampini PCType coarse_pc_type; 4667c8587f34SStefano Zampini KSPType coarse_ksp_type; 4668f9eb5b7dSStefano Zampini PetscBool multilevel_requested,multilevel_allowed; 46694f3a063dSStefano Zampini PetscBool isredundant,isbddc,isnn,coarse_reuse; 46706e683305SStefano Zampini Mat t_coarse_mat_is; 46716e683305SStefano Zampini PetscInt void_procs,ncoarse_ml,ncoarse_ds,ncoarse; 46726e683305SStefano Zampini PetscMPIInt all_procs; 467374e2c79eSStefano Zampini PetscBool csin_ml,csin_ds,csin,csin_type_simple,redist; 467468457ee5SStefano Zampini PetscBool compute_vecs = PETSC_FALSE; 467522bc73bbSStefano Zampini PetscScalar *array; 46769881197aSStefano Zampini PetscErrorCode ierr; 4677fdc09c96SStefano Zampini 4678c8587f34SStefano Zampini PetscFunctionBegin; 4679c8587f34SStefano Zampini /* Assign global numbering to coarse dofs */ 468068457ee5SStefano 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 */ 4681fa7f1dd8SStefano Zampini PetscInt ocoarse_size; 46825a75c04eSSatish Balay compute_vecs = PETSC_TRUE; 4683fa7f1dd8SStefano Zampini ocoarse_size = pcbddc->coarse_size; 4684f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->global_primal_indices);CHKERRQ(ierr); 4685f4ddd8eeSStefano Zampini ierr = PCBDDCComputePrimalNumbering(pc,&pcbddc->coarse_size,&pcbddc->global_primal_indices);CHKERRQ(ierr); 4686f4ddd8eeSStefano Zampini /* see if we can avoid some work */ 4687fa7f1dd8SStefano Zampini if (pcbddc->coarse_ksp) { /* coarse ksp has already been created */ 468851bea450SStefano Zampini /* if the coarse size is different or we are using adaptive selection, better to not reuse the coarse matrix */ 468951bea450SStefano Zampini if (ocoarse_size != pcbddc->coarse_size || pcbddc->adaptive_selection) { 4690dc4bcba2SStefano Zampini PC pc; 4691dc4bcba2SStefano Zampini PetscBool isbddc; 4692dc4bcba2SStefano Zampini 4693dc4bcba2SStefano Zampini /* temporary workaround since PCBDDC does not have a reset method so far */ 4694dc4bcba2SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&pc);CHKERRQ(ierr); 4695dc4bcba2SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc,PCBDDC,&isbddc);CHKERRQ(ierr); 4696dc4bcba2SStefano Zampini if (isbddc) { 4697dc4bcba2SStefano Zampini ierr = PCDestroy(&pc);CHKERRQ(ierr); 4698dc4bcba2SStefano Zampini } 4699727cdba6SStefano Zampini ierr = KSPReset(pcbddc->coarse_ksp);CHKERRQ(ierr); 4700fa7f1dd8SStefano Zampini coarse_reuse = PETSC_FALSE; 4701fa7f1dd8SStefano Zampini } else { /* we can safely reuse already computed coarse matrix */ 4702fa7f1dd8SStefano Zampini coarse_reuse = PETSC_TRUE; 4703f4ddd8eeSStefano Zampini } 4704fa7f1dd8SStefano Zampini } else { /* there's no coarse ksp, so we need to create the coarse matrix too */ 4705fa7f1dd8SStefano Zampini coarse_reuse = PETSC_FALSE; 4706f4ddd8eeSStefano Zampini } 470770cf5478SStefano Zampini /* reset any subassembling information */ 470870cf5478SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling);CHKERRQ(ierr); 47096e683305SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 47106e683305SStefano Zampini } else { /* primal space is unchanged, so we can reuse coarse matrix */ 4711fa7f1dd8SStefano Zampini coarse_reuse = PETSC_TRUE; 4712f4ddd8eeSStefano Zampini } 4713c8587f34SStefano Zampini 47146e683305SStefano Zampini /* count "active" (i.e. with positive local size) and "void" processes */ 47152b510759SStefano Zampini im_active = !!(pcis->n); 47162b510759SStefano Zampini ierr = MPI_Allreduce(&im_active,&active_procs,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 47176e683305SStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)pc),&all_procs);CHKERRQ(ierr); 47186e683305SStefano Zampini void_procs = all_procs-active_procs; 47196e683305SStefano Zampini csin_type_simple = PETSC_TRUE; 472074e2c79eSStefano Zampini redist = PETSC_FALSE; 472122bc73bbSStefano Zampini if (pcbddc->current_level && void_procs) { 47226e683305SStefano Zampini csin_ml = PETSC_TRUE; 47236e683305SStefano Zampini ncoarse_ml = void_procs; 4724779c1cceSStefano Zampini /* it has no sense to redistribute on a set of processors larger than the number of active processes */ 4725779c1cceSStefano Zampini if (pcbddc->redistribute_coarse > 0 && pcbddc->redistribute_coarse < active_procs) { 47266e683305SStefano Zampini csin_ds = PETSC_TRUE; 472718a45a71SStefano Zampini ncoarse_ds = pcbddc->redistribute_coarse; 472818a45a71SStefano Zampini redist = PETSC_TRUE; 472918a45a71SStefano Zampini } else { 47306e683305SStefano Zampini csin_ds = PETSC_TRUE; 4731779c1cceSStefano Zampini ncoarse_ds = active_procs; 4732779c1cceSStefano Zampini redist = PETSC_TRUE; 473318a45a71SStefano Zampini } 47346e683305SStefano Zampini } else { 47356e683305SStefano Zampini csin_ml = PETSC_FALSE; 47366e683305SStefano Zampini ncoarse_ml = all_procs; 47376e683305SStefano Zampini if (void_procs) { 47386e683305SStefano Zampini csin_ds = PETSC_TRUE; 47396e683305SStefano Zampini ncoarse_ds = void_procs; 47406e683305SStefano Zampini csin_type_simple = PETSC_FALSE; 47416e683305SStefano Zampini } else { 4742779c1cceSStefano Zampini if (pcbddc->redistribute_coarse > 0 && pcbddc->redistribute_coarse < all_procs) { 474374e2c79eSStefano Zampini csin_ds = PETSC_TRUE; 474474e2c79eSStefano Zampini ncoarse_ds = pcbddc->redistribute_coarse; 474574e2c79eSStefano Zampini redist = PETSC_TRUE; 474674e2c79eSStefano Zampini } else { 47476e683305SStefano Zampini csin_ds = PETSC_FALSE; 47486e683305SStefano Zampini ncoarse_ds = all_procs; 47496e683305SStefano Zampini } 47506e683305SStefano Zampini } 475174e2c79eSStefano Zampini } 47526e683305SStefano Zampini 47536e683305SStefano Zampini /* 47546e683305SStefano Zampini test if we can go multilevel: three conditions must be satisfied: 47556e683305SStefano Zampini - we have not exceeded the number of levels requested 47566e683305SStefano Zampini - we can actually subassemble the active processes 47576e683305SStefano Zampini - we can find a suitable number of MPI processes where we can place the subassembled problem 47586e683305SStefano Zampini */ 47596e683305SStefano Zampini multilevel_allowed = PETSC_FALSE; 47606e683305SStefano Zampini multilevel_requested = PETSC_FALSE; 47616e683305SStefano Zampini if (pcbddc->current_level < pcbddc->max_levels) { 47626e683305SStefano Zampini multilevel_requested = PETSC_TRUE; 47636e683305SStefano Zampini if (active_procs/pcbddc->coarsening_ratio < 2 || ncoarse_ml/pcbddc->coarsening_ratio < 2) { 4764f9eb5b7dSStefano Zampini multilevel_allowed = PETSC_FALSE; 47652b510759SStefano Zampini } else { 4766f9eb5b7dSStefano Zampini multilevel_allowed = PETSC_TRUE; 4767c8587f34SStefano Zampini } 4768c8587f34SStefano Zampini } 47696e683305SStefano Zampini /* determine number of process partecipating to coarse solver */ 47706e683305SStefano Zampini if (multilevel_allowed) { 47716e683305SStefano Zampini ncoarse = ncoarse_ml; 47726e683305SStefano Zampini csin = csin_ml; 477358da7f69SStefano Zampini redist = PETSC_FALSE; 47746e683305SStefano Zampini } else { 47756e683305SStefano Zampini ncoarse = ncoarse_ds; 47766e683305SStefano Zampini csin = csin_ds; 47776e683305SStefano Zampini } 4778e7931f94SStefano Zampini 4779abbbba34SStefano Zampini /* creates temporary l2gmap and IS for coarse indexes */ 4780abbbba34SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),pcbddc->local_primal_size,pcbddc->global_primal_indices,PETSC_COPY_VALUES,&coarse_is);CHKERRQ(ierr); 4781abbbba34SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(coarse_is,&coarse_islg);CHKERRQ(ierr); 4782abbbba34SStefano Zampini 4783abbbba34SStefano Zampini /* creates temporary MATIS object for coarse matrix */ 478422bc73bbSStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,pcbddc->local_primal_size,pcbddc->local_primal_size,NULL,&coarse_submat_dense);CHKERRQ(ierr); 478522bc73bbSStefano Zampini ierr = MatDenseGetArray(coarse_submat_dense,&array);CHKERRQ(ierr); 478622bc73bbSStefano Zampini ierr = PetscMemcpy(array,coarse_submat_vals,sizeof(*coarse_submat_vals)*pcbddc->local_primal_size*pcbddc->local_primal_size);CHKERRQ(ierr); 478722bc73bbSStefano Zampini ierr = MatDenseRestoreArray(coarse_submat_dense,&array);CHKERRQ(ierr); 4788b1ecc7b1SStefano Zampini #if 0 4789b9b85e73SStefano Zampini { 4790b9b85e73SStefano Zampini PetscViewer viewer; 4791b9b85e73SStefano Zampini char filename[256]; 4792b1ecc7b1SStefano Zampini sprintf(filename,"local_coarse_mat%d_level%d.m",PetscGlobalRank,pcbddc->current_level); 4793b9b85e73SStefano Zampini ierr = PetscViewerASCIIOpen(PETSC_COMM_SELF,filename,&viewer);CHKERRQ(ierr); 4794b9b85e73SStefano Zampini ierr = PetscViewerSetFormat(viewer,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 4795b9b85e73SStefano Zampini ierr = MatView(coarse_submat_dense,viewer);CHKERRQ(ierr); 4796b9b85e73SStefano Zampini ierr = PetscViewerDestroy(&viewer);CHKERRQ(ierr); 4797b9b85e73SStefano Zampini } 4798b9b85e73SStefano Zampini #endif 4799e176bc59SStefano 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); 48006e683305SStefano Zampini ierr = MatISSetLocalMat(t_coarse_mat_is,coarse_submat_dense);CHKERRQ(ierr); 48016e683305SStefano Zampini ierr = MatAssemblyBegin(t_coarse_mat_is,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 48026e683305SStefano Zampini ierr = MatAssemblyEnd(t_coarse_mat_is,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4803abbbba34SStefano Zampini ierr = MatDestroy(&coarse_submat_dense);CHKERRQ(ierr); 4804abbbba34SStefano Zampini 48056e683305SStefano Zampini /* compute dofs splitting and neumann boundaries for coarse dofs */ 48066e683305SStefano Zampini if (multilevel_allowed && (pcbddc->n_ISForDofsLocal || pcbddc->NeumannBoundariesLocal) ) { /* protects from unneded computations */ 48076e683305SStefano Zampini PetscInt *tidxs,*tidxs2,nout,tsize,i; 48086e683305SStefano Zampini const PetscInt *idxs; 48096e683305SStefano Zampini ISLocalToGlobalMapping tmap; 48106e683305SStefano Zampini 48116e683305SStefano Zampini /* create map between primal indices (in local representative ordering) and local primal numbering */ 48120be93d54SStefano Zampini ierr = ISLocalToGlobalMappingCreate(PETSC_COMM_SELF,1,pcbddc->local_primal_size,pcbddc->primal_indices_local_idxs,PETSC_COPY_VALUES,&tmap);CHKERRQ(ierr); 48136e683305SStefano Zampini /* allocate space for temporary storage */ 4814854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->local_primal_size,&tidxs);CHKERRQ(ierr); 4815854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->local_primal_size,&tidxs2);CHKERRQ(ierr); 48166e683305SStefano Zampini /* allocate for IS array */ 48176e683305SStefano Zampini nisdofs = pcbddc->n_ISForDofsLocal; 48186e683305SStefano Zampini nisneu = !!pcbddc->NeumannBoundariesLocal; 48196e683305SStefano Zampini nis = nisdofs + nisneu; 4820854ce69bSBarry Smith ierr = PetscMalloc1(nis,&isarray);CHKERRQ(ierr); 48216e683305SStefano Zampini /* dofs splitting */ 48226e683305SStefano Zampini for (i=0;i<nisdofs;i++) { 48236e683305SStefano Zampini /* ierr = ISView(pcbddc->ISForDofsLocal[i],0);CHKERRQ(ierr); */ 48246e683305SStefano Zampini ierr = ISGetLocalSize(pcbddc->ISForDofsLocal[i],&tsize);CHKERRQ(ierr); 48256e683305SStefano Zampini ierr = ISGetIndices(pcbddc->ISForDofsLocal[i],&idxs);CHKERRQ(ierr); 48266e683305SStefano Zampini ierr = ISGlobalToLocalMappingApply(tmap,IS_GTOLM_DROP,tsize,idxs,&nout,tidxs);CHKERRQ(ierr); 48276e683305SStefano Zampini ierr = ISRestoreIndices(pcbddc->ISForDofsLocal[i],&idxs);CHKERRQ(ierr); 48286e683305SStefano Zampini ierr = ISLocalToGlobalMappingApply(coarse_islg,nout,tidxs,tidxs2);CHKERRQ(ierr); 48296e683305SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pcbddc->ISForDofsLocal[i]),nout,tidxs2,PETSC_COPY_VALUES,&isarray[i]);CHKERRQ(ierr); 48306e683305SStefano Zampini /* ierr = ISView(isarray[i],0);CHKERRQ(ierr); */ 48316e683305SStefano Zampini } 48326e683305SStefano Zampini /* neumann boundaries */ 48336e683305SStefano Zampini if (pcbddc->NeumannBoundariesLocal) { 48346e683305SStefano Zampini /* ierr = ISView(pcbddc->NeumannBoundariesLocal,0);CHKERRQ(ierr); */ 48356e683305SStefano Zampini ierr = ISGetLocalSize(pcbddc->NeumannBoundariesLocal,&tsize);CHKERRQ(ierr); 48366e683305SStefano Zampini ierr = ISGetIndices(pcbddc->NeumannBoundariesLocal,&idxs);CHKERRQ(ierr); 48376e683305SStefano Zampini ierr = ISGlobalToLocalMappingApply(tmap,IS_GTOLM_DROP,tsize,idxs,&nout,tidxs);CHKERRQ(ierr); 48386e683305SStefano Zampini ierr = ISRestoreIndices(pcbddc->NeumannBoundariesLocal,&idxs);CHKERRQ(ierr); 48396e683305SStefano Zampini ierr = ISLocalToGlobalMappingApply(coarse_islg,nout,tidxs,tidxs2);CHKERRQ(ierr); 48406e683305SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pcbddc->NeumannBoundariesLocal),nout,tidxs2,PETSC_COPY_VALUES,&isarray[nisdofs]);CHKERRQ(ierr); 48416e683305SStefano Zampini /* ierr = ISView(isarray[nisdofs],0);CHKERRQ(ierr); */ 48426e683305SStefano Zampini } 48436e683305SStefano Zampini /* free memory */ 48446e683305SStefano Zampini ierr = PetscFree(tidxs);CHKERRQ(ierr); 48456e683305SStefano Zampini ierr = PetscFree(tidxs2);CHKERRQ(ierr); 48466e683305SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&tmap);CHKERRQ(ierr); 48476e683305SStefano Zampini } else { 48486e683305SStefano Zampini nis = 0; 48496e683305SStefano Zampini nisdofs = 0; 48506e683305SStefano Zampini nisneu = 0; 48516e683305SStefano Zampini isarray = NULL; 48526e683305SStefano Zampini } 48536e683305SStefano Zampini /* destroy no longer needed map */ 48546e683305SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&coarse_islg);CHKERRQ(ierr); 48556e683305SStefano Zampini 48566e683305SStefano Zampini /* restrict on coarse candidates (if needed) */ 48576e683305SStefano Zampini coarse_mat_is = NULL; 48586e683305SStefano Zampini if (csin) { 48596e683305SStefano Zampini if (!pcbddc->coarse_subassembling_init ) { /* creates subassembling init pattern if not present */ 486074e2c79eSStefano Zampini if (redist) { 486174e2c79eSStefano Zampini PetscMPIInt rank; 4862779c1cceSStefano Zampini PetscInt spc,n_spc_p1,dest[1],destsize; 486374e2c79eSStefano Zampini 486474e2c79eSStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)pc),&rank);CHKERRQ(ierr); 486558da7f69SStefano Zampini spc = active_procs/ncoarse; 486658da7f69SStefano Zampini n_spc_p1 = active_procs%ncoarse; 4867779c1cceSStefano Zampini if (im_active) { 4868779c1cceSStefano Zampini destsize = 1; 486974e2c79eSStefano Zampini if (rank > n_spc_p1*(spc+1)-1) { 487074e2c79eSStefano Zampini dest[0] = n_spc_p1+(rank-(n_spc_p1*(spc+1)))/spc; 487174e2c79eSStefano Zampini } else { 487274e2c79eSStefano Zampini dest[0] = rank/(spc+1); 487374e2c79eSStefano Zampini } 487474e2c79eSStefano Zampini } else { 4875779c1cceSStefano Zampini destsize = 0; 48766e683305SStefano Zampini } 4877779c1cceSStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),destsize,dest,PETSC_COPY_VALUES,&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 4878779c1cceSStefano Zampini } else if (csin_type_simple) { 48796e683305SStefano Zampini PetscMPIInt rank; 48806e683305SStefano Zampini PetscInt issize,isidx; 4881779c1cceSStefano Zampini 48826e683305SStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)pc),&rank);CHKERRQ(ierr); 48836e683305SStefano Zampini if (im_active) { 48846e683305SStefano Zampini issize = 1; 48856e683305SStefano Zampini isidx = (PetscInt)rank; 48866e683305SStefano Zampini } else { 48876e683305SStefano Zampini issize = 0; 48886e683305SStefano Zampini isidx = -1; 48896e683305SStefano Zampini } 48906e683305SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),issize,&isidx,PETSC_COPY_VALUES,&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 4891779c1cceSStefano Zampini } else { /* get a suitable subassembling pattern from MATIS code */ 4892b0c7d250SStefano Zampini ierr = MatISGetSubassemblingPattern(t_coarse_mat_is,ncoarse,pcbddc->coarse_adj_red,&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 48936e683305SStefano Zampini } 4894779c1cceSStefano Zampini 4895779c1cceSStefano Zampini /* we need to shift on coarse candidates either if we are not redistributing or we are redistributing and we have enough void processes */ 4896779c1cceSStefano Zampini if (!redist || ncoarse <= void_procs) { 4897779c1cceSStefano Zampini PetscInt ncoarse_cand,tissize,*nisindices; 4898779c1cceSStefano Zampini PetscInt *coarse_candidates; 4899779c1cceSStefano Zampini const PetscInt* tisindices; 4900779c1cceSStefano Zampini 4901779c1cceSStefano Zampini /* get coarse candidates' ranks in pc communicator */ 4902779c1cceSStefano Zampini ierr = PetscMalloc1(all_procs,&coarse_candidates);CHKERRQ(ierr); 4903779c1cceSStefano Zampini ierr = MPI_Allgather(&im_active,1,MPIU_INT,coarse_candidates,1,MPIU_INT,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 4904779c1cceSStefano Zampini for (i=0,ncoarse_cand=0;i<all_procs;i++) { 4905779c1cceSStefano Zampini if (!coarse_candidates[i]) { 4906779c1cceSStefano Zampini coarse_candidates[ncoarse_cand++]=i; 4907779c1cceSStefano Zampini } 4908779c1cceSStefano Zampini } 4909779c1cceSStefano Zampini if (ncoarse_cand < ncoarse) SETERRQ2(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"This should not happen! %d < %d",ncoarse_cand,ncoarse); 4910779c1cceSStefano Zampini 4911779c1cceSStefano Zampini 49126e683305SStefano Zampini if (pcbddc->dbg_flag) { 49136e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 49146e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Subassembling pattern init (before shift)\n");CHKERRQ(ierr); 49156e683305SStefano Zampini ierr = ISView(pcbddc->coarse_subassembling_init,pcbddc->dbg_viewer);CHKERRQ(ierr); 49166e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Coarse candidates\n");CHKERRQ(ierr); 4917779c1cceSStefano Zampini for (i=0;i<ncoarse_cand;i++) { 49186e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"%d ",coarse_candidates[i]);CHKERRQ(ierr); 49196e683305SStefano Zampini } 49206e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"\n");CHKERRQ(ierr); 49216e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 49226e683305SStefano Zampini } 49236e683305SStefano Zampini /* shift the pattern on coarse candidates */ 49246e683305SStefano Zampini ierr = ISGetLocalSize(pcbddc->coarse_subassembling_init,&tissize);CHKERRQ(ierr); 49256e683305SStefano Zampini ierr = ISGetIndices(pcbddc->coarse_subassembling_init,&tisindices);CHKERRQ(ierr); 4926854ce69bSBarry Smith ierr = PetscMalloc1(tissize,&nisindices);CHKERRQ(ierr); 49276e683305SStefano Zampini for (i=0;i<tissize;i++) nisindices[i] = coarse_candidates[tisindices[i]]; 49286e683305SStefano Zampini ierr = ISRestoreIndices(pcbddc->coarse_subassembling_init,&tisindices);CHKERRQ(ierr); 49296e683305SStefano Zampini ierr = ISGeneralSetIndices(pcbddc->coarse_subassembling_init,tissize,nisindices,PETSC_OWN_POINTER);CHKERRQ(ierr); 49306e683305SStefano Zampini ierr = PetscFree(coarse_candidates);CHKERRQ(ierr); 49316e683305SStefano Zampini } 49326e683305SStefano Zampini if (pcbddc->dbg_flag) { 49336e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 49346e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Subassembling pattern init\n");CHKERRQ(ierr); 49356e683305SStefano Zampini ierr = ISView(pcbddc->coarse_subassembling_init,pcbddc->dbg_viewer);CHKERRQ(ierr); 