11cf9b237SStefano Zampini #include <../src/mat/impls/aij/seq/aij.h> 2ab5c6b0cSJed Brown #include <../src/ksp/pc/impls/bddc/bddc.h> 3ab5c6b0cSJed Brown #include <../src/ksp/pc/impls/bddc/bddcprivate.h> 4674ae819SStefano Zampini #include <petscblaslapack.h> 5674ae819SStefano Zampini 6906d46d4SStefano Zampini static PetscErrorCode PCBDDCMatMultTranspose_Private(Mat A, Vec x, Vec y); 7906d46d4SStefano Zampini static PetscErrorCode PCBDDCMatMult_Private(Mat A, Vec x, Vec y); 8906d46d4SStefano Zampini 91cf9b237SStefano Zampini /* TODO: add reuse flag */ 101cf9b237SStefano Zampini #undef __FUNCT__ 111cf9b237SStefano Zampini #define __FUNCT__ "MatSeqAIJCompress" 121cf9b237SStefano Zampini PetscErrorCode MatSeqAIJCompress(Mat A, Mat *B) 131cf9b237SStefano Zampini { 141cf9b237SStefano Zampini Mat Bt; 151cf9b237SStefano Zampini PetscScalar *a,*bdata; 161cf9b237SStefano Zampini const PetscInt *ii,*ij; 171cf9b237SStefano Zampini PetscInt m,n,i,nnz,*bii,*bij; 181cf9b237SStefano Zampini PetscBool flg_row; 191cf9b237SStefano Zampini PetscErrorCode ierr; 201cf9b237SStefano Zampini 211cf9b237SStefano Zampini PetscFunctionBegin; 221cf9b237SStefano Zampini ierr = MatGetSize(A,&n,&m);CHKERRQ(ierr); 231cf9b237SStefano Zampini ierr = MatGetRowIJ(A,0,PETSC_FALSE,PETSC_FALSE,&n,&ii,&ij,&flg_row);CHKERRQ(ierr); 241cf9b237SStefano Zampini ierr = MatSeqAIJGetArray(A,&a);CHKERRQ(ierr); 251cf9b237SStefano Zampini nnz = n; 261cf9b237SStefano Zampini for (i=0;i<ii[n];i++) { 271cf9b237SStefano Zampini if (PetscLikely(PetscAbsScalar(a[i]) > PETSC_SMALL)) nnz++; 281cf9b237SStefano Zampini } 291cf9b237SStefano Zampini ierr = PetscMalloc1(n+1,&bii);CHKERRQ(ierr); 301cf9b237SStefano Zampini ierr = PetscMalloc1(nnz,&bij);CHKERRQ(ierr); 311cf9b237SStefano Zampini ierr = PetscMalloc1(nnz,&bdata);CHKERRQ(ierr); 321cf9b237SStefano Zampini nnz = 0; 331cf9b237SStefano Zampini bii[0] = 0; 341cf9b237SStefano Zampini for (i=0;i<n;i++) { 351cf9b237SStefano Zampini PetscInt j; 361cf9b237SStefano Zampini for (j=ii[i];j<ii[i+1];j++) { 371cf9b237SStefano Zampini PetscScalar entry = a[j]; 381cf9b237SStefano Zampini if (PetscLikely(PetscAbsScalar(entry) > PETSC_SMALL) || ij[j] == i) { 391cf9b237SStefano Zampini bij[nnz] = ij[j]; 401cf9b237SStefano Zampini bdata[nnz] = entry; 411cf9b237SStefano Zampini nnz++; 421cf9b237SStefano Zampini } 431cf9b237SStefano Zampini } 441cf9b237SStefano Zampini bii[i+1] = nnz; 451cf9b237SStefano Zampini } 461cf9b237SStefano Zampini ierr = MatSeqAIJRestoreArray(A,&a);CHKERRQ(ierr); 471cf9b237SStefano Zampini ierr = MatCreateSeqAIJWithArrays(PetscObjectComm((PetscObject)A),n,m,bii,bij,bdata,&Bt);CHKERRQ(ierr); 481cf9b237SStefano Zampini ierr = MatRestoreRowIJ(A,0,PETSC_FALSE,PETSC_FALSE,&n,&ii,&ij,&flg_row);CHKERRQ(ierr); 491cf9b237SStefano Zampini { 501cf9b237SStefano Zampini Mat_SeqAIJ *b = (Mat_SeqAIJ*)(Bt->data); 511cf9b237SStefano Zampini b->free_a = PETSC_TRUE; 521cf9b237SStefano Zampini b->free_ij = PETSC_TRUE; 531cf9b237SStefano Zampini } 541cf9b237SStefano Zampini *B = Bt; 551cf9b237SStefano Zampini PetscFunctionReturn(0); 561cf9b237SStefano Zampini } 571cf9b237SStefano Zampini 58674ae819SStefano Zampini #undef __FUNCT__ 594f1b2e48SStefano Zampini #define __FUNCT__ "MatDetectDisconnectedComponents" 604f1b2e48SStefano Zampini PetscErrorCode MatDetectDisconnectedComponents(Mat A, PetscBool filter, PetscInt *ncc, IS* cc[]) 614f1b2e48SStefano Zampini { 624f1b2e48SStefano Zampini Mat B; 634f1b2e48SStefano Zampini IS is_dummy,*cc_n; 644f1b2e48SStefano Zampini ISLocalToGlobalMapping l2gmap_dummy; 654f1b2e48SStefano Zampini PCBDDCGraph graph; 664f1b2e48SStefano Zampini PetscInt i,n; 674f1b2e48SStefano Zampini PetscInt *xadj,*adjncy; 684f1b2e48SStefano Zampini PetscInt *xadj_filtered,*adjncy_filtered; 694f1b2e48SStefano Zampini PetscBool flg_row,isseqaij; 704f1b2e48SStefano Zampini PetscErrorCode ierr; 714f1b2e48SStefano Zampini 724f1b2e48SStefano Zampini PetscFunctionBegin; 734f1b2e48SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)A,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 744f1b2e48SStefano Zampini if (!isseqaij && filter) { 751cf9b237SStefano Zampini PetscBool isseqdense; 761cf9b237SStefano Zampini 771cf9b237SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)A,MATSEQDENSE,&isseqdense);CHKERRQ(ierr); 781cf9b237SStefano Zampini if (!isseqdense) { 794f1b2e48SStefano Zampini ierr = MatConvert(A,MATSEQAIJ,MAT_INITIAL_MATRIX,&B);CHKERRQ(ierr); 801cf9b237SStefano Zampini } else { /* TODO: rectangular case and LDA */ 811cf9b237SStefano Zampini PetscScalar *array; 821cf9b237SStefano Zampini PetscReal chop=1.e-6; 831cf9b237SStefano Zampini 841cf9b237SStefano Zampini ierr = MatDuplicate(A,MAT_COPY_VALUES,&B);CHKERRQ(ierr); 851cf9b237SStefano Zampini ierr = MatDenseGetArray(B,&array);CHKERRQ(ierr); 861cf9b237SStefano Zampini ierr = MatGetSize(B,&n,NULL);CHKERRQ(ierr); 871cf9b237SStefano Zampini for (i=0;i<n;i++) { 881cf9b237SStefano Zampini PetscInt j; 891cf9b237SStefano Zampini for (j=i+1;j<n;j++) { 901cf9b237SStefano Zampini PetscReal thresh = chop*(PetscAbsScalar(array[i*(n+1)])+PetscAbsScalar(array[j*(n+1)])); 911cf9b237SStefano Zampini if (PetscAbsScalar(array[i*n+j]) < thresh) array[i*n+j] = 0.; 921cf9b237SStefano Zampini if (PetscAbsScalar(array[j*n+i]) < thresh) array[j*n+i] = 0.; 931cf9b237SStefano Zampini } 941cf9b237SStefano Zampini } 951cf9b237SStefano Zampini ierr = MatDenseRestoreArray(B,&array);CHKERRQ(ierr); 961cf9b237SStefano Zampini ierr = MatConvert(B,MATSEQAIJ,MAT_REUSE_MATRIX,&B);CHKERRQ(ierr); 971cf9b237SStefano Zampini } 984f1b2e48SStefano Zampini } else { 994f1b2e48SStefano Zampini B = A; 1004f1b2e48SStefano Zampini } 1014f1b2e48SStefano Zampini ierr = MatGetRowIJ(B,0,PETSC_TRUE,PETSC_FALSE,&n,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 1024f1b2e48SStefano Zampini 1034f1b2e48SStefano Zampini /* if filter is true, then removes entries lower than PETSC_SMALL in magnitude */ 1044f1b2e48SStefano Zampini if (filter) { 1054f1b2e48SStefano Zampini PetscScalar *data; 1064f1b2e48SStefano Zampini PetscInt j,cum; 1074f1b2e48SStefano Zampini 1084f1b2e48SStefano Zampini ierr = PetscCalloc2(n+1,&xadj_filtered,xadj[n],&adjncy_filtered);CHKERRQ(ierr); 1094f1b2e48SStefano Zampini ierr = MatSeqAIJGetArray(B,&data);CHKERRQ(ierr); 1104f1b2e48SStefano Zampini cum = 0; 1114f1b2e48SStefano Zampini for (i=0;i<n;i++) { 1124f1b2e48SStefano Zampini PetscInt t; 1134f1b2e48SStefano Zampini 1144f1b2e48SStefano Zampini for (j=xadj[i];j<xadj[i+1];j++) { 1154f1b2e48SStefano Zampini if (PetscUnlikely(PetscAbsScalar(data[j]) < PETSC_SMALL)) { 1164f1b2e48SStefano Zampini continue; 1174f1b2e48SStefano Zampini } 1184f1b2e48SStefano Zampini adjncy_filtered[cum+xadj_filtered[i]++] = adjncy[j]; 1194f1b2e48SStefano Zampini } 1204f1b2e48SStefano Zampini t = xadj_filtered[i]; 1214f1b2e48SStefano Zampini xadj_filtered[i] = cum; 1224f1b2e48SStefano Zampini cum += t; 1234f1b2e48SStefano Zampini } 1244f1b2e48SStefano Zampini ierr = MatSeqAIJRestoreArray(B,&data);CHKERRQ(ierr); 1254f1b2e48SStefano Zampini } else { 1264f1b2e48SStefano Zampini xadj_filtered = NULL; 1274f1b2e48SStefano Zampini adjncy_filtered = NULL; 1284f1b2e48SStefano Zampini } 1294f1b2e48SStefano Zampini 1304f1b2e48SStefano Zampini /* compute local connected components using PCBDDCGraph */ 1314f1b2e48SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,n,0,1,&is_dummy);CHKERRQ(ierr); 1324f1b2e48SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(is_dummy,&l2gmap_dummy);CHKERRQ(ierr); 1334f1b2e48SStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 1344f1b2e48SStefano Zampini ierr = PCBDDCGraphCreate(&graph);CHKERRQ(ierr); 1354f1b2e48SStefano Zampini ierr = PCBDDCGraphInit(graph,l2gmap_dummy,n);CHKERRQ(ierr); 1364f1b2e48SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&l2gmap_dummy);CHKERRQ(ierr); 1374f1b2e48SStefano Zampini if (xadj_filtered) { 1384f1b2e48SStefano Zampini graph->xadj = xadj_filtered; 1394f1b2e48SStefano Zampini graph->adjncy = adjncy_filtered; 1404f1b2e48SStefano Zampini } else { 1414f1b2e48SStefano Zampini graph->xadj = xadj; 1424f1b2e48SStefano Zampini graph->adjncy = adjncy; 1434f1b2e48SStefano Zampini } 1444f1b2e48SStefano Zampini ierr = PCBDDCGraphSetUp(graph,1,NULL,NULL,0,NULL,NULL);CHKERRQ(ierr); 1454f1b2e48SStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(graph);CHKERRQ(ierr); 1464f1b2e48SStefano Zampini /* partial clean up */ 1474f1b2e48SStefano Zampini ierr = PetscFree2(xadj_filtered,adjncy_filtered);CHKERRQ(ierr); 1484f1b2e48SStefano Zampini ierr = MatRestoreRowIJ(B,0,PETSC_TRUE,PETSC_FALSE,&n,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 1491cf9b237SStefano Zampini if (A != B) { 1504f1b2e48SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 1514f1b2e48SStefano Zampini } 1524f1b2e48SStefano Zampini 1534f1b2e48SStefano Zampini /* get back data */ 1541cf9b237SStefano Zampini if (ncc) *ncc = graph->ncc; 1551cf9b237SStefano Zampini if (cc) { 1564f1b2e48SStefano Zampini ierr = PetscMalloc1(graph->ncc,&cc_n);CHKERRQ(ierr); 1574f1b2e48SStefano Zampini for (i=0;i<graph->ncc;i++) { 1584f1b2e48SStefano 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); 1594f1b2e48SStefano Zampini } 1604f1b2e48SStefano Zampini *cc = cc_n; 1611cf9b237SStefano Zampini } 1624f1b2e48SStefano Zampini /* clean up graph */ 1634f1b2e48SStefano Zampini graph->xadj = 0; 1644f1b2e48SStefano Zampini graph->adjncy = 0; 1654f1b2e48SStefano Zampini ierr = PCBDDCGraphDestroy(&graph);CHKERRQ(ierr); 1664f1b2e48SStefano Zampini PetscFunctionReturn(0); 1674f1b2e48SStefano Zampini } 1684f1b2e48SStefano Zampini 1694f1b2e48SStefano Zampini #undef __FUNCT__ 1705408967cSStefano Zampini #define __FUNCT__ "PCBDDCBenignCheck" 1715408967cSStefano Zampini PetscErrorCode PCBDDCBenignCheck(PC pc, IS zerodiag) 1725408967cSStefano Zampini { 1735408967cSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 1745408967cSStefano Zampini PC_IS* pcis = (PC_IS*)(pc->data); 175dee84bffSStefano Zampini IS dirIS = NULL; 1764f1b2e48SStefano Zampini PetscInt i; 1775408967cSStefano Zampini PetscErrorCode ierr; 1785408967cSStefano Zampini 1795408967cSStefano Zampini PetscFunctionBegin; 180dee84bffSStefano Zampini ierr = PCBDDCGraphGetDirichletDofs(pcbddc->mat_graph,&dirIS);CHKERRQ(ierr); 1815408967cSStefano Zampini if (zerodiag) { 1825408967cSStefano Zampini Mat A; 1835408967cSStefano Zampini Vec vec3_N; 1845408967cSStefano Zampini PetscScalar *vals; 1855408967cSStefano Zampini const PetscInt *idxs; 1864f1b2e48SStefano Zampini PetscInt nz; 1875408967cSStefano Zampini 1885408967cSStefano Zampini /* p0 */ 1895408967cSStefano Zampini ierr = VecSet(pcis->vec1_N,0.);CHKERRQ(ierr); 1905408967cSStefano Zampini ierr = PetscMalloc1(pcis->n,&vals);CHKERRQ(ierr); 1915408967cSStefano Zampini ierr = ISGetLocalSize(zerodiag,&nz);CHKERRQ(ierr); 1925408967cSStefano Zampini ierr = ISGetIndices(zerodiag,&idxs);CHKERRQ(ierr); 1934f1b2e48SStefano Zampini for (i=0;i<nz;i++) vals[i] = 1.; 1945408967cSStefano Zampini ierr = VecSetValues(pcis->vec1_N,nz,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 1955408967cSStefano Zampini ierr = VecAssemblyBegin(pcis->vec1_N);CHKERRQ(ierr); 1965408967cSStefano Zampini ierr = VecAssemblyEnd(pcis->vec1_N);CHKERRQ(ierr); 1975408967cSStefano Zampini /* v_I */ 1985408967cSStefano Zampini ierr = VecSetRandom(pcis->vec2_N,NULL);CHKERRQ(ierr); 1995408967cSStefano Zampini for (i=0;i<nz;i++) vals[i] = 0.; 2005408967cSStefano Zampini ierr = VecSetValues(pcis->vec2_N,nz,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 2015408967cSStefano Zampini ierr = ISRestoreIndices(zerodiag,&idxs);CHKERRQ(ierr); 2025408967cSStefano Zampini ierr = ISGetIndices(pcis->is_B_local,&idxs);CHKERRQ(ierr); 2035408967cSStefano Zampini for (i=0;i<pcis->n_B;i++) vals[i] = 0.; 2045408967cSStefano Zampini ierr = VecSetValues(pcis->vec2_N,pcis->n_B,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 2055408967cSStefano Zampini ierr = ISRestoreIndices(pcis->is_B_local,&idxs);CHKERRQ(ierr); 2065408967cSStefano Zampini if (dirIS) { 2075408967cSStefano Zampini PetscInt n; 2085408967cSStefano Zampini 2095408967cSStefano Zampini ierr = ISGetLocalSize(dirIS,&n);CHKERRQ(ierr); 2105408967cSStefano Zampini ierr = ISGetIndices(dirIS,&idxs);CHKERRQ(ierr); 2115408967cSStefano Zampini for (i=0;i<n;i++) vals[i] = 0.; 2125408967cSStefano Zampini ierr = VecSetValues(pcis->vec2_N,n,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 2135408967cSStefano Zampini ierr = ISRestoreIndices(dirIS,&idxs);CHKERRQ(ierr); 2145408967cSStefano Zampini } 2155408967cSStefano Zampini ierr = VecAssemblyBegin(pcis->vec2_N);CHKERRQ(ierr); 2165408967cSStefano Zampini ierr = VecAssemblyEnd(pcis->vec2_N);CHKERRQ(ierr); 2175408967cSStefano Zampini ierr = VecDuplicate(pcis->vec1_N,&vec3_N);CHKERRQ(ierr); 2185408967cSStefano Zampini ierr = VecSet(vec3_N,0.);CHKERRQ(ierr); 2195408967cSStefano Zampini ierr = MatISGetLocalMat(pc->mat,&A);CHKERRQ(ierr); 2205408967cSStefano Zampini ierr = MatMult(A,pcis->vec1_N,vec3_N);CHKERRQ(ierr); 2215408967cSStefano Zampini ierr = VecDot(vec3_N,pcis->vec2_N,&vals[0]);CHKERRQ(ierr); 2224f1b2e48SStefano Zampini if (PetscAbsScalar(vals[0]) > PETSC_SMALL) { 2235408967cSStefano 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])); 2245408967cSStefano Zampini } 2255408967cSStefano Zampini ierr = PetscFree(vals);CHKERRQ(ierr); 2265408967cSStefano Zampini ierr = VecDestroy(&vec3_N);CHKERRQ(ierr); 2275408967cSStefano Zampini } 228dee84bffSStefano Zampini ierr = ISDestroy(&dirIS);CHKERRQ(ierr); 2295408967cSStefano Zampini 2305408967cSStefano Zampini /* check PCBDDCBenignGetOrSetP0 */ 2315408967cSStefano Zampini ierr = VecSetRandom(pcis->vec1_global,NULL);CHKERRQ(ierr); 2324f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) pcbddc->benign_p0[i] = -PetscGlobalRank-i; 2335408967cSStefano Zampini ierr = PCBDDCBenignGetOrSetP0(pc,pcis->vec1_global,PETSC_FALSE);CHKERRQ(ierr); 2344f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) pcbddc->benign_p0[i] = 1; 2355408967cSStefano Zampini ierr = PCBDDCBenignGetOrSetP0(pc,pcis->vec1_global,PETSC_TRUE);CHKERRQ(ierr); 2364f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 2374f1b2e48SStefano Zampini if ((PetscInt)PetscRealPart(pcbddc->benign_p0[i]) != -PetscGlobalRank-i) { 2384f1b2e48SStefano 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); 2394f1b2e48SStefano Zampini } 2405408967cSStefano Zampini } 2415408967cSStefano Zampini PetscFunctionReturn(0); 2425408967cSStefano Zampini } 2435408967cSStefano Zampini 2445408967cSStefano Zampini #undef __FUNCT__ 245339f8db1SStefano Zampini #define __FUNCT__ "PCBDDCBenignDetectSaddlePoint" 246339f8db1SStefano Zampini PetscErrorCode PCBDDCBenignDetectSaddlePoint(PC pc, IS *zerodiaglocal) 247339f8db1SStefano Zampini { 248339f8db1SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 2494f1b2e48SStefano Zampini IS pressures,zerodiag,*zerodiag_subs; 250b0f5fe93SStefano Zampini PetscInt nz,n; 2514f1b2e48SStefano Zampini PetscBool sorted,have_null,has_null_pressures,recompute_zerodiag; 252339f8db1SStefano Zampini PetscErrorCode ierr; 253339f8db1SStefano Zampini 254339f8db1SStefano Zampini PetscFunctionBegin; 2559f47a83aSStefano Zampini ierr = PetscSFDestroy(&pcbddc->benign_sf);CHKERRQ(ierr); 2569f47a83aSStefano Zampini ierr = MatDestroy(&pcbddc->benign_B0);CHKERRQ(ierr); 257339f8db1SStefano Zampini ierr = MatDestroy(&pcbddc->benign_original_mat);CHKERRQ(ierr); 2584f1b2e48SStefano Zampini /* if a local info on dofs is present, assumes the last field is represented by "pressures" 2594f1b2e48SStefano Zampini otherwise, it uses only zerodiagonal dofs (ok if the pressure block is all zero; it could fail if it is not) 2604f1b2e48SStefano Zampini Checks if all the pressure dofs in each subdomain have a zero diagonal 2614f1b2e48SStefano Zampini If not, a change of basis on pressures is not needed 2624f1b2e48SStefano Zampini since the local Schur complements are SPD 2634f1b2e48SStefano Zampini */ 2644f1b2e48SStefano Zampini has_null_pressures = PETSC_TRUE; 2654f1b2e48SStefano Zampini have_null = PETSC_TRUE; 26640fa8d13SStefano Zampini if (pcbddc->n_ISForDofsLocal) { 2674f1b2e48SStefano Zampini PetscInt npl,*idxs,p = pcbddc->n_ISForDofsLocal-1; 2684f1b2e48SStefano Zampini 2694f1b2e48SStefano Zampini /* Dofs splitting for BDDC cannot have PETSC_COMM_SELF, so create a sequential IS */ 2704f1b2e48SStefano Zampini ierr = ISGetLocalSize(pcbddc->ISForDofsLocal[p],&npl);CHKERRQ(ierr); 2714f1b2e48SStefano Zampini ierr = ISGetIndices(pcbddc->ISForDofsLocal[p],(const PetscInt**)&idxs);CHKERRQ(ierr); 2724f1b2e48SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,npl,idxs,PETSC_COPY_VALUES,&pressures);CHKERRQ(ierr); 273ba14f8e3SStefano Zampini ierr = ISRestoreIndices(pcbddc->ISForDofsLocal[p],(const PetscInt**)&idxs);CHKERRQ(ierr); 27440fa8d13SStefano Zampini ierr = ISSorted(pressures,&sorted);CHKERRQ(ierr); 27540fa8d13SStefano Zampini if (!sorted) { 27640fa8d13SStefano Zampini ierr = ISSort(pressures);CHKERRQ(ierr); 27740fa8d13SStefano Zampini } 27840fa8d13SStefano Zampini } else { 27940fa8d13SStefano Zampini pressures = NULL; 28040fa8d13SStefano Zampini } 28197d764eeSStefano Zampini /* pcis has not been setup yet, so get the local size from the subdomain matrix */ 28297d764eeSStefano Zampini ierr = MatGetLocalSize(pcbddc->local_mat,&n,NULL);CHKERRQ(ierr); 2834f1b2e48SStefano Zampini /* TODO: add check for shared dofs and raise error */ 28497d764eeSStefano Zampini ierr = MatFindZeroDiagonals(pcbddc->local_mat,&zerodiag);CHKERRQ(ierr); 285339f8db1SStefano Zampini ierr = ISSorted(zerodiag,&sorted);CHKERRQ(ierr); 286339f8db1SStefano Zampini if (!sorted) { 287339f8db1SStefano Zampini ierr = ISSort(zerodiag);CHKERRQ(ierr); 288339f8db1SStefano Zampini } 289339f8db1SStefano Zampini ierr = ISGetLocalSize(zerodiag,&nz);CHKERRQ(ierr); 2904f1b2e48SStefano Zampini if (!nz) { 2914f1b2e48SStefano Zampini if (n) have_null = PETSC_FALSE; 2924f1b2e48SStefano Zampini has_null_pressures = PETSC_FALSE; 29340fa8d13SStefano Zampini ierr = ISDestroy(&zerodiag);CHKERRQ(ierr); 29440fa8d13SStefano Zampini } 2954f1b2e48SStefano Zampini recompute_zerodiag = PETSC_FALSE; 2964f1b2e48SStefano Zampini /* in case disconnected subdomains info is present, split the pressures accordingly (otherwise the benign trick could fail) */ 2974f1b2e48SStefano Zampini zerodiag_subs = NULL; 2984f1b2e48SStefano Zampini pcbddc->benign_n = 0; 2994f1b2e48SStefano Zampini if (has_null_pressures) { 3004f1b2e48SStefano Zampini IS *subs; 3014f1b2e48SStefano Zampini PetscInt nsubs,i; 3024f1b2e48SStefano Zampini 3034f1b2e48SStefano Zampini subs = pcbddc->local_subs; 3044f1b2e48SStefano Zampini nsubs = pcbddc->n_local_subs; 3054f1b2e48SStefano Zampini if (nsubs > 1) { 3064f1b2e48SStefano Zampini ierr = PetscCalloc1(nsubs,&zerodiag_subs);CHKERRQ(ierr); 3074f1b2e48SStefano Zampini for (i=0;i<nsubs;i++) { 3084f1b2e48SStefano Zampini ISLocalToGlobalMapping l2g; 3094f1b2e48SStefano Zampini IS t_zerodiag_subs; 3104f1b2e48SStefano Zampini PetscInt nl; 3114f1b2e48SStefano Zampini 3124f1b2e48SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(subs[i],&l2g);CHKERRQ(ierr); 3134f1b2e48SStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(l2g,IS_GTOLM_DROP,zerodiag,&t_zerodiag_subs);CHKERRQ(ierr); 3144f1b2e48SStefano Zampini ierr = ISGetLocalSize(t_zerodiag_subs,&nl);CHKERRQ(ierr); 3154f1b2e48SStefano Zampini if (nl) { 3164f1b2e48SStefano Zampini PetscBool valid = PETSC_TRUE; 3174f1b2e48SStefano Zampini 3184f1b2e48SStefano Zampini if (pressures) { 3194f1b2e48SStefano Zampini IS t_pressure_subs; 3204f1b2e48SStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(l2g,IS_GTOLM_DROP,pressures,&t_pressure_subs);CHKERRQ(ierr); 3214f1b2e48SStefano Zampini ierr = ISEqual(t_pressure_subs,t_zerodiag_subs,&valid);CHKERRQ(ierr); 3224f1b2e48SStefano Zampini ierr = ISDestroy(&t_pressure_subs);CHKERRQ(ierr); 3234f1b2e48SStefano Zampini } 3244f1b2e48SStefano Zampini if (valid) { 3254f1b2e48SStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(l2g,t_zerodiag_subs,&zerodiag_subs[pcbddc->benign_n]);CHKERRQ(ierr); 3264f1b2e48SStefano Zampini pcbddc->benign_n++; 3274f1b2e48SStefano Zampini } else { 3284f1b2e48SStefano Zampini recompute_zerodiag = PETSC_TRUE; 3294f1b2e48SStefano Zampini } 3304f1b2e48SStefano Zampini } 3314f1b2e48SStefano Zampini ierr = ISDestroy(&t_zerodiag_subs);CHKERRQ(ierr); 3324f1b2e48SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&l2g);CHKERRQ(ierr); 3334f1b2e48SStefano Zampini } 3344f1b2e48SStefano Zampini } else { /* there's just one subdomain (or zero if they have not been detected */ 3354f1b2e48SStefano Zampini PetscBool valid = PETSC_TRUE; 3364f1b2e48SStefano Zampini if (pressures) { 3374f1b2e48SStefano Zampini ierr = ISEqual(pressures,zerodiag,&valid);CHKERRQ(ierr); 3384f1b2e48SStefano Zampini } 3394f1b2e48SStefano Zampini if (valid) { 3404f1b2e48SStefano Zampini pcbddc->benign_n = 1; 341*ca92afb2SStefano Zampini ierr = PetscMalloc1(pcbddc->benign_n,&zerodiag_subs);CHKERRQ(ierr); 3424f1b2e48SStefano Zampini ierr = PetscObjectReference((PetscObject)zerodiag);CHKERRQ(ierr); 3434f1b2e48SStefano Zampini zerodiag_subs[0] = zerodiag; 3444f1b2e48SStefano Zampini } 3454f1b2e48SStefano Zampini } 3464f1b2e48SStefano Zampini } 3474f1b2e48SStefano Zampini 3484f1b2e48SStefano Zampini if (!pcbddc->benign_n) { 3494f1b2e48SStefano Zampini ierr = ISDestroy(&zerodiag);CHKERRQ(ierr); 3504f1b2e48SStefano Zampini recompute_zerodiag = PETSC_FALSE; 3514f1b2e48SStefano Zampini has_null_pressures = PETSC_FALSE; 3524f1b2e48SStefano Zampini have_null = PETSC_FALSE; 3534f1b2e48SStefano Zampini } 3544f1b2e48SStefano Zampini 3554f1b2e48SStefano Zampini /* final check for null pressures */ 3564f1b2e48SStefano Zampini if (zerodiag && pressures) { 3574f1b2e48SStefano Zampini PetscInt nz,np; 3584f1b2e48SStefano Zampini ierr = ISGetLocalSize(zerodiag,&nz);CHKERRQ(ierr); 3594f1b2e48SStefano Zampini ierr = ISGetLocalSize(pressures,&np);CHKERRQ(ierr); 3604f1b2e48SStefano Zampini if (nz != np) have_null = PETSC_FALSE; 3614f1b2e48SStefano Zampini } 3624f1b2e48SStefano Zampini 3634f1b2e48SStefano Zampini if (recompute_zerodiag) { 3644f1b2e48SStefano Zampini ierr = ISDestroy(&zerodiag);CHKERRQ(ierr); 3654f1b2e48SStefano Zampini if (pcbddc->benign_n == 1) { 3664f1b2e48SStefano Zampini ierr = PetscObjectReference((PetscObject)zerodiag_subs[0]);CHKERRQ(ierr); 3674f1b2e48SStefano Zampini zerodiag = zerodiag_subs[0]; 3684f1b2e48SStefano Zampini } else { 3694f1b2e48SStefano Zampini PetscInt i,nzn,*new_idxs; 3704f1b2e48SStefano Zampini 3714f1b2e48SStefano Zampini nzn = 0; 3724f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 3734f1b2e48SStefano Zampini PetscInt ns; 3744f1b2e48SStefano Zampini ierr = ISGetLocalSize(zerodiag_subs[i],&ns);CHKERRQ(ierr); 3754f1b2e48SStefano Zampini nzn += ns; 3764f1b2e48SStefano Zampini } 3774f1b2e48SStefano Zampini ierr = PetscMalloc1(nzn,&new_idxs);CHKERRQ(ierr); 3784f1b2e48SStefano Zampini nzn = 0; 3794f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 3804f1b2e48SStefano Zampini PetscInt ns,*idxs; 3814f1b2e48SStefano Zampini ierr = ISGetLocalSize(zerodiag_subs[i],&ns);CHKERRQ(ierr); 3824f1b2e48SStefano Zampini ierr = ISGetIndices(zerodiag_subs[i],(const PetscInt**)&idxs);CHKERRQ(ierr); 3834f1b2e48SStefano Zampini ierr = PetscMemcpy(new_idxs+nzn,idxs,ns*sizeof(PetscInt));CHKERRQ(ierr); 3844f1b2e48SStefano Zampini ierr = ISRestoreIndices(zerodiag_subs[i],(const PetscInt**)&idxs);CHKERRQ(ierr); 3854f1b2e48SStefano Zampini nzn += ns; 3864f1b2e48SStefano Zampini } 3874f1b2e48SStefano Zampini ierr = PetscSortInt(nzn,new_idxs);CHKERRQ(ierr); 3884f1b2e48SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,nzn,new_idxs,PETSC_OWN_POINTER,&zerodiag);CHKERRQ(ierr); 3894f1b2e48SStefano Zampini } 3904f1b2e48SStefano Zampini have_null = PETSC_FALSE; 3914f1b2e48SStefano Zampini } 3924f1b2e48SStefano Zampini 3934f1b2e48SStefano Zampini if (has_null_pressures) { 3944f1b2e48SStefano Zampini