49366e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 49376e683305SStefano Zampini } 4938779c1cceSStefano Zampini } 49396e683305SStefano Zampini /* get temporary coarse mat in IS format restricted on coarse procs (plus additional index sets of isarray) */ 494053a05cb3SStefano Zampini if (multilevel_allowed) { /* we need to keep tracking of void processes for future placements */ 494153a05cb3SStefano 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); 494253a05cb3SStefano Zampini } else { /* this is the last level, so use just receiving processes in subcomm */ 494353a05cb3SStefano 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); 494453a05cb3SStefano Zampini } 49456e683305SStefano Zampini } else { 49466e683305SStefano Zampini if (pcbddc->dbg_flag) { 49476e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 49486e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Subassembling pattern init not needed\n");CHKERRQ(ierr); 49496e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 49506e683305SStefano Zampini } 49516e683305SStefano Zampini ierr = PetscObjectReference((PetscObject)t_coarse_mat_is);CHKERRQ(ierr); 49526e683305SStefano Zampini coarse_mat_is = t_coarse_mat_is; 49536e683305SStefano Zampini } 49546e683305SStefano Zampini 49556e683305SStefano Zampini /* create local to global scatters for coarse problem */ 495668457ee5SStefano Zampini if (compute_vecs) { 49576e683305SStefano Zampini PetscInt lrows; 49586e683305SStefano Zampini ierr = VecDestroy(&pcbddc->coarse_vec);CHKERRQ(ierr); 49596e683305SStefano Zampini if (coarse_mat_is) { 49606e683305SStefano Zampini ierr = MatGetLocalSize(coarse_mat_is,&lrows,NULL);CHKERRQ(ierr); 49616e683305SStefano Zampini } else { 49626e683305SStefano Zampini lrows = 0; 49636e683305SStefano Zampini } 49646e683305SStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pc),&pcbddc->coarse_vec);CHKERRQ(ierr); 49656e683305SStefano Zampini ierr = VecSetSizes(pcbddc->coarse_vec,lrows,PETSC_DECIDE);CHKERRQ(ierr); 49666e683305SStefano Zampini ierr = VecSetType(pcbddc->coarse_vec,VECSTANDARD);CHKERRQ(ierr); 49676e683305SStefano Zampini ierr = VecScatterDestroy(&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 49686e683305SStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_P,NULL,pcbddc->coarse_vec,coarse_is,&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 49696e683305SStefano Zampini } 49706e683305SStefano Zampini ierr = ISDestroy(&coarse_is);CHKERRQ(ierr); 49716e683305SStefano Zampini ierr = MatDestroy(&t_coarse_mat_is);CHKERRQ(ierr); 4972c8587f34SStefano Zampini 4973f9eb5b7dSStefano Zampini /* set defaults for coarse KSP and PC */ 4974f9eb5b7dSStefano Zampini if (multilevel_allowed) { 4975f9eb5b7dSStefano Zampini coarse_ksp_type = KSPRICHARDSON; 4976f9eb5b7dSStefano Zampini coarse_pc_type = PCBDDC; 4977f9eb5b7dSStefano Zampini } else { 4978f9eb5b7dSStefano Zampini coarse_ksp_type = KSPPREONLY; 4979f9eb5b7dSStefano Zampini coarse_pc_type = PCREDUNDANT; 4980c8587f34SStefano Zampini } 4981c8587f34SStefano Zampini 49826e683305SStefano Zampini /* print some info if requested */ 49836e683305SStefano Zampini if (pcbddc->dbg_flag) { 49846e683305SStefano Zampini if (!multilevel_allowed) { 49856e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 49866e683305SStefano Zampini if (multilevel_requested) { 49876e683305SStefano 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); 49886e683305SStefano Zampini } else if (pcbddc->max_levels) { 49896e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Maximum number of requested levels reached (%d)\n",pcbddc->max_levels);CHKERRQ(ierr); 49906e683305SStefano Zampini } 49916e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 49926e683305SStefano Zampini } 49936e683305SStefano Zampini } 49946e683305SStefano Zampini 4995f9eb5b7dSStefano Zampini /* create the coarse KSP object only once with defaults */ 49966e683305SStefano Zampini if (coarse_mat_is) { 49976e683305SStefano Zampini MatReuse coarse_mat_reuse; 49986a1308c2SStefano Zampini PetscViewer dbg_viewer = NULL; 49996e683305SStefano Zampini if (pcbddc->dbg_flag) { 50006e683305SStefano Zampini dbg_viewer = PETSC_VIEWER_STDOUT_(PetscObjectComm((PetscObject)coarse_mat_is)); 50016e683305SStefano Zampini ierr = PetscViewerASCIIAddTab(dbg_viewer,2*pcbddc->current_level);CHKERRQ(ierr); 50026e683305SStefano Zampini } 5003f9eb5b7dSStefano Zampini if (!pcbddc->coarse_ksp) { 5004312be037SStefano Zampini char prefix[256],str_level[16]; 5005e604994aSStefano Zampini size_t len; 50066e683305SStefano Zampini ierr = KSPCreate(PetscObjectComm((PetscObject)coarse_mat_is),&pcbddc->coarse_ksp);CHKERRQ(ierr); 5007422a814eSBarry Smith ierr = KSPSetErrorIfNotConverged(pcbddc->coarse_ksp,pc->erroriffailure);CHKERRQ(ierr); 5008c8587f34SStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->coarse_ksp,(PetscObject)pc,1);CHKERRQ(ierr); 5009f9eb5b7dSStefano Zampini ierr = KSPSetTolerances(pcbddc->coarse_ksp,PETSC_DEFAULT,PETSC_DEFAULT,PETSC_DEFAULT,1);CHKERRQ(ierr); 50105f76c7aeSStefano Zampini ierr = KSPSetOperators(pcbddc->coarse_ksp,coarse_mat_is,coarse_mat_is);CHKERRQ(ierr); 5011c8587f34SStefano Zampini ierr = KSPSetType(pcbddc->coarse_ksp,coarse_ksp_type);CHKERRQ(ierr); 50126e683305SStefano Zampini ierr = KSPSetNormType(pcbddc->coarse_ksp,KSP_NORM_NONE);CHKERRQ(ierr); 5013c8587f34SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&pc_temp);CHKERRQ(ierr); 5014c8587f34SStefano Zampini ierr = PCSetType(pc_temp,coarse_pc_type);CHKERRQ(ierr); 5015e604994aSStefano Zampini /* prefix */ 5016e604994aSStefano Zampini ierr = PetscStrcpy(prefix,"");CHKERRQ(ierr); 5017e604994aSStefano Zampini ierr = PetscStrcpy(str_level,"");CHKERRQ(ierr); 5018e604994aSStefano Zampini if (!pcbddc->current_level) { 5019e604994aSStefano Zampini ierr = PetscStrcpy(prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 5020e604994aSStefano Zampini ierr = PetscStrcat(prefix,"pc_bddc_coarse_");CHKERRQ(ierr); 5021c8587f34SStefano Zampini } else { 5022e604994aSStefano Zampini ierr = PetscStrlen(((PetscObject)pc)->prefix,&len);CHKERRQ(ierr); 5023312be037SStefano Zampini if (pcbddc->current_level>1) len -= 3; /* remove "lX_" with X level number */ 5024312be037SStefano Zampini if (pcbddc->current_level>10) len -= 1; /* remove another char from level number */ 502534d6797cSStefano Zampini ierr = PetscStrncpy(prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 5026312be037SStefano Zampini sprintf(str_level,"l%d_",(int)(pcbddc->current_level)); 5027e604994aSStefano Zampini ierr = PetscStrcat(prefix,str_level);CHKERRQ(ierr); 5028e604994aSStefano Zampini } 5029e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->coarse_ksp,prefix);CHKERRQ(ierr); 50303e3c6dadSStefano Zampini /* propagate BDDC info to the next level (these are dummy calls if pc_temp is not of type PCBDDC) */ 50313e3c6dadSStefano Zampini ierr = PCBDDCSetLevel(pc_temp,pcbddc->current_level+1);CHKERRQ(ierr); 50323e3c6dadSStefano Zampini ierr = PCBDDCSetCoarseningRatio(pc_temp,pcbddc->coarsening_ratio);CHKERRQ(ierr); 50333e3c6dadSStefano Zampini ierr = PCBDDCSetLevels(pc_temp,pcbddc->max_levels);CHKERRQ(ierr); 5034f9eb5b7dSStefano Zampini /* allow user customization */ 5035f9eb5b7dSStefano Zampini ierr = KSPSetFromOptions(pcbddc->coarse_ksp);CHKERRQ(ierr); 50363e3c6dadSStefano Zampini } 50373e3c6dadSStefano Zampini /* propagate BDDC info to the next level (these are dummy calls if pc_temp is not of type PCBDDC) */ 503851bea450SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&pc_temp);CHKERRQ(ierr); 50393e3c6dadSStefano Zampini if (nisdofs) { 50403e3c6dadSStefano Zampini ierr = PCBDDCSetDofsSplitting(pc_temp,nisdofs,isarray);CHKERRQ(ierr); 50413e3c6dadSStefano Zampini for (i=0;i<nisdofs;i++) { 50423e3c6dadSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 50433e3c6dadSStefano Zampini } 50443e3c6dadSStefano Zampini } 50453e3c6dadSStefano Zampini if (nisneu) { 50463e3c6dadSStefano Zampini ierr = PCBDDCSetNeumannBoundaries(pc_temp,isarray[nisdofs]);CHKERRQ(ierr); 50473e3c6dadSStefano Zampini ierr = ISDestroy(&isarray[nisdofs]);CHKERRQ(ierr); 5048312be037SStefano Zampini } 5049f9eb5b7dSStefano Zampini 5050f9eb5b7dSStefano Zampini /* get some info after set from options */ 5051f9eb5b7dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCNN,&isnn);CHKERRQ(ierr); 5052f9eb5b7dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCBDDC,&isbddc);CHKERRQ(ierr); 50534f3a063dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCREDUNDANT,&isredundant);CHKERRQ(ierr); 50546e683305SStefano Zampini if (isbddc && !multilevel_allowed) { /* multilevel can only be requested via pc_bddc_set_levels */ 5055f9eb5b7dSStefano Zampini ierr = PCSetType(pc_temp,coarse_pc_type);CHKERRQ(ierr); 5056f9eb5b7dSStefano Zampini isbddc = PETSC_FALSE; 5057f9eb5b7dSStefano Zampini } 505839f0f02cSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 50594f3a063dSStefano Zampini if (isredundant) { 50604f3a063dSStefano Zampini KSP inner_ksp; 50614f3a063dSStefano Zampini PC inner_pc; 50624f3a063dSStefano Zampini ierr = PCRedundantGetKSP(pc_temp,&inner_ksp);CHKERRQ(ierr); 50634f3a063dSStefano Zampini ierr = KSPGetPC(inner_ksp,&inner_pc);CHKERRQ(ierr); 50644f3a063dSStefano Zampini ierr = PCFactorSetReuseFill(inner_pc,PETSC_TRUE);CHKERRQ(ierr); 50654f3a063dSStefano Zampini } 5066f9eb5b7dSStefano Zampini 5067f9eb5b7dSStefano Zampini /* assemble coarse matrix */ 5068fa7f1dd8SStefano Zampini if (coarse_reuse) { 506981d9aea3SBarry Smith ierr = KSPGetOperators(pcbddc->coarse_ksp,&coarse_mat,NULL);CHKERRQ(ierr); 5070fa7f1dd8SStefano