IS zerodiagc; 3954f1b2e48SStefano Zampini const PetscInt *idxs,*idxsc; 3964f1b2e48SStefano Zampini PetscInt i,s,*nnz; 3974f1b2e48SStefano Zampini 3984f1b2e48SStefano Zampini ierr = ISGetLocalSize(zerodiag,&nz);CHKERRQ(ierr); 399339f8db1SStefano Zampini ierr = ISComplement(zerodiag,0,n,&zerodiagc);CHKERRQ(ierr); 400339f8db1SStefano Zampini ierr = ISGetIndices(zerodiagc,&idxsc);CHKERRQ(ierr); 401339f8db1SStefano Zampini /* local change of basis for pressures */ 402339f8db1SStefano Zampini ierr = MatDestroy(&pcbddc->benign_change);CHKERRQ(ierr); 40397d764eeSStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)pcbddc->local_mat),&pcbddc->benign_change);CHKERRQ(ierr); 404339f8db1SStefano Zampini ierr = MatSetType(pcbddc->benign_change,MATAIJ);CHKERRQ(ierr); 405339f8db1SStefano Zampini ierr = MatSetSizes(pcbddc->benign_change,n,n,PETSC_DECIDE,PETSC_DECIDE);CHKERRQ(ierr); 406339f8db1SStefano Zampini ierr = PetscMalloc1(n,&nnz);CHKERRQ(ierr); 4074f1b2e48SStefano Zampini for (i=0;i<n-nz;i++) nnz[idxsc[i]] = 1; /* identity on velocities plus pressure dofs for non-singular subdomains */ 4084f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 4094f1b2e48SStefano Zampini PetscInt nzs,j; 4104f1b2e48SStefano Zampini 4114f1b2e48SStefano Zampini ierr = ISGetLocalSize(zerodiag_subs[i],&nzs);CHKERRQ(ierr); 4124f1b2e48SStefano Zampini ierr = ISGetIndices(zerodiag_subs[i],&idxs);CHKERRQ(ierr); 4134f1b2e48SStefano Zampini for (j=0;j<nzs-1;j++) nnz[idxs[j]] = 2; /* change on pressures */ 4144f1b2e48SStefano Zampini nnz[idxs[nzs-1]] = nzs; /* last local pressure dof in subdomain */ 4154f1b2e48SStefano Zampini ierr = ISRestoreIndices(zerodiag_subs[i],&idxs);CHKERRQ(ierr); 4164f1b2e48SStefano Zampini } 417339f8db1SStefano Zampini ierr = MatSeqAIJSetPreallocation(pcbddc->benign_change,0,nnz);CHKERRQ(ierr); 418339f8db1SStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 419339f8db1SStefano Zampini /* set identity on velocities */ 420339f8db1SStefano Zampini for (i=0;i<n-nz;i++) { 421339f8db1SStefano Zampini ierr = MatSetValue(pcbddc->benign_change,idxsc[i],idxsc[i],1.,INSERT_VALUES);CHKERRQ(ierr); 422339f8db1SStefano Zampini } 4234f1b2e48SStefano Zampini ierr = ISRestoreIndices(zerodiagc,&idxsc);CHKERRQ(ierr); 4244f1b2e48SStefano Zampini ierr = ISDestroy(&zerodiagc);CHKERRQ(ierr); 4259f47a83aSStefano Zampini ierr = PetscFree3(pcbddc->benign_p0_lidx,pcbddc->benign_p0_gidx,pcbddc->benign_p0);CHKERRQ(ierr); 4264f1b2e48SStefano 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); 427339f8db1SStefano Zampini /* set change on pressures */ 4284f1b2e48SStefano Zampini for (s=0;s<pcbddc->benign_n;s++) { 4294f1b2e48SStefano Zampini PetscScalar *array; 4304f1b2e48SStefano Zampini PetscInt nzs; 4314f1b2e48SStefano Zampini 4324f1b2e48SStefano Zampini ierr = ISGetLocalSize(zerodiag_subs[s],&nzs);CHKERRQ(ierr); 4334f1b2e48SStefano Zampini ierr = ISGetIndices(zerodiag_subs[s],&idxs);CHKERRQ(ierr); 4344f1b2e48SStefano Zampini for (i=0;i<nzs-1;i++) { 435339f8db1SStefano Zampini PetscScalar vals[2]; 436339f8db1SStefano Zampini PetscInt cols[2]; 437339f8db1SStefano Zampini 438339f8db1SStefano Zampini cols[0] = idxs[i]; 4394f1b2e48SStefano Zampini cols[1] = idxs[nzs-1]; 440339f8db1SStefano Zampini vals[0] = 1.; 441b0f5fe93SStefano Zampini vals[1] = 1.; 4424f1b2e48SStefano Zampini ierr = MatSetValues(pcbddc->benign_change,1,cols,2,cols,vals,INSERT_VALUES);CHKERRQ(ierr); 443339f8db1SStefano Zampini } 4444f1b2e48SStefano Zampini ierr = PetscMalloc1(nzs,&array);CHKERRQ(ierr); 4454f1b2e48SStefano Zampini for (i=0;i<nzs-1;i++) array[i] = -1.; 4464f1b2e48SStefano Zampini array[nzs-1] = 1.; 4474f1b2e48SStefano Zampini ierr = MatSetValues(pcbddc->benign_change,1,idxs+nzs-1,nzs,idxs,array,INSERT_VALUES);CHKERRQ(ierr); 4484f1b2e48SStefano Zampini /* store local idxs for p0 */ 4494f1b2e48SStefano Zampini pcbddc->benign_p0_lidx[s] = idxs[nzs-1]; 4504f1b2e48SStefano Zampini ierr = ISRestoreIndices(zerodiag_subs[s],&idxs);CHKERRQ(ierr); 451339f8db1SStefano Zampini ierr = PetscFree(array);CHKERRQ(ierr); 4524f1b2e48SStefano Zampini } 453339f8db1SStefano Zampini ierr = MatAssemblyBegin(pcbddc->benign_change,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 454339f8db1SStefano Zampini ierr = MatAssemblyEnd(pcbddc->benign_change,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 455339f8db1SStefano Zampini /* TODO: need optimization? */ 45697d764eeSStefano Zampini ierr = MatPtAP(pcbddc->local_mat,pcbddc->benign_change,MAT_INITIAL_MATRIX,2.0,&pcbddc->benign_original_mat);CHKERRQ(ierr); 457339f8db1SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 45897d764eeSStefano Zampini ierr = MatSeqAIJCompress(pcbddc->benign_original_mat,&pcbddc->local_mat);CHKERRQ(ierr); 45997d764eeSStefano Zampini ierr = MatDestroy(&pcbddc->benign_original_mat);CHKERRQ(ierr); 4604f1b2e48SStefano Zampini /* store global idxs for p0 */ 4614f1b2e48SStefano Zampini ierr = ISLocalToGlobalMappingApply(pc->pmat->rmap->mapping,pcbddc->benign_n,pcbddc->benign_p0_lidx,pcbddc->benign_p0_gidx);CHKERRQ(ierr); 462339f8db1SStefano Zampini } 463*ca92afb2SStefano Zampini pcbddc->benign_zerodiag_subs = zerodiag_subs; 4644f1b2e48SStefano Zampini ierr = ISDestroy(&pressures);CHKERRQ(ierr); 465b0f5fe93SStefano Zampini 466b0f5fe93SStefano Zampini /* determines if the coarse solver will be singular or not */ 467b0f5fe93SStefano Zampini ierr = MPI_Allreduce(&have_null,&pcbddc->benign_null,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 468339f8db1SStefano Zampini *zerodiaglocal = zerodiag; 469339f8db1SStefano Zampini PetscFunctionReturn(0); 470339f8db1SStefano Zampini } 471339f8db1SStefano Zampini 472339f8db1SStefano Zampini #undef __FUNCT__ 473015636ebSStefano Zampini #define __FUNCT__ "PCBDDCBenignGetOrSetP0" 474015636ebSStefano Zampini PetscErrorCode PCBDDCBenignGetOrSetP0(PC pc, Vec v, PetscBool get) 475efc2fbd9SStefano Zampini { 476efc2fbd9SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 477efc2fbd9SStefano Zampini PetscErrorCode ierr; 478efc2fbd9SStefano Zampini 479efc2fbd9SStefano Zampini PetscFunctionBegin; 480efc2fbd9SStefano Zampini if (!pcbddc->benign_sf) { 481efc2fbd9SStefano Zampini ierr = PetscSFCreate(PetscObjectComm((PetscObject)pc),&pcbddc->benign_sf);CHKERRQ(ierr); 4824f1b2e48SStefano Zampini ierr = PetscSFSetGraphLayout(pcbddc->benign_sf,pc->pmat->rmap,pcbddc->benign_n,NULL,PETSC_OWN_POINTER,pcbddc->benign_p0_gidx);CHKERRQ(ierr); 483efc2fbd9SStefano Zampini } 484015636ebSStefano Zampini if (get) { /* use SF to get values */ 485efc2fbd9SStefano Zampini PetscScalar *array; 486efc2fbd9SStefano Zampini 487efc2fbd9SStefano Zampini ierr = VecGetArrayRead(v,(const PetscScalar**)&array);CHKERRQ(ierr); 4884f1b2e48SStefano Zampini ierr = PetscSFBcastBegin(pcbddc->benign_sf,MPIU_SCALAR,array,pcbddc->benign_p0);CHKERRQ(ierr); 4894f1b2e48SStefano Zampini ierr = PetscSFBcastEnd(pcbddc->benign_sf,MPIU_SCALAR,array,pcbddc->benign_p0);CHKERRQ(ierr); 490efc2fbd9SStefano Zampini ierr = VecRestoreArrayRead(v,(const PetscScalar**)&array);CHKERRQ(ierr); 4914f1b2e48SStefano Zampini } else { /* use VecSetValues (not scalable, I should try to find a better solution (defining a new MPI_OP for reduction) */ 4924f1b2e48SStefano Zampini ierr = VecSetValues(v,pcbddc->benign_n,pcbddc->benign_p0_gidx,pcbddc->benign_p0,INSERT_VALUES);CHKERRQ(ierr); 493efc2fbd9SStefano Zampini ierr = VecAssemblyBegin(v);CHKERRQ(ierr); 494efc2fbd9SStefano Zampini ierr = VecAssemblyEnd(v);CHKERRQ(ierr); 495efc2fbd9SStefano Zampini } 496efc2fbd9SStefano Zampini PetscFunctionReturn(0); 497efc2fbd9SStefano Zampini } 498efc2fbd9SStefano Zampini 499efc2fbd9SStefano Zampini #undef __FUNCT__ 500c263805aSStefano Zampini #define __FUNCT__ "PCBDDCBenignPopOrPushB0" 501c263805aSStefano Zampini PetscErrorCode PCBDDCBenignPopOrPushB0(PC pc, PetscBool pop) 502c263805aSStefano Zampini { 503c263805aSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 504c263805aSStefano Zampini PetscErrorCode ierr; 505c263805aSStefano Zampini 506c263805aSStefano Zampini PetscFunctionBegin; 507c263805aSStefano Zampini /* TODO: add error checking 508c263805aSStefano Zampini - avoid nested pop (or push) calls. 509c263805aSStefano Zampini - cannot push before pop. 5101c604dc7SStefano Zampini - cannot call this if pcbddc->local_mat is NULL 511c263805aSStefano Zampini */ 5124f1b2e48SStefano Zampini if (!pcbddc->benign_n) { 513efc2fbd9SStefano Zampini PetscFunctionReturn(0); 514efc2fbd9SStefano Zampini } 515c263805aSStefano Zampini if (pop) { 5164f1b2e48SStefano Zampini IS is_p0; 5174f1b2e48SStefano Zampini MatReuse reuse; 518c263805aSStefano Zampini 519c263805aSStefano Zampini /* extract B_0 */ 5204f1b2e48SStefano Zampini reuse = MAT_INITIAL_MATRIX; 5214f1b2e48SStefano Zampini if (pcbddc->benign_B0) { 5224f1b2e48SStefano Zampini reuse = MAT_REUSE_MATRIX; 5234f1b2e48SStefano Zampini } 5244f1b2e48SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,pcbddc->benign_n,pcbddc->benign_p0_lidx,PETSC_COPY_VALUES,&is_p0);CHKERRQ(ierr); 5254f1b2e48SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,is_p0,NULL,reuse,&pcbddc->benign_B0);CHKERRQ(ierr); 526c263805aSStefano Zampini /* remove rows and cols from local problem */ 527c263805aSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_KEEP_NONZERO_PATTERN,PETSC_TRUE);CHKERRQ(ierr); 52897d764eeSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_FALSE);CHKERRQ(ierr); 5294f1b2e48SStefano Zampini ierr = MatZeroRowsColumnsIS(pcbddc->local_mat,is_p0,1.0,NULL,NULL);CHKERRQ(ierr); 5304f1b2e48SStefano Zampini ierr = ISDestroy(&is_p0);CHKERRQ(ierr); 531c263805aSStefano Zampini } else { /* push */ 5324f1b2e48SStefano Zampini PetscInt i; 5334f1b2e48SStefano Zampini 5344f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 5354f1b2e48SStefano Zampini PetscScalar *B0_vals; 5364f1b2e48SStefano Zampini PetscInt *B0_cols,B0_ncol; 5374f1b2e48SStefano Zampini 5384f1b2e48SStefano Zampini ierr = MatGetRow(pcbddc->benign_B0,i,&B0_ncol,(const PetscInt**)&B0_cols,(const PetscScalar**)&B0_vals);CHKERRQ(ierr); 5394f1b2e48SStefano Zampini ierr = MatSetValues(pcbddc->local_mat,1,pcbddc->benign_p0_lidx+i,B0_ncol,B0_cols,B0_vals,INSERT_VALUES);CHKERRQ(ierr); 5407b034428SStefano Zampini ierr = MatSetValues(pcbddc->local_mat,B0_ncol,B0_cols,1,pcbddc->benign_p0_lidx+i,B0_vals,INSERT_VALUES);CHKERRQ(ierr); 5414f1b2e48SStefano Zampini ierr = MatSetValue(pcbddc->local_mat,pcbddc->benign_p0_lidx[i],pcbddc->benign_p0_lidx[i],0.0,INSERT_VALUES);CHKERRQ(ierr); 5424f1b2e48SStefano Zampini ierr = MatRestoreRow(pcbddc->benign_B0,i,&B0_ncol,(const PetscInt**)&B0_cols,(const PetscScalar**)&B0_vals);CHKERRQ(ierr); 5434f1b2e48SStefano Zampini } 544c263805aSStefano Zampini ierr = MatAssemblyBegin(pcbddc->local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 545c263805aSStefano Zampini ierr = MatAssemblyEnd(pcbddc->local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 546c263805aSStefano Zampini } 547c263805aSStefano Zampini PetscFunctionReturn(0); 548c263805aSStefano Zampini } 549c263805aSStefano Zampini 550c263805aSStefano Zampini #undef __FUNCT__ 551b1b3d7a2SStefano Zampini #define __FUNCT__ "PCBDDCAdaptiveSelection" 55208122e43SStefano Zampini PetscErrorCode PCBDDCAdaptiveSelection(PC pc) 553b1b3d7a2SStefano Zampini { 554b1b3d7a2SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 55508122e43SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 55608122e43SStefano Zampini PetscBLASInt B_dummyint,B_neigs,B_ierr,B_lwork; 55708122e43SStefano Zampini PetscBLASInt *B_iwork,*B_ifail; 55808122e43SStefano Zampini PetscScalar *work,lwork; 55908122e43SStefano Zampini PetscScalar *St,*S,*eigv; 56008122e43SStefano Zampini PetscScalar *Sarray,*Starray; 56108122e43SStefano Zampini PetscReal *eigs,thresh; 5621b968477SStefano Zampini PetscInt i,nmax,nmin,nv,cum,mss,cum2,cumarray,maxneigs; 563f6f667cfSStefano Zampini PetscBool allocated_S_St; 56408122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 56508122e43SStefano Zampini PetscReal *rwork; 56608122e43SStefano Zampini #endif 567b1b3d7a2SStefano Zampini PetscErrorCode ierr; 568b1b3d7a2SStefano Zampini 569b1b3d7a2SStefano Zampini PetscFunctionBegin; 57008122e43SStefano Zampini if (!sub_schurs->use_mumps) { 57108122e43SStefano Zampini SETERRQ(PetscObjectComm((PetscObject)pc),PETSC_ERR_SUP,"Adaptive selection of constraints requires MUMPS"); 57208122e43SStefano Zampini } 57308122e43SStefano Zampini 57406a4e24aSStefano Zampini if (sub_schurs->n_subs && (!sub_schurs->is_hermitian || !sub_schurs->is_posdef)) { 57506a4e24aSStefano 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); 57606a4e24aSStefano Zampini } 57706a4e24aSStefano Zampini 578fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 579fd14bc51SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 580fd14bc51SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 581fd14bc51SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Check adaptive selection of constraints\n");CHKERRQ(ierr); 5821575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 583fd14bc51SStefano Zampini } 584fd14bc51SStefano Zampini 585e496cd5dSStefano Zampini if (pcbddc->dbg_flag) { 586e496cd5dSStefano 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); 587e496cd5dSStefano Zampini } 588e496cd5dSStefano Zampini 58908122e43SStefano Zampini /* max size of subsets */ 59008122e43SStefano Zampini mss = 0; 59108122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 59208122e43SStefano Zampini PetscInt subset_size; 593862806e4SStefano Zampini 59408122e43SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 59508122e43SStefano Zampini mss = PetscMax(mss,subset_size); 59608122e43SStefano Zampini } 59708122e43SStefano Zampini 59808122e43SStefano Zampini /* min/max and threshold */ 59908122e43SStefano Zampini nmax = pcbddc->adaptive_nmax > 0 ? pcbddc->adaptive_nmax : mss; 600f6f667cfSStefano Zampini nmin = pcbddc->adaptive_nmin > 0 ? pcbddc->adaptive_nmin : 0; 60108122e43SStefano Zampini nmax = PetscMax(nmin,nmax); 602f6f667cfSStefano Zampini allocated_S_St = PETSC_FALSE; 603f6f667cfSStefano Zampini if (nmin) { 604f6f667cfSStefano Zampini allocated_S_St = PETSC_TRUE; 605f6f667cfSStefano Zampini } 60608122e43SStefano Zampini 60708122e43SStefano Zampini /* allocate lapack workspace */ 60808122e43SStefano Zampini cum = cum2 = 0; 60908122e43SStefano Zampini maxneigs = 0; 61008122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 61108122e43SStefano Zampini PetscInt n,subset_size; 612f6f667cfSStefano Zampini 61308122e43SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 61408122e43SStefano Zampini n = PetscMin(subset_size,nmax); 6159162d606SStefano Zampini cum += subset_size; 6169162d606SStefano Zampini cum2 += subset_size*n; 61708122e43SStefano Zampini maxneigs = PetscMax(maxneigs,n); 61808122e43SStefano Zampini } 61908122e43SStefano Zampini if (mss) { 6209ab7bb16SStefano Zampini if (sub_schurs->is_hermitian && sub_schurs->is_posdef) { 62108122e43SStefano Zampini PetscBLASInt B_itype = 1; 62208122e43SStefano Zampini PetscBLASInt B_N = mss; 6234c6709b3SStefano Zampini PetscReal zero = 0.0; 6244c6709b3SStefano Zampini PetscReal eps = 0.0; /* dlamch? */ 62508122e43SStefano Zampini 62608122e43SStefano Zampini B_lwork = -1; 62708122e43SStefano Zampini S = NULL; 62808122e43SStefano Zampini St = NULL; 629a58a30b4SStefano Zampini eigs = NULL; 630a58a30b4SStefano Zampini eigv = NULL; 631a58a30b4SStefano Zampini B_iwork = NULL; 632a58a30b4SStefano Zampini B_ifail = NULL; 633d1710679SStefano Zampini #if defined(PETSC_USE_COMPLEX) 634d1710679SStefano Zampini rwork = NULL; 635d1710679SStefano Zampini #endif 6368bec7fa6SStefano Zampini thresh = 1.0; 63708122e43SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 63808122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 63908122e43SStefano 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)); 64008122e43SStefano Zampini #else 64108122e43SStefano 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)); 64208122e43SStefano Zampini #endif 64308122e43SStefano Zampini if (B_ierr != 0) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to SYGVX Lapack routine %d",(int)B_ierr); 64408122e43SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 64508122e43SStefano Zampini } else { 64608122e43SStefano Zampini /* TODO */ 64708122e43SStefano Zampini } 64808122e43SStefano Zampini } else { 64908122e43SStefano Zampini lwork = 0; 65008122e43SStefano Zampini } 65108122e43SStefano Zampini 65208122e43SStefano Zampini nv = 0; 653d62866d3SStefano 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) */ 654d62866d3SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_vertices,&nv);CHKERRQ(ierr); 65508122e43SStefano Zampini } 6564c6709b3SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lwork),&B_lwork);CHKERRQ(ierr); 657f6f667cfSStefano Zampini if (allocated_S_St) { 658f6f667cfSStefano Zampini ierr = PetscMalloc2(mss*mss,&S,mss*mss,&St);CHKERRQ(ierr); 659f6f667cfSStefano Zampini } 660f6f667cfSStefano Zampini ierr = PetscMalloc5(mss*mss,&eigv,mss,&eigs,B_lwork,&work,5*mss,&B_iwork,mss,&B_ifail);CHKERRQ(ierr); 66108122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 66208122e43SStefano Zampini ierr = PetscMalloc1(7*mss,&rwork);CHKERRQ(ierr); 66308122e43SStefano Zampini #endif 6649162d606SStefano Zampini ierr = PetscMalloc5(nv+sub_schurs->n_subs,&pcbddc->adaptive_constraints_n, 6659162d606SStefano Zampini nv+sub_schurs->n_subs+1,&pcbddc->adaptive_constraints_idxs_ptr, 6669162d606SStefano Zampini nv+sub_schurs->n_subs+1,&pcbddc->adaptive_constraints_data_ptr, 66708122e43SStefano Zampini nv+cum,&pcbddc->adaptive_constraints_idxs, 6689162d606SStefano Zampini nv+cum2,&pcbddc->adaptive_constraints_data);CHKERRQ(ierr); 66908122e43SStefano Zampini ierr = PetscMemzero(pcbddc->adaptive_constraints_n,(nv+sub_schurs->n_subs)*sizeof(PetscInt));CHKERRQ(ierr); 67008122e43SStefano Zampini 67108122e43SStefano Zampini maxneigs = 0; 67208122e43SStefano Zampini cum = cum2 = cumarray = 0; 6739162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr[0] = 0; 6749162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr[0] = 0; 675d62866d3SStefano Zampini if (sub_schurs->is_vertices && pcbddc->use_vertices) { 67608122e43SStefano Zampini const PetscInt *idxs; 67708122e43SStefano Zampini 678d62866d3SStefano Zampini ierr = ISGetIndices(sub_schurs->is_vertices,&idxs);CHKERRQ(ierr); 67908122e43SStefano Zampini for (cum=0;cum<nv;cum++) { 68008122e43SStefano Zampini pcbddc->adaptive_constraints_n[cum] = 1; 68108122e43SStefano Zampini pcbddc->adaptive_constraints_idxs[cum] = idxs[cum]; 68208122e43SStefano Zampini pcbddc->adaptive_constraints_data[cum] = 1.0; 6839162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr[cum+1] = pcbddc->adaptive_constraints_idxs_ptr[cum]+1; 6849162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr[cum+1] = pcbddc->adaptive_constraints_data_ptr[cum]+1; 68508122e43SStefano Zampini } 68608122e43SStefano Zampini cum2 = cum; 687d62866d3SStefano Zampini ierr = ISRestoreIndices(sub_schurs->is_vertices,&idxs);CHKERRQ(ierr); 68808122e43SStefano Zampini } 68908122e43SStefano Zampini 69008122e43SStefano Zampini if (mss) { /* multilevel */ 69108122e43SStefano Zampini ierr = MatSeqAIJGetArray(sub_schurs->sum_S_Ej_inv_all,&Sarray);CHKERRQ(ierr); 69208122e43SStefano Zampini ierr = MatSeqAIJGetArray(sub_schurs->sum_S_Ej_tilda_all,&Starray);CHKERRQ(ierr); 69308122e43SStefano Zampini } 69408122e43SStefano Zampini 695ffd830a3SStefano Zampini thresh = pcbddc->adaptive_threshold; 69608122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 69708122e43SStefano Zampini const PetscInt *idxs; 698f6f667cfSStefano Zampini PetscReal infty = PETSC_MAX_REAL; 699862806e4SStefano Zampini PetscInt j,subset_size,eigs_start = 0; 70008122e43SStefano Zampini PetscBLASInt B_N; 701aff50787SStefano Zampini PetscBool same_data = PETSC_FALSE; 70208122e43SStefano Zampini 703862806e4SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 704ffd830a3SStefano Zampini ierr = ISGetIndices(sub_schurs->is_subs[i],&idxs);CHKERRQ(ierr); 705f6f667cfSStefano Zampini ierr = PetscBLASIntCast(subset_size,&B_N);CHKERRQ(ierr); 706f6f667cfSStefano Zampini if (allocated_S_St) { /* S and S_t should be copied since we could need them later */ 7079ab7bb16SStefano Zampini if (sub_schurs->is_hermitian) { 708aff50787SStefano Zampini PetscInt j,k; 709aff50787SStefano Zampini if (sub_schurs->n_subs == 1) { /* zeroing memory to use PetscMemcmp later */ 710aff50787SStefano Zampini ierr = PetscMemzero(S,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 711aff50787SStefano Zampini ierr = PetscMemzero(St,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 71208122e43SStefano Zampini } 71308122e43SStefano Zampini for (j=0;j<subset_size;j++) { 714aff50787SStefano Zampini for (k=j;k<subset_size;k++) { 715aff50787SStefano Zampini S [j*subset_size+k] = Sarray [cumarray+j*subset_size+k]; 716aff50787SStefano Zampini St[j*subset_size+k] = Starray[cumarray+j*subset_size+k]; 717aff50787SStefano Zampini } 71808122e43SStefano Zampini } 71908122e43SStefano Zampini } else { 72008122e43SStefano Zampini ierr = PetscMemcpy(S,Sarray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 72108122e43SStefano Zampini ierr = PetscMemcpy(St,Starray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 72208122e43SStefano Zampini } 7238bec7fa6SStefano Zampini } else { 724f6f667cfSStefano Zampini S = Sarray + cumarray; 725f6f667cfSStefano Zampini St = Starray + cumarray; 7268bec7fa6SStefano Zampini } 72708122e43SStefano Zampini 728aff50787SStefano Zampini /* see if we can save some work */ 729aff50787SStefano Zampini if (sub_schurs->n_subs == 1) { 730aff50787SStefano Zampini ierr = PetscMemcmp(S,St,subset_size*subset_size*sizeof(PetscScalar),&same_data);CHKERRQ(ierr); 731aff50787SStefano Zampini } 732aff50787SStefano Zampini 733aff50787SStefano Zampini if (same_data) { /* there's no need of constraints here, deluxe scaling is enough */ 734aff50787SStefano Zampini B_neigs = 0; 735aff50787SStefano Zampini } else { 7369ab7bb16SStefano Zampini if (sub_schurs->is_hermitian && sub_schurs->is_posdef) { 73708122e43SStefano Zampini PetscBLASInt B_itype = 1; 738f6f667cfSStefano Zampini PetscBLASInt B_IL, B_IU; 7394c6709b3SStefano Zampini PetscReal eps = -1.0; /* dlamch? */ 7409552c7c7SStefano Zampini PetscInt nmin_s; 74108122e43SStefano Zampini 742fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 7438bec7fa6SStefano 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]]); 744fd14bc51SStefano Zampini } 745d16cbb6bSStefano Zampini 74608122e43SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 747d16cbb6bSStefano Zampini if (thresh > 1.+PETSC_SMALL) { 748d16cbb6bSStefano Zampini 749d16cbb6bSStefano Zampini /* ask for eigenvalues larger than thresh */ 75008122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 751f6f667cfSStefano 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)); 75208122e43SStefano Zampini #else 753f6f667cfSStefano 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)); 75408122e43SStefano Zampini #endif 755d16cbb6bSStefano Zampini } else { 756d16cbb6bSStefano Zampini B_IU = PetscMax(1,PetscMin(B_N,nmax)); 757d16cbb6bSStefano Zampini B_IL = 1; 758d16cbb6bSStefano Zampini #if defined(PETSC_USE_COMPLEX) 759d16cbb6bSStefano 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)); 760d16cbb6bSStefano Zampini #else 761d16cbb6bSStefano 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)); 762d16cbb6bSStefano Zampini #endif 763d16cbb6bSStefano Zampini } 76408122e43SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 76508122e43SStefano Zampini if (B_ierr) { 76608122e43SStefano Zampini if (B_ierr < 0 ) { 76708122e43SStefano Zampini SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYGVX Lapack routine: illegal value for argument %d",-(int)B_ierr); 76808122e43SStefano Zampini } else if (B_ierr <= B_N) { 76908122e43SStefano Zampini SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYGVX Lapack routine: %d eigenvalues failed to converge",(int)B_ierr); 77008122e43SStefano Zampini } else { 7719552c7c7SStefano 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); 77208122e43SStefano Zampini } 77308122e43SStefano Zampini } 77408122e43SStefano Zampini 77508122e43SStefano Zampini if (B_neigs > nmax) { 776fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 777fd14bc51SStefano Zampini PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," found %d eigs, more than maximum required %d.