Zampini ierr = PetscObjectReference((PetscObject)coarse_mat);CHKERRQ(ierr); 50716e683305SStefano Zampini coarse_mat_reuse = MAT_REUSE_MATRIX; 5072fa7f1dd8SStefano Zampini } else { 50736e683305SStefano Zampini coarse_mat_reuse = MAT_INITIAL_MATRIX; 5074fa7f1dd8SStefano Zampini } 5075c8587f34SStefano Zampini if (isbddc || isnn) { 507622bc73bbSStefano Zampini if (pcbddc->coarsening_ratio > 1) { 507770cf5478SStefano Zampini if (!pcbddc->coarse_subassembling) { /* subassembling info is not present */ 5078b0c7d250SStefano Zampini ierr = MatISGetSubassemblingPattern(coarse_mat_is,active_procs/pcbddc->coarsening_ratio,pcbddc->coarse_adj_red,&pcbddc->coarse_subassembling);CHKERRQ(ierr); 507922b6e8a2SStefano Zampini if (pcbddc->dbg_flag) { 50806e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 50816e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Subassembling pattern\n");CHKERRQ(ierr); 50826e683305SStefano Zampini ierr = ISView(pcbddc->coarse_subassembling,dbg_viewer);CHKERRQ(ierr); 50836e683305SStefano Zampini ierr = PetscViewerFlush(dbg_viewer);CHKERRQ(ierr); 508422b6e8a2SStefano Zampini } 508570cf5478SStefano Zampini } 508653a05cb3SStefano Zampini ierr = MatISSubassemble(coarse_mat_is,pcbddc->coarse_subassembling,0,PETSC_FALSE,PETSC_FALSE,coarse_mat_reuse,&coarse_mat,0,NULL);CHKERRQ(ierr); 508770cf5478SStefano Zampini } else { 508822bc73bbSStefano Zampini ierr = PetscObjectReference((PetscObject)coarse_mat_is);CHKERRQ(ierr); 508922bc73bbSStefano Zampini coarse_mat = coarse_mat_is; 509022bc73bbSStefano Zampini } 509122bc73bbSStefano Zampini } else { 50922e1e5fa4SStefano Zampini ierr = MatISGetMPIXAIJ(coarse_mat_is,coarse_mat_reuse,&coarse_mat);CHKERRQ(ierr); 5093c8587f34SStefano Zampini } 5094c8587f34SStefano Zampini ierr = MatDestroy(&coarse_mat_is);CHKERRQ(ierr); 5095c8587f34SStefano Zampini 50963301b35fSStefano Zampini /* propagate symmetry info of coarse matrix */ 50975a16e3a0SStefano Zampini ierr = MatSetOption(coarse_mat,MAT_STRUCTURALLY_SYMMETRIC,PETSC_TRUE);CHKERRQ(ierr); 50983301b35fSStefano Zampini if (pc->pmat->symmetric_set) { 50993301b35fSStefano Zampini ierr = MatSetOption(coarse_mat,MAT_SYMMETRIC,pc->pmat->symmetric);CHKERRQ(ierr); 51003301b35fSStefano Zampini } 51013301b35fSStefano Zampini if (pc->pmat->hermitian_set) { 51023301b35fSStefano Zampini ierr = MatSetOption(coarse_mat,MAT_HERMITIAN,pc->pmat->hermitian);CHKERRQ(ierr); 51033301b35fSStefano Zampini } 51043301b35fSStefano Zampini if (pc->pmat->spd_set) { 51053301b35fSStefano Zampini ierr = MatSetOption(coarse_mat,MAT_SPD,pc->pmat->spd);CHKERRQ(ierr); 51063301b35fSStefano Zampini } 51076e683305SStefano Zampini /* set operators */ 51085f76c7aeSStefano Zampini ierr = KSPSetOperators(pcbddc->coarse_ksp,coarse_mat,coarse_mat);CHKERRQ(ierr); 51096e683305SStefano Zampini if (pcbddc->dbg_flag) { 51106e683305SStefano Zampini ierr = PetscViewerASCIISubtractTab(dbg_viewer,2*pcbddc->current_level);CHKERRQ(ierr); 51116e683305SStefano Zampini } 51126e683305SStefano Zampini } else { /* processes non partecipating to coarse solver (if any) */ 51136e683305SStefano Zampini coarse_mat = 0; 51146e683305SStefano Zampini } 51156e683305SStefano Zampini ierr = PetscFree(isarray);CHKERRQ(ierr); 5116b1ecc7b1SStefano Zampini #if 0 5117b9b85e73SStefano Zampini { 5118b9b85e73SStefano Zampini PetscViewer viewer; 5119b9b85e73SStefano Zampini char filename[256]; 5120b1ecc7b1SStefano Zampini sprintf(filename,"coarse_mat_level%d.m",pcbddc->current_level); 5121b1ecc7b1SStefano Zampini ierr = PetscViewerASCIIOpen(PetscObjectComm((PetscObject)coarse_mat),filename,&viewer);CHKERRQ(ierr); 5122b9b85e73SStefano Zampini ierr = PetscViewerSetFormat(viewer,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 5123b9b85e73SStefano Zampini ierr = MatView(coarse_mat,viewer);CHKERRQ(ierr); 5124b9b85e73SStefano Zampini ierr = PetscViewerDestroy(&viewer);CHKERRQ(ierr); 5125b9b85e73SStefano Zampini } 5126b9b85e73SStefano Zampini #endif 5127c8587f34SStefano Zampini 5128c8587f34SStefano Zampini /* Compute coarse null space (special handling by BDDC only) */ 5129298c0119SStefano Zampini #if 0 5130c8587f34SStefano Zampini if (pcbddc->NullSpace) { 5131c8587f34SStefano Zampini ierr = PCBDDCNullSpaceAssembleCoarse(pc,coarse_mat,&CoarseNullSpace);CHKERRQ(ierr); 513298a51de6SStefano Zampini } 5133298c0119SStefano Zampini #endif 5134b0f5fe93SStefano Zampini /* hack */ 513598a51de6SStefano Zampini if (pcbddc->coarse_ksp) { 513698a51de6SStefano Zampini Vec crhs,csol; 513704708bb6SStefano Zampini 5138f347579bSStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&csol);CHKERRQ(ierr); 5139f347579bSStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&crhs);CHKERRQ(ierr); 5140f347579bSStefano Zampini if (!csol) { 51412a7a6963SBarry Smith ierr = MatCreateVecs(coarse_mat,&((pcbddc->coarse_ksp)->vec_sol),NULL);CHKERRQ(ierr); 5142f9eb5b7dSStefano Zampini } 5143f347579bSStefano Zampini if (!crhs) { 51442a7a6963SBarry Smith ierr = MatCreateVecs(coarse_mat,NULL,&((pcbddc->coarse_ksp)->vec_rhs));CHKERRQ(ierr); 5145f347579bSStefano Zampini } 5146b0f5fe93SStefano Zampini } 5147b0f5fe93SStefano Zampini 5148b0f5fe93SStefano Zampini /* compute null space for coarse solver if the benign trick has been requested */ 5149b0f5fe93SStefano Zampini if (pcbddc->benign_null) { 5150b0f5fe93SStefano Zampini 5151b0f5fe93SStefano Zampini ierr = VecSet(pcbddc->vec1_P,0.);CHKERRQ(ierr); 5152*4f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 5153*4f1b2e48SStefano Zampini ierr = VecSetValue(pcbddc->vec1_P,pcbddc->local_primal_size-pcbddc->benign_n+i,1.0,INSERT_VALUES);CHKERRQ(ierr); 5154*4f1b2e48SStefano Zampini } 5155b0f5fe93SStefano Zampini ierr = VecAssemblyBegin(pcbddc->vec1_P);CHKERRQ(ierr); 5156b0f5fe93SStefano Zampini ierr = VecAssemblyEnd(pcbddc->vec1_P);CHKERRQ(ierr); 5157b0f5fe93SStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 5158b0f5fe93SStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 5159b0f5fe93SStefano Zampini if (coarse_mat) { 5160b0f5fe93SStefano Zampini Vec nullv; 5161b0f5fe93SStefano Zampini PetscScalar *array,*array2; 5162b0f5fe93SStefano Zampini PetscInt nl; 5163b0f5fe93SStefano Zampini 5164b0f5fe93SStefano Zampini ierr = MatCreateVecs(coarse_mat,&nullv,NULL);CHKERRQ(ierr); 5165b0f5fe93SStefano Zampini ierr = VecGetLocalSize(nullv,&nl);CHKERRQ(ierr); 5166b0f5fe93SStefano Zampini ierr = VecGetArrayRead(pcbddc->coarse_vec,(const PetscScalar**)&array);CHKERRQ(ierr); 5167b0f5fe93SStefano Zampini ierr = VecGetArray(nullv,&array2);CHKERRQ(ierr); 5168b0f5fe93SStefano Zampini ierr = PetscMemcpy(array2,array,nl*sizeof(*array));CHKERRQ(ierr); 5169b0f5fe93SStefano Zampini ierr = VecRestoreArray(nullv,&array2);CHKERRQ(ierr); 5170b0f5fe93SStefano Zampini ierr = VecRestoreArrayRead(pcbddc->coarse_vec,(const PetscScalar**)&array);CHKERRQ(ierr); 5171b0f5fe93SStefano Zampini ierr = VecNormalize(nullv,NULL);CHKERRQ(ierr); 5172b0f5fe93SStefano Zampini ierr = MatNullSpaceCreate(PetscObjectComm((PetscObject)coarse_mat),PETSC_FALSE,1,&nullv,&CoarseNullSpace);CHKERRQ(ierr); 5173b0f5fe93SStefano Zampini ierr = VecDestroy(&nullv);CHKERRQ(ierr); 5174b0f5fe93SStefano Zampini } 5175b0f5fe93SStefano Zampini } 5176b0f5fe93SStefano Zampini 5177b0f5fe93SStefano Zampini if (pcbddc->coarse_ksp) { 5178b0f5fe93SStefano Zampini PetscBool ispreonly; 5179b0f5fe93SStefano Zampini 5180b0f5fe93SStefano Zampini if (CoarseNullSpace) { 5181b0f5fe93SStefano Zampini PetscBool isnull; 5182b0f5fe93SStefano Zampini ierr = MatNullSpaceTest(CoarseNullSpace,coarse_mat,&isnull);CHKERRQ(ierr); 5183b0f5fe93SStefano Zampini if (isnull) { 5184b0f5fe93SStefano Zampini if (isbddc) { 5185b0f5fe93SStefano Zampini ierr = PCBDDCSetNullSpace(pc_temp,CoarseNullSpace);CHKERRQ(ierr); 5186b0f5fe93SStefano Zampini } else { 5187b0f5fe93SStefano Zampini ierr = MatSetNullSpace(coarse_mat,CoarseNullSpace);CHKERRQ(ierr); 5188b0f5fe93SStefano Zampini } 5189b0f5fe93SStefano Zampini } else { 5190b0f5fe93SStefano Zampini ierr = MatNullSpaceDestroy(&CoarseNullSpace);CHKERRQ(ierr); 5191b0f5fe93SStefano Zampini } 5192b0f5fe93SStefano Zampini } 5193b0f5fe93SStefano Zampini /* setup coarse ksp */ 5194b0f5fe93SStefano Zampini ierr = KSPSetUp(pcbddc->coarse_ksp);CHKERRQ(ierr); 5195cbcc2c2aSStefano Zampini /* Check coarse problem if in debug mode or if solving with an iterative method */ 5196cbcc2c2aSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcbddc->coarse_ksp,KSPPREONLY,&ispreonly);CHKERRQ(ierr); 51976e683305SStefano Zampini if (pcbddc->dbg_flag || (!ispreonly && pcbddc->use_coarse_estimates) ) { 5198c8587f34SStefano Zampini KSP check_ksp; 51992b510759SStefano Zampini KSPType check_ksp_type; 5200c8587f34SStefano Zampini PC check_pc; 52016e683305SStefano Zampini Vec check_vec,coarse_vec; 52026a1308c2SStefano Zampini PetscReal abs_infty_error,infty_error,lambda_min=1.0,lambda_max=1.0; 52032b510759SStefano Zampini PetscInt its; 52046e683305SStefano Zampini PetscBool compute_eigs; 52056e683305SStefano Zampini PetscReal *eigs_r,*eigs_c; 52066e683305SStefano Zampini PetscInt neigs; 52078e185a42SStefano Zampini const char *prefix; 5208c8587f34SStefano Zampini 52092b510759SStefano Zampini /* Create ksp object suitable for estimation of extreme eigenvalues */ 52106e683305SStefano Zampini ierr = KSPCreate(PetscObjectComm((PetscObject)pcbddc->coarse_ksp),&check_ksp);CHKERRQ(ierr); 5211422a814eSBarry Smith ierr = KSPSetErrorIfNotConverged(pcbddc->coarse_ksp,pc->erroriffailure);CHKERRQ(ierr); 521223ee1639SBarry Smith ierr = KSPSetOperators(check_ksp,coarse_mat,coarse_mat);CHKERRQ(ierr); 5213f4ddd8eeSStefano Zampini ierr = KSPSetTolerances(check_ksp,1.e-12,1.e-12,PETSC_DEFAULT,pcbddc->coarse_size);CHKERRQ(ierr); 52142b510759SStefano Zampini if (ispreonly) { 52152b510759SStefano Zampini check_ksp_type = KSPPREONLY; 52166e683305SStefano Zampini compute_eigs = PETSC_FALSE; 52172b510759SStefano Zampini } else { 5218cbcc2c2aSStefano Zampini check_ksp_type = KSPGMRES; 52196e683305SStefano Zampini compute_eigs = PETSC_TRUE; 5220c8587f34SStefano Zampini } 5221c8587f34SStefano Zampini ierr = KSPSetType(check_ksp,check_ksp_type);CHKERRQ(ierr); 52226e683305SStefano Zampini ierr = KSPSetComputeSingularValues(check_ksp,compute_eigs);CHKERRQ(ierr); 52236e683305SStefano Zampini ierr = KSPSetComputeEigenvalues(check_ksp,compute_eigs);CHKERRQ(ierr); 52246e683305SStefano Zampini ierr = KSPGMRESSetRestart(check_ksp,pcbddc->coarse_size+1);CHKERRQ(ierr); 5225a7dc3881SStefano Zampini ierr = KSPGetOptionsPrefix(pcbddc->coarse_ksp,&prefix);CHKERRQ(ierr); 5226a7dc3881SStefano Zampini ierr = KSPSetOptionsPrefix(check_ksp,prefix);CHKERRQ(ierr); 5227a7dc3881SStefano Zampini ierr = KSPAppendOptionsPrefix(check_ksp,"check_");CHKERRQ(ierr); 5228a7dc3881SStefano Zampini ierr = KSPSetFromOptions(check_ksp);CHKERRQ(ierr); 5229c8587f34SStefano Zampini ierr = KSPSetUp(check_ksp);CHKERRQ(ierr); 5230c8587f34SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&check_pc);CHKERRQ(ierr); 5231c8587f34SStefano Zampini ierr = KSPSetPC(check_ksp,check_pc);CHKERRQ(ierr); 5232c8587f34SStefano Zampini /* create random vec */ 52336e683305SStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&coarse_vec);CHKERRQ(ierr); 52346e683305SStefano Zampini ierr = VecDuplicate(coarse_vec,&check_vec);CHKERRQ(ierr); 5235c8587f34SStefano Zampini ierr = VecSetRandom(check_vec,NULL);CHKERRQ(ierr); 5236c8587f34SStefano Zampini if (CoarseNullSpace) { 5237c8587f34SStefano Zampini ierr = MatNullSpaceRemove(CoarseNullSpace,check_vec);CHKERRQ(ierr); 5238c8587f34SStefano Zampini } 52396e683305SStefano Zampini ierr = MatMult(coarse_mat,check_vec,coarse_vec);CHKERRQ(ierr); 5240c8587f34SStefano Zampini /* solve coarse problem */ 52416e683305SStefano Zampini ierr = KSPSolve(check_ksp,coarse_vec,coarse_vec);CHKERRQ(ierr); 5242c8587f34SStefano Zampini if (CoarseNullSpace) { 52436e683305SStefano Zampini ierr = MatNullSpaceRemove(CoarseNullSpace,coarse_vec);CHKERRQ(ierr); 5244c8587f34SStefano Zampini } 5245cbcc2c2aSStefano Zampini /* set eigenvalue estimation if preonly has not been requested */ 52466e683305SStefano Zampini if (compute_eigs) { 5247854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->coarse_size+1,&eigs_r);CHKERRQ(ierr); 5248854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->coarse_size+1,&eigs_c);CHKERRQ(ierr); 52496e683305SStefano Zampini ierr = KSPComputeEigenvalues(check_ksp,pcbddc->coarse_size+1,eigs_r,eigs_c,&neigs);CHKERRQ(ierr); 52506e683305SStefano Zampini lambda_max = eigs_r[neigs-1]; 52516e683305SStefano Zampini lambda_min = eigs_r[0]; 52526e683305SStefano Zampini if (pcbddc->use_coarse_estimates) { 52536e683305SStefano Zampini if (lambda_max>lambda_min) { 5254cbcc2c2aSStefano Zampini ierr = KSPChebyshevSetEigenvalues(pcbddc->coarse_ksp,lambda_max,lambda_min);CHKERRQ(ierr); 5255cbcc2c2aSStefano Zampini ierr = KSPRichardsonSetScale(pcbddc->coarse_ksp,2.0/(lambda_max+lambda_min));CHKERRQ(ierr); 5256cbcc2c2aSStefano Zampini } 5257c8587f34SStefano Zampini } 5258c8587f34SStefano Zampini } 5259cbcc2c2aSStefano Zampini 5260c8587f34SStefano Zampini /* check coarse problem residual error */ 52616e683305SStefano Zampini if (pcbddc->dbg_flag) { 52626e683305SStefano Zampini PetscViewer dbg_viewer = PETSC_VIEWER_STDOUT_(PetscObjectComm((PetscObject)pcbddc->coarse_ksp)); 52636e683305SStefano Zampini ierr = PetscViewerASCIIAddTab(dbg_viewer,2*(pcbddc->current_level+1));CHKERRQ(ierr); 52646e683305SStefano Zampini ierr = VecAXPY(check_vec,-1.0,coarse_vec);CHKERRQ(ierr); 5265c8587f34SStefano Zampini ierr = VecNorm(check_vec,NORM_INFINITY,&infty_error);CHKERRQ(ierr); 52666e683305SStefano Zampini ierr = MatMult(coarse_mat,check_vec,coarse_vec);CHKERRQ(ierr); 52676e683305SStefano Zampini ierr = VecNorm(coarse_vec,NORM_INFINITY,&abs_infty_error);CHKERRQ(ierr); 5268c8587f34SStefano Zampini ierr = VecDestroy(&check_vec);CHKERRQ(ierr); 5269779c1cceSStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem details (use estimates %d)\n",pcbddc->use_coarse_estimates);CHKERRQ(ierr); 52706e683305SStefano Zampini ierr = PetscObjectPrintClassNamePrefixType((PetscObject)(pcbddc->coarse_ksp),dbg_viewer);CHKERRQ(ierr); 52716e683305SStefano Zampini ierr = PetscObjectPrintClassNamePrefixType((PetscObject)(check_pc),dbg_viewer);CHKERRQ(ierr); 52726e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem exact infty_error : %1.6e\n",infty_error);CHKERRQ(ierr); 52736e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem residual infty_error: %1.6e\n",abs_infty_error);CHKERRQ(ierr); 5274b0f5fe93SStefano Zampini if (CoarseNullSpace) { 5275b0f5fe93SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem is singular\n");CHKERRQ(ierr); 5276b0f5fe93SStefano Zampini } 52776e683305SStefano Zampini if (compute_eigs) { 52786e683305SStefano Zampini PetscReal lambda_max_s,lambda_min_s; 5279deec49d1SStefano Zampini ierr = KSPGetType(check_ksp,&check_ksp_type);CHKERRQ(ierr); 5280c8587f34SStefano Zampini ierr = KSPGetIterationNumber(check_ksp,&its);CHKERRQ(ierr); 52816e683305SStefano Zampini ierr = KSPComputeExtremeSingularValues(check_ksp,&lambda_max_s,&lambda_min_s);CHKERRQ(ierr); 52826e683305SStefano 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); 52836e683305SStefano Zampini for (i=0;i<neigs;i++) { 52846e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"%1.6e %1.6ei\n",eigs_r[i],eigs_c[i]);CHKERRQ(ierr); 5285c8587f34SStefano Zampini } 52866e683305SStefano Zampini } 52876e683305SStefano Zampini ierr = PetscViewerFlush(dbg_viewer);CHKERRQ(ierr); 52886e683305SStefano Zampini ierr = PetscViewerASCIISubtractTab(dbg_viewer,2*(pcbddc->current_level+1));CHKERRQ(ierr); 52896e683305SStefano Zampini } 5290c8587f34SStefano Zampini ierr = KSPDestroy(&check_ksp);CHKERRQ(ierr); 52916e683305SStefano Zampini if (compute_eigs) { 52926e683305SStefano Zampini ierr = PetscFree(eigs_r);CHKERRQ(ierr); 52936e683305SStefano Zampini ierr = PetscFree(eigs_c);CHKERRQ(ierr); 5294c8587f34SStefano Zampini } 52956e683305SStefano Zampini } 52966e683305SStefano Zampini } 5297cbcc2c2aSStefano Zampini /* print additional info */ 5298cbcc2c2aSStefano Zampini if (pcbddc->dbg_flag) { 52996e683305SStefano Zampini /* waits until all processes reaches this point */ 53006e683305SStefano Zampini ierr = PetscBarrier((PetscObject)pc);CHKERRQ(ierr); 5301cbcc2c2aSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Coarse solver setup completed at level %d\n",pcbddc->current_level);CHKERRQ(ierr); 5302cbcc2c2aSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 5303cbcc2c2aSStefano Zampini } 5304cbcc2c2aSStefano Zampini 53052b510759SStefano Zampini /* free memory */ 5306c8587f34SStefano Zampini ierr = MatNullSpaceDestroy(&CoarseNullSpace);CHKERRQ(ierr); 5307fdc635d7SStefano Zampini ierr = MatDestroy(&coarse_mat);CHKERRQ(ierr); 5308c8587f34SStefano Zampini PetscFunctionReturn(0); 5309c8587f34SStefano Zampini } 5310674ae819SStefano Zampini 5311f34684f1SStefano Zampini #undef __FUNCT__ 5312f34684f1SStefano Zampini #define __FUNCT__ "PCBDDCComputePrimalNumbering" 5313f34684f1SStefano Zampini PetscErrorCode PCBDDCComputePrimalNumbering(PC pc,PetscInt* coarse_size_n,PetscInt** local_primal_indices_n) 5314f34684f1SStefano Zampini { 5315f34684f1SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 5316f34684f1SStefano Zampini PC_IS* pcis = (PC_IS*)pc->data; 5317f34684f1SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 5318dc456d91SStefano Zampini IS subset,subset_mult,subset_n; 5319dc456d91SStefano Zampini PetscInt local_size,coarse_size=0; 532073be2a3aSStefano Zampini PetscInt *local_primal_indices=NULL; 5321dc456d91SStefano Zampini const PetscInt *t_local_primal_indices; 5322f34684f1SStefano Zampini PetscErrorCode ierr; 5323f34684f1SStefano Zampini 5324f34684f1SStefano Zampini PetscFunctionBegin; 5325f34684f1SStefano Zampini /* Compute global number of coarse dofs */ 53260e6343abSStefano Zampini if (pcbddc->local_primal_size && !pcbddc->local_primal_ref_node) { 53270e6343abSStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"BDDC ConstraintsSetUp should be called first"); 5328727cdba6SStefano Zampini } 5329dc456d91SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)(pc->pmat)),pcbddc->local_primal_size_cc,pcbddc->local_primal_ref_node,PETSC_COPY_VALUES,&subset_n);CHKERRQ(ierr); 53303bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(pcis->mapping,subset_n,&subset);CHKERRQ(ierr); 5331dc456d91SStefano Zampini ierr = ISDestroy(&subset_n);CHKERRQ(ierr); 5332dc456d91SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)(pc->pmat)),pcbddc->local_primal_size_cc,pcbddc->local_primal_ref_mult,PETSC_COPY_VALUES,&subset_mult);CHKERRQ(ierr); 5333dc456d91SStefano Zampini ierr = PCBDDCSubsetNumbering(subset,subset_mult,&coarse_size,&subset_n);CHKERRQ(ierr); 5334dc456d91SStefano Zampini ierr = ISDestroy(&subset);CHKERRQ(ierr); 5335dc456d91SStefano Zampini ierr = ISDestroy(&subset_mult);CHKERRQ(ierr); 5336dc456d91SStefano Zampini ierr = ISGetLocalSize(subset_n,&local_size);CHKERRQ(ierr); 53370e6343abSStefano Zampini if (local_size != pcbddc->local_primal_size) { 53380e6343abSStefano Zampini SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Invalid number of local primal indices