\n",B_neigs,nmax); 778fd14bc51SStefano Zampini } 779f6f667cfSStefano Zampini eigs_start = B_neigs -nmax; 78008122e43SStefano Zampini B_neigs = nmax; 78108122e43SStefano Zampini } 78208122e43SStefano Zampini 7839552c7c7SStefano Zampini nmin_s = PetscMin(nmin,B_N); 7849552c7c7SStefano Zampini if (B_neigs < nmin_s) { 78508122e43SStefano Zampini PetscBLASInt B_neigs2; 78608122e43SStefano Zampini 787f6f667cfSStefano Zampini B_IU = B_N - B_neigs; 788f6f667cfSStefano Zampini B_IL = B_N - nmin_s + 1; 789fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 790fd14bc51SStefano 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); 791fd14bc51SStefano Zampini } 7929ab7bb16SStefano Zampini if (sub_schurs->is_hermitian) { 79308122e43SStefano Zampini PetscInt j; 79408122e43SStefano Zampini for (j=0;j<subset_size;j++) { 79508122e43SStefano Zampini ierr = PetscMemcpy(S+j*(subset_size+1),Sarray+cumarray+j*(subset_size+1),(subset_size-j)*sizeof(PetscScalar));CHKERRQ(ierr); 79608122e43SStefano Zampini } 79708122e43SStefano Zampini for (j=0;j<subset_size;j++) { 79808122e43SStefano Zampini ierr = PetscMemcpy(St+j*(subset_size+1),Starray+cumarray+j*(subset_size+1),(subset_size-j)*sizeof(PetscScalar));CHKERRQ(ierr); 79908122e43SStefano Zampini } 80008122e43SStefano Zampini } else { 80108122e43SStefano Zampini ierr = PetscMemcpy(S,Sarray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 80208122e43SStefano Zampini ierr = PetscMemcpy(St,Starray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 80308122e43SStefano Zampini } 80408122e43SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 80508122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 806f6f667cfSStefano 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)); 80708122e43SStefano Zampini #else 808f6f667cfSStefano 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)); 80908122e43SStefano Zampini #endif 81008122e43SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 81108122e43SStefano Zampini B_neigs += B_neigs2; 81208122e43SStefano Zampini } 81308122e43SStefano Zampini if (B_ierr) { 81408122e43SStefano Zampini if (B_ierr < 0 ) { 81508122e43SStefano Zampini SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYGVX Lapack routine: illegal value for argument %d",-(int)B_ierr); 81608122e43SStefano Zampini } else if (B_ierr <= B_N) { 81708122e43SStefano Zampini SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYGVX Lapack routine: %d eigenvalues failed to converge",(int)B_ierr); 81808122e43SStefano Zampini } else { 8199552c7c7SStefano 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); 82008122e43SStefano Zampini } 82108122e43SStefano Zampini } 822fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 823ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," -> Got %d eigs\n",B_neigs);CHKERRQ(ierr); 82408122e43SStefano Zampini for (j=0;j<B_neigs;j++) { 82508122e43SStefano Zampini if (eigs[j] == 0.0) { 826ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," Inf\n");CHKERRQ(ierr); 82708122e43SStefano Zampini } else { 828ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," %1.6e\n",eigs[j+eigs_start]);CHKERRQ(ierr); 829fd14bc51SStefano Zampini } 83008122e43SStefano Zampini } 83108122e43SStefano Zampini } 83208122e43SStefano Zampini } else { 83308122e43SStefano Zampini /* TODO */ 83408122e43SStefano Zampini } 835aff50787SStefano Zampini } 8368bec7fa6SStefano Zampini maxneigs = PetscMax(B_neigs,maxneigs); 8378bec7fa6SStefano Zampini pcbddc->adaptive_constraints_n[i+nv] = B_neigs; 8389162d606SStefano Zampini if (B_neigs) { 8399162d606SStefano 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); 840fd14bc51SStefano Zampini 841fd14bc51SStefano Zampini if (pcbddc->dbg_flag > 1) { 8429552c7c7SStefano Zampini PetscInt ii; 8439552c7c7SStefano Zampini for (ii=0;ii<B_neigs;ii++) { 844ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," -> Eigenvector %d/%d (%d)\n",ii,B_neigs,B_N);CHKERRQ(ierr); 8459552c7c7SStefano Zampini for (j=0;j<B_N;j++) { 846ac47001eSStefano Zampini #if defined(PETSC_USE_COMPLEX) 847ac47001eSStefano Zampini PetscReal r = PetscRealPart(pcbddc->adaptive_constraints_data[ii*subset_size+j+pcbddc->adaptive_constraints_data_ptr[cum]]); 848ac47001eSStefano Zampini PetscReal c = PetscImaginaryPart(pcbddc->adaptive_constraints_data[ii*subset_size+j+pcbddc->adaptive_constraints_data_ptr[cum]]); 849ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," %1.4e + %1.4e i\n",r,c);CHKERRQ(ierr); 850ac47001eSStefano Zampini #else 851ac47001eSStefano 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); 852ac47001eSStefano Zampini #endif 8539552c7c7SStefano Zampini } 8549552c7c7SStefano Zampini } 855fd14bc51SStefano Zampini } 85608122e43SStefano Zampini #if 0 8579162d606SStefano Zampini for (j=0;j<B_neigs;j++) { 85808122e43SStefano Zampini PetscBLASInt Blas_N,Blas_one = 1.0; 85908122e43SStefano Zampini PetscScalar norm; 86008122e43SStefano Zampini ierr = PetscBLASIntCast(subset_size,&Blas_N);CHKERRQ(ierr); 8619162d606SStefano Zampini PetscStackCallBLAS("BLASdot",norm = BLASdot_(&Blas_N,pcbddc->adaptive_constraints_data+pcbddc->adaptive_constraints_data_ptr[cum]+j*subset_size, 8629162d606SStefano Zampini &Blas_one,pcbddc->adaptive_constraints_data+pcbddc->adaptive_constraints_data_ptr[cum]+j*subset_size,&Blas_one)); 86308122e43SStefano Zampini if (pcbddc->adaptive_constraints_data[cum2] > 0.0) { 86408122e43SStefano Zampini norm = 1.0/PetscSqrtReal(PetscRealPart(norm)); 86508122e43SStefano Zampini } else { 86608122e43SStefano Zampini norm = -1.0/PetscSqrtReal(PetscRealPart(norm)); 86708122e43SStefano Zampini } 8689162d606SStefano Zampini PetscStackCallBLAS("BLASscal",BLASscal_(&Blas_N,&norm,pcbddc->adaptive_constraints_data+pcbddc->adaptive_constraints_data_ptr[cum]+j*subset_size,&Blas_one)); 869b1b3d7a2SStefano Zampini } 870b1b3d7a2SStefano Zampini #endif 8719162d606SStefano Zampini ierr = PetscMemcpy(pcbddc->adaptive_constraints_idxs+pcbddc->adaptive_constraints_idxs_ptr[cum],idxs,subset_size*sizeof(PetscInt));CHKERRQ(ierr); 8729162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr[cum+1] = pcbddc->adaptive_constraints_idxs_ptr[cum] + subset_size; 8739162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr[cum+1] = pcbddc->adaptive_constraints_data_ptr[cum] + subset_size*B_neigs; 8749162d606SStefano Zampini cum++; 87508122e43SStefano Zampini } 87608122e43SStefano Zampini ierr = ISRestoreIndices(sub_schurs->is_subs[i],&idxs);CHKERRQ(ierr); 87708122e43SStefano Zampini /* shift for next computation */ 87808122e43SStefano Zampini cumarray += subset_size*subset_size; 87908122e43SStefano Zampini } 880fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 881fd14bc51SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 882fd14bc51SStefano Zampini } 88308122e43SStefano Zampini 88408122e43SStefano Zampini if (mss) { 88508122e43SStefano Zampini ierr = MatSeqAIJRestoreArray(sub_schurs->sum_S_Ej_inv_all,&Sarray);CHKERRQ(ierr); 88608122e43SStefano Zampini ierr = MatSeqAIJRestoreArray(sub_schurs->sum_S_Ej_tilda_all,&Starray);CHKERRQ(ierr); 887f6f667cfSStefano Zampini /* destroy matrices (junk) */ 888f6f667cfSStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_inv_all);CHKERRQ(ierr); 889f6f667cfSStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_tilda_all);CHKERRQ(ierr); 89008122e43SStefano Zampini } 891f6f667cfSStefano Zampini if (allocated_S_St) { 892f6f667cfSStefano Zampini ierr = PetscFree2(S,St);CHKERRQ(ierr); 893f6f667cfSStefano Zampini } 894f6f667cfSStefano Zampini ierr = PetscFree5(eigv,eigs,work,B_iwork,B_ifail);CHKERRQ(ierr); 89508122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 89608122e43SStefano Zampini ierr = PetscFree(rwork);CHKERRQ(ierr); 89708122e43SStefano Zampini #endif 89808122e43SStefano Zampini if (pcbddc->dbg_flag) { 8991b968477SStefano Zampini PetscInt maxneigs_r; 90008122e43SStefano Zampini ierr = MPI_Allreduce(&maxneigs,&maxneigs_r,1,MPIU_INT,MPI_MAX,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 9019b28b941SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Maximum number of constraints per cc %d\n",maxneigs_r);CHKERRQ(ierr); 90208122e43SStefano Zampini } 90308122e43SStefano Zampini PetscFunctionReturn(0); 90408122e43SStefano Zampini } 905b1b3d7a2SStefano Zampini 906674ae819SStefano Zampini #undef __FUNCT__ 907c8587f34SStefano Zampini #define __FUNCT__ "PCBDDCSetUpSolvers" 908c8587f34SStefano Zampini PetscErrorCode PCBDDCSetUpSolvers(PC pc) 909c8587f34SStefano Zampini { 910c8587f34SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 9118629588bSStefano Zampini PetscScalar *coarse_submat_vals; 912c8587f34SStefano Zampini PetscErrorCode ierr; 913c8587f34SStefano Zampini 914c8587f34SStefano Zampini PetscFunctionBegin; 915f4ddd8eeSStefano Zampini /* Setup local scatters R_to_B and (optionally) R_to_D */ 9165e8657edSStefano Zampini /* PCBDDCSetUpLocalWorkVectors should be called first! */ 917c8587f34SStefano Zampini ierr = PCBDDCSetUpLocalScatters(pc);CHKERRQ(ierr); 918c8587f34SStefano Zampini 919684f6988SStefano Zampini /* Setup local neumann solver ksp_R */ 9200fccc4e9SStefano Zampini /* PCBDDCSetUpLocalScatters should be called first! */ 921684f6988SStefano Zampini ierr = PCBDDCSetUpLocalSolvers(pc,PETSC_FALSE,PETSC_TRUE);CHKERRQ(ierr); 922c8587f34SStefano Zampini 923c8587f34SStefano Zampini /* Change global null space passed in by the user if change of basis has been requested */ 924b9b85e73SStefano Zampini if (pcbddc->NullSpace && pcbddc->ChangeOfBasisMatrix) { 925c8587f34SStefano Zampini ierr = PCBDDCNullSpaceAdaptGlobal(pc);CHKERRQ(ierr); 926c8587f34SStefano Zampini } 927c8587f34SStefano Zampini 9288629588bSStefano Zampini /* 9298629588bSStefano Zampini Setup local correction and local part of coarse basis. 9308629588bSStefano Zampini Gives back the dense local part of the coarse matrix in column major ordering 9318629588bSStefano Zampini */ 93247f4ddc3SStefano Zampini ierr = PCBDDCSetUpCorrection(pc,&coarse_submat_vals);CHKERRQ(ierr); 9338629588bSStefano Zampini 9348629588bSStefano Zampini /* Compute total number of coarse nodes and setup coarse solver */ 9358629588bSStefano Zampini ierr = PCBDDCSetUpCoarseSolver(pc,coarse_submat_vals);CHKERRQ(ierr); 9368629588bSStefano Zampini 9378629588bSStefano Zampini /* free */ 9388629588bSStefano Zampini ierr = PetscFree(coarse_submat_vals);CHKERRQ(ierr); 939c8587f34SStefano Zampini PetscFunctionReturn(0); 940c8587f34SStefano Zampini } 941c8587f34SStefano Zampini 942c8587f34SStefano Zampini #undef __FUNCT__ 943674ae819SStefano Zampini #define __FUNCT__ "PCBDDCResetCustomization" 944674ae819SStefano Zampini PetscErrorCode PCBDDCResetCustomization(PC pc) 945674ae819SStefano Zampini { 946674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 947674ae819SStefano Zampini PetscErrorCode ierr; 948674ae819SStefano Zampini 949674ae819SStefano Zampini PetscFunctionBegin; 950674ae819SStefano Zampini ierr = PCBDDCGraphResetCSR(pcbddc->mat_graph);CHKERRQ(ierr); 951674ae819SStefano Zampini ierr = ISDestroy(&pcbddc->user_primal_vertices);CHKERRQ(ierr); 95230368db7SStefano Zampini ierr = ISDestroy(&pcbddc->user_primal_vertices_local);CHKERRQ(ierr); 953674ae819SStefano Zampini ierr = MatNullSpaceDestroy(&pcbddc->NullSpace);CHKERRQ(ierr); 954674ae819SStefano Zampini ierr = ISDestroy(&pcbddc->NeumannBoundaries);CHKERRQ(ierr); 955785d1243SStefano Zampini ierr = ISDestroy(&pcbddc->NeumannBoundariesLocal);CHKERRQ(ierr); 956674ae819SStefano Zampini ierr = ISDestroy(&pcbddc->DirichletBoundaries);CHKERRQ(ierr); 957f4ddd8eeSStefano Zampini ierr = MatNullSpaceDestroy(&pcbddc->onearnullspace);CHKERRQ(ierr); 958f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->onearnullvecs_state);CHKERRQ(ierr); 959785d1243SStefano Zampini ierr = ISDestroy(&pcbddc->DirichletBoundariesLocal);CHKERRQ(ierr); 96063602bcaSStefano Zampini ierr = PCBDDCSetDofsSplitting(pc,0,NULL);CHKERRQ(ierr); 96163602bcaSStefano Zampini ierr = PCBDDCSetDofsSplittingLocal(pc,0,NULL);CHKERRQ(ierr); 962674ae819SStefano Zampini PetscFunctionReturn(0); 963674ae819SStefano Zampini } 964674ae819SStefano Zampini 965674ae819SStefano Zampini #undef __FUNCT__ 966674ae819SStefano Zampini #define __FUNCT__ "PCBDDCResetTopography" 967674ae819SStefano Zampini PetscErrorCode PCBDDCResetTopography(PC pc) 968674ae819SStefano Zampini { 969674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 9704f1b2e48SStefano Zampini PetscInt i; 971674ae819SStefano Zampini PetscErrorCode ierr; 972674ae819SStefano Zampini 973674ae819SStefano Zampini PetscFunctionBegin; 974b9b85e73SStefano Zampini ierr = MatDestroy(&pcbddc->user_ChangeOfBasisMatrix);CHKERRQ(ierr); 975674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 976674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 977674ae819SStefano Zampini ierr = PCBDDCGraphReset(pcbddc->mat_graph);CHKERRQ(ierr); 9784f1b2e48SStefano Zampini for (i=0;i<pcbddc->n_local_subs;i++) { 9794f1b2e48SStefano Zampini ierr = ISDestroy(&pcbddc->local_subs[i]);CHKERRQ(ierr); 9804f1b2e48SStefano Zampini } 9814f1b2e48SStefano Zampini ierr = PetscFree(pcbddc->local_subs);CHKERRQ(ierr); 982b96c3477SStefano Zampini ierr = PCBDDCSubSchursReset(pcbddc->sub_schurs);CHKERRQ(ierr); 983674ae819SStefano Zampini PetscFunctionReturn(0); 984674ae819SStefano Zampini } 985674ae819SStefano Zampini 986674ae819SStefano Zampini #undef __FUNCT__ 987674ae819SStefano Zampini #define __FUNCT__ "PCBDDCResetSolvers" 988674ae819SStefano Zampini PetscErrorCode PCBDDCResetSolvers(PC pc) 989674ae819SStefano Zampini { 990674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 991674ae819SStefano Zampini PetscErrorCode ierr; 992674ae819SStefano Zampini 993674ae819SStefano Zampini PetscFunctionBegin; 994674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->coarse_vec);CHKERRQ(ierr); 99558da7f69SStefano Zampini if (pcbddc->coarse_phi_B) { 996*ca92afb2SStefano Zampini PetscScalar *array; 99706656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&array);CHKERRQ(ierr); 99806656605SStefano Zampini ierr = PetscFree(array);CHKERRQ(ierr); 99958da7f69SStefano Zampini } 1000674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_B);CHKERRQ(ierr); 1001674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_D);CHKERRQ(ierr); 100215aaf578SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_B);CHKERRQ(ierr); 100315aaf578SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_D);CHKERRQ(ierr); 1004674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec1_P);CHKERRQ(ierr); 1005674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec1_C);CHKERRQ(ierr); 1006674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat2);CHKERRQ(ierr); 100706656605SStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat1);CHKERRQ(ierr); 1008674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec1_R);CHKERRQ(ierr); 1009674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec2_R);CHKERRQ(ierr); 10108ce42a96SStefano Zampini ierr = ISDestroy(&pcbddc->is_R_local);CHKERRQ(ierr); 1011674ae819SStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_B);CHKERRQ(ierr); 1012674ae819SStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_D);CHKERRQ(ierr); 1013674ae819SStefano Zampini ierr = VecScatterDestroy(&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 1014f4ddd8eeSStefano Zampini ierr = KSPDestroy(&pcbddc->ksp_D);CHKERRQ(ierr); 1015f4ddd8eeSStefano Zampini ierr = KSPDestroy(&pcbddc->ksp_R);CHKERRQ(ierr); 1016f4ddd8eeSStefano Zampini ierr = KSPDestroy(&pcbddc->coarse_ksp);CHKERRQ(ierr); 1017f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 1018727cdba6SStefano Zampini ierr = PetscFree(pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 10190e6343abSStefano Zampini ierr = PetscFree2(pcbddc->local_primal_ref_node,pcbddc->local_primal_ref_mult);CHKERRQ(ierr); 1020f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->global_primal_indices);CHKERRQ(ierr); 102170cf5478SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling);CHKERRQ(ierr); 10226e683305SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 102381d14e9dSStefano Zampini ierr = MatDestroy(&pcbddc->benign_change);CHKERRQ(ierr); 10240369aaf7SStefano Zampini ierr = VecDestroy(&pcbddc->benign_vec);CHKERRQ(ierr); 10258b9f24d4SStefano Zampini ierr = MatDestroy(&pcbddc->benign_original_mat);CHKERRQ(ierr); 10264f1b2e48SStefano Zampini ierr = MatDestroy(&pcbddc->benign_B0);CHKERRQ(ierr); 10278b9f24d4SStefano Zampini ierr = PetscSFDestroy(&pcbddc->benign_sf);CHKERRQ(ierr); 1028*ca92afb2SStefano Zampini if (pcbddc->benign_zerodiag_subs) { 1029*ca92afb2SStefano Zampini PetscInt i; 1030*ca92afb2SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 1031*ca92afb2SStefano Zampini ierr = ISDestroy(&pcbddc->benign_zerodiag_subs[i]);CHKERRQ(ierr); 1032*ca92afb2SStefano Zampini } 1033*ca92afb2SStefano Zampini ierr = PetscFree(pcbddc->benign_zerodiag_subs);CHKERRQ(ierr); 1034*ca92afb2SStefano Zampini } 10354f1b2e48SStefano Zampini ierr = PetscFree3(pcbddc->benign_p0_lidx,pcbddc->benign_p0_gidx,pcbddc->benign_p0);CHKERRQ(ierr); 1036674ae819SStefano Zampini PetscFunctionReturn(0); 1037674ae819SStefano Zampini } 1038674ae819SStefano Zampini 1039674ae819SStefano Zampini #undef __FUNCT__ 1040f4ddd8eeSStefano Zampini #define __FUNCT__ "PCBDDCSetUpLocalWorkVectors" 1041f4ddd8eeSStefano Zampini PetscErrorCode PCBDDCSetUpLocalWorkVectors(PC pc) 10426bfb1811SStefano Zampini { 10436bfb1811SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 10446bfb1811SStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 10456bfb1811SStefano Zampini VecType impVecType; 10464f1b2e48SStefano Zampini PetscInt n_constraints,n_R,old_size; 10476bfb1811SStefano Zampini PetscErrorCode ierr; 10486bfb1811SStefano Zampini 10496bfb1811SStefano Zampini PetscFunctionBegin; 1050f4ddd8eeSStefano Zampini if (!pcbddc->ConstraintMatrix) { 1051019a44ceSStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"BDDC Constraint matrix has not been created"); 1052f4ddd8eeSStefano Zampini } 1053e7b262bdSStefano Zampini /* get sizes */ 10544f1b2e48SStefano Zampini n_constraints = pcbddc->local_primal_size - pcbddc->benign_n - pcbddc->n_vertices; 1055b371cd4fSStefano Zampini n_R = pcis->n - pcbddc->n_vertices; 10566bfb1811SStefano Zampini ierr = VecGetType(pcis->vec1_N,&impVecType);CHKERRQ(ierr); 1057e7b262bdSStefano Zampini /* local work vectors (try to avoid unneeded work)*/ 1058e7b262bdSStefano Zampini /* R nodes */ 1059e7b262bdSStefano Zampini old_size = -1; 1060e7b262bdSStefano Zampini if (pcbddc->vec1_R) { 1061e7b262bdSStefano Zampini ierr = VecGetSize(pcbddc->vec1_R,&old_size);CHKERRQ(ierr); 1062e7b262bdSStefano Zampini } 1063e7b262bdSStefano Zampini if (n_R != old_size) { 1064e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec1_R);CHKERRQ(ierr); 1065e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec2_R);CHKERRQ(ierr); 10666bfb1811SStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_R);CHKERRQ(ierr); 10676bfb1811SStefano Zampini ierr = VecSetSizes(pcbddc->vec1_R,PETSC_DECIDE,n_R);CHKERRQ(ierr); 10686bfb1811SStefano Zampini ierr = VecSetType(pcbddc->vec1_R,impVecType);CHKERRQ(ierr); 10696bfb1811SStefano Zampini ierr = VecDuplicate(pcbddc->vec1_R,&pcbddc->vec2_R);CHKERRQ(ierr); 1070e7b262bdSStefano Zampini } 1071e7b262bdSStefano Zampini /* local primal dofs */ 1072e7b262bdSStefano Zampini old_size = -1; 1073e7b262bdSStefano Zampini if (pcbddc->vec1_P) { 1074e7b262bdSStefano Zampini ierr = VecGetSize(pcbddc->vec1_P,&old_size);CHKERRQ(ierr); 1075e7b262bdSStefano Zampini } 1076e9189074SStefano Zampini if (pcbddc->local_primal_size != old_size) { 1077e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec1_P);CHKERRQ(ierr); 107883b7ccabSStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_P);CHKERRQ(ierr); 1079e9189074SStefano Zampini ierr = VecSetSizes(pcbddc->vec1_P,PETSC_DECIDE,pcbddc->local_primal_size);CHKERRQ(ierr); 10806bfb1811SStefano Zampini ierr = VecSetType(pcbddc->vec1_P,impVecType);CHKERRQ(ierr); 1081e7b262bdSStefano Zampini } 1082e7b262bdSStefano Zampini /* local explicit constraints */ 1083e7b262bdSStefano Zampini old_size = -1; 1084e7b262bdSStefano Zampini if (pcbddc->vec1_C) { 1085e7b262bdSStefano Zampini ierr = VecGetSize(pcbddc->vec1_C,&old_size);CHKERRQ(ierr); 1086e7b262bdSStefano Zampini } 1087e7b262bdSStefano Zampini if (n_constraints && n_constraints != old_size) { 1088e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec1_C);CHKERRQ(ierr); 108983b7ccabSStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_C);CHKERRQ(ierr); 109083b7ccabSStefano Zampini ierr = VecSetSizes(pcbddc->vec1_C,PETSC_DECIDE,n_constraints);CHKERRQ(ierr); 109183b7ccabSStefano Zampini ierr = VecSetType(pcbddc->vec1_C,impVecType);CHKERRQ(ierr); 109283b7ccabSStefano Zampini } 10936bfb1811SStefano Zampini PetscFunctionReturn(0); 10946bfb1811SStefano Zampini } 10956bfb1811SStefano Zampini 10966bfb1811SStefano Zampini #undef __FUNCT__ 109747f4ddc3SStefano Zampini #define __FUNCT__ "PCBDDCSetUpCorrection" 109847f4ddc3SStefano Zampini PetscErrorCode PCBDDCSetUpCorrection(PC pc, PetscScalar **coarse_submat_vals_n) 109988ebb749SStefano Zampini { 110025084f0cSStefano Zampini PetscErrorCode ierr; 110125084f0cSStefano Zampini /* pointers to pcis and pcbddc */ 110288ebb749SStefano Zampini PC_IS* pcis = (PC_IS*)pc->data; 110388ebb749SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 1104d62866d3SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 110525084f0cSStefano Zampini /* submatrices of local problem */ 110680677318SStefano Zampini Mat A_RV,A_VR,A_VV,local_auxmat2_R; 1107019a44ceSStefano Zampini /* submatrices of benign trick */ 1108d16cbb6bSStefano Zampini Mat B0_V = NULL; 110906656605SStefano Zampini /* submatrices of local coarse problem */ 111006656605SStefano Zampini Mat S_VV,S_CV,S_VC,S_CC; 111125084f0cSStefano Zampini /* working matrices */ 111206656605SStefano Zampini Mat C_CR; 111325084f0cSStefano Zampini /* additional working stuff */ 111406656605SStefano Zampini PC pc_R; 11154f1b2e48SStefano Zampini Mat F; 111606656605SStefano Zampini PetscBool isLU,isCHOL,isILU; 111706656605SStefano Zampini 111825084f0cSStefano Zampini PetscScalar *coarse_submat_vals; /* TODO: use a PETSc matrix */ 111906656605SStefano Zampini PetscScalar *work; 112006656605SStefano Zampini PetscInt *idx_V_B; 1121ffd830a3SStefano Zampini PetscInt lda_rhs,n,n_vertices,n_constraints,*p0_lidx_I; 112206656605SStefano Zampini PetscInt i,n_R,n_D,n_B; 1123ffd830a3SStefano Zampini 112425084f0cSStefano Zampini /* some shortcuts to scalars */ 112506656605SStefano Zampini PetscScalar one=1.0,m_one=-1.0; 112688ebb749SStefano Zampini 112788ebb749SStefano Zampini PetscFunctionBegin; 1128ffd830a3SStefano Zampini if (!pcbddc->symmetric_primal && pcbddc->benign_n) { 1129ffd830a3SStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Non-symmetric primal basis computation with benign trick not yet implemented"); 1130ffd830a3SStefano Zampini } 1131ffd830a3SStefano Zampini 1132ffd830a3SStefano Zampini /* Set Non-overlapping dimensions */ 1133b371cd4fSStefano Zampini n_vertices = pcbddc->n_vertices; 11344f1b2e48SStefano Zampini n_constraints = pcbddc->local_primal_size - pcbddc->benign_n - n_vertices; 1135b371cd4fSStefano Zampini n_B = pcis->n_B; 1136b371cd4fSStefano Zampini n_D = pcis->n - n_B; 113788ebb749SStefano Zampini n_R = pcis->n - n_vertices; 113888ebb749SStefano Zampini 113988ebb749SStefano Zampini /* vertices in boundary numbering */ 1140785e854fSJed Brown ierr = PetscMalloc1(n_vertices,&idx_V_B);CHKERRQ(ierr); 11410e6343abSStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,n_vertices,pcbddc->local_primal_ref_node,&i,idx_V_B);CHKERRQ(ierr); 114288ebb749SStefano Zampini if (i != n_vertices) { 114322d5777bSStefano Zampini SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Error in boundary numbering for BDDC vertices! %d != %d\n",n_vertices,i); 114488ebb749SStefano Zampini } 114588ebb749SStefano Zampini 114606656605SStefano Zampini /* Subdomain contribution (Non-overlapping) to coarse matrix */ 1147019a44ceSStefano Zampini ierr = PetscCalloc1(pcbddc->local_primal_size*pcbddc->local_primal_size,&coarse_submat_vals);CHKERRQ(ierr); 114806656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_vertices,n_vertices,coarse_submat_vals,&S_VV);CHKERRQ(ierr); 114906656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_VV,pcbddc->local_primal_size);CHKERRQ(ierr); 115006656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_constraints,n_vertices,coarse_submat_vals+n_vertices,&S_CV);CHKERRQ(ierr); 115106656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_CV,pcbddc->local_primal_size);CHKERRQ(ierr); 115206656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_vertices,n_constraints,coarse_submat_vals+pcbddc->local_primal_size*n_vertices,&S_VC);CHKERRQ(ierr); 115306656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_VC,pcbddc->local_primal_size);CHKERRQ(ierr); 115406656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_constraints,n_constraints,coarse_submat_vals+(pcbddc->local_primal_size+1)*n_vertices,&S_CC);CHKERRQ(ierr); 115506656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_CC,pcbddc->local_primal_size);CHKERRQ(ierr); 115606656605SStefano Zampini 115706656605SStefano Zampini /* determine if can use MatSolve routines instead of calling KSPSolve on ksp_R */ 115806656605SStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_R);CHKERRQ(ierr); 115906656605SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_R,PCLU,&isLU);CHKERRQ(ierr); 116006656605SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_R,PCILU,&isILU);CHKERRQ(ierr); 116106656605SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_R,PCCHOLESKY,&isCHOL);CHKERRQ(ierr); 1162ffd830a3SStefano Zampini lda_rhs = n_R; 116306656605SStefano Zampini if (isLU || isILU || isCHOL) { 116406656605SStefano Zampini ierr = PCFactorGetMatrix(pc_R,&F);CHKERRQ(ierr); 1165d62866d3SStefano Zampini } else if (sub_schurs->reuse_mumps) { 1166d62866d3SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 1167d62866d3SStefano Zampini MatFactorType type; 1168d62866d3SStefano Zampini 11696816873aSStefano Zampini F = reuse_mumps->F; 11706816873aSStefano Zampini ierr = MatGetFactorType(F,&type);CHKERRQ(ierr); 1171d62866d3SStefano Zampini if (type == MAT_FACTOR_CHOLESKY) isCHOL = PETSC_TRUE; 1172ffd830a3SStefano Zampini ierr = MatGetSize(F,&lda_rhs,NULL);CHKERRQ(ierr); 117306656605SStefano Zampini } else { 117406656605SStefano Zampini F = NULL; 117506656605SStefano Zampini } 117606656605SStefano Zampini 1177ffd830a3SStefano Zampini /* allocate workspace */ 1178ffd830a3SStefano Zampini n = 0; 1179ffd830a3SStefano Zampini if (n_constraints) { 1180ffd830a3SStefano Zampini n += lda_rhs*n_constraints; 1181ffd830a3SStefano Zampini } 1182ffd830a3SStefano Zampini if (n_vertices) { 1183ffd830a3SStefano Zampini n = PetscMax(2*lda_rhs*n_vertices,n); 1184ffd830a3SStefano Zampini n = PetscMax((lda_rhs+n_B)*n_vertices,n); 1185ffd830a3SStefano Zampini } 1186ffd830a3SStefano Zampini ierr = PetscMalloc1(n,&work);CHKERRQ(ierr); 1187ffd830a3SStefano Zampini 118888ebb749SStefano Zampini /* Precompute stuffs needed for preprocessing and application of BDDC*/ 118988ebb749SStefano Zampini if (n_constraints) { 119006656605SStefano Zampini Mat M1,M2,M3; 119180677318SStefano Zampini Mat auxmat; 119206656605SStefano Zampini IS is_aux; 119380677318SStefano Zampini PetscScalar *array,*array2; 119406656605SStefano Zampini 1195f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat1);CHKERRQ(ierr); 119680677318SStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat2);CHKERRQ(ierr); 119788ebb749SStefano Zampini 119825084f0cSStefano Zampini /* Extract constraints on R nodes: C_{CR} */ 119925084f0cSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,n_constraints,n_vertices,1,&is_aux);CHKERRQ(ierr); 12008ce42a96SStefano Zampini ierr = MatGetSubMatrix(pcbddc->ConstraintMatrix,is_aux,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&C_CR);CHKERRQ(ierr); 120180677318SStefano Zampini ierr = MatGetSubMatrix(pcbddc->ConstraintMatrix,is_aux,pcis->is_B_local,MAT_INITIAL_MATRIX,&auxmat);CHKERRQ(ierr); 120288ebb749SStefano Zampini 120380677318SStefano Zampini /* Assemble local_auxmat2_R = (- A_{RR}^{-1} C^T_{CR}) needed by BDDC setup */ 120480677318SStefano Zampini /* Assemble pcbddc->local_auxmat2 = R_to_B (- A_{RR}^{-1} C^T_{CR}) needed by BDDC application */ 1205ffd830a3SStefano Zampini ierr = PetscMemzero(work,lda_rhs*n_constraints*sizeof(PetscScalar));CHKERRQ(ierr); 120688ebb749SStefano Zampini for (i=0;i<n_constraints;i++) { 120706656605SStefano Zampini const PetscScalar *row_cmat_values; 120806656605SStefano Zampini const PetscInt *row_cmat_indices; 120906656605SStefano Zampini PetscInt size_of_constraint,j; 121088ebb749SStefano Zampini 121106656605SStefano Zampini ierr = MatGetRow(C_CR,i,&size_of_constraint,&row_cmat_indices,&row_cmat_values);CHKERRQ(ierr); 121206656605SStefano Zampini for (j=0;j<size_of_constraint;j++) { 1213ffd830a3SStefano Zampini work[row_cmat_indices[j]+i*lda_rhs] = -row_cmat_values[j]; 121406656605SStefano Zampini } 121506656605SStefano Zampini ierr = MatRestoreRow(C_CR,i,&size_of_constraint,&row_cmat_indices,&row_cmat_values);CHKERRQ(ierr); 121606656605SStefano Zampini } 1217ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_constraints,NULL,&local_auxmat2_R);CHKERRQ(ierr); 121806656605SStefano Zampini if (F) { 121906656605SStefano Zampini Mat B; 122006656605SStefano Zampini 1221ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_constraints,work,&B);CHKERRQ(ierr); 122280677318SStefano Zampini ierr = MatMatSolve(F,B,local_auxmat2_R);CHKERRQ(ierr); 122306656605SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 122406656605SStefano Zampini } else { 122580677318SStefano Zampini PetscScalar *marr; 122680677318SStefano Zampini 122780677318SStefano Zampini ierr = MatDenseGetArray(local_auxmat2_R,&marr);CHKERRQ(ierr); 122806656605SStefano Zampini for (i=0;i<n_constraints;i++) { 1229ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*lda_rhs);CHKERRQ(ierr); 1230ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,marr+i*lda_rhs);CHKERRQ(ierr); 123106656605SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 123206656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 123306656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 123406656605SStefano Zampini } 123580677318SStefano Zampini ierr = MatDenseRestoreArray(local_auxmat2_R,&marr);CHKERRQ(ierr); 123606656605SStefano Zampini } 123780677318SStefano Zampini if (!pcbddc->switch_static) { 123880677318SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,n_constraints,NULL,&pcbddc->local_auxmat2);CHKERRQ(ierr); 123980677318SStefano Zampini ierr = MatDenseGetArray(pcbddc->local_auxmat2,&array);CHKERRQ(ierr); 124080677318SStefano Zampini ierr = MatDenseGetArray(local_auxmat2_R,&array2);CHKERRQ(ierr); 124180677318SStefano Zampini for (i=0;i<n_constraints;i++) { 1242ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,array2+i*lda_rhs);CHKERRQ(ierr); 124380677318SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,array+i*n_B);CHKERRQ(ierr); 124480677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 124580677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 124680677318SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 124780677318SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 124880677318SStefano Zampini } 124980677318SStefano Zampini ierr = MatDenseRestoreArray(local_auxmat2_R,&array2);CHKERRQ(ierr); 125080677318SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->local_auxmat2,&array);CHKERRQ(ierr); 125180677318SStefano Zampini ierr = MatMatMult(auxmat,pcbddc->local_auxmat2,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&M3);CHKERRQ(ierr); 125280677318SStefano Zampini } else { 1253ffd830a3SStefano Zampini if (lda_rhs != n_R) { 1254ffd830a3SStefano Zampini IS dummy; 1255ffd830a3SStefano Zampini 1256ffd830a3SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,n_R,0,1,&dummy);CHKERRQ(ierr); 1257ffd830a3SStefano Zampini ierr = MatGetSubMatrix(local_auxmat2_R,dummy,dummy,MAT_INITIAL_MATRIX,&pcbddc->local_auxmat2);CHKERRQ(ierr); 1258ffd830a3SStefano Zampini ierr = ISDestroy(&dummy);CHKERRQ(ierr); 1259ffd830a3SStefano Zampini } else { 126080677318SStefano Zampini ierr = PetscObjectReference((PetscObject)local_auxmat2_R);CHKERRQ(ierr); 126180677318SStefano Zampini pcbddc->local_auxmat2 = local_auxmat2_R; 1262ffd830a3SStefano Zampini } 126325084f0cSStefano Zampini ierr = MatMatMult(C_CR,pcbddc->local_auxmat2,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&M3);CHKERRQ(ierr); 126480677318SStefano Zampini } 126580677318SStefano Zampini ierr = ISDestroy(&is_aux);CHKERRQ(ierr); 126680677318SStefano Zampini /* Assemble explicitly S_CC = ( C_{CR} A_{RR}^{-1} C^T_{CR} )^{-1} */ 126780677318SStefano Zampini ierr = MatScale(M3,m_one);CHKERRQ(ierr); 126806656605SStefano Zampini ierr = MatDuplicate(M3,MAT_DO_NOT_COPY_VALUES,&M1);CHKERRQ(ierr); 126906656605SStefano Zampini ierr = MatDuplicate(M3,MAT_DO_NOT_COPY_VALUES,&M2);CHKERRQ(ierr); 127080677318SStefano Zampini if (isCHOL) { 127180677318SStefano Zampini ierr = MatCholeskyFactor(M3,NULL,NULL);CHKERRQ(ierr); 127280677318SStefano Zampini } else { 127325084f0cSStefano Zampini ierr = MatLUFactor(M3,NULL,NULL,NULL);CHKERRQ(ierr); 127480677318SStefano Zampini } 127580677318SStefano Zampini ierr = VecSet(pcbddc->vec1_C,one);CHKERRQ(ierr); 127606656605SStefano Zampini ierr = MatDiagonalSet(M2,pcbddc->vec1_C,INSERT_VALUES);CHKERRQ(ierr); 127725084f0cSStefano Zampini ierr = MatMatSolve(M3,M2,M1);CHKERRQ(ierr); 127825084f0cSStefano Zampini ierr = MatDestroy(&M2);CHKERRQ(ierr); 127925084f0cSStefano Zampini ierr = MatDestroy(&M3);CHKERRQ(ierr); 128080677318SStefano Zampini /* Assemble local_auxmat1 = S_CC*C_{CB} needed by BDDC application in KSP and in preproc */ 128180677318SStefano Zampini ierr = MatMatMult(M1,auxmat,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&pcbddc->local_auxmat1);CHKERRQ(ierr); 128280677318SStefano Zampini ierr = MatDestroy(&auxmat);CHKERRQ(ierr); 128306656605SStefano Zampini ierr = MatCopy(M1,S_CC,SAME_NONZERO_PATTERN);CHKERRQ(ierr); /* S_CC can have a different LDA, MatMatSolve doesn't support it */ 128406656605SStefano Zampini ierr = MatDestroy(&M1);CHKERRQ(ierr); 1285f4ddd8eeSStefano Zampini } 128688ebb749SStefano Zampini /* Get submatrices from subdomain matrix */ 12874f1b2e48SStefano Zampini if (pcbddc->benign_n) { 1288d16cbb6bSStefano Zampini IS dummy; 1289d16cbb6bSStefano Zampini Mat B0_R; 1290d16cbb6bSStefano Zampini PetscReal norm; 1291d16cbb6bSStefano Zampini 12924f1b2e48SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcbddc->benign_n,0,1,&dummy);CHKERRQ(ierr); 12934f1b2e48SStefano Zampini ierr = MatGetSubMatrix(pcbddc->benign_B0,dummy,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&B0_R);CHKERRQ(ierr); 1294d16cbb6bSStefano Zampini ierr = MatNorm(B0_R,NORM_INFINITY,&norm);CHKERRQ(ierr); 1295d16cbb6bSStefano Zampini if (norm > PETSC_SMALL) { 1296d16cbb6bSStefano Zampini SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_SUP,"Benign trick can not be applied! ||B0_R|| = %f (should be numerically 0.)",norm); 1297d16cbb6bSStefano Zampini } 1298d16cbb6bSStefano Zampini ierr = MatDestroy(&B0_R);CHKERRQ(ierr); 1299d16cbb6bSStefano Zampini ierr = ISDestroy(&dummy);CHKERRQ(ierr); 1300d16cbb6bSStefano Zampini } 1301d16cbb6bSStefano Zampini 130288ebb749SStefano Zampini if (n_vertices) { 130306656605SStefano Zampini IS is_aux; 13043a50541eSStefano Zampini 13056816873aSStefano Zampini if (sub_schurs->reuse_mumps) { /* is_R_local is not sorted, ISComplement doesn't like it */ 13066816873aSStefano Zampini IS tis; 13076816873aSStefano Zampini 13086816873aSStefano Zampini ierr = ISDuplicate(pcbddc->is_R_local,&tis);CHKERRQ(ierr); 13096816873aSStefano Zampini ierr = ISSort(tis);CHKERRQ(ierr); 13106816873aSStefano Zampini ierr = ISComplement(tis,0,pcis->n,&is_aux);CHKERRQ(ierr); 13116816873aSStefano Zampini ierr = ISDestroy(&tis);CHKERRQ(ierr); 13126816873aSStefano Zampini } else { 13133a50541eSStefano Zampini ierr = ISComplement(pcbddc->is_R_local,0,pcis->n,&is_aux);CHKERRQ(ierr); 13146816873aSStefano Zampini } 13159577ea80SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,pcbddc->is_R_local,is_aux,MAT_INITIAL_MATRIX,&A_RV);CHKERRQ(ierr); 13169577ea80SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,is_aux,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&A_VR);CHKERRQ(ierr); 131704708bb6SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,is_aux,is_aux,MAT_INITIAL_MATRIX,&A_VV);CHKERRQ(ierr); 13184f1b2e48SStefano Zampini if (pcbddc->benign_n) { 1319019a44ceSStefano Zampini IS dummy; 1320019a44ceSStefano Zampini 13214f1b2e48SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcbddc->benign_n,0,1,&dummy);CHKERRQ(ierr); 13224f1b2e48SStefano Zampini ierr = MatGetSubMatrix(pcbddc->benign_B0,dummy,is_aux,MAT_INITIAL_MATRIX,&B0_V);CHKERRQ(ierr); 1323019a44ceSStefano Zampini ierr = ISDestroy(&dummy);CHKERRQ(ierr); 1324019a44ceSStefano Zampini } 132525084f0cSStefano Zampini ierr = ISDestroy(&is_aux);CHKERRQ(ierr); 132688ebb749SStefano Zampini } 132788ebb749SStefano Zampini 132888ebb749SStefano Zampini /* Matrix of coarse basis functions (local) */ 1329f4ddd8eeSStefano Zampini if (pcbddc->coarse_phi_B) { 133006656605SStefano Zampini PetscInt on_B,on_primal,on_D=n_D; 133106656605SStefano Zampini if (pcbddc->coarse_phi_D) { 133206656605SStefano Zampini ierr = MatGetSize(pcbddc->coarse_phi_D,&on_D,NULL);CHKERRQ(ierr); 133306656605SStefano Zampini } 1334f4ddd8eeSStefano Zampini ierr = MatGetSize(pcbddc->coarse_phi_B,&on_B,&on_primal);CHKERRQ(ierr); 133506656605SStefano Zampini if (on_B != n_B || on_primal != pcbddc->local_primal_size || on_D != n_D) { 133606656605SStefano Zampini PetscScalar *marray; 133706656605SStefano Zampini 133806656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&marray);CHKERRQ(ierr); 133906656605SStefano Zampini ierr = PetscFree(marray);CHKERRQ(ierr); 1340f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_B);CHKERRQ(ierr); 1341f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_B);CHKERRQ(ierr); 1342f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_D);CHKERRQ(ierr); 1343f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_D);CHKERRQ(ierr); 1344f4ddd8eeSStefano Zampini } 1345f4ddd8eeSStefano Zampini } 134606656605SStefano Zampini 1347f4ddd8eeSStefano Zampini if (!pcbddc->coarse_phi_B) { 134806656605SStefano Zampini PetscScalar *marray; 134988ebb749SStefano Zampini 135006656605SStefano Zampini n = n_B*pcbddc->local_primal_size; 13518eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 135206656605SStefano Zampini n += n_D*pcbddc->local_primal_size; 135388ebb749SStefano Zampini } 13543301b35fSStefano Zampini if (!pcbddc->symmetric_primal) { 135506656605SStefano Zampini n *= 2; 135688ebb749SStefano Zampini } 135706656605SStefano Zampini ierr = PetscCalloc1(n,&marray);CHKERRQ(ierr); 135806656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,pcbddc->local_primal_size,marray,&pcbddc->coarse_phi_B);CHKERRQ(ierr); 135906656605SStefano Zampini n = n_B*pcbddc->local_primal_size; 13608eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 136106656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_D,pcbddc->local_primal_size,marray+n,&pcbddc->coarse_phi_D);CHKERRQ(ierr); 136206656605SStefano Zampini n += n_D*pcbddc->local_primal_size; 136388ebb749SStefano Zampini } 13643301b35fSStefano Zampini if (!pcbddc->symmetric_primal) { 136506656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,pcbddc->local_primal_size,marray+n,&pcbddc->coarse_psi_B);CHKERRQ(ierr); 13668eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 136706656605SStefano Zampini n = n_B*pcbddc->local_primal_size; 136806656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_D,pcbddc->local_primal_size,marray+n,&pcbddc->coarse_psi_D);CHKERRQ(ierr); 136988ebb749SStefano Zampini } 137088ebb749SStefano Zampini } else { 1371c0553b1fSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->coarse_phi_B);CHKERRQ(ierr); 1372c0553b1fSStefano Zampini pcbddc->coarse_psi_B = pcbddc->coarse_phi_B; 13731b968477SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 1374c0553b1fSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->coarse_phi_D);CHKERRQ(ierr); 1375c0553b1fSStefano Zampini pcbddc->coarse_psi_D = pcbddc->coarse_phi_D; 1376c0553b1fSStefano Zampini } 137788ebb749SStefano Zampini } 137806656605SStefano Zampini } 1379019a44ceSStefano Zampini 138006656605SStefano Zampini /* We are now ready to evaluate coarse basis functions and subdomain contribution to coarse problem */ 13814f1b2e48SStefano Zampini p0_lidx_I = NULL; 13824f1b2e48SStefano Zampini if (pcbddc->benign_n && (pcbddc->switch_static || pcbddc->dbg_flag)) { 1383d12edf2fSStefano Zampini const PetscInt *idxs; 1384d12edf2fSStefano Zampini 1385d12edf2fSStefano Zampini ierr = ISGetIndices(pcis->is_I_local,&idxs);CHKERRQ(ierr); 13864f1b2e48SStefano Zampini ierr = PetscMalloc1(pcbddc->benign_n,&p0_lidx_I);CHKERRQ(ierr); 13874f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 13884f1b2e48SStefano Zampini ierr = PetscFindInt(pcbddc->benign_p0_lidx[i],pcis->n-pcis->n_B,idxs,&p0_lidx_I[i]);CHKERRQ(ierr); 13894f1b2e48SStefano Zampini } 1390d12edf2fSStefano Zampini ierr = ISRestoreIndices(pcis->is_I_local,&idxs);CHKERRQ(ierr); 1391d12edf2fSStefano Zampini } 1392d16cbb6bSStefano Zampini 139306656605SStefano Zampini /* vertices */ 139406656605SStefano Zampini if (n_vertices) { 139516f15bc4SStefano Zampini 1396ffd830a3SStefano Zampini ierr = MatConvert(A_VV,MATDENSE,MAT_REUSE_MATRIX,&A_VV);CHKERRQ(ierr); 139704708bb6SStefano Zampini 139816f15bc4SStefano Zampini if (n_R) { 139906656605SStefano Zampini Mat A_RRmA_RV,S_VVt; /* S_VVt with LDA=N */ 140006656605SStefano Zampini PetscBLASInt B_N,B_one = 1; 140116f15bc4SStefano Zampini PetscScalar *x,*y; 140204708bb6SStefano Zampini PetscBool isseqaij; 140306656605SStefano Zampini 140421eccb56SStefano Zampini ierr = MatScale(A_RV,m_one);CHKERRQ(ierr); 1405ffd830a3SStefano Zampini if (lda_rhs == n_R) { 1406ffd830a3SStefano Zampini ierr = MatConvert(A_RV,MATDENSE,MAT_REUSE_MATRIX,&A_RV);CHKERRQ(ierr); 1407ffd830a3SStefano Zampini } else { 1408*ca92afb2SStefano Zampini PetscScalar *av,*array; 1409*ca92afb2SStefano Zampini const PetscInt *xadj,*adjncy; 1410*ca92afb2SStefano Zampini PetscInt n; 1411*ca92afb2SStefano Zampini PetscBool flg_row; 1412ffd830a3SStefano Zampini 1413*ca92afb2SStefano Zampini array = work+lda_rhs*n_vertices; 1414*ca92afb2SStefano Zampini ierr = PetscMemzero(array,lda_rhs*n_vertices*sizeof(PetscScalar));CHKERRQ(ierr); 1415*ca92afb2SStefano Zampini ierr = MatConvert(A_RV,MATSEQAIJ,MAT_REUSE_MATRIX,&A_RV);CHKERRQ(ierr); 1416*ca92afb2SStefano Zampini ierr = MatGetRowIJ(A_RV,0,PETSC_FALSE,PETSC_FALSE,&n,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 1417*ca92afb2SStefano Zampini ierr = MatSeqAIJGetArray(A_RV,&av);CHKERRQ(ierr); 1418*ca92afb2SStefano Zampini for (i=0;i<n;i++) { 1419*ca92afb2SStefano Zampini PetscInt j; 1420*ca92afb2SStefano Zampini for (j=xadj[i];j<xadj[i+1];j++) array[lda_rhs*adjncy[j]+i] = av[j]; 1421ffd830a3SStefano Zampini } 1422*ca92afb2SStefano Zampini ierr = MatRestoreRowIJ(A_RV,0,PETSC_FALSE,PETSC_FALSE,&n,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 1423*ca92afb2SStefano Zampini ierr = MatDestroy(&A_RV);CHKERRQ(ierr); 1424*ca92afb2SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_vertices,array,&A_RV);CHKERRQ(ierr); 1425ffd830a3SStefano Zampini } 1426ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_vertices,work,&A_RRmA_RV);CHKERRQ(ierr); 14276816873aSStefano Zampini if (F) { /* TODO could be optimized for symmetric problems */ 142806656605SStefano Zampini ierr = MatMatSolve(F,A_RV,A_RRmA_RV);CHKERRQ(ierr); 142906656605SStefano Zampini } else { 143006656605SStefano Zampini ierr = MatDenseGetArray(A_RV,&y);CHKERRQ(ierr); 143106656605SStefano Zampini for (i=0;i<n_vertices;i++) { 1432ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,y+i*lda_rhs);CHKERRQ(ierr); 1433ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,work+i*lda_rhs);CHKERRQ(ierr); 143406656605SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 143506656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 143606656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 143706656605SStefano Zampini } 143806656605SStefano Zampini ierr = MatDenseRestoreArray(A_RV,&y);CHKERRQ(ierr); 143906656605SStefano Zampini } 144080677318SStefano Zampini ierr = MatDestroy(&A_RV);CHKERRQ(ierr); 1441ffd830a3SStefano Zampini /* S_VV and S_CV */ 144206656605SStefano Zampini if (n_constraints) { 144306656605SStefano Zampini Mat B; 144480677318SStefano Zampini 1445ffd830a3SStefano Zampini ierr = PetscMemzero(work+lda_rhs*n_vertices,n_B*n_vertices*sizeof(PetscScalar));CHKERRQ(ierr); 144680677318SStefano Zampini for (i=0;i<n_vertices;i++) { 1447ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*lda_rhs);CHKERRQ(ierr); 1448ffd830a3SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,work+lda_rhs*n_vertices+i*n_B);CHKERRQ(ierr); 144980677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 145080677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 145180677318SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 145280677318SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 145380677318SStefano Zampini } 1454ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,n_vertices,work+lda_rhs*n_vertices,&B);CHKERRQ(ierr); 145580677318SStefano Zampini ierr = MatMatMult(pcbddc->local_auxmat1,B,MAT_REUSE_MATRIX,PETSC_DEFAULT,&S_CV);CHKERRQ(ierr); 145680677318SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 1457ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_vertices,work+lda_rhs*n_vertices,&B);CHKERRQ(ierr); 145880677318SStefano Zampini ierr = MatMatMult(local_auxmat2_R,S_CV,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B);CHKERRQ(ierr); 145906656605SStefano Zampini ierr = MatScale(S_CV,m_one);CHKERRQ(ierr); 1460ffd830a3SStefano Zampini ierr = PetscBLASIntCast(lda_rhs*n_vertices,&B_N);CHKERRQ(ierr); 1461ffd830a3SStefano Zampini PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&B_N,&one,work+lda_rhs*n_vertices,&B_one,work,&B_one)); 146206656605SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 146306656605SStefano Zampini } 146404708bb6SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)A_VR,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 146504708bb6SStefano Zampini if (!isseqaij) { /* MatMatMult with SEQ(S)BAIJ below will raise an error */ 146604708bb6SStefano Zampini ierr = MatConvert(A_VR,MATSEQAIJ,MAT_REUSE_MATRIX,&A_VR);CHKERRQ(ierr); 146704708bb6SStefano Zampini } 1468ffd830a3SStefano Zampini if (lda_rhs != n_R) { 1469ffd830a3SStefano Zampini ierr = MatDestroy(&A_RRmA_RV);CHKERRQ(ierr); 1470ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_vertices,work,&A_RRmA_RV);CHKERRQ(ierr); 1471ffd830a3SStefano Zampini ierr = MatSeqDenseSetLDA(A_RRmA_RV,lda_rhs);CHKERRQ(ierr); 1472ffd830a3SStefano Zampini } 147306656605SStefano Zampini ierr = MatMatMult(A_VR,A_RRmA_RV,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&S_VVt);CHKERRQ(ierr); 147480677318SStefano Zampini ierr = MatDestroy(&A_RRmA_RV);CHKERRQ(ierr); 147506656605SStefano Zampini ierr = PetscBLASIntCast(n_vertices*n_vertices,&B_N);CHKERRQ(ierr); 147606656605SStefano Zampini ierr = MatDenseGetArray(A_VV,&x);CHKERRQ(ierr); 147706656605SStefano Zampini ierr = MatDenseGetArray(S_VVt,&y);CHKERRQ(ierr); 147806656605SStefano Zampini PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&B_N,&one,x,&B_one,y,&B_one)); 147906656605SStefano Zampini ierr = MatDenseRestoreArray(A_VV,&x);CHKERRQ(ierr); 148006656605SStefano Zampini ierr = MatDenseRestoreArray(S_VVt,&y);CHKERRQ(ierr); 148106656605SStefano Zampini ierr = MatCopy(S_VVt,S_VV,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 1482d16cbb6bSStefano Zampini ierr = MatDestroy(&S_VVt);CHKERRQ(ierr); 1483019a44ceSStefano Zampini } else { 1484d16cbb6bSStefano Zampini ierr = MatCopy(A_VV,S_VV,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 1485d16cbb6bSStefano Zampini } 14864f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 1487019a44ceSStefano Zampini const PetscScalar *vals; 1488019a44ceSStefano Zampini const PetscInt *idxs; 14894f1b2e48SStefano Zampini PetscInt n,j,primal_idx; 1490019a44ceSStefano Zampini 14914f1b2e48SStefano Zampini ierr = MatGetRow(B0_V,i,&n,&idxs,&vals);CHKERRQ(ierr); 14924f1b2e48SStefano Zampini primal_idx = pcbddc->local_primal_size - pcbddc->benign_n + i; 1493d16cbb6bSStefano Zampini for (j=0;j<n;j++) { 14944f1b2e48SStefano Zampini coarse_submat_vals[primal_idx*pcbddc->local_primal_size+idxs[j]] = vals[j]; 14954f1b2e48SStefano Zampini coarse_submat_vals[idxs[j]*pcbddc->local_primal_size+primal_idx] = vals[j]; 1496019a44ceSStefano Zampini } 14974f1b2e48SStefano Zampini ierr = MatRestoreRow(B0_V,i,&n,&idxs,&vals);CHKERRQ(ierr); 149816f15bc4SStefano Zampini } 149921eccb56SStefano Zampini ierr = MatDestroy(&A_VV);CHKERRQ(ierr); 1500d16cbb6bSStefano Zampini 150106656605SStefano Zampini /* coarse basis functions */ 150206656605SStefano Zampini for (i=0;i<n_vertices;i++) { 150316f15bc4SStefano Zampini PetscScalar *y; 150416f15bc4SStefano Zampini 1505ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+lda_rhs*i);CHKERRQ(ierr); 150606656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 150706656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*i);CHKERRQ(ierr); 150806656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 150906656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 151006656605SStefano Zampini y[n_B*i+idx_V_B[i]] = 1.0; 151106656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 151206656605SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 151306656605SStefano Zampini 151406656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 15154f1b2e48SStefano Zampini PetscInt j; 15164f1b2e48SStefano Zampini 151706656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 151806656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*i);CHKERRQ(ierr); 151906656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 152006656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 152106656605SStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 15224f1b2e48SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) y[n_D*i+p0_lidx_I[j]] = 0.0; 152306656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 152406656605SStefano Zampini } 152506656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 152606656605SStefano Zampini } 152704708bb6SStefano Zampini /* if n_R == 0 the object is not destroyed */ 152804708bb6SStefano Zampini ierr = MatDestroy(&A_RV);CHKERRQ(ierr); 152906656605SStefano Zampini } 153006656605SStefano Zampini 153106656605SStefano Zampini if (n_constraints) { 153206656605SStefano Zampini Mat B; 153306656605SStefano Zampini 1534ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_constraints,work,&B);CHKERRQ(ierr); 153506656605SStefano Zampini ierr = MatScale(S_CC,m_one);CHKERRQ(ierr); 153680677318SStefano Zampini ierr = MatMatMult(local_auxmat2_R,S_CC,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B);CHKERRQ(ierr); 153706656605SStefano Zampini ierr = MatScale(S_CC,m_one);CHKERRQ(ierr); 153806656605SStefano Zampini if (n_vertices) { 153980677318SStefano Zampini if (isCHOL) { /* if we can solve the interior problem with cholesky, we should also be fine with transposing here */ 154080677318SStefano Zampini ierr = MatTranspose(S_CV,MAT_REUSE_MATRIX,&S_VC);CHKERRQ(ierr); 154180677318SStefano Zampini } else { 154280677318SStefano Zampini Mat S_VCt; 154380677318SStefano Zampini 1544ffd830a3SStefano Zampini if (lda_rhs != n_R) { 1545ffd830a3SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 1546ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_vertices,work,&B);CHKERRQ(ierr); 1547ffd830a3SStefano Zampini ierr = MatSeqDenseSetLDA(B,lda_rhs);CHKERRQ(ierr); 1548ffd830a3SStefano Zampini } 154980677318SStefano Zampini ierr = MatMatMult(A_VR,B,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&S_VCt);CHKERRQ(ierr); 155080677318SStefano Zampini ierr = MatCopy(S_VCt,S_VC,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 155180677318SStefano Zampini ierr = MatDestroy(&S_VCt);CHKERRQ(ierr); 155280677318SStefano Zampini } 155306656605SStefano Zampini } 155406656605SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 155506656605SStefano Zampini /* coarse basis functions */ 155606656605SStefano Zampini for (i=0;i<n_constraints;i++) { 155706656605SStefano Zampini PetscScalar *y; 155806656605SStefano Zampini 1559ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+lda_rhs*i);CHKERRQ(ierr); 156006656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 156106656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*(i+n_vertices));CHKERRQ(ierr); 156206656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 156306656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 156406656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 156506656605SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 156606656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 15674f1b2e48SStefano Zampini PetscInt j; 15684f1b2e48SStefano Zampini 156906656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 157006656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*(i+n_vertices));CHKERRQ(ierr); 157106656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 157206656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 157306656605SStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 15744f1b2e48SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) y[n_D*i+p0_lidx_I[j]] = 0.0; 157506656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 157606656605SStefano Zampini } 157706656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 157806656605SStefano Zampini } 157906656605SStefano Zampini } 158080677318SStefano Zampini if (n_constraints) { 158180677318SStefano Zampini ierr = MatDestroy(&local_auxmat2_R);CHKERRQ(ierr); 158280677318SStefano Zampini } 15834f1b2e48SStefano Zampini ierr = PetscFree(p0_lidx_I);CHKERRQ(ierr); 1584019a44ceSStefano Zampini ierr = MatDestroy(&B0_V);CHKERRQ(ierr); 1585019a44ceSStefano Zampini 158606656605SStefano Zampini /* compute other basis functions for non-symmetric problems */ 15873301b35fSStefano Zampini if (!pcbddc->symmetric_primal) { 1588ffd830a3SStefano Zampini Mat B_V=NULL,B_C=NULL; 1589ffd830a3SStefano Zampini PetscScalar *marray; 159006656605SStefano Zampini 159106656605SStefano Zampini if (n_constraints) { 1592ffd830a3SStefano Zampini Mat S_CCT,C_CRT; 159306656605SStefano Zampini 1594ffd830a3SStefano Zampini ierr = MatTranspose(C_CR,MAT_INITIAL_MATRIX,&C_CRT);CHKERRQ(ierr); 159506656605SStefano Zampini ierr = MatTranspose(S_CC,MAT_INITIAL_MATRIX,&S_CCT);CHKERRQ(ierr); 1596ffd830a3SStefano Zampini ierr = MatMatMult(C_CRT,S_CCT,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&B_C);CHKERRQ(ierr); 159716f15bc4SStefano Zampini ierr = MatDestroy(&S_CCT);CHKERRQ(ierr); 159806656605SStefano Zampini if (n_vertices) { 1599ffd830a3SStefano Zampini Mat S_VCT; 160006656605SStefano Zampini 160106656605SStefano Zampini ierr = MatTranspose(S_VC,MAT_INITIAL_MATRIX,&S_VCT);CHKERRQ(ierr); 1602ffd830a3SStefano Zampini ierr = MatMatMult(C_CRT,S_VCT,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&B_V);CHKERRQ(ierr); 160316f15bc4SStefano Zampini ierr = MatDestroy(&S_VCT);CHKERRQ(ierr); 160406656605SStefano Zampini } 1605ffd830a3SStefano Zampini ierr = MatDestroy(&C_CRT);CHKERRQ(ierr); 160606656605SStefano Zampini } 160716f15bc4SStefano Zampini if (n_vertices && n_R) { 1608ffd830a3SStefano Zampini PetscScalar *av,*marray; 1609ffd830a3SStefano Zampini const PetscInt *xadj,*adjncy; 1610ffd830a3SStefano Zampini PetscInt n; 1611ffd830a3SStefano Zampini PetscBool flg_row; 161206656605SStefano Zampini 1613ffd830a3SStefano Zampini /* B_V = B_V - A_VR^T */ 1614ffd830a3SStefano Zampini ierr = MatConvert(A_VR,MATSEQAIJ,MAT_REUSE_MATRIX,&A_VR);CHKERRQ(ierr); 1615ffd830a3SStefano Zampini ierr = MatGetRowIJ(A_VR,0,PETSC_FALSE,PETSC_FALSE,&n,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 1616ffd830a3SStefano Zampini ierr = MatSeqAIJGetArray(A_VR,&av);CHKERRQ(ierr); 1617ffd830a3SStefano Zampini ierr = MatDenseGetArray(B_V,&marray);CHKERRQ(ierr); 1618ffd830a3SStefano Zampini for (i=0;i<n;i++) { 1619ffd830a3SStefano Zampini PetscInt j; 1620ffd830a3SStefano Zampini for (j=xadj[i];j<xadj[i+1];j++) marray[i*n_R + adjncy[j]] -= av[j]; 1621ffd830a3SStefano Zampini } 1622ffd830a3SStefano Zampini ierr = MatDenseRestoreArray(B_V,&marray);CHKERRQ(ierr); 1623ffd830a3SStefano Zampini ierr = MatRestoreRowIJ(A_VR,0,PETSC_FALSE,PETSC_FALSE,&n,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 1624ffd830a3SStefano Zampini ierr = MatDestroy(&A_VR);CHKERRQ(ierr); 162506656605SStefano Zampini } 162606656605SStefano Zampini 1627ffd830a3SStefano Zampini /* currently there's no support for MatTransposeMatSolve(F,B,X) */ 1628ffd830a3SStefano Zampini ierr = MatDenseGetArray(B_V,&marray);CHKERRQ(ierr); 1629ffd830a3SStefano Zampini for (i=0;i<n_vertices;i++) { 1630ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,marray+i*n_R);CHKERRQ(ierr); 1631ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,work+i*n_R);CHKERRQ(ierr); 163206656605SStefano Zampini ierr = KSPSolveTranspose(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 163306656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 163406656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 163506656605SStefano Zampini } 1636ffd830a3SStefano Zampini ierr = MatDenseRestoreArray(B_V,&marray);CHKERRQ(ierr); 1637ffd830a3SStefano Zampini ierr = MatDenseGetArray(B_C,&marray);CHKERRQ(ierr); 1638ffd830a3SStefano Zampini for (i=n_vertices;i<n_constraints+n_vertices;i++) { 1639ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,marray+(i-n_vertices)*n_R);CHKERRQ(ierr); 1640ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,work+i*n_R);CHKERRQ(ierr); 1641ffd830a3SStefano Zampini ierr = KSPSolveTranspose(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 1642ffd830a3SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 1643ffd830a3SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 164406656605SStefano Zampini } 1645ffd830a3SStefano Zampini ierr = MatDenseRestoreArray(B_C,&marray);CHKERRQ(ierr); 164606656605SStefano Zampini /* coarse basis functions */ 164706656605SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 164806656605SStefano Zampini PetscScalar *y; 164906656605SStefano Zampini 1650ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*n_R);CHKERRQ(ierr); 165106656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_psi_B,&y);CHKERRQ(ierr); 165206656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*i);CHKERRQ(ierr); 165306656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 165406656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 165506656605SStefano Zampini if (i<n_vertices) { 165606656605SStefano Zampini y[n_B*i+idx_V_B[i]] = 1.0; 165706656605SStefano Zampini } 165806656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_psi_B,&y);CHKERRQ(ierr); 165906656605SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 166006656605SStefano Zampini 166106656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 166206656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_psi_D,&y);CHKERRQ(ierr); 166306656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*i);CHKERRQ(ierr); 166406656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 166506656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 166606656605SStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 166706656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_psi_D,&y);CHKERRQ(ierr); 166806656605SStefano Zampini } 166906656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 167006656605SStefano Zampini } 1671ffd830a3SStefano Zampini ierr = MatDestroy(&B_V);CHKERRQ(ierr); 1672ffd830a3SStefano Zampini ierr = MatDestroy(&B_C);CHKERRQ(ierr); 167306656605SStefano Zampini } 1674d62866d3SStefano Zampini /* free memory */ 167588ebb749SStefano Zampini ierr = PetscFree(idx_V_B);CHKERRQ(ierr); 167606656605SStefano Zampini ierr = MatDestroy(&S_VV);CHKERRQ(ierr); 167706656605SStefano Zampini ierr = MatDestroy(&S_CV);CHKERRQ(ierr); 167806656605SStefano Zampini ierr = MatDestroy(&S_VC);CHKERRQ(ierr); 167906656605SStefano Zampini ierr = MatDestroy(&S_CC);CHKERRQ(ierr); 1680d62866d3SStefano Zampini ierr = PetscFree(work);CHKERRQ(ierr); 1681d62866d3SStefano Zampini if (n_vertices) { 1682d62866d3SStefano Zampini ierr = MatDestroy(&A_VR);CHKERRQ(ierr); 1683d62866d3SStefano Zampini } 1684d62866d3SStefano Zampini if (n_constraints) { 1685d62866d3SStefano Zampini ierr = MatDestroy(&C_CR);CHKERRQ(ierr); 1686d62866d3SStefano Zampini } 168788ebb749SStefano Zampini /* Checking coarse_sub_mat and coarse basis functios */ 168888ebb749SStefano Zampini /* Symmetric case : It should be \Phi^{(j)^T} A^{(j)} \Phi^{(j)}=coarse_sub_mat */ 168988ebb749SStefano Zampini /* Non-symmetric case : It should be \Psi^{(j)^T} A^{(j)} \Phi^{(j)}=coarse_sub_mat */ 1690d12edf2fSStefano Zampini if (pcbddc->dbg_flag) { 169188ebb749SStefano Zampini Mat coarse_sub_mat; 169225084f0cSStefano Zampini Mat AUXMAT,TM1,TM2,TM3,TM4; 169388ebb749SStefano Zampini Mat coarse_phi_D,coarse_phi_B; 169488ebb749SStefano Zampini Mat coarse_psi_D,coarse_psi_B; 169588ebb749SStefano Zampini Mat A_II,A_BB,A_IB,A_BI; 16968bec7fa6SStefano Zampini Mat C_B,CPHI; 16978bec7fa6SStefano Zampini IS is_dummy; 16988bec7fa6SStefano Zampini Vec mones; 169988ebb749SStefano Zampini MatType checkmattype=MATSEQAIJ; 170088ebb749SStefano Zampini PetscReal real_value; 170188ebb749SStefano Zampini 170288ebb749SStefano Zampini ierr = MatConvert(pcis->A_II,checkmattype,MAT_INITIAL_MATRIX,&A_II);CHKERRQ(ierr); 170388ebb749SStefano Zampini ierr = MatConvert(pcis->A_IB,checkmattype,MAT_INITIAL_MATRIX,&A_IB);CHKERRQ(ierr); 170488ebb749SStefano Zampini ierr = MatConvert(pcis->A_BI,checkmattype,MAT_INITIAL_MATRIX,&A_BI);CHKERRQ(ierr); 170588ebb749SStefano Zampini ierr = MatConvert(pcis->A_BB,checkmattype,MAT_INITIAL_MATRIX,&A_BB);CHKERRQ(ierr); 170688ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_phi_D,checkmattype,MAT_INITIAL_MATRIX,&coarse_phi_D);CHKERRQ(ierr); 170788ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_phi_B,checkmattype,MAT_INITIAL_MATRIX,&coarse_phi_B);CHKERRQ(ierr); 1708ffd830a3SStefano Zampini if (!pcbddc->symmetric_primal) { 170988ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_psi_D,checkmattype,MAT_INITIAL_MATRIX,&coarse_psi_D);CHKERRQ(ierr); 171088ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_psi_B,checkmattype,MAT_INITIAL_MATRIX,&coarse_psi_B);CHKERRQ(ierr); 171188ebb749SStefano Zampini } 171288ebb749SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,pcbddc->local_primal_size,pcbddc->local_primal_size,coarse_submat_vals,&coarse_sub_mat);CHKERRQ(ierr); 171388ebb749SStefano Zampini 171425084f0cSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 17153301b35fSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Check coarse sub mat computation (symmetric %d)\n",pcbddc->symmetric_primal);CHKERRQ(ierr); 171625084f0cSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1717ffd830a3SStefano Zampini if (!pcbddc->symmetric_primal) { 171888ebb749SStefano Zampini ierr = MatMatMult(A_II,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 171988ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM1);CHKERRQ(ierr); 172088ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 172188ebb749SStefano Zampini ierr = MatMatMult(A_BB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 172288ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM2);CHKERRQ(ierr); 172388ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 172488ebb749SStefano Zampini ierr = MatMatMult(A_IB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 172588ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM3);CHKERRQ(ierr); 172688ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 172788ebb749SStefano Zampini ierr = MatMatMult(A_BI,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 172888ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM4);CHKERRQ(ierr); 172988ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 173088ebb749SStefano Zampini } else { 173188ebb749SStefano Zampini ierr = MatPtAP(A_II,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&TM1);CHKERRQ(ierr); 173288ebb749SStefano Zampini ierr = MatPtAP(A_BB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&TM2);CHKERRQ(ierr); 173388ebb749SStefano Zampini ierr = MatMatMult(A_IB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 173488ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_phi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM3);CHKERRQ(ierr); 173588ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 173688ebb749SStefano Zampini ierr = MatMatMult(A_BI,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 173788ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_phi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM4);CHKERRQ(ierr); 173888ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 173988ebb749SStefano Zampini } 174088ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM2,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 174188ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM3,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 174288ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM4,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 174388ebb749SStefano Zampini ierr = MatConvert(TM1,MATSEQDENSE,MAT_REUSE_MATRIX,&TM1);CHKERRQ(ierr); 17444f1b2e48SStefano Zampini if (pcbddc->benign_n) { 1745d12edf2fSStefano Zampini Mat B0_I,B0_B,B0_BPHI,B0_IPHI; 1746d12edf2fSStefano Zampini PetscScalar *data,*data2; 17474f1b2e48SStefano Zampini PetscInt j; 1748d12edf2fSStefano Zampini 17494f1b2e48SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcbddc->benign_n,0,1,&is_dummy);CHKERRQ(ierr); 17504f1b2e48SStefano Zampini ierr = MatGetSubMatrix(pcbddc->benign_B0,is_dummy,pcis->is_B_local,MAT_INITIAL_MATRIX,&B0_B); 17514f1b2e48SStefano Zampini ierr = MatGetSubMatrix(pcbddc->benign_B0,is_dummy,pcis->is_I_local,MAT_INITIAL_MATRIX,&B0_I); 1752d12edf2fSStefano Zampini ierr = MatMatMult(B0_B,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&B0_BPHI);CHKERRQ(ierr); 1753d12edf2fSStefano Zampini ierr = MatConvert(B0_BPHI,MATSEQDENSE,MAT_REUSE_MATRIX,&B0_BPHI);CHKERRQ(ierr); 1754d12edf2fSStefano Zampini ierr = MatDenseGetArray(TM1,&data);CHKERRQ(ierr); 1755d12edf2fSStefano Zampini ierr = MatDenseGetArray(B0_BPHI,&data2);CHKERRQ(ierr); 17564f1b2e48SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) { 17574f1b2e48SStefano Zampini PetscInt primal_idx = pcbddc->local_primal_size - pcbddc->benign_n + j; 1758d12edf2fSStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 17594f1b2e48SStefano Zampini data[primal_idx*pcbddc->local_primal_size+i] += data2[i*pcbddc->benign_n+j]; 17604f1b2e48SStefano Zampini data[i*pcbddc->local_primal_size+primal_idx] += data2[i*pcbddc->benign_n+j]; 17614f1b2e48SStefano Zampini } 1762d12edf2fSStefano Zampini } 1763d12edf2fSStefano Zampini ierr = MatDenseRestoreArray(TM1,&data);CHKERRQ(ierr); 1764d12edf2fSStefano Zampini ierr = MatDenseRestoreArray(B0_BPHI,&data2);CHKERRQ(ierr); 1765d12edf2fSStefano Zampini ierr = MatDestroy(&B0_B);CHKERRQ(ierr); 1766d12edf2fSStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 1767d12edf2fSStefano Zampini ierr = MatDestroy(&B0_BPHI);CHKERRQ(ierr); 1768d12edf2fSStefano Zampini ierr = MatMatMult(B0_I,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&B0_IPHI);CHKERRQ(ierr); 1769d12edf2fSStefano Zampini ierr = MatDestroy(&B0_I);CHKERRQ(ierr); 1770d12edf2fSStefano Zampini ierr = MatNorm(B0_IPHI,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 1771d12edf2fSStefano Zampini ierr = MatDestroy(&B0_IPHI);CHKERRQ(ierr); 1772d12edf2fSStefano Zampini } 1773d12edf2fSStefano Zampini #if 0 1774d12edf2fSStefano Zampini { 1775d12edf2fSStefano Zampini PetscViewer viewer; 1776d12edf2fSStefano Zampini char filename[256]; 1777ffd830a3SStefano Zampini sprintf(filename,"details_local_coarse_mat%d_level%d.m",PetscGlobalRank,pcbddc->current_level); 1778d12edf2fSStefano Zampini ierr = PetscViewerASCIIOpen(PETSC_COMM_SELF,filename,&viewer);CHKERRQ(ierr); 1779d12edf2fSStefano Zampini ierr = PetscViewerSetFormat(viewer,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 1780ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)coarse_sub_mat,"computed");CHKERRQ(ierr); 1781ffd830a3SStefano Zampini ierr = MatView(coarse_sub_mat,viewer);CHKERRQ(ierr); 1782ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)TM1,"projected");CHKERRQ(ierr); 1783d12edf2fSStefano Zampini ierr = MatView(TM1,viewer);CHKERRQ(ierr); 1784ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)pcbddc->coarse_phi_B,"phi_B");CHKERRQ(ierr); 1785ffd830a3SStefano Zampini ierr = MatView(pcbddc->coarse_phi_B,viewer);CHKERRQ(ierr); 1786ffd830a3SStefano Zampini if (pcbddc->coarse_phi_D) { 1787ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)pcbddc->coarse_phi_D,"phi_D");CHKERRQ(ierr); 1788ffd830a3SStefano Zampini ierr = MatView(pcbddc->coarse_phi_D,viewer);CHKERRQ(ierr); 1789ffd830a3SStefano Zampini } 1790ffd830a3SStefano Zampini if (pcbddc->coarse_psi_B) { 1791ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)pcbddc->coarse_psi_B,"psi_B");CHKERRQ(ierr); 1792ffd830a3SStefano Zampini ierr = MatView(pcbddc->coarse_psi_B,viewer);CHKERRQ(ierr); 1793ffd830a3SStefano Zampini } 1794ffd830a3SStefano Zampini if (pcbddc->coarse_phi_B) { 1795ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)pcbddc->coarse_psi_D,"psi_D");CHKERRQ(ierr); 1796ffd830a3SStefano Zampini ierr = MatView(pcbddc->coarse_psi_D,viewer);CHKERRQ(ierr); 1797ffd830a3SStefano Zampini } 1798d12edf2fSStefano Zampini ierr = PetscViewerDestroy(&viewer);CHKERRQ(ierr); 1799d12edf2fSStefano Zampini } 1800d12edf2fSStefano Zampini #endif 180181d9aea3SBarry Smith ierr = MatAXPY(TM1,m_one,coarse_sub_mat,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 18028bec7fa6SStefano Zampini ierr = MatNorm(TM1,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 18031575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 180406656605SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d matrix error % 1.14e\n",PetscGlobalRank,real_value);CHKERRQ(ierr); 18058bec7fa6SStefano Zampini 18068bec7fa6SStefano Zampini /* check constraints */ 18074f1b2e48SStefano Zampini if (!pcbddc->benign_n) { /* TODO: add benign case */ 18088bec7fa6SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcbddc->local_primal_size,0,1,&is_dummy);CHKERRQ(ierr); 18098bec7fa6SStefano Zampini ierr = MatGetSubMatrix(pcbddc->ConstraintMatrix,is_dummy,pcis->is_B_local,MAT_INITIAL_MATRIX,&C_B); 18108bec7fa6SStefano Zampini ierr = MatMatMult(C_B,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&CPHI);CHKERRQ(ierr); 18118bec7fa6SStefano Zampini ierr = MatCreateVecs(CPHI,&mones,NULL);CHKERRQ(ierr); 18128bec7fa6SStefano Zampini ierr = VecSet(mones,-1.0);CHKERRQ(ierr); 18138bec7fa6SStefano Zampini ierr = MatDiagonalSet(CPHI,mones,ADD_VALUES);CHKERRQ(ierr); 