computed %d != %d",local_size,pcbddc->local_primal_size); 53390e6343abSStefano Zampini } 5340dc456d91SStefano Zampini ierr = PetscMalloc1(local_size,&local_primal_indices);CHKERRQ(ierr); 5341dc456d91SStefano Zampini ierr = ISGetIndices(subset_n,&t_local_primal_indices);CHKERRQ(ierr); 5342dc456d91SStefano Zampini ierr = PetscMemcpy(local_primal_indices,t_local_primal_indices,local_size*sizeof(PetscInt));CHKERRQ(ierr); 5343dc456d91SStefano Zampini ierr = ISRestoreIndices(subset_n,&t_local_primal_indices);CHKERRQ(ierr); 5344dc456d91SStefano Zampini ierr = ISDestroy(&subset_n);CHKERRQ(ierr); 5345f34684f1SStefano Zampini 5346f34684f1SStefano Zampini /* check numbering */ 5347f34684f1SStefano Zampini if (pcbddc->dbg_flag) { 5348019a44ceSStefano Zampini PetscScalar coarsesum,*array,*array2; 5349dc456d91SStefano Zampini PetscInt i; 5350b9b85e73SStefano Zampini PetscBool set_error = PETSC_FALSE,set_error_reduced = PETSC_FALSE; 5351f34684f1SStefano Zampini 5352f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 5353f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 5354f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Check coarse indices\n");CHKERRQ(ierr); 53550fccc4e9SStefano Zampini ierr = PetscViewerASCIISynchronizedAllow(pcbddc->dbg_viewer,PETSC_TRUE);CHKERRQ(ierr); 5356019a44ceSStefano Zampini /* counter */ 5357019a44ceSStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 5358019a44ceSStefano Zampini ierr = VecSet(pcis->vec1_N,1.0);CHKERRQ(ierr); 5359019a44ceSStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 5360019a44ceSStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 5361019a44ceSStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_global,pcis->vec2_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 5362019a44ceSStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_global,pcis->vec2_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 5363019a44ceSStefano Zampini 5364f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_N,0.0);CHKERRQ(ierr); 5365f34684f1SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 5366727cdba6SStefano Zampini ierr = VecSetValue(pcis->vec1_N,pcbddc->primal_indices_local_idxs[i],1.0,INSERT_VALUES);CHKERRQ(ierr); 5367f34684f1SStefano Zampini } 5368f34684f1SStefano Zampini ierr = VecAssemblyBegin(pcis->vec1_N);CHKERRQ(ierr); 5369f34684f1SStefano Zampini ierr = VecAssemblyEnd(pcis->vec1_N);CHKERRQ(ierr); 5370f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 5371e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 5372e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 5373e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_global,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 5374e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_global,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 5375f34684f1SStefano Zampini ierr = VecGetArray(pcis->vec1_N,&array);CHKERRQ(ierr); 5376019a44ceSStefano Zampini ierr = VecGetArray(pcis->vec2_N,&array2);CHKERRQ(ierr); 5377f34684f1SStefano Zampini for (i=0;i<pcis->n;i++) { 5378019a44ceSStefano Zampini if (array[i] != 0.0 && array[i] != array2[i]) { 537975c01103SStefano Zampini PetscInt owned = (PetscInt)PetscRealPart(array[i]); 538075c01103SStefano Zampini PetscInt neigh = (PetscInt)PetscRealPart(array2[i]); 5381b9b85e73SStefano Zampini set_error = PETSC_TRUE; 5382019a44ceSStefano 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); 5383f34684f1SStefano Zampini } 5384f34684f1SStefano Zampini } 5385019a44ceSStefano Zampini ierr = VecRestoreArray(pcis->vec2_N,&array2);CHKERRQ(ierr); 5386b9b85e73SStefano Zampini ierr = MPI_Allreduce(&set_error,&set_error_reduced,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 5387f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 5388f34684f1SStefano Zampini for (i=0;i<pcis->n;i++) { 5389f34684f1SStefano Zampini if (PetscRealPart(array[i]) > 0.0) array[i] = 1.0/PetscRealPart(array[i]); 5390f34684f1SStefano Zampini } 5391f34684f1SStefano Zampini ierr = VecRestoreArray(pcis->vec1_N,&array);CHKERRQ(ierr); 5392f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 5393e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 5394e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 5395f34684f1SStefano Zampini ierr = VecSum(pcis->vec1_global,&coarsesum);CHKERRQ(ierr); 5396f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Size of coarse problem is %d (%lf)\n",coarse_size,PetscRealPart(coarsesum));CHKERRQ(ierr); 5397b9b85e73SStefano Zampini if (pcbddc->dbg_flag > 1 || set_error_reduced) { 5398ca8b9ea9SStefano Zampini PetscInt *gidxs; 5399ca8b9ea9SStefano Zampini 5400ca8b9ea9SStefano Zampini ierr = PetscMalloc1(pcbddc->local_primal_size,&gidxs);CHKERRQ(ierr); 54013bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingApply(pcis->mapping,pcbddc->local_primal_size,pcbddc->primal_indices_local_idxs,gidxs);CHKERRQ(ierr); 5402f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Distribution of local primal indices\n");CHKERRQ(ierr); 5403f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 5404f34684f1SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d\n",PetscGlobalRank);CHKERRQ(ierr); 5405f34684f1SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 54064bc2dc4bSStefano 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); 5407f34684f1SStefano Zampini } 5408f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 5409ca8b9ea9SStefano Zampini ierr = PetscFree(gidxs);CHKERRQ(ierr); 5410f34684f1SStefano Zampini } 5411f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 5412302440fdSBarry Smith if (set_error_reduced) SETERRQ(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"BDDC Numbering of coarse dofs failed"); 5413f34684f1SStefano Zampini } 54148bec7fa6SStefano Zampini /* ierr = PetscPrintf(PetscObjectComm((PetscObject)pc),"Size of coarse problem is %d\n",coarse_size);CHKERRQ(ierr); */ 5415f34684f1SStefano Zampini /* get back data */ 5416f34684f1SStefano Zampini *coarse_size_n = coarse_size; 5417f34684f1SStefano Zampini *local_primal_indices_n = local_primal_indices; 5418674ae819SStefano Zampini PetscFunctionReturn(0); 5419674ae819SStefano Zampini } 5420674ae819SStefano Zampini 5421e456f2a8SStefano Zampini #undef __FUNCT__ 5422e456f2a8SStefano Zampini #define __FUNCT__ "PCBDDCGlobalToLocal" 5423a7dc3881SStefano Zampini PetscErrorCode PCBDDCGlobalToLocal(VecScatter g2l_ctx,Vec gwork, Vec lwork, IS globalis, IS* localis) 5424e456f2a8SStefano Zampini { 5425e456f2a8SStefano Zampini IS localis_t; 5426a7dc3881SStefano Zampini PetscInt i,lsize,*idxs,n; 5427e456f2a8SStefano Zampini PetscScalar *vals; 5428e456f2a8SStefano Zampini PetscErrorCode ierr; 5429e456f2a8SStefano Zampini 5430e456f2a8SStefano Zampini PetscFunctionBegin; 5431a7dc3881SStefano Zampini /* get indices in local ordering exploiting local to global map */ 5432e456f2a8SStefano Zampini ierr = ISGetLocalSize(globalis,&lsize);CHKERRQ(ierr); 5433854ce69bSBarry Smith ierr = PetscMalloc1(lsize,&vals);CHKERRQ(ierr); 5434e456f2a8SStefano Zampini for (i=0;i<lsize;i++) vals[i] = 1.0; 5435e456f2a8SStefano Zampini ierr = ISGetIndices(globalis,(const PetscInt**)&idxs);CHKERRQ(ierr); 5436a7dc3881SStefano Zampini ierr = VecSet(gwork,0.0);CHKERRQ(ierr); 5437a7dc3881SStefano Zampini ierr = VecSet(lwork,0.0);CHKERRQ(ierr); 54381035eff8SStefano Zampini if (idxs) { /* multilevel guard */ 5439a7dc3881SStefano Zampini ierr = VecSetValues(gwork,lsize,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 54401035eff8SStefano Zampini } 5441a7dc3881SStefano Zampini ierr = VecAssemblyBegin(gwork);CHKERRQ(ierr); 5442e456f2a8SStefano Zampini ierr = ISRestoreIndices(globalis,(const PetscInt**)&idxs);CHKERRQ(ierr); 5443e456f2a8SStefano Zampini ierr = PetscFree(vals);CHKERRQ(ierr); 5444a7dc3881SStefano Zampini ierr = VecAssemblyEnd(gwork);CHKERRQ(ierr); 5445a7dc3881SStefano Zampini /* now compute set in local ordering */ 5446a7dc3881SStefano Zampini ierr = VecScatterBegin(g2l_ctx,gwork,lwork,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 5447a7dc3881SStefano Zampini ierr = VecScatterEnd(g2l_ctx,gwork,lwork,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 5448a7dc3881SStefano Zampini ierr = VecGetArrayRead(lwork,(const PetscScalar**)&vals);CHKERRQ(ierr); 5449a7dc3881SStefano Zampini ierr = VecGetSize(lwork,&n);CHKERRQ(ierr); 5450a7dc3881SStefano Zampini for (i=0,lsize=0;i<n;i++) { 5451ff92baa0SMatthew G. Knepley if (PetscRealPart(vals[i]) > 0.5) { 5452e456f2a8SStefano Zampini lsize++; 5453e456f2a8SStefano Zampini } 5454e456f2a8SStefano Zampini } 5455854ce69bSBarry Smith ierr = PetscMalloc1(lsize,&idxs);CHKERRQ(ierr); 5456a7dc3881SStefano Zampini for (i=0,lsize=0;i<n;i++) { 5457ff92baa0SMatthew G. Knepley if (PetscRealPart(vals[i]) > 0.5) { 5458e456f2a8SStefano Zampini idxs[lsize++] = i; 5459e456f2a8SStefano Zampini } 5460e456f2a8SStefano Zampini } 5461a7dc3881SStefano Zampini ierr = VecRestoreArrayRead(lwork,(const PetscScalar**)&vals);CHKERRQ(ierr); 5462a7dc3881SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)gwork),lsize,idxs,PETSC_OWN_POINTER,&localis_t);CHKERRQ(ierr); 5463e456f2a8SStefano Zampini *localis = localis_t; 5464e456f2a8SStefano Zampini PetscFunctionReturn(0); 5465e456f2a8SStefano Zampini } 5466906d46d4SStefano Zampini 5467906d46d4SStefano Zampini /* the next two functions will be called in KSPMatMult if a change of basis has been requested */ 5468906d46d4SStefano Zampini #undef __FUNCT__ 5469906d46d4SStefano Zampini #define __FUNCT__ "PCBDDCMatMult_Private" 5470906d46d4SStefano Zampini static PetscErrorCode PCBDDCMatMult_Private(Mat A, Vec x, Vec y) 5471906d46d4SStefano Zampini { 5472906d46d4SStefano Zampini PCBDDCChange_ctx change_ctx; 5473906d46d4SStefano Zampini PetscErrorCode ierr; 5474906d46d4SStefano Zampini 5475906d46d4SStefano Zampini PetscFunctionBegin; 5476906d46d4SStefano Zampini ierr = MatShellGetContext(A,&change_ctx);CHKERRQ(ierr); 5477906d46d4SStefano Zampini ierr = MatMult(change_ctx->global_change,x,change_ctx->work[0]);CHKERRQ(ierr); 5478906d46d4SStefano Zampini ierr = MatMult(change_ctx->original_mat,change_ctx->work[0],change_ctx->work[1]);CHKERRQ(ierr); 5479906d46d4SStefano Zampini ierr = MatMultTranspose(change_ctx->global_change,change_ctx->work[1],y);CHKERRQ(ierr); 5480906d46d4SStefano Zampini PetscFunctionReturn(0); 5481906d46d4SStefano Zampini } 5482906d46d4SStefano Zampini 5483906d46d4SStefano Zampini #undef __FUNCT__ 5484906d46d4SStefano Zampini #define __FUNCT__ "PCBDDCMatMultTranspose_Private" 5485906d46d4SStefano Zampini static PetscErrorCode PCBDDCMatMultTranspose_Private(Mat A, Vec x, Vec y) 5486906d46d4SStefano Zampini { 5487906d46d4SStefano Zampini PCBDDCChange_ctx change_ctx; 5488906d46d4SStefano Zampini PetscErrorCode ierr; 5489906d46d4SStefano Zampini 5490906d46d4SStefano Zampini PetscFunctionBegin; 5491906d46d4SStefano Zampini ierr = MatShellGetContext(A,&change_ctx);CHKERRQ(ierr); 5492906d46d4SStefano Zampini ierr = MatMult(change_ctx->global_change,x,change_ctx->work[0]);CHKERRQ(ierr); 5493906d46d4SStefano Zampini ierr = MatMultTranspose(change_ctx->original_mat,change_ctx->work[0],change_ctx->work[1]);CHKERRQ(ierr); 5494906d46d4SStefano Zampini ierr = MatMultTranspose(change_ctx->global_change,change_ctx->work[1],y);CHKERRQ(ierr); 5495906d46d4SStefano Zampini PetscFunctionReturn(0); 5496906d46d4SStefano Zampini } 5497b96c3477SStefano Zampini 5498b96c3477SStefano Zampini #undef __FUNCT__ 5499b96c3477SStefano Zampini #define __FUNCT__ "PCBDDCSetUpSubSchurs" 550008122e43SStefano Zampini PetscErrorCode PCBDDCSetUpSubSchurs(PC pc) 5501b96c3477SStefano Zampini { 5502a64f4aa4SStefano Zampini PC_IS *pcis=(PC_IS*)pc->data; 5503b96c3477SStefano Zampini PC_BDDC *pcbddc=(PC_BDDC*)pc->data; 5504b96c3477SStefano Zampini PCBDDCSubSchurs sub_schurs=pcbddc->sub_schurs; 5505a64f4aa4SStefano Zampini Mat S_j; 5506b96c3477SStefano Zampini PetscInt *used_xadj,*used_adjncy; 5507b96c3477SStefano Zampini PetscBool free_used_adj; 5508b96c3477SStefano Zampini PetscErrorCode ierr; 5509b96c3477SStefano Zampini 5510b96c3477SStefano Zampini PetscFunctionBegin; 5511b96c3477SStefano Zampini /* decide the adjacency to be used for determining internal problems for local schur on subsets */ 5512b96c3477SStefano Zampini free_used_adj = PETSC_FALSE; 551308122e43SStefano Zampini if (pcbddc->sub_schurs_layers == -1) { 5514b96c3477SStefano Zampini used_xadj = NULL; 5515b96c3477SStefano Zampini used_adjncy = NULL; 5516b96c3477SStefano Zampini } else { 551708122e43SStefano Zampini if (pcbddc->sub_schurs_use_useradj && pcbddc->mat_graph->xadj) { 551808122e43SStefano Zampini used_xadj = pcbddc->mat_graph->xadj; 551908122e43SStefano Zampini used_adjncy = pcbddc->mat_graph->adjncy; 552008122e43SStefano Zampini } else if (pcbddc->computed_rowadj) { 5521b96c3477SStefano Zampini used_xadj = pcbddc->mat_graph->xadj; 5522b96c3477SStefano Zampini used_adjncy = pcbddc->mat_graph->adjncy; 5523b96c3477SStefano Zampini } else { 55242fffb893SStefano Zampini PetscBool flg_row=PETSC_FALSE; 5525b96c3477SStefano Zampini const PetscInt *xadj,*adjncy; 5526b96c3477SStefano Zampini PetscInt nvtxs; 5527b96c3477SStefano Zampini 55282fffb893SStefano Zampini ierr = MatGetRowIJ(pcbddc->local_mat,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 55292fffb893SStefano Zampini if (flg_row) { 5530b96c3477SStefano Zampini ierr = PetscMalloc2(nvtxs+1,&used_xadj,xadj[nvtxs],&used_adjncy);CHKERRQ(ierr); 5531b96c3477SStefano Zampini ierr = PetscMemcpy(used_xadj,xadj,(nvtxs+1)*sizeof(*xadj));CHKERRQ(ierr); 5532b96c3477SStefano Zampini ierr = PetscMemcpy(used_adjncy,adjncy,(xadj[nvtxs])*sizeof(*adjncy));CHKERRQ(ierr); 5533b96c3477SStefano Zampini free_used_adj = PETSC_TRUE; 55342fffb893SStefano Zampini } else { 55352fffb893SStefano Zampini pcbddc->sub_schurs_layers = -1; 55362fffb893SStefano Zampini used_xadj = NULL; 55372fffb893SStefano Zampini used_adjncy = NULL; 55382fffb893SStefano Zampini } 55392fffb893SStefano Zampini ierr = MatRestoreRowIJ(pcbddc->local_mat,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 5540b96c3477SStefano Zampini } 5541b96c3477SStefano Zampini } 5542d5574798SStefano Zampini 5543d5574798SStefano Zampini /* setup sub_schurs data */ 5544a64f4aa4SStefano Zampini ierr = MatCreateSchurComplement(pcis->A_II,pcis->A_II,pcis->A_IB,pcis->A_BI,pcis->A_BB,&S_j);CHKERRQ(ierr); 5545a64f4aa4SStefano Zampini if (!sub_schurs->use_mumps) { 5546a64f4aa4SStefano Zampini /* pcbddc->ksp_D up to date only if not using MUMPS */ 5547a64f4aa4SStefano Zampini ierr = MatSchurComplementSetKSP(S_j,pcbddc->ksp_D);CHKERRQ(ierr); 554806a4e24aSStefano 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); 5549a64f4aa4SStefano Zampini } else { 55506816873aSStefano Zampini PetscBool reuse_solvers = (PetscBool)!pcbddc->use_change_of_basis; 555104708bb6SStefano Zampini PetscBool isseqaij; 55525feab87aSStefano Zampini if (!pcbddc->use_vertices && reuse_solvers) { 55535feab87aSStefano Zampini PetscInt n_vertices; 55545feab87aSStefano Zampini 55555feab87aSStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_vertices,&n_vertices);CHKERRQ(ierr); 55562034aafcSStefano Zampini reuse_solvers = (PetscBool)!n_vertices; 55575feab87aSStefano Zampini } 555804708bb6SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcbddc->local_mat,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 555904708bb6SStefano Zampini if (!isseqaij) { 556004708bb6SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 556104708bb6SStefano Zampini if (matis->A == pcbddc->local_mat) { 556204708bb6SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 556304708bb6SStefano Zampini ierr = MatConvert(matis->A,MATSEQAIJ,MAT_INITIAL_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 556404708bb6SStefano Zampini } else { 556504708bb6SStefano Zampini ierr = MatConvert(pcbddc->local_mat,MATSEQAIJ,MAT_REUSE_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 556604708bb6SStefano Zampini } 556704708bb6SStefano Zampini } 556806a4e24aSStefano 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); 5569a64f4aa4SStefano Zampini } 5570a64f4aa4SStefano Zampini ierr = MatDestroy(&S_j);CHKERRQ(ierr); 5571b96c3477SStefano Zampini 5572b96c3477SStefano Zampini /* free adjacency */ 5573b96c3477SStefano Zampini if (free_used_adj) { 5574b96c3477SStefano Zampini ierr = PetscFree2(used_xadj,used_adjncy);CHKERRQ(ierr); 5575b96c3477SStefano Zampini } 5576b96c3477SStefano Zampini PetscFunctionReturn(0); 5577b96c3477SStefano Zampini } 5578b96c3477SStefano Zampini 5579b96c3477SStefano Zampini #undef __FUNCT__ 5580b96c3477SStefano Zampini #define __FUNCT__ "PCBDDCInitSubSchurs" 558108122e43SStefano Zampini PetscErrorCode PCBDDCInitSubSchurs(PC pc) 5582b96c3477SStefano Zampini { 5583b96c3477SStefano Zampini PC_IS *pcis=(PC_IS*)pc->data; 5584b96c3477SStefano Zampini PC_BDDC *pcbddc=(PC_BDDC*)pc->data; 5585b96c3477SStefano Zampini PCBDDCSubSchurs sub_schurs=pcbddc->sub_schurs; 5586b96c3477SStefano Zampini PCBDDCGraph graph; 5587b96c3477SStefano Zampini PetscErrorCode ierr; 5588b96c3477SStefano Zampini 5589b96c3477SStefano Zampini PetscFunctionBegin; 5590b96c3477SStefano Zampini /* attach interface graph for determining subsets */ 559108122e43SStefano Zampini if (pcbddc->sub_schurs_rebuild) { /* in case rebuild has been requested, it uses a graph generated only by the neighbouring information */ 55923301b35fSStefano Zampini IS verticesIS,verticescomm; 55933301b35fSStefano Zampini PetscInt vsize,*idxs; 5594b96c3477SStefano Zampini 5595b96c3477SStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,NULL,NULL,NULL,NULL,&verticesIS);CHKERRQ(ierr); 55963301b35fSStefano Zampini ierr = ISGetSize(verticesIS,&vsize);CHKERRQ(ierr); 55973301b35fSStefano Zampini ierr = ISGetIndices(verticesIS,(const PetscInt**)&idxs);CHKERRQ(ierr); 55983301b35fSStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),vsize,idxs,PETSC_COPY_VALUES,&verticescomm);CHKERRQ(ierr); 55993301b35fSStefano Zampini ierr = ISRestoreIndices(verticesIS,(const PetscInt**)&idxs);CHKERRQ(ierr); 56003301b35fSStefano Zampini ierr = ISDestroy(&verticesIS);CHKERRQ(ierr); 5601b96c3477SStefano Zampini ierr = PCBDDCGraphCreate(&graph);CHKERRQ(ierr); 56027fb0e2dbSStefano Zampini ierr = PCBDDCGraphInit(graph,pcbddc->mat_graph->l2gmap,pcbddc->mat_graph->nvtxs_global);CHKERRQ(ierr); 56033301b35fSStefano Zampini ierr = PCBDDCGraphSetUp(graph,0,NULL,pcbddc->DirichletBoundariesLocal,0,NULL,verticescomm);CHKERRQ(ierr); 56043301b35fSStefano Zampini ierr = ISDestroy(&verticescomm);CHKERRQ(ierr); 5605b96c3477SStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(graph);CHKERRQ(ierr); 5606b96c3477SStefano Zampini /* 5607b96c3477SStefano Zampini if (pcbddc->dbg_flag) { 5608b96c3477SStefano Zampini ierr = PCBDDCGraphASCIIView(graph,pcbddc->dbg_flag,pcbddc->dbg_viewer);CHKERRQ(ierr); 5609b96c3477SStefano Zampini } 5610b96c3477SStefano Zampini */ 5611b96c3477SStefano Zampini } else { 5612b96c3477SStefano Zampini graph = pcbddc->mat_graph; 5613b96c3477SStefano Zampini } 5614b96c3477SStefano Zampini 5615b96c3477SStefano Zampini /* sub_schurs init */ 5616a64f4aa4SStefano Zampini ierr = PCBDDCSubSchursInit(sub_schurs,pcis->is_I_local,pcis->is_B_local,graph,pcis->BtoNmap);CHKERRQ(ierr); 