18148bec7fa6SStefano Zampini ierr = MatNorm(CPHI,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 1815bdae7319SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d phi constraints error % 1.14e\n",PetscGlobalRank,real_value);CHKERRQ(ierr); 1816ffd830a3SStefano Zampini if (!pcbddc->symmetric_primal) { 1817bdae7319SStefano Zampini ierr = MatMatMult(C_B,coarse_psi_B,MAT_REUSE_MATRIX,1.0,&CPHI);CHKERRQ(ierr); 1818bdae7319SStefano Zampini ierr = VecSet(mones,-1.0);CHKERRQ(ierr); 1819bdae7319SStefano Zampini ierr = MatDiagonalSet(CPHI,mones,ADD_VALUES);CHKERRQ(ierr); 1820bdae7319SStefano Zampini ierr = MatNorm(CPHI,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 1821bdae7319SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d psi constraints error % 1.14e\n",PetscGlobalRank,real_value);CHKERRQ(ierr); 182288ebb749SStefano Zampini } 18238bec7fa6SStefano Zampini ierr = MatDestroy(&C_B);CHKERRQ(ierr); 18248bec7fa6SStefano Zampini ierr = MatDestroy(&CPHI);CHKERRQ(ierr); 18258bec7fa6SStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 18268bec7fa6SStefano Zampini ierr = VecDestroy(&mones);CHKERRQ(ierr); 1827d12edf2fSStefano Zampini } 182825084f0cSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 182988ebb749SStefano Zampini ierr = MatDestroy(&A_II);CHKERRQ(ierr); 183088ebb749SStefano Zampini ierr = MatDestroy(&A_BB);CHKERRQ(ierr); 183188ebb749SStefano Zampini ierr = MatDestroy(&A_IB);CHKERRQ(ierr); 183288ebb749SStefano Zampini ierr = MatDestroy(&A_BI);CHKERRQ(ierr); 183388ebb749SStefano Zampini ierr = MatDestroy(&TM1);CHKERRQ(ierr); 183488ebb749SStefano Zampini ierr = MatDestroy(&TM2);CHKERRQ(ierr); 183588ebb749SStefano Zampini ierr = MatDestroy(&TM3);CHKERRQ(ierr); 183688ebb749SStefano Zampini ierr = MatDestroy(&TM4);CHKERRQ(ierr); 183788ebb749SStefano Zampini ierr = MatDestroy(&coarse_phi_D);CHKERRQ(ierr); 183888ebb749SStefano Zampini ierr = MatDestroy(&coarse_phi_B);CHKERRQ(ierr); 1839ffd830a3SStefano Zampini if (!pcbddc->symmetric_primal) { 184088ebb749SStefano Zampini ierr = MatDestroy(&coarse_psi_D);CHKERRQ(ierr); 184188ebb749SStefano Zampini ierr = MatDestroy(&coarse_psi_B);CHKERRQ(ierr); 184288ebb749SStefano Zampini } 184388ebb749SStefano Zampini ierr = MatDestroy(&coarse_sub_mat);CHKERRQ(ierr); 184488ebb749SStefano Zampini } 18458629588bSStefano Zampini /* get back data */ 18468629588bSStefano Zampini *coarse_submat_vals_n = coarse_submat_vals; 184788ebb749SStefano Zampini PetscFunctionReturn(0); 184888ebb749SStefano Zampini } 184988ebb749SStefano Zampini 185088ebb749SStefano Zampini #undef __FUNCT__ 1851d65f70fdSStefano Zampini #define __FUNCT__ "MatGetSubMatrixUnsorted" 1852d648f858SStefano Zampini PetscErrorCode MatGetSubMatrixUnsorted(Mat A, IS isrow, IS iscol, Mat* B) 1853aa0d41d4SStefano Zampini { 1854d65f70fdSStefano Zampini Mat *work_mat; 1855d65f70fdSStefano Zampini IS isrow_s,iscol_s; 1856d65f70fdSStefano Zampini PetscBool rsorted,csorted; 1857d65f70fdSStefano Zampini PetscInt rsize,*idxs_perm_r,csize,*idxs_perm_c; 1858aa0d41d4SStefano Zampini PetscErrorCode ierr; 1859aa0d41d4SStefano Zampini 1860aa0d41d4SStefano Zampini PetscFunctionBegin; 1861d65f70fdSStefano Zampini ierr = ISSorted(isrow,&rsorted);CHKERRQ(ierr); 1862d65f70fdSStefano Zampini ierr = ISSorted(iscol,&csorted);CHKERRQ(ierr); 1863d65f70fdSStefano Zampini ierr = ISGetLocalSize(isrow,&rsize);CHKERRQ(ierr); 1864d65f70fdSStefano Zampini ierr = ISGetLocalSize(iscol,&csize);CHKERRQ(ierr); 1865aa0d41d4SStefano Zampini 1866d65f70fdSStefano Zampini if (!rsorted) { 1867906d46d4SStefano Zampini const PetscInt *idxs; 1868906d46d4SStefano Zampini PetscInt *idxs_sorted,i; 1869aa0d41d4SStefano Zampini 1870d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_perm_r);CHKERRQ(ierr); 1871d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_sorted);CHKERRQ(ierr); 1872d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 1873d65f70fdSStefano Zampini idxs_perm_r[i] = i; 1874aa0d41d4SStefano Zampini } 1875d65f70fdSStefano Zampini ierr = ISGetIndices(isrow,&idxs);CHKERRQ(ierr); 1876d65f70fdSStefano Zampini ierr = PetscSortIntWithPermutation(rsize,idxs,idxs_perm_r);CHKERRQ(ierr); 1877d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 1878d65f70fdSStefano Zampini idxs_sorted[i] = idxs[idxs_perm_r[i]]; 1879aa0d41d4SStefano Zampini } 1880d65f70fdSStefano Zampini ierr = ISRestoreIndices(isrow,&idxs);CHKERRQ(ierr); 1881d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,rsize,idxs_sorted,PETSC_OWN_POINTER,&isrow_s);CHKERRQ(ierr); 1882d65f70fdSStefano Zampini } else { 1883d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)isrow);CHKERRQ(ierr); 1884d65f70fdSStefano Zampini isrow_s = isrow; 1885aa0d41d4SStefano Zampini } 1886906d46d4SStefano Zampini 1887d65f70fdSStefano Zampini if (!csorted) { 1888d65f70fdSStefano Zampini if (isrow == iscol) { 1889d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)isrow_s);CHKERRQ(ierr); 1890d65f70fdSStefano Zampini iscol_s = isrow_s; 1891d65f70fdSStefano Zampini } else { 1892d65f70fdSStefano Zampini const PetscInt *idxs; 1893d65f70fdSStefano Zampini PetscInt *idxs_sorted,i; 1894906d46d4SStefano Zampini 1895d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_perm_c);CHKERRQ(ierr); 1896d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_sorted);CHKERRQ(ierr); 1897d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 1898d65f70fdSStefano Zampini idxs_perm_c[i] = i; 1899d65f70fdSStefano Zampini } 1900d65f70fdSStefano Zampini ierr = ISGetIndices(iscol,&idxs);CHKERRQ(ierr); 1901d65f70fdSStefano Zampini ierr = PetscSortIntWithPermutation(csize,idxs,idxs_perm_c);CHKERRQ(ierr); 1902d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 1903d65f70fdSStefano Zampini idxs_sorted[i] = idxs[idxs_perm_c[i]]; 1904d65f70fdSStefano Zampini } 1905d65f70fdSStefano Zampini ierr = ISRestoreIndices(iscol,&idxs);CHKERRQ(ierr); 1906d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,csize,idxs_sorted,PETSC_OWN_POINTER,&iscol_s);CHKERRQ(ierr); 1907d65f70fdSStefano Zampini } 1908d65f70fdSStefano Zampini } else { 1909d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)iscol);CHKERRQ(ierr); 1910d65f70fdSStefano Zampini iscol_s = iscol; 1911d65f70fdSStefano Zampini } 1912d65f70fdSStefano Zampini 1913d648f858SStefano Zampini ierr = MatGetSubMatrices(A,1,&isrow_s,&iscol_s,MAT_INITIAL_MATRIX,&work_mat);CHKERRQ(ierr); 1914d65f70fdSStefano Zampini 1915d65f70fdSStefano Zampini if (!rsorted || !csorted) { 1916906d46d4SStefano Zampini Mat new_mat; 1917d65f70fdSStefano Zampini IS is_perm_r,is_perm_c; 1918906d46d4SStefano Zampini 1919d65f70fdSStefano Zampini if (!rsorted) { 1920d65f70fdSStefano Zampini PetscInt *idxs_r,i; 1921d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_r);CHKERRQ(ierr); 1922d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 1923d65f70fdSStefano Zampini idxs_r[idxs_perm_r[i]] = i; 1924906d46d4SStefano Zampini } 1925d65f70fdSStefano Zampini ierr = PetscFree(idxs_perm_r);CHKERRQ(ierr); 1926d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,rsize,idxs_r,PETSC_OWN_POINTER,&is_perm_r);CHKERRQ(ierr); 1927d65f70fdSStefano Zampini } else { 1928d65f70fdSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,rsize,0,1,&is_perm_r);CHKERRQ(ierr); 1929906d46d4SStefano Zampini } 1930d65f70fdSStefano Zampini ierr = ISSetPermutation(is_perm_r);CHKERRQ(ierr); 1931d65f70fdSStefano Zampini 1932d65f70fdSStefano Zampini if (!csorted) { 1933d65f70fdSStefano Zampini if (isrow_s == iscol_s) { 1934d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)is_perm_r);CHKERRQ(ierr); 1935d65f70fdSStefano Zampini is_perm_c = is_perm_r; 1936d65f70fdSStefano Zampini } else { 1937d65f70fdSStefano Zampini PetscInt *idxs_c,i; 1938d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_c);CHKERRQ(ierr); 1939d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 1940d65f70fdSStefano Zampini idxs_c[idxs_perm_c[i]] = i; 1941d65f70fdSStefano Zampini } 1942d65f70fdSStefano Zampini ierr = PetscFree(idxs_perm_c);CHKERRQ(ierr); 1943d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,csize,idxs_c,PETSC_OWN_POINTER,&is_perm_c);CHKERRQ(ierr); 1944d65f70fdSStefano Zampini } 1945d65f70fdSStefano Zampini } else { 1946d65f70fdSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,csize,0,1,&is_perm_c);CHKERRQ(ierr); 1947d65f70fdSStefano Zampini } 1948d65f70fdSStefano Zampini ierr = ISSetPermutation(is_perm_c);CHKERRQ(ierr); 1949d65f70fdSStefano Zampini 1950d65f70fdSStefano Zampini ierr = MatPermute(work_mat[0],is_perm_r,is_perm_c,&new_mat);CHKERRQ(ierr); 1951d65f70fdSStefano Zampini ierr = MatDestroy(&work_mat[0]);CHKERRQ(ierr); 1952d65f70fdSStefano Zampini work_mat[0] = new_mat; 1953d65f70fdSStefano Zampini ierr = ISDestroy(&is_perm_r);CHKERRQ(ierr); 1954d65f70fdSStefano Zampini ierr = ISDestroy(&is_perm_c);CHKERRQ(ierr); 1955d65f70fdSStefano Zampini } 1956d65f70fdSStefano Zampini 1957d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)work_mat[0]);CHKERRQ(ierr); 1958d65f70fdSStefano Zampini *B = work_mat[0]; 1959d65f70fdSStefano Zampini ierr = MatDestroyMatrices(1,&work_mat);CHKERRQ(ierr); 1960d65f70fdSStefano Zampini ierr = ISDestroy(&isrow_s);CHKERRQ(ierr); 1961d65f70fdSStefano Zampini ierr = ISDestroy(&iscol_s);CHKERRQ(ierr); 1962d65f70fdSStefano Zampini PetscFunctionReturn(0); 1963d65f70fdSStefano Zampini } 1964d65f70fdSStefano Zampini 1965d65f70fdSStefano Zampini #undef __FUNCT__ 19665e8657edSStefano Zampini #define __FUNCT__ "PCBDDCComputeLocalMatrix" 19675e8657edSStefano Zampini PetscErrorCode PCBDDCComputeLocalMatrix(PC pc, Mat ChangeOfBasisMatrix) 1968aa0d41d4SStefano Zampini { 1969aa0d41d4SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 19705e8657edSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 1971d65f70fdSStefano Zampini Mat new_mat; 19725e8657edSStefano Zampini IS is_local,is_global; 1973d65f70fdSStefano Zampini PetscInt local_size; 1974d65f70fdSStefano Zampini PetscBool isseqaij; 1975aa0d41d4SStefano Zampini PetscErrorCode ierr; 1976aa0d41d4SStefano Zampini 1977aa0d41d4SStefano Zampini PetscFunctionBegin; 1978aa0d41d4SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 19795e8657edSStefano Zampini ierr = MatGetSize(matis->A,&local_size,NULL);CHKERRQ(ierr); 19805e8657edSStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)matis->A),local_size,0,1,&is_local);CHKERRQ(ierr); 1981b087196eSStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(pc->pmat->rmap->mapping,is_local,&is_global);CHKERRQ(ierr); 1982aa0d41d4SStefano Zampini ierr = ISDestroy(&is_local);CHKERRQ(ierr); 1983d648f858SStefano Zampini ierr = MatGetSubMatrixUnsorted(ChangeOfBasisMatrix,is_global,is_global,&new_mat);CHKERRQ(ierr); 1984aa0d41d4SStefano Zampini ierr = ISDestroy(&is_global);CHKERRQ(ierr); 1985906d46d4SStefano Zampini 1986906d46d4SStefano Zampini /* check */ 1987906d46d4SStefano Zampini if (pcbddc->dbg_flag) { 1988906d46d4SStefano Zampini Vec x,x_change; 1989906d46d4SStefano Zampini PetscReal error; 1990906d46d4SStefano Zampini 19915e8657edSStefano Zampini ierr = MatCreateVecs(ChangeOfBasisMatrix,&x,&x_change);CHKERRQ(ierr); 1992906d46d4SStefano Zampini ierr = VecSetRandom(x,NULL);CHKERRQ(ierr); 19935e8657edSStefano Zampini ierr = MatMult(ChangeOfBasisMatrix,x,x_change);CHKERRQ(ierr); 1994e176bc59SStefano Zampini ierr = VecScatterBegin(matis->cctx,x,matis->x,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1995e176bc59SStefano Zampini ierr = VecScatterEnd(matis->cctx,x,matis->x,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1996d65f70fdSStefano Zampini ierr = MatMult(new_mat,matis->x,matis->y);CHKERRQ(ierr); 1997e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,matis->y,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1998e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,matis->y,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1999906d46d4SStefano Zampini ierr = VecAXPY(x,-1.0,x_change);CHKERRQ(ierr); 2000906d46d4SStefano Zampini ierr = VecNorm(x,NORM_INFINITY,&error);CHKERRQ(ierr); 2001906d46d4SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 2002906d46d4SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Error global vs local change on N: %1.6e\n",error);CHKERRQ(ierr); 2003906d46d4SStefano Zampini ierr = VecDestroy(&x);CHKERRQ(ierr); 2004906d46d4SStefano Zampini ierr = VecDestroy(&x_change);CHKERRQ(ierr); 2005906d46d4SStefano Zampini } 2006906d46d4SStefano Zampini 200722d5777bSStefano Zampini /* TODO: HOW TO WORK WITH BAIJ and SBAIJ and SEQDENSE? */ 20089b28b3ffSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)matis->A,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 200922d5777bSStefano Zampini if (isseqaij) { 20101cf9b237SStefano Zampini Mat M; 20111cf9b237SStefano Zampini 20121cf9b237SStefano Zampini ierr = MatPtAP(matis->A,new_mat,MAT_INITIAL_MATRIX,2.0,&M);CHKERRQ(ierr); 20131cf9b237SStefano Zampini ierr = MatSeqAIJCompress(M,&pcbddc->local_mat);CHKERRQ(ierr); 20141cf9b237SStefano Zampini ierr = MatDestroy(&M);CHKERRQ(ierr); 2015aa0d41d4SStefano Zampini } else { 20161cf9b237SStefano Zampini Mat work_mat,M; 20171cf9b237SStefano Zampini 2018aa0d41d4SStefano Zampini ierr = MatConvert(matis->A,MATSEQAIJ,MAT_INITIAL_MATRIX,&work_mat);CHKERRQ(ierr); 20191cf9b237SStefano Zampini ierr = MatPtAP(work_mat,new_mat,MAT_INITIAL_MATRIX,2.0,&M);CHKERRQ(ierr); 20201cf9b237SStefano Zampini ierr = MatSeqAIJCompress(M,&pcbddc->local_mat);CHKERRQ(ierr); 20211cf9b237SStefano Zampini ierr = MatDestroy(&M);CHKERRQ(ierr); 2022aa0d41d4SStefano Zampini ierr = MatDestroy(&work_mat);CHKERRQ(ierr); 2023aa0d41d4SStefano Zampini } 20243301b35fSStefano Zampini if (matis->A->symmetric_set) { 20253301b35fSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_SYMMETRIC,matis->A->symmetric);CHKERRQ(ierr); 2026e496cd5dSStefano Zampini #if !defined(PETSC_USE_COMPLEX) 20273301b35fSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_HERMITIAN,matis->A->symmetric);CHKERRQ(ierr); 2028e496cd5dSStefano Zampini #endif 20293301b35fSStefano Zampini } 203045a1bb75SStefano Zampini /* 203145a1bb75SStefano Zampini ierr = PetscViewerSetFormat(PETSC_VIEWER_STDOUT_SELF,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 2032d65f70fdSStefano Zampini ierr = MatView(new_mat,(PetscViewer)0);CHKERRQ(ierr); 203345a1bb75SStefano Zampini */ 2034d65f70fdSStefano Zampini ierr = MatDestroy(&new_mat);CHKERRQ(ierr); 2035aa0d41d4SStefano Zampini PetscFunctionReturn(0); 2036aa0d41d4SStefano Zampini } 2037aa0d41d4SStefano Zampini 2038aa0d41d4SStefano Zampini #undef __FUNCT__ 2039a64d13efSStefano Zampini #define __FUNCT__ "PCBDDCSetUpLocalScatters" 20408ce42a96SStefano Zampini PetscErrorCode PCBDDCSetUpLocalScatters(PC pc) 2041a64d13efSStefano Zampini { 2042a64d13efSStefano Zampini PC_IS* pcis = (PC_IS*)(pc->data); 2043a64d13efSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 2044d62866d3SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 204553892102SStefano Zampini PetscInt *idx_R_local=NULL; 20463a50541eSStefano Zampini PetscInt n_vertices,i,j,n_R,n_D,n_B; 20473a50541eSStefano Zampini PetscInt vbs,bs; 20486816873aSStefano Zampini PetscBT bitmask=NULL; 2049a64d13efSStefano Zampini PetscErrorCode ierr; 2050a64d13efSStefano Zampini 2051a64d13efSStefano Zampini PetscFunctionBegin; 2052b23d619eSStefano Zampini /* 2053b23d619eSStefano Zampini No need to setup local scatters if 2054b23d619eSStefano Zampini - primal space is unchanged 2055b23d619eSStefano Zampini AND 2056b23d619eSStefano Zampini - we actually have locally some primal dofs (could not be true in multilevel or for isolated subdomains) 2057b23d619eSStefano Zampini AND 2058b23d619eSStefano Zampini - we are not in debugging mode (this is needed since there are Synchronized prints at the end of the subroutine 2059b23d619eSStefano Zampini */ 2060b23d619eSStefano Zampini if (!pcbddc->new_primal_space_local && pcbddc->local_primal_size && !pcbddc->dbg_flag) { 2061f4ddd8eeSStefano Zampini PetscFunctionReturn(0); 2062f4ddd8eeSStefano Zampini } 2063f4ddd8eeSStefano Zampini /* destroy old objects */ 2064f4ddd8eeSStefano Zampini ierr = ISDestroy(&pcbddc->is_R_local);CHKERRQ(ierr); 2065f4ddd8eeSStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_B);CHKERRQ(ierr); 2066f4ddd8eeSStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_D);CHKERRQ(ierr); 2067a64d13efSStefano Zampini /* Set Non-overlapping dimensions */ 2068b371cd4fSStefano Zampini n_B = pcis->n_B; 2069b371cd4fSStefano Zampini n_D = pcis->n - n_B; 2070b371cd4fSStefano Zampini n_vertices = pcbddc->n_vertices; 20713a50541eSStefano Zampini 2072a64d13efSStefano Zampini /* Dohrmann's notation: dofs splitted in R (Remaining: all dofs but the vertices) and V (Vertices) */ 20736816873aSStefano Zampini 207453892102SStefano Zampini /* create auxiliary bitmask and allocate workspace */ 20756816873aSStefano Zampini if (!sub_schurs->reuse_mumps) { 2076854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n-n_vertices,&idx_R_local);CHKERRQ(ierr); 2077a64d13efSStefano Zampini ierr = PetscBTCreate(pcis->n,&bitmask);CHKERRQ(ierr); 2078a64d13efSStefano Zampini for (i=0;i<n_vertices;i++) { 20790e6343abSStefano Zampini ierr = PetscBTSet(bitmask,pcbddc->local_primal_ref_node[i]);CHKERRQ(ierr); 2080a64d13efSStefano Zampini } 2081a64d13efSStefano Zampini 2082a64d13efSStefano Zampini for (i=0, n_R=0; i<pcis->n; i++) { 20834641a718SStefano Zampini if (!PetscBTLookup(bitmask,i)) { 20846816873aSStefano Zampini idx_R_local[n_R++] = i; 2085a64d13efSStefano Zampini } 2086a64d13efSStefano Zampini } 208753892102SStefano Zampini } else { /* A different ordering (already computed) is present if we are reusing MUMPS Schur solver */ 20886816873aSStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 20896816873aSStefano Zampini 209053892102SStefano Zampini ierr = ISGetIndices(reuse_mumps->is_R,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 209153892102SStefano Zampini ierr = ISGetLocalSize(reuse_mumps->is_R,&n_R);CHKERRQ(ierr); 20926816873aSStefano Zampini } 20933a50541eSStefano Zampini 20943a50541eSStefano Zampini /* Block code */ 20953a50541eSStefano Zampini vbs = 1; 20963a50541eSStefano Zampini ierr = MatGetBlockSize(pcbddc->local_mat,&bs);CHKERRQ(ierr); 20973a50541eSStefano Zampini if (bs>1 && !(n_vertices%bs)) { 20983a50541eSStefano Zampini PetscBool is_blocked = PETSC_TRUE; 20993a50541eSStefano Zampini PetscInt *vary; 2100d3df7717SStefano Zampini if (!sub_schurs->reuse_mumps) { 2101785e854fSJed Brown ierr = PetscMalloc1(pcis->n/bs,&vary);CHKERRQ(ierr); 21023a50541eSStefano Zampini ierr = PetscMemzero(vary,pcis->n/bs*sizeof(PetscInt));CHKERRQ(ierr); 2103d3df7717SStefano Zampini /* Verify that the vertex indices correspond to each element in a block (code taken from sbaij2.c) */ 2104d3df7717SStefano 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 */ 21050e6343abSStefano Zampini for (i=0; i<n_vertices; i++) vary[pcbddc->local_primal_ref_node[i]/bs]++; 2106d3df7717SStefano Zampini for (i=0; i<pcis->n/bs; i++) { 21073a50541eSStefano Zampini if (vary[i]!=0 && vary[i]!=bs) { 21083a50541eSStefano Zampini is_blocked = PETSC_FALSE; 21093a50541eSStefano Zampini break; 21103a50541eSStefano Zampini } 21113a50541eSStefano Zampini } 2112d3df7717SStefano Zampini ierr = PetscFree(vary);CHKERRQ(ierr); 2113d3df7717SStefano Zampini } else { 2114d3df7717SStefano Zampini /* Verify directly the R set */ 2115d3df7717SStefano Zampini for (i=0; i<n_R/bs; i++) { 2116d3df7717SStefano Zampini PetscInt j,node=idx_R_local[bs*i]; 2117d3df7717SStefano Zampini for (j=1; j<bs; j++) { 2118d3df7717SStefano Zampini if (node != idx_R_local[bs*i+j]-j) { 2119d3df7717SStefano Zampini is_blocked = PETSC_FALSE; 2120d3df7717SStefano Zampini break; 2121d3df7717SStefano Zampini } 2122d3df7717SStefano Zampini } 2123d3df7717SStefano Zampini } 2124d3df7717SStefano Zampini } 21253a50541eSStefano Zampini if (is_blocked) { /* build compressed IS for R nodes (complement of vertices) */ 21263a50541eSStefano Zampini vbs = bs; 21273a50541eSStefano Zampini for (i=0;i<n_R/vbs;i++) { 21283a50541eSStefano Zampini idx_R_local[i] = idx_R_local[vbs*i]/vbs; 21293a50541eSStefano Zampini } 21303a50541eSStefano Zampini } 21313a50541eSStefano Zampini } 21323a50541eSStefano Zampini ierr = ISCreateBlock(PETSC_COMM_SELF,vbs,n_R/vbs,idx_R_local,PETSC_COPY_VALUES,&pcbddc->is_R_local);CHKERRQ(ierr); 213353892102SStefano Zampini if (sub_schurs->reuse_mumps) { 213453892102SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 213553892102SStefano Zampini 213653892102SStefano Zampini ierr = ISRestoreIndices(reuse_mumps->is_R,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 213753892102SStefano Zampini ierr = ISDestroy(&reuse_mumps->is_R);CHKERRQ(ierr); 213853892102SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->is_R_local);CHKERRQ(ierr); 213953892102SStefano Zampini reuse_mumps->is_R = pcbddc->is_R_local; 214053892102SStefano Zampini } else { 21413a50541eSStefano Zampini ierr = PetscFree(idx_R_local);CHKERRQ(ierr); 214253892102SStefano Zampini } 2143a64d13efSStefano Zampini 2144a64d13efSStefano Zampini /* print some info if requested */ 2145a64d13efSStefano Zampini if (pcbddc->dbg_flag) { 2146a64d13efSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 2147a64d13efSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 21481575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 2149a64d13efSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d local dimensions\n",PetscGlobalRank);CHKERRQ(ierr); 2150a64d13efSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"local_size = %d, dirichlet_size = %d, boundary_size = %d\n",pcis->n,n_D,n_B);CHKERRQ(ierr); 21514f1b2e48SStefano 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); 2152a64d13efSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 2153a64d13efSStefano Zampini } 2154a64d13efSStefano Zampini 2155a64d13efSStefano Zampini /* VecScatters pcbddc->R_to_B and (optionally) pcbddc->R_to_D */ 21566816873aSStefano Zampini if (!sub_schurs->reuse_mumps) { 21576816873aSStefano Zampini IS is_aux1,is_aux2; 21586816873aSStefano Zampini PetscInt *aux_array1,*aux_array2,*is_indices,*idx_R_local; 21596816873aSStefano Zampini 21603a50541eSStefano Zampini ierr = ISGetIndices(pcbddc->is_R_local,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 2161854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n_B-n_vertices,&aux_array1);CHKERRQ(ierr); 2162854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n_B-n_vertices,&aux_array2);CHKERRQ(ierr); 2163a64d13efSStefano Zampini ierr = ISGetIndices(pcis->is_I_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 21644641a718SStefano Zampini for (i=0; i<n_D; i++) { 21654641a718SStefano Zampini ierr = PetscBTSet(bitmask,is_indices[i]);CHKERRQ(ierr); 21664641a718SStefano Zampini } 2167a64d13efSStefano Zampini ierr = ISRestoreIndices(pcis->is_I_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 2168a64d13efSStefano Zampini for (i=0, j=0; i<n_R; i++) { 21694641a718SStefano Zampini if (!PetscBTLookup(bitmask,idx_R_local[i])) { 21704641a718SStefano Zampini aux_array1[j++] = i; 2171a64d13efSStefano Zampini } 2172a64d13efSStefano Zampini } 2173a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array1,PETSC_OWN_POINTER,&is_aux1);CHKERRQ(ierr); 2174a64d13efSStefano Zampini ierr = ISGetIndices(pcis->is_B_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 2175a64d13efSStefano Zampini for (i=0, j=0; i<n_B; i++) { 21764641a718SStefano Zampini if (!PetscBTLookup(bitmask,is_indices[i])) { 21774641a718SStefano Zampini aux_array2[j++] = i; 2178a64d13efSStefano Zampini } 2179a64d13efSStefano Zampini } 2180a64d13efSStefano Zampini ierr = ISRestoreIndices(pcis->is_B_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 2181a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array2,PETSC_OWN_POINTER,&is_aux2);CHKERRQ(ierr); 2182a64d13efSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,is_aux1,pcis->vec1_B,is_aux2,&pcbddc->R_to_B);CHKERRQ(ierr); 2183a64d13efSStefano Zampini ierr = ISDestroy(&is_aux1);CHKERRQ(ierr); 2184a64d13efSStefano Zampini ierr = ISDestroy(&is_aux2);CHKERRQ(ierr); 2185a64d13efSStefano Zampini 21868eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 2187785e854fSJed Brown ierr = PetscMalloc1(n_D,&aux_array1);CHKERRQ(ierr); 2188a64d13efSStefano Zampini for (i=0, j=0; i<n_R; i++) { 21894641a718SStefano Zampini if (PetscBTLookup(bitmask,idx_R_local[i])) { 21904641a718SStefano Zampini aux_array1[j++] = i; 2191a64d13efSStefano Zampini } 2192a64d13efSStefano Zampini } 2193a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array1,PETSC_OWN_POINTER,&is_aux1);CHKERRQ(ierr); 2194a64d13efSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,is_aux1,pcis->vec1_D,(IS)0,&pcbddc->R_to_D);CHKERRQ(ierr); 2195a64d13efSStefano Zampini ierr = ISDestroy(&is_aux1);CHKERRQ(ierr); 2196a64d13efSStefano Zampini } 21974641a718SStefano Zampini ierr = PetscBTDestroy(&bitmask);CHKERRQ(ierr); 21983a50541eSStefano Zampini ierr = ISRestoreIndices(pcbddc->is_R_local,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 2199d62866d3SStefano Zampini } else { 220053892102SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 22016816873aSStefano Zampini IS tis; 22026816873aSStefano Zampini PetscInt schur_size; 22036816873aSStefano Zampini 220453892102SStefano Zampini ierr = ISGetLocalSize(reuse_mumps->is_B,&schur_size);CHKERRQ(ierr); 22056816873aSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,schur_size,n_D,1,&tis);CHKERRQ(ierr); 220653892102SStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,tis,pcis->vec1_B,reuse_mumps->is_B,&pcbddc->R_to_B);CHKERRQ(ierr); 22076816873aSStefano Zampini ierr = ISDestroy(&tis);CHKERRQ(ierr); 22086816873aSStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 22096816873aSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,n_D,0,1,&tis);CHKERRQ(ierr); 22106816873aSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,tis,pcis->vec1_D,(IS)0,&pcbddc->R_to_D);CHKERRQ(ierr); 22116816873aSStefano Zampini ierr = ISDestroy(&tis);CHKERRQ(ierr); 2212d62866d3SStefano Zampini } 2213d62866d3SStefano Zampini } 2214a64d13efSStefano Zampini PetscFunctionReturn(0); 2215a64d13efSStefano Zampini } 2216a64d13efSStefano Zampini 2217304d26faSStefano Zampini 2218304d26faSStefano Zampini #undef __FUNCT__ 2219304d26faSStefano Zampini #define __FUNCT__ "PCBDDCSetUpLocalSolvers" 2220684f6988SStefano Zampini PetscErrorCode PCBDDCSetUpLocalSolvers(PC pc, PetscBool dirichlet, PetscBool neumann) 2221304d26faSStefano Zampini { 2222304d26faSStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 2223304d26faSStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 2224304d26faSStefano Zampini PC pc_temp; 2225304d26faSStefano Zampini Mat A_RR; 2226f4ddd8eeSStefano Zampini MatReuse reuse; 2227304d26faSStefano Zampini PetscScalar m_one = -1.0; 2228304d26faSStefano Zampini PetscReal value; 222904708bb6SStefano Zampini PetscInt n_D,n_R; 22309577ea80SStefano Zampini PetscBool use_exact,use_exact_reduced,issbaij; 2231304d26faSStefano Zampini PetscErrorCode ierr; 2232e604994aSStefano Zampini /* prefixes stuff */ 2233312be037SStefano Zampini char dir_prefix[256],neu_prefix[256],str_level[16]; 2234e604994aSStefano Zampini size_t len; 2235304d26faSStefano Zampini 2236304d26faSStefano Zampini PetscFunctionBegin; 2237304d26faSStefano Zampini 2238e604994aSStefano Zampini /* compute prefixes */ 2239e604994aSStefano Zampini ierr = PetscStrcpy(dir_prefix,"");CHKERRQ(ierr); 2240e604994aSStefano Zampini ierr = PetscStrcpy(neu_prefix,"");CHKERRQ(ierr); 2241e604994aSStefano Zampini if (!pcbddc->current_level) { 2242e604994aSStefano Zampini ierr = PetscStrcpy(dir_prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 2243e604994aSStefano Zampini ierr = PetscStrcpy(neu_prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 2244e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,"pc_bddc_dirichlet_");CHKERRQ(ierr); 2245e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,"pc_bddc_neumann_");CHKERRQ(ierr); 2246e604994aSStefano Zampini } else { 2247e604994aSStefano Zampini ierr = PetscStrcpy(str_level,"");CHKERRQ(ierr); 2248312be037SStefano Zampini sprintf(str_level,"l%d_",(int)(pcbddc->current_level)); 2249e604994aSStefano Zampini ierr = PetscStrlen(((PetscObject)pc)->prefix,&len);CHKERRQ(ierr); 2250e604994aSStefano Zampini len -= 15; /* remove "pc_bddc_coarse_" */ 2251312be037SStefano Zampini if (pcbddc->current_level>1) len -= 3; /* remove "lX_" with X level number */ 2252312be037SStefano Zampini if (pcbddc->current_level>10) len -= 1; /* remove another char from level number */ 225334d6797cSStefano Zampini ierr = PetscStrncpy(dir_prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 225434d6797cSStefano Zampini ierr = PetscStrncpy(neu_prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 2255e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,"pc_bddc_dirichlet_");CHKERRQ(ierr); 2256e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,"pc_bddc_neumann_");CHKERRQ(ierr); 2257e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,str_level);CHKERRQ(ierr); 2258e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,str_level);CHKERRQ(ierr); 2259e604994aSStefano Zampini } 2260e604994aSStefano Zampini 2261304d26faSStefano Zampini /* DIRICHLET PROBLEM */ 2262684f6988SStefano Zampini if (dirichlet) { 2263d5574798SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 22643301b35fSStefano Zampini if (pcbddc->local_mat->symmetric_set) { 22653301b35fSStefano Zampini ierr = MatSetOption(pcis->A_II,MAT_SYMMETRIC,pcbddc->local_mat->symmetric_set);CHKERRQ(ierr); 2266964fefecSStefano Zampini } 2267ac78edfcSStefano Zampini /* Matrix for Dirichlet problem is pcis->A_II */ 2268964fefecSStefano Zampini n_D = pcis->n - pcis->n_B; 2269304d26faSStefano Zampini if (!pcbddc->ksp_D) { /* create object if not yet build */ 2270304d26faSStefano Zampini ierr = KSPCreate(PETSC_COMM_SELF,&pcbddc->ksp_D);CHKERRQ(ierr); 2271304d26faSStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->ksp_D,(PetscObject)pc,1);CHKERRQ(ierr); 2272304d26faSStefano Zampini /* default */ 2273304d26faSStefano Zampini ierr = KSPSetType(pcbddc->ksp_D,KSPPREONLY);CHKERRQ(ierr); 2274e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->ksp_D,dir_prefix);CHKERRQ(ierr); 22759577ea80SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcis->A_II,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 2276304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_D,&pc_temp);CHKERRQ(ierr); 22779577ea80SStefano Zampini if (issbaij) { 22789577ea80SStefano Zampini ierr = PCSetType(pc_temp,PCCHOLESKY);CHKERRQ(ierr); 22799577ea80SStefano Zampini } else { 2280304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCLU);CHKERRQ(ierr); 22819577ea80SStefano Zampini } 2282304d26faSStefano Zampini /* Allow user's customization */ 2283304d26faSStefano Zampini ierr = KSPSetFromOptions(pcbddc->ksp_D);CHKERRQ(ierr); 2284304d26faSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 2285304d26faSStefano Zampini } 2286d1e9a80fSBarry Smith ierr = KSPSetOperators(pcbddc->ksp_D,pcis->A_II,pcis->A_II);CHKERRQ(ierr); 2287d62866d3SStefano Zampini if (sub_schurs->reuse_mumps) { 2288d62866d3SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 2289d62866d3SStefano Zampini 2290d62866d3SStefano Zampini ierr = KSPSetPC(pcbddc->ksp_D,reuse_mumps->interior_solver);CHKERRQ(ierr); 2291d5574798SStefano Zampini } 2292304d26faSStefano Zampini /* umfpack interface has a bug when matrix dimension is zero. TODO solve from umfpack interface */ 2293304d26faSStefano Zampini if (!n_D) { 2294304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_D,&pc_temp);CHKERRQ(ierr); 2295304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCNONE);CHKERRQ(ierr); 2296304d26faSStefano Zampini } 2297304d26faSStefano Zampini /* Set Up KSP for Dirichlet problem of BDDC */ 2298304d26faSStefano Zampini ierr = KSPSetUp(pcbddc->ksp_D);CHKERRQ(ierr); 2299304d26faSStefano Zampini /* set ksp_D into pcis data */ 2300304d26faSStefano Zampini ierr = KSPDestroy(&pcis->ksp_D);CHKERRQ(ierr); 2301304d26faSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->ksp_D);CHKERRQ(ierr); 2302304d26faSStefano Zampini pcis->ksp_D = pcbddc->ksp_D; 2303684f6988SStefano Zampini } 2304304d26faSStefano Zampini 2305304d26faSStefano Zampini /* NEUMANN PROBLEM */ 2306684f6988SStefano Zampini A_RR = 0; 2307684f6988SStefano Zampini if (neumann) { 2308d62866d3SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 230904708bb6SStefano Zampini PetscInt ibs,mbs; 231004708bb6SStefano Zampini PetscBool issbaij; 231104708bb6SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 2312f4ddd8eeSStefano Zampini /* Matrix for Neumann problem is A_RR -> we need to create/reuse it at this point */ 23138ce42a96SStefano Zampini ierr = ISGetSize(pcbddc->is_R_local,&n_R);CHKERRQ(ierr); 2314f4ddd8eeSStefano Zampini if (pcbddc->ksp_R) { /* already created ksp */ 2315f4ddd8eeSStefano Zampini PetscInt nn_R; 231681d9aea3SBarry Smith ierr = KSPGetOperators(pcbddc->ksp_R,NULL,&A_RR);CHKERRQ(ierr); 2317f4ddd8eeSStefano Zampini ierr = PetscObjectReference((PetscObject)A_RR);CHKERRQ(ierr); 2318f4ddd8eeSStefano Zampini ierr = MatGetSize(A_RR,&nn_R,NULL);CHKERRQ(ierr); 2319f4ddd8eeSStefano Zampini if (nn_R != n_R) { /* old ksp is not reusable, so reset it */ 2320f4ddd8eeSStefano Zampini ierr = KSPReset(pcbddc->ksp_R);CHKERRQ(ierr); 2321f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 2322f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 2323f4ddd8eeSStefano Zampini } else { /* same sizes, but nonzero pattern depend on primal vertices so it can be changed */ 2324727cdba6SStefano Zampini if (pcbddc->new_primal_space_local) { /* we are not sure the matrix will have the same nonzero pattern */ 2325f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 2326f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 2327f4ddd8eeSStefano Zampini } else { /* safe to reuse the matrix */ 2328f4ddd8eeSStefano Zampini reuse = MAT_REUSE_MATRIX; 2329f4ddd8eeSStefano Zampini } 2330f4ddd8eeSStefano Zampini } 2331f4ddd8eeSStefano Zampini /* last check */ 2332d1e9a80fSBarry Smith if (pc->flag == DIFFERENT_NONZERO_PATTERN) { 2333f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 2334f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 2335f4ddd8eeSStefano Zampini } 2336f4ddd8eeSStefano Zampini } else { /* first time, so we need to create the matrix */ 2337f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 2338f4ddd8eeSStefano Zampini } 2339f4ddd8eeSStefano Zampini /* extract A_RR */ 2340af732b37SStefano Zampini ierr = MatGetBlockSize(pcbddc->local_mat,&mbs);CHKERRQ(ierr); 2341af732b37SStefano Zampini ierr = ISGetBlockSize(pcbddc->is_R_local,&ibs);CHKERRQ(ierr); 234204708bb6SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcbddc->local_mat,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 234304708bb6SStefano Zampini if (ibs != mbs) { /* need to convert to SEQAIJ to extract any submatrix with is_R_local */ 234404708bb6SStefano Zampini if (matis->A == pcbddc->local_mat) { 234504708bb6SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 234604708bb6SStefano Zampini ierr = MatConvert(matis->A,MATSEQAIJ,MAT_INITIAL_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 2347af732b37SStefano Zampini } else { 234804708bb6SStefano Zampini ierr = MatConvert(pcbddc->local_mat,MATSEQAIJ,MAT_REUSE_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 23496816873aSStefano Zampini } 235004708bb6SStefano Zampini } else if (issbaij) { /* need to convert to BAIJ to get offdiagonal blocks */ 235104708bb6SStefano Zampini if (matis->A == pcbddc->local_mat) { 235204708bb6SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 235304708bb6SStefano Zampini ierr = MatConvert(matis->A,MATSEQBAIJ,MAT_INITIAL_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 235404708bb6SStefano Zampini } else { 235504708bb6SStefano Zampini ierr = MatConvert(pcbddc->local_mat,MATSEQBAIJ,MAT_REUSE_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 235604708bb6SStefano Zampini } 235704708bb6SStefano Zampini } 2358f4ddd8eeSStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,pcbddc->is_R_local,pcbddc->is_R_local,reuse,&A_RR);CHKERRQ(ierr); 23593301b35fSStefano Zampini if (pcbddc->local_mat->symmetric_set) { 23603301b35fSStefano Zampini ierr = MatSetOption(A_RR,MAT_SYMMETRIC,pcbddc->local_mat->symmetric_set);CHKERRQ(ierr); 23616816873aSStefano Zampini } 2362f4ddd8eeSStefano Zampini if (!pcbddc->ksp_R) { /* create object if not present */ 2363304d26faSStefano Zampini ierr = KSPCreate(PETSC_COMM_SELF,&pcbddc->ksp_R);CHKERRQ(ierr); 2364304d26faSStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->ksp_R,(PetscObject)pc,1);CHKERRQ(ierr); 2365304d26faSStefano Zampini /* default */ 2366304d26faSStefano Zampini ierr = KSPSetType(pcbddc->ksp_R,KSPPREONLY);CHKERRQ(ierr); 2367e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->ksp_R,neu_prefix);CHKERRQ(ierr); 2368304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_temp);CHKERRQ(ierr); 23699577ea80SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)A_RR,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 23709577ea80SStefano Zampini if (issbaij) { 23719577ea80SStefano Zampini ierr = PCSetType(pc_temp,PCCHOLESKY);CHKERRQ(ierr); 23729577ea80SStefano Zampini } else { 2373304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCLU);CHKERRQ(ierr); 23749577ea80SStefano Zampini } 2375304d26faSStefano Zampini /* Allow user's customization */ 2376304d26faSStefano Zampini ierr = KSPSetFromOptions(pcbddc->ksp_R);CHKERRQ(ierr); 2377304d26faSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 2378304d26faSStefano Zampini } 2379304d26faSStefano Zampini /* umfpack interface has a bug when matrix dimension is zero. TODO solve from umfpack interface */ 2380304d26faSStefano Zampini if (!n_R) { 2381304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_temp);CHKERRQ(ierr); 2382304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCNONE);CHKERRQ(ierr); 2383304d26faSStefano Zampini } 2384d62866d3SStefano Zampini /* Reuse MUMPS solver if it is present */ 2385d62866d3SStefano Zampini if (sub_schurs->reuse_mumps) { 2386d62866d3SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 2387d62866d3SStefano Zampini 2388d62866d3SStefano Zampini ierr = KSPSetPC(pcbddc->ksp_R,reuse_mumps->correction_solver);CHKERRQ(ierr); 2389d62866d3SStefano Zampini } 2390ffd830a3SStefano Zampini ierr = KSPSetOperators(pcbddc->ksp_R,A_RR,A_RR);CHKERRQ(ierr); 2391304d26faSStefano Zampini /* Set Up KSP for Neumann problem of BDDC */ 2392304d26faSStefano Zampini ierr = KSPSetUp(pcbddc->ksp_R);CHKERRQ(ierr); 2393684f6988SStefano Zampini } 23946816873aSStefano Zampini /* free Neumann problem's matrix */ 23956816873aSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 2396304d26faSStefano Zampini 2397304d26faSStefano Zampini /* check Dirichlet and Neumann solvers and adapt them if a nullspace correction is needed */ 23980fccc4e9SStefano Zampini if (pcbddc->NullSpace || pcbddc->dbg_flag) { 2399684f6988SStefano Zampini if (pcbddc->dbg_flag) { 2400684f6988SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 24011575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 2402684f6988SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 2403684f6988SStefano Zampini } 2404684f6988SStefano Zampini if (dirichlet) { /* Dirichlet */ 24050fccc4e9SStefano Zampini ierr = VecSetRandom(pcis->vec1_D,NULL);CHKERRQ(ierr); 24060fccc4e9SStefano Zampini ierr = MatMult(pcis->A_II,pcis->vec1_D,pcis->vec2_D);CHKERRQ(ierr); 24070fccc4e9SStefano Zampini ierr = KSPSolve(pcbddc->ksp_D,pcis->vec2_D,pcis->vec2_D);CHKERRQ(ierr); 24080fccc4e9SStefano Zampini ierr = VecAXPY(pcis->vec1_D,m_one,pcis->vec2_D);CHKERRQ(ierr); 24090fccc4e9SStefano Zampini ierr = VecNorm(pcis->vec1_D,NORM_INFINITY,&value);CHKERRQ(ierr); 2410304d26faSStefano Zampini /* need to be adapted? */ 2411b8ffe317SStefano Zampini use_exact = (PetscAbsReal(value) > 1.e-4 ? PETSC_FALSE : PETSC_TRUE); 2412b8ffe317SStefano Zampini ierr = MPI_Allreduce(&use_exact,&use_exact_reduced,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 2413b8ffe317SStefano Zampini ierr = PCBDDCSetUseExactDirichlet(pc,use_exact_reduced);CHKERRQ(ierr); 2414304d26faSStefano Zampini /* print info */ 2415304d26faSStefano Zampini if (pcbddc->dbg_flag) { 2416e604994aSStefano 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); 2417304d26faSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 2418304d26faSStefano Zampini } 2419b8ffe317SStefano Zampini if (pcbddc->NullSpace && !use_exact_reduced && !pcbddc->switch_static) { 2420298c0119SStefano Zampini ierr = PCBDDCNullSpaceAssembleCorrection(pc,PETSC_TRUE,pcis->is_I_local);CHKERRQ(ierr); 2421304d26faSStefano Zampini } 2422684f6988SStefano Zampini } 2423684f6988SStefano Zampini if (neumann) { /* Neumann */ 24246816873aSStefano Zampini ierr = KSPGetOperators(pcbddc->ksp_R,&A_RR,NULL);CHKERRQ(ierr); 24250fccc4e9SStefano Zampini ierr = VecSetRandom(pcbddc->vec1_R,NULL);CHKERRQ(ierr); 24260fccc4e9SStefano Zampini ierr = MatMult(A_RR,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 24270fccc4e9SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec2_R,pcbddc->vec2_R);CHKERRQ(ierr); 24280fccc4e9SStefano Zampini ierr = VecAXPY(pcbddc->vec1_R,m_one,pcbddc->vec2_R);CHKERRQ(ierr); 24290fccc4e9SStefano Zampini ierr = VecNorm(pcbddc->vec1_R,NORM_INFINITY,&value);CHKERRQ(ierr); 2430304d26faSStefano Zampini /* need to be adapted? */ 2431b8ffe317SStefano Zampini use_exact = (PetscAbsReal(value) > 1.e-4 ? PETSC_FALSE : PETSC_TRUE); 2432b8ffe317SStefano Zampini ierr = MPI_Allreduce(&use_exact,&use_exact_reduced,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 2433304d26faSStefano Zampini /* print info */ 2434304d26faSStefano Zampini if (pcbddc->dbg_flag) { 2435e604994aSStefano 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); 2436304d26faSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 2437304d26faSStefano Zampini } 2438b8ffe317SStefano Zampini if (pcbddc->NullSpace && !use_exact_reduced) { /* is it the right logic? */ 2439298c0119SStefano Zampini ierr = PCBDDCNullSpaceAssembleCorrection(pc,PETSC_FALSE,pcbddc->is_R_local);CHKERRQ(ierr); 2440304d26faSStefano Zampini } 24410fccc4e9SStefano Zampini } 2442684f6988SStefano Zampini } 2443304d26faSStefano Zampini PetscFunctionReturn(0); 2444304d26faSStefano Zampini } 2445304d26faSStefano Zampini 2446304d26faSStefano Zampini #undef __FUNCT__ 2447ba15a52eSStefano Zampini #define __FUNCT__ "PCBDDCSolveSubstructureCorrection" 244880677318SStefano Zampini static PetscErrorCode PCBDDCSolveSubstructureCorrection(PC pc, Vec inout_B, Vec inout_D, PetscBool applytranspose) 2449674ae819SStefano Zampini { 2450674ae819SStefano Zampini PetscErrorCode ierr; 2451674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 2452be83ff47SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 2453674ae819SStefano Zampini 2454674ae819SStefano Zampini PetscFunctionBegin; 2455be83ff47SStefano Zampini if (!sub_schurs->reuse_mumps) { 245680677318SStefano Zampini ierr = VecSet(pcbddc->vec1_R,0.);CHKERRQ(ierr); 245720c7b377SStefano Zampini } 245880677318SStefano Zampini if (!pcbddc->switch_static) { 245980677318SStefano Zampini if (applytranspose && pcbddc->local_auxmat1) { 246080677318SStefano Zampini ierr = MatMultTranspose(pcbddc->local_auxmat2,inout_B,pcbddc->vec1_C);CHKERRQ(ierr); 246180677318SStefano Zampini ierr = MatMultTransposeAdd(pcbddc->local_auxmat1,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr); 246220c7b377SStefano Zampini } 2463be83ff47SStefano Zampini if (!sub_schurs->reuse_mumps) { 246480677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 246580677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 246620c7b377SStefano Zampini } else { 2467be83ff47SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 2468be83ff47SStefano Zampini 246953892102SStefano Zampini ierr = VecScatterBegin(reuse_mumps->correction_scatter_B,inout_B,reuse_mumps->rhs_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 247053892102SStefano Zampini ierr = VecScatterEnd(reuse_mumps->correction_scatter_B,inout_B,reuse_mumps->rhs_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 247120c7b377SStefano Zampini } 2472be83ff47SStefano Zampini } else { 247380677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 247480677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 247580677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,inout_D,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 247680677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,inout_D,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 247780677318SStefano Zampini if (applytranspose && pcbddc->local_auxmat1) { 247880677318SStefano Zampini ierr = MatMultTranspose(pcbddc->local_auxmat2,pcbddc->vec1_R,pcbddc->vec1_C);CHKERRQ(ierr); 247980677318SStefano Zampini ierr = MatMultTransposeAdd(pcbddc->local_auxmat1,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr); 248080677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 248180677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2482674ae819SStefano Zampini } 2483674ae819SStefano Zampini } 2484be83ff47SStefano Zampini if (!sub_schurs->reuse_mumps) { 248580677318SStefano Zampini if (applytranspose) { 248680677318SStefano Zampini ierr = KSPSolveTranspose(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 248780677318SStefano Zampini } else { 248880677318SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 248980677318SStefano Zampini } 2490be83ff47SStefano Zampini } else { 2491be83ff47SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 2492be83ff47SStefano Zampini 2493be83ff47SStefano Zampini if (applytranspose) { 24945a05ddb0SStefano Zampini ierr = MatFactorSolveSchurComplementTranspose(reuse_mumps->F,reuse_mumps->rhs_B,reuse_mumps->sol_B);CHKERRQ(ierr); 2495be83ff47SStefano Zampini } else { 24965a05ddb0SStefano Zampini ierr = MatFactorSolveSchurComplement(reuse_mumps->F,reuse_mumps->rhs_B,reuse_mumps->sol_B);CHKERRQ(ierr); 2497be83ff47SStefano Zampini } 2498be83ff47SStefano Zampini } 249980677318SStefano Zampini ierr = VecSet(inout_B,0.);CHKERRQ(ierr); 250080677318SStefano Zampini if (!pcbddc->switch_static) { 2501be83ff47SStefano Zampini if (!sub_schurs->reuse_mumps) { 250280677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 250380677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2504be83ff47SStefano Zampini } else { 2505be83ff47SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 2506be83ff47SStefano Zampini 250753892102SStefano Zampini ierr = VecScatterBegin(reuse_mumps->correction_scatter_B,reuse_mumps->sol_B,inout_B,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 250853892102SStefano Zampini ierr = VecScatterEnd(reuse_mumps->correction_scatter_B,reuse_mumps->sol_B,inout_B,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2509be83ff47SStefano Zampini } 251080677318SStefano Zampini if (!applytranspose && pcbddc->local_auxmat1) { 251180677318SStefano Zampini ierr = MatMult(pcbddc->local_auxmat1,inout_B,pcbddc->vec1_C);CHKERRQ(ierr); 251280677318SStefano Zampini ierr = MatMultAdd(pcbddc->local_auxmat2,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr); 251380677318SStefano Zampini } 251480677318SStefano Zampini } else { 251580677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 251680677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 251780677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 251880677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 251980677318SStefano Zampini if (!applytranspose && pcbddc->local_auxmat1) { 252080677318SStefano Zampini ierr = MatMult(pcbddc->local_auxmat1,inout_B,pcbddc->vec1_C);CHKERRQ(ierr); 252180677318SStefano Zampini ierr = MatMultAdd(pcbddc->local_auxmat2,pcbddc->vec1_C,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 252280677318SStefano Zampini } 252380677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 