5617a64f4aa4SStefano Zampini 5618b96c3477SStefano Zampini /* free graph struct */ 561908122e43SStefano Zampini if (pcbddc->sub_schurs_rebuild) { 5620b96c3477SStefano Zampini ierr = PCBDDCGraphDestroy(&graph);CHKERRQ(ierr); 5621b96c3477SStefano Zampini } 5622b96c3477SStefano Zampini PetscFunctionReturn(0); 5623b96c3477SStefano Zampini } 5624fa34dd3eSStefano Zampini 5625fa34dd3eSStefano Zampini #undef __FUNCT__ 5626fa34dd3eSStefano Zampini #define __FUNCT__ "PCBDDCCheckOperator" 5627fa34dd3eSStefano Zampini PetscErrorCode PCBDDCCheckOperator(PC pc) 5628fa34dd3eSStefano Zampini { 5629fa34dd3eSStefano Zampini PC_IS *pcis=(PC_IS*)pc->data; 5630fa34dd3eSStefano Zampini PC_BDDC *pcbddc=(PC_BDDC*)pc->data; 5631fa34dd3eSStefano Zampini PetscErrorCode ierr; 5632fa34dd3eSStefano Zampini 5633fa34dd3eSStefano Zampini PetscFunctionBegin; 5634fa34dd3eSStefano Zampini if (pcbddc->n_vertices == pcbddc->local_primal_size) { 5635fa34dd3eSStefano Zampini IS zerodiag = NULL; 5636*4f1b2e48SStefano Zampini Mat S_j,B0_B=NULL; 5637fa34dd3eSStefano Zampini Vec dummy_vec=NULL,vec_check_B,vec_scale_P; 5638*4f1b2e48SStefano Zampini PetscScalar *p0_check,*array,*array2; 563975c01103SStefano Zampini PetscReal norm; 5640fa34dd3eSStefano Zampini PetscInt i; 5641fa34dd3eSStefano Zampini 5642fa34dd3eSStefano Zampini /* B0 and B0_B */ 5643fa34dd3eSStefano Zampini if (zerodiag) { 5644fa34dd3eSStefano Zampini IS dummy; 5645fa34dd3eSStefano Zampini 5646*4f1b2e48SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcbddc->benign_n,0,1,&dummy);CHKERRQ(ierr); 5647*4f1b2e48SStefano Zampini ierr = MatGetSubMatrix(pcbddc->benign_B0,dummy,pcis->is_B_local,MAT_INITIAL_MATRIX,&B0_B);CHKERRQ(ierr); 5648fa34dd3eSStefano Zampini ierr = MatCreateVecs(B0_B,NULL,&dummy_vec);CHKERRQ(ierr); 5649fa34dd3eSStefano Zampini ierr = ISDestroy(&dummy);CHKERRQ(ierr); 5650fa34dd3eSStefano Zampini } 5651fa34dd3eSStefano Zampini /* I need a primal vector to scale primal nodes since BDDC sums contibutions */ 5652fa34dd3eSStefano Zampini ierr = VecDuplicate(pcbddc->vec1_P,&vec_scale_P);CHKERRQ(ierr); 5653fa34dd3eSStefano Zampini ierr = VecSet(pcbddc->vec1_P,1.0);CHKERRQ(ierr); 5654fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 5655fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 5656fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,vec_scale_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 5657fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,vec_scale_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 5658fa34dd3eSStefano Zampini ierr = VecReciprocal(vec_scale_P);CHKERRQ(ierr); 5659fa34dd3eSStefano Zampini /* S_j */ 5660fa34dd3eSStefano Zampini ierr = MatCreateSchurComplement(pcis->A_II,pcis->A_II,pcis->A_IB,pcis->A_BI,pcis->A_BB,&S_j);CHKERRQ(ierr); 5661fa34dd3eSStefano Zampini ierr = MatSchurComplementSetKSP(S_j,pcbddc->ksp_D);CHKERRQ(ierr); 5662fa34dd3eSStefano Zampini 5663fa34dd3eSStefano Zampini /* mimic vector in \widetilde{W}_\Gamma */ 5664fa34dd3eSStefano Zampini ierr = VecSetRandom(pcis->vec1_N,NULL);CHKERRQ(ierr); 5665fa34dd3eSStefano Zampini /* continuous in primal space */ 5666fa34dd3eSStefano Zampini ierr = VecSetRandom(pcbddc->coarse_vec,NULL);CHKERRQ(ierr); 5667fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,pcbddc->vec1_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 5668fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,pcbddc->vec1_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 5669fa34dd3eSStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 5670*4f1b2e48SStefano Zampini ierr = PetscCalloc1(pcbddc->benign_n,&p0_check);CHKERRQ(ierr); 5671*4f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) p0_check[i] = array[pcbddc->local_primal_size-pcbddc->benign_n+i]; 5672fa34dd3eSStefano Zampini ierr = VecSetValues(pcis->vec1_N,pcbddc->local_primal_size,pcbddc->local_primal_ref_node,array,INSERT_VALUES);CHKERRQ(ierr); 5673fa34dd3eSStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 5674fa34dd3eSStefano Zampini ierr = VecAssemblyBegin(pcis->vec1_N);CHKERRQ(ierr); 5675fa34dd3eSStefano Zampini ierr = VecAssemblyEnd(pcis->vec1_N);CHKERRQ(ierr); 5676fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcis->N_to_B,pcis->vec1_N,pcis->vec2_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 5677fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcis->N_to_B,pcis->vec1_N,pcis->vec2_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 5678fa34dd3eSStefano Zampini ierr = VecDuplicate(pcis->vec2_B,&vec_check_B);CHKERRQ(ierr); 5679fa34dd3eSStefano Zampini ierr = VecCopy(pcis->vec2_B,vec_check_B);CHKERRQ(ierr); 5680fa34dd3eSStefano Zampini 5681fa34dd3eSStefano Zampini /* assemble rhs for coarse problem */ 5682fa34dd3eSStefano Zampini /* widetilde{S}_\Gamma w_\Gamma + \widetilde{B0}^T_B p0 */ 5683fa34dd3eSStefano Zampini /* local with Schur */ 5684fa34dd3eSStefano Zampini ierr = MatMult(S_j,pcis->vec2_B,pcis->vec1_B);CHKERRQ(ierr); 5685fa34dd3eSStefano Zampini if (zerodiag) { 5686fa34dd3eSStefano Zampini ierr = VecGetArray(dummy_vec,&array);CHKERRQ(ierr); 5687*4f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) array[i] = p0_check[i]; 5688fa34dd3eSStefano Zampini ierr = VecRestoreArray(dummy_vec,&array);CHKERRQ(ierr); 5689fa34dd3eSStefano Zampini ierr = MatMultTransposeAdd(B0_B,dummy_vec,pcis->vec1_B,pcis->vec1_B);CHKERRQ(ierr); 5690fa34dd3eSStefano Zampini } 5691fa34dd3eSStefano Zampini /* sum on primal nodes the local contributions */ 5692fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcis->N_to_B,pcis->vec1_B,pcis->vec1_N,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 5693fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcis->N_to_B,pcis->vec1_B,pcis->vec1_N,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 5694fa34dd3eSStefano Zampini ierr = VecGetArray(pcis->vec1_N,&array);CHKERRQ(ierr); 5695fa34dd3eSStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array2);CHKERRQ(ierr); 5696fa34dd3eSStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) array2[i] = array[pcbddc->local_primal_ref_node[i]]; 5697fa34dd3eSStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array2);CHKERRQ(ierr); 5698fa34dd3eSStefano Zampini ierr = VecRestoreArray(pcis->vec1_N,&array);CHKERRQ(ierr); 5699fa34dd3eSStefano Zampini ierr = VecSet(pcbddc->coarse_vec,0.);CHKERRQ(ierr); 5700fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 5701fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 5702fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,pcbddc->vec1_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 5703fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,pcbddc->vec1_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 5704fa34dd3eSStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 5705fa34dd3eSStefano Zampini /* scale primal nodes (BDDC sums contibutions) */ 5706fa34dd3eSStefano Zampini ierr = VecPointwiseMult(pcbddc->vec1_P,vec_scale_P,pcbddc->vec1_P);CHKERRQ(ierr); 5707fa34dd3eSStefano Zampini ierr = VecSetValues(pcis->vec1_N,pcbddc->local_primal_size,pcbddc->local_primal_ref_node,array,INSERT_VALUES);CHKERRQ(ierr); 5708fa34dd3eSStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 5709fa34dd3eSStefano Zampini ierr = VecAssemblyBegin(pcis->vec1_N);CHKERRQ(ierr); 5710fa34dd3eSStefano Zampini ierr = VecAssemblyEnd(pcis->vec1_N);CHKERRQ(ierr); 5711fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcis->N_to_B,pcis->vec1_N,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 5712fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcis->N_to_B,pcis->vec1_N,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 5713fa34dd3eSStefano Zampini /* global: \widetilde{B0}_B w_\Gamma */ 5714fa34dd3eSStefano Zampini if (zerodiag) { 5715fa34dd3eSStefano Zampini ierr = MatMult(B0_B,pcis->vec2_B,dummy_vec);CHKERRQ(ierr); 5716fa34dd3eSStefano Zampini ierr = VecGetArray(dummy_vec,&array);CHKERRQ(ierr); 5717*4f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) pcbddc->benign_p0[i] = array[i]; 5718fa34dd3eSStefano Zampini ierr = VecRestoreArray(dummy_vec,&array);CHKERRQ(ierr); 5719fa34dd3eSStefano Zampini } 5720fa34dd3eSStefano Zampini /* BDDC */ 5721fa34dd3eSStefano Zampini ierr = VecSet(pcis->vec1_D,0.);CHKERRQ(ierr); 5722fa34dd3eSStefano Zampini ierr = PCBDDCApplyInterfacePreconditioner(pc,PETSC_FALSE);CHKERRQ(ierr); 5723fa34dd3eSStefano Zampini 5724fa34dd3eSStefano Zampini ierr = VecCopy(pcis->vec1_B,pcis->vec2_B);CHKERRQ(ierr); 5725fa34dd3eSStefano Zampini ierr = VecAXPY(pcis->vec1_B,-1.0,vec_check_B);CHKERRQ(ierr); 5726fa34dd3eSStefano Zampini ierr = VecNorm(pcis->vec1_B,NORM_INFINITY,&norm);CHKERRQ(ierr); 5727fa34dd3eSStefano Zampini PetscPrintf(PETSC_COMM_SELF,"[%d] BDDC local error is %1.4e\n",PetscGlobalRank,norm); 5728*4f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 5729*4f1b2e48SStefano Zampini PetscPrintf(PETSC_COMM_SELF,"[%d] BDDC p0[%d] error is %1.4e\n",PetscGlobalRank,i,PetscAbsScalar(pcbddc->benign_p0[i]-p0_check[i])); 5730fa34dd3eSStefano Zampini } 5731*4f1b2e48SStefano Zampini ierr = PetscFree(p0_check);CHKERRQ(ierr); 5732fa34dd3eSStefano Zampini ierr = VecDestroy(&vec_scale_P);CHKERRQ(ierr); 5733fa34dd3eSStefano Zampini ierr = VecDestroy(&vec_check_B);CHKERRQ(ierr); 5734fa34dd3eSStefano Zampini ierr = VecDestroy(&dummy_vec);CHKERRQ(ierr); 5735fa34dd3eSStefano Zampini ierr = MatDestroy(&S_j);CHKERRQ(ierr); 5736fa34dd3eSStefano Zampini ierr = MatDestroy(&B0_B);CHKERRQ(ierr); 5737fa34dd3eSStefano Zampini } 5738fa34dd3eSStefano Zampini PetscFunctionReturn(0); 5739fa34dd3eSStefano Zampini } 5740