252480677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 252580677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 252680677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2527674ae819SStefano Zampini } 2528674ae819SStefano Zampini PetscFunctionReturn(0); 2529674ae819SStefano Zampini } 2530674ae819SStefano Zampini 2531dc359a40SStefano Zampini /* parameter apply transpose determines if the interface preconditioner should be applied transposed or not */ 2532674ae819SStefano Zampini #undef __FUNCT__ 2533674ae819SStefano Zampini #define __FUNCT__ "PCBDDCApplyInterfacePreconditioner" 2534dc359a40SStefano Zampini PetscErrorCode PCBDDCApplyInterfacePreconditioner(PC pc, PetscBool applytranspose) 2535674ae819SStefano Zampini { 2536674ae819SStefano Zampini PetscErrorCode ierr; 2537674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 2538674ae819SStefano Zampini PC_IS* pcis = (PC_IS*) (pc->data); 2539674ae819SStefano Zampini const PetscScalar zero = 0.0; 2540674ae819SStefano Zampini 2541674ae819SStefano Zampini PetscFunctionBegin; 2542dc359a40SStefano Zampini /* Application of PSI^T or PHI^T (depending on applytranspose, see comment above) */ 2543dc359a40SStefano Zampini if (applytranspose) { 2544674ae819SStefano Zampini ierr = MatMultTranspose(pcbddc->coarse_phi_B,pcis->vec1_B,pcbddc->vec1_P);CHKERRQ(ierr); 25458eeda7d8SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultTransposeAdd(pcbddc->coarse_phi_D,pcis->vec1_D,pcbddc->vec1_P,pcbddc->vec1_P);CHKERRQ(ierr); } 2546dc359a40SStefano Zampini } else { 2547674ae819SStefano Zampini ierr = MatMultTranspose(pcbddc->coarse_psi_B,pcis->vec1_B,pcbddc->vec1_P);CHKERRQ(ierr); 2548674ae819SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultTransposeAdd(pcbddc->coarse_psi_D,pcis->vec1_D,pcbddc->vec1_P,pcbddc->vec1_P);CHKERRQ(ierr); } 254915aaf578SStefano Zampini } 2550efc2fbd9SStefano Zampini 2551efc2fbd9SStefano Zampini /* add p0 to the last value of vec1_P holding the coarse dof relative to p0 */ 25524f1b2e48SStefano Zampini if (pcbddc->benign_n) { 2553efc2fbd9SStefano Zampini PetscScalar *array; 25544f1b2e48SStefano Zampini PetscInt j; 2555efc2fbd9SStefano Zampini 2556efc2fbd9SStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 25574f1b2e48SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) array[pcbddc->local_primal_size-pcbddc->benign_n+j] += pcbddc->benign_p0[j]; 2558efc2fbd9SStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 2559efc2fbd9SStefano Zampini } 2560efc2fbd9SStefano Zampini 256112edc857SStefano Zampini /* start communications from local primal nodes to rhs of coarse solver */ 256212edc857SStefano Zampini ierr = VecSet(pcbddc->coarse_vec,zero);CHKERRQ(ierr); 256312edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataBegin(pc,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 256412edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataEnd(pc,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 256512edc857SStefano Zampini 25669f00e9b4SStefano Zampini /* Coarse solution -> rhs and sol updated inside PCBDDCScattarCoarseDataBegin/End */ 256712edc857SStefano Zampini /* TODO remove null space when doing multilevel */ 256812edc857SStefano Zampini if (pcbddc->coarse_ksp) { 256951694757SStefano Zampini Mat coarse_mat; 2570964fefecSStefano Zampini Vec rhs,sol; 257151694757SStefano Zampini MatNullSpace nullsp; 2572964fefecSStefano Zampini 2573964fefecSStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&rhs);CHKERRQ(ierr); 2574964fefecSStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&sol);CHKERRQ(ierr); 257551694757SStefano Zampini ierr = KSPGetOperators(pcbddc->coarse_ksp,&coarse_mat,NULL);CHKERRQ(ierr); 257651694757SStefano Zampini ierr = MatGetNullSpace(coarse_mat,&nullsp);CHKERRQ(ierr); 257751694757SStefano Zampini if (nullsp) { 257851694757SStefano Zampini ierr = MatNullSpaceRemove(nullsp,rhs);CHKERRQ(ierr); 257951694757SStefano Zampini } 258012edc857SStefano Zampini if (applytranspose) { 2581964fefecSStefano Zampini ierr = KSPSolveTranspose(pcbddc->coarse_ksp,rhs,sol);CHKERRQ(ierr); 258212edc857SStefano Zampini } else { 2583964fefecSStefano Zampini ierr = KSPSolve(pcbddc->coarse_ksp,rhs,sol);CHKERRQ(ierr); 258412edc857SStefano Zampini } 258551694757SStefano Zampini if (nullsp) { 258651694757SStefano Zampini ierr = MatNullSpaceRemove(nullsp,sol);CHKERRQ(ierr); 258751694757SStefano Zampini } 258812edc857SStefano Zampini } 2589674ae819SStefano Zampini 2590674ae819SStefano Zampini /* Local solution on R nodes */ 259180677318SStefano Zampini if (pcis->n) { /* in/out pcbddc->vec1_B,pcbddc->vec1_D */ 259280677318SStefano Zampini ierr = PCBDDCSolveSubstructureCorrection(pc,pcis->vec1_B,pcis->vec1_D,applytranspose);CHKERRQ(ierr); 25939f00e9b4SStefano Zampini } 2594674ae819SStefano Zampini 25959f00e9b4SStefano Zampini /* communications from coarse sol to local primal nodes */ 25969f00e9b4SStefano Zampini ierr = PCBDDCScatterCoarseDataBegin(pc,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 259712edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataEnd(pc,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2598674ae819SStefano Zampini 2599674ae819SStefano Zampini /* Sum contributions from two levels */ 2600dc359a40SStefano Zampini if (applytranspose) { 2601dc359a40SStefano Zampini ierr = MatMultAdd(pcbddc->coarse_psi_B,pcbddc->vec1_P,pcis->vec1_B,pcis->vec1_B);CHKERRQ(ierr); 2602dc359a40SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultAdd(pcbddc->coarse_psi_D,pcbddc->vec1_P,pcis->vec1_D,pcis->vec1_D);CHKERRQ(ierr); } 2603dc359a40SStefano Zampini } else { 2604674ae819SStefano Zampini ierr = MatMultAdd(pcbddc->coarse_phi_B,pcbddc->vec1_P,pcis->vec1_B,pcis->vec1_B);CHKERRQ(ierr); 26058eeda7d8SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultAdd(pcbddc->coarse_phi_D,pcbddc->vec1_P,pcis->vec1_D,pcis->vec1_D);CHKERRQ(ierr); } 2606dc359a40SStefano Zampini } 2607efc2fbd9SStefano Zampini /* store p0 */ 26084f1b2e48SStefano Zampini if (pcbddc->benign_n) { 2609efc2fbd9SStefano Zampini PetscScalar *array; 26104f1b2e48SStefano Zampini PetscInt j; 2611efc2fbd9SStefano Zampini 2612efc2fbd9SStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 26134f1b2e48SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) pcbddc->benign_p0[j] = array[pcbddc->local_primal_size-pcbddc->benign_n+j]; 2614efc2fbd9SStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 2615efc2fbd9SStefano Zampini } 2616674ae819SStefano Zampini PetscFunctionReturn(0); 2617674ae819SStefano Zampini } 2618674ae819SStefano Zampini 2619674ae819SStefano Zampini #undef __FUNCT__ 2620674ae819SStefano Zampini #define __FUNCT__ "PCBDDCScatterCoarseDataBegin" 262112edc857SStefano Zampini PetscErrorCode PCBDDCScatterCoarseDataBegin(PC pc,InsertMode imode, ScatterMode smode) 2622674ae819SStefano Zampini { 2623674ae819SStefano Zampini PetscErrorCode ierr; 2624674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 262558da7f69SStefano Zampini PetscScalar *array; 262612edc857SStefano Zampini Vec from,to; 2627674ae819SStefano Zampini 2628674ae819SStefano Zampini PetscFunctionBegin; 262912edc857SStefano Zampini if (smode == SCATTER_REVERSE) { /* from global to local -> get data from coarse solution */ 263012edc857SStefano Zampini from = pcbddc->coarse_vec; 263112edc857SStefano Zampini to = pcbddc->vec1_P; 263212edc857SStefano Zampini if (pcbddc->coarse_ksp) { /* get array from coarse processes */ 263312edc857SStefano Zampini Vec tvec; 263458da7f69SStefano Zampini 263558da7f69SStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 263658da7f69SStefano Zampini ierr = VecResetArray(tvec);CHKERRQ(ierr); 263712edc857SStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 263858da7f69SStefano Zampini ierr = VecGetArray(tvec,&array);CHKERRQ(ierr); 263958da7f69SStefano Zampini ierr = VecPlaceArray(from,array);CHKERRQ(ierr); 264058da7f69SStefano Zampini ierr = VecRestoreArray(tvec,&array);CHKERRQ(ierr); 264112edc857SStefano Zampini } 264212edc857SStefano Zampini } else { /* from local to global -> put data in coarse right hand side */ 264312edc857SStefano Zampini from = pcbddc->vec1_P; 264412edc857SStefano Zampini to = pcbddc->coarse_vec; 264512edc857SStefano Zampini } 264612edc857SStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,from,to,imode,smode);CHKERRQ(ierr); 2647674ae819SStefano Zampini PetscFunctionReturn(0); 2648674ae819SStefano Zampini } 2649674ae819SStefano Zampini 2650674ae819SStefano Zampini #undef __FUNCT__ 2651674ae819SStefano Zampini #define __FUNCT__ "PCBDDCScatterCoarseDataEnd" 265212edc857SStefano Zampini PetscErrorCode PCBDDCScatterCoarseDataEnd(PC pc, InsertMode imode, ScatterMode smode) 2653674ae819SStefano Zampini { 2654674ae819SStefano Zampini PetscErrorCode ierr; 2655674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 265658da7f69SStefano Zampini PetscScalar *array; 265712edc857SStefano Zampini Vec from,to; 2658674ae819SStefano Zampini 2659674ae819SStefano Zampini PetscFunctionBegin; 266012edc857SStefano Zampini if (smode == SCATTER_REVERSE) { /* from global to local -> get data from coarse solution */ 266112edc857SStefano Zampini from = pcbddc->coarse_vec; 266212edc857SStefano Zampini to = pcbddc->vec1_P; 266312edc857SStefano Zampini } else { /* from local to global -> put data in coarse right hand side */ 266412edc857SStefano Zampini from = pcbddc->vec1_P; 266512edc857SStefano Zampini to = pcbddc->coarse_vec; 266612edc857SStefano Zampini } 266712edc857SStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,from,to,imode,smode);CHKERRQ(ierr); 266812edc857SStefano Zampini if (smode == SCATTER_FORWARD) { 266912edc857SStefano Zampini if (pcbddc->coarse_ksp) { /* get array from coarse processes */ 267012edc857SStefano Zampini Vec tvec; 267158da7f69SStefano Zampini 267212edc857SStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 267358da7f69SStefano Zampini ierr = VecGetArray(to,&array);CHKERRQ(ierr); 267458da7f69SStefano Zampini ierr = VecPlaceArray(tvec,array);CHKERRQ(ierr); 267558da7f69SStefano Zampini ierr = VecRestoreArray(to,&array);CHKERRQ(ierr); 267658da7f69SStefano Zampini } 267758da7f69SStefano Zampini } else { 267858da7f69SStefano Zampini if (pcbddc->coarse_ksp) { /* restore array of pcbddc->coarse_vec */ 267958da7f69SStefano Zampini ierr = VecResetArray(from);CHKERRQ(ierr); 268012edc857SStefano Zampini } 268112edc857SStefano Zampini } 2682674ae819SStefano Zampini PetscFunctionReturn(0); 2683674ae819SStefano Zampini } 2684674ae819SStefano Zampini 2685984c4197SStefano Zampini /* uncomment for testing purposes */ 2686984c4197SStefano Zampini /* #define PETSC_MISSING_LAPACK_GESVD 1 */ 2687674ae819SStefano Zampini #undef __FUNCT__ 2688674ae819SStefano Zampini #define __FUNCT__ "PCBDDCConstraintsSetUp" 2689674ae819SStefano Zampini PetscErrorCode PCBDDCConstraintsSetUp(PC pc) 2690674ae819SStefano Zampini { 2691674ae819SStefano Zampini PetscErrorCode ierr; 2692674ae819SStefano Zampini PC_IS* pcis = (PC_IS*)(pc->data); 2693674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 2694674ae819SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 2695984c4197SStefano Zampini /* one and zero */ 2696984c4197SStefano Zampini PetscScalar one=1.0,zero=0.0; 2697984c4197SStefano Zampini /* space to store constraints and their local indices */ 26989162d606SStefano Zampini PetscScalar *constraints_data; 26999162d606SStefano Zampini PetscInt *constraints_idxs,*constraints_idxs_B; 27009162d606SStefano Zampini PetscInt *constraints_idxs_ptr,*constraints_data_ptr; 27019162d606SStefano Zampini PetscInt *constraints_n; 2702984c4197SStefano Zampini /* iterators */ 2703b3d85658SStefano Zampini PetscInt i,j,k,total_counts,total_counts_cc,cum; 2704984c4197SStefano Zampini /* BLAS integers */ 2705e310c8b4SStefano Zampini PetscBLASInt lwork,lierr; 2706e310c8b4SStefano Zampini PetscBLASInt Blas_N,Blas_M,Blas_K,Blas_one=1; 2707c4303822SStefano Zampini PetscBLASInt Blas_LDA,Blas_LDB,Blas_LDC; 2708727cdba6SStefano Zampini /* reuse */ 27090e6343abSStefano Zampini PetscInt olocal_primal_size,olocal_primal_size_cc; 27100e6343abSStefano Zampini PetscInt *olocal_primal_ref_node,*olocal_primal_ref_mult; 2711984c4197SStefano Zampini /* change of basis */ 2712b3d85658SStefano Zampini PetscBool qr_needed; 27139162d606SStefano Zampini PetscBT change_basis,qr_needed_idx; 2714984c4197SStefano Zampini /* auxiliary stuff */ 271564efe560SStefano Zampini PetscInt *nnz,*is_indices; 27168a0068c3SStefano Zampini PetscInt ncc; 2717984c4197SStefano Zampini /* some quantities */ 271845a1bb75SStefano Zampini PetscInt n_vertices,total_primal_vertices,valid_constraints; 2719a58a30b4SStefano Zampini PetscInt size_of_constraint,max_size_of_constraint=0,max_constraints,temp_constraints; 2720984c4197SStefano Zampini 2721674ae819SStefano Zampini PetscFunctionBegin; 27228e61c736SStefano Zampini /* Destroy Mat objects computed previously */ 27238e61c736SStefano Zampini ierr = MatDestroy(&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 27248e61c736SStefano Zampini ierr = MatDestroy(&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 2725088faed8SStefano Zampini /* save info on constraints from previous setup (if any) */ 2726088faed8SStefano Zampini olocal_primal_size = pcbddc->local_primal_size; 27270e6343abSStefano Zampini olocal_primal_size_cc = pcbddc->local_primal_size_cc; 27280e6343abSStefano Zampini ierr = PetscMalloc2(olocal_primal_size_cc,&olocal_primal_ref_node,olocal_primal_size_cc,&olocal_primal_ref_mult);CHKERRQ(ierr); 27290e6343abSStefano Zampini ierr = PetscMemcpy(olocal_primal_ref_node,pcbddc->local_primal_ref_node,olocal_primal_size_cc*sizeof(PetscInt));CHKERRQ(ierr); 27300e6343abSStefano Zampini ierr = PetscMemcpy(olocal_primal_ref_mult,pcbddc->local_primal_ref_mult,olocal_primal_size_cc*sizeof(PetscInt));CHKERRQ(ierr); 27310e6343abSStefano Zampini ierr = PetscFree2(pcbddc->local_primal_ref_node,pcbddc->local_primal_ref_mult);CHKERRQ(ierr); 2732088faed8SStefano Zampini ierr = PetscFree(pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 2733cf5a6209SStefano Zampini 2734cf5a6209SStefano Zampini /* print some info */ 2735cf5a6209SStefano Zampini if (pcbddc->dbg_flag) { 2736cf5a6209SStefano Zampini IS vertices; 2737cf5a6209SStefano Zampini PetscInt nv,nedges,nfaces; 2738cf5a6209SStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,&nfaces,NULL,&nedges,NULL,&vertices);CHKERRQ(ierr); 2739cf5a6209SStefano Zampini ierr = ISGetSize(vertices,&nv);CHKERRQ(ierr); 2740cf5a6209SStefano Zampini ierr = ISDestroy(&vertices);CHKERRQ(ierr); 27411575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 2742cf5a6209SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"--------------------------------------------------------------\n");CHKERRQ(ierr); 2743cf5a6209SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate vertices (%d)\n",PetscGlobalRank,nv,pcbddc->use_vertices);CHKERRQ(ierr); 2744fd14bc51SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate edges (%d)\n",PetscGlobalRank,nedges,pcbddc->use_edges);CHKERRQ(ierr); 2745fd14bc51SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate faces (%d)\n",PetscGlobalRank,nfaces,pcbddc->use_faces);CHKERRQ(ierr); 2746cf5a6209SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 27471575c14dSBarry Smith ierr = PetscViewerASCIIPopSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 2748cf5a6209SStefano Zampini } 2749cf5a6209SStefano Zampini 2750cf5a6209SStefano Zampini if (!pcbddc->adaptive_selection) { 27519162d606SStefano Zampini IS ISForVertices,*ISForFaces,*ISForEdges; 2752cf5a6209SStefano Zampini MatNullSpace nearnullsp; 2753cf5a6209SStefano Zampini const Vec *nearnullvecs; 2754cf5a6209SStefano Zampini Vec *localnearnullsp; 2755cf5a6209SStefano Zampini PetscScalar *array; 2756cf5a6209SStefano Zampini PetscInt n_ISForFaces,n_ISForEdges,nnsp_size; 2757cf5a6209SStefano Zampini PetscBool nnsp_has_cnst; 2758674ae819SStefano Zampini /* LAPACK working arrays for SVD or POD */ 2759b3d85658SStefano Zampini PetscBool skip_lapack,boolforchange; 2760674ae819SStefano Zampini PetscScalar *work; 2761674ae819SStefano Zampini PetscReal *singular_vals; 2762674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX) 2763674ae819SStefano Zampini PetscReal *rwork; 2764674ae819SStefano Zampini #endif 2765674ae819SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 2766674ae819SStefano Zampini PetscScalar *temp_basis,*correlation_mat; 2767674ae819SStefano Zampini #else 2768964fefecSStefano Zampini PetscBLASInt dummy_int=1; 2769964fefecSStefano Zampini PetscScalar dummy_scalar=1.; 2770674ae819SStefano Zampini #endif 2771674ae819SStefano Zampini 2772674ae819SStefano Zampini /* Get index sets for faces, edges and vertices from graph */ 2773d06fc5fdSStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,&n_ISForFaces,&ISForFaces,&n_ISForEdges,&ISForEdges,&ISForVertices);CHKERRQ(ierr); 2774d06fc5fdSStefano Zampini /* free unneeded index sets */ 2775d06fc5fdSStefano Zampini if (!pcbddc->use_vertices) { 2776d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForVertices);CHKERRQ(ierr); 2777674ae819SStefano Zampini } 2778d06fc5fdSStefano Zampini if (!pcbddc->use_edges) { 2779d06fc5fdSStefano Zampini for (i=0;i<n_ISForEdges;i++) { 2780d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForEdges[i]);CHKERRQ(ierr); 2781d06fc5fdSStefano Zampini } 2782d06fc5fdSStefano Zampini ierr = PetscFree(ISForEdges);CHKERRQ(ierr); 2783d06fc5fdSStefano Zampini n_ISForEdges = 0; 2784d06fc5fdSStefano Zampini } 2785d06fc5fdSStefano Zampini if (!pcbddc->use_faces) { 2786d06fc5fdSStefano Zampini for (i=0;i<n_ISForFaces;i++) { 2787d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForFaces[i]);CHKERRQ(ierr); 2788d06fc5fdSStefano Zampini } 2789d06fc5fdSStefano Zampini ierr = PetscFree(ISForFaces);CHKERRQ(ierr); 2790d06fc5fdSStefano Zampini n_ISForFaces = 0; 2791d06fc5fdSStefano Zampini } 279270022509SStefano Zampini 279370022509SStefano Zampini #if defined(PETSC_USE_DEBUG) 279470022509SStefano Zampini /* HACK: when solving singular problems not using vertices, a change of basis is mandatory. 279570022509SStefano Zampini Also use_change_of_basis should be consistent among processors */ 279670022509SStefano Zampini if (pcbddc->NullSpace) { 279770022509SStefano Zampini PetscBool tbool[2],gbool[2]; 279870022509SStefano Zampini 279970022509SStefano Zampini if (!ISForVertices && !pcbddc->user_ChangeOfBasisMatrix) { 2800b8ffe317SStefano Zampini pcbddc->use_change_of_basis = PETSC_TRUE; 2801d06fc5fdSStefano Zampini if (!ISForEdges) { 2802d06fc5fdSStefano Zampini pcbddc->use_change_on_faces = PETSC_TRUE; 2803d06fc5fdSStefano Zampini } 2804b8ffe317SStefano Zampini } 2805d06fc5fdSStefano Zampini tbool[0] = pcbddc->use_change_of_basis; 2806d06fc5fdSStefano Zampini tbool[1] = pcbddc->use_change_on_faces; 2807d06fc5fdSStefano Zampini ierr = MPI_Allreduce(tbool,gbool,2,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 2808d06fc5fdSStefano Zampini pcbddc->use_change_of_basis = gbool[0]; 2809d06fc5fdSStefano Zampini pcbddc->use_change_on_faces = gbool[1]; 281098a51de6SStefano Zampini } 281170022509SStefano Zampini #endif 281208122e43SStefano Zampini 2813674ae819SStefano Zampini /* check if near null space is attached to global mat */ 2814674ae819SStefano Zampini ierr = MatGetNearNullSpace(pc->pmat,&nearnullsp);CHKERRQ(ierr); 2815674ae819SStefano Zampini if (nearnullsp) { 2816674ae819SStefano Zampini ierr = MatNullSpaceGetVecs(nearnullsp,&nnsp_has_cnst,&nnsp_size,&nearnullvecs);CHKERRQ(ierr); 2817f4ddd8eeSStefano Zampini /* remove any stored info */ 2818f4ddd8eeSStefano Zampini ierr = MatNullSpaceDestroy(&pcbddc->onearnullspace);CHKERRQ(ierr); 2819f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->onearnullvecs_state);CHKERRQ(ierr); 2820f4ddd8eeSStefano Zampini /* store information for BDDC solver reuse */ 2821f4ddd8eeSStefano Zampini ierr = PetscObjectReference((PetscObject)nearnullsp);CHKERRQ(ierr); 2822f4ddd8eeSStefano Zampini pcbddc->onearnullspace = nearnullsp; 2823473ba861SJed Brown ierr = PetscMalloc1(nnsp_size,&pcbddc->onearnullvecs_state);CHKERRQ(ierr); 2824f4ddd8eeSStefano Zampini for (i=0;i<nnsp_size;i++) { 2825f4ddd8eeSStefano Zampini ierr = PetscObjectStateGet((PetscObject)nearnullvecs[i],&pcbddc->onearnullvecs_state[i]);CHKERRQ(ierr); 2826f4ddd8eeSStefano Zampini } 2827984c4197SStefano Zampini } else { /* if near null space is not provided BDDC uses constants by default */ 2828984c4197SStefano Zampini nnsp_size = 0; 2829674ae819SStefano Zampini nnsp_has_cnst = PETSC_TRUE; 2830674ae819SStefano Zampini } 2831984c4197SStefano Zampini /* get max number of constraints on a single cc */ 2832984c4197SStefano Zampini max_constraints = nnsp_size; 2833984c4197SStefano Zampini if (nnsp_has_cnst) max_constraints++; 2834984c4197SStefano Zampini 2835674ae819SStefano Zampini /* 2836674ae819SStefano Zampini Evaluate maximum storage size needed by the procedure 28379162d606SStefano Zampini - Indices for connected component i stored at "constraints_idxs + constraints_idxs_ptr[i]" 28389162d606SStefano Zampini - Values for constraints on connected component i stored at "constraints_data + constraints_data_ptr[i]" 28399162d606SStefano Zampini There can be multiple constraints per connected component 2840674ae819SStefano Zampini */ 2841674ae819SStefano Zampini n_vertices = 0; 2842674ae819SStefano Zampini if (ISForVertices) { 2843674ae819SStefano Zampini ierr = ISGetSize(ISForVertices,&n_vertices);CHKERRQ(ierr); 2844674ae819SStefano Zampini } 28459162d606SStefano Zampini ncc = n_vertices+n_ISForFaces+n_ISForEdges; 28469162d606SStefano Zampini ierr = PetscMalloc3(ncc+1,&constraints_idxs_ptr,ncc+1,&constraints_data_ptr,ncc,&constraints_n);CHKERRQ(ierr); 28479162d606SStefano Zampini 28489162d606SStefano Zampini total_counts = n_ISForFaces+n_ISForEdges; 28499162d606SStefano Zampini total_counts *= max_constraints; 2850674ae819SStefano Zampini total_counts += n_vertices; 28514641a718SStefano Zampini ierr = PetscBTCreate(total_counts,&change_basis);CHKERRQ(ierr); 28529162d606SStefano Zampini 2853674ae819SStefano Zampini total_counts = 0; 2854674ae819SStefano Zampini max_size_of_constraint = 0; 2855674ae819SStefano Zampini for (i=0;i<n_ISForEdges+n_ISForFaces;i++) { 28569162d606SStefano Zampini IS used_is; 2857674ae819SStefano Zampini if (i<n_ISForEdges) { 28589162d606SStefano Zampini used_is = ISForEdges[i]; 2859674ae819SStefano Zampini } else { 28609162d606SStefano Zampini used_is = ISForFaces[i-n_ISForEdges]; 2861674ae819SStefano Zampini } 28629162d606SStefano Zampini ierr = ISGetSize(used_is,&j);CHKERRQ(ierr); 2863674ae819SStefano Zampini total_counts += j; 2864674ae819SStefano Zampini max_size_of_constraint = PetscMax(j,max_size_of_constraint); 2865674ae819SStefano Zampini } 28669162d606SStefano 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); 28679162d606SStefano Zampini 2868984c4197SStefano Zampini /* get local part of global near null space vectors */ 2869785e854fSJed Brown ierr = PetscMalloc1(nnsp_size,&localnearnullsp);CHKERRQ(ierr); 2870984c4197SStefano Zampini for (k=0;k<nnsp_size;k++) { 2871984c4197SStefano Zampini ierr = VecDuplicate(pcis->vec1_N,&localnearnullsp[k]);CHKERRQ(ierr); 2872e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,nearnullvecs[k],localnearnullsp[k],INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2873e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,nearnullvecs[k],localnearnullsp[k],INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2874984c4197SStefano Zampini } 2875674ae819SStefano Zampini 2876242a89d7SStefano Zampini /* whether or not to skip lapack calls */ 2877242a89d7SStefano Zampini skip_lapack = PETSC_TRUE; 2878a773dcb8SStefano Zampini if (n_ISForFaces+n_ISForEdges && max_constraints > 1 && !pcbddc->use_nnsp_true) skip_lapack = PETSC_FALSE; 2879242a89d7SStefano Zampini 2880984c4197SStefano Zampini /* First we issue queries to allocate optimal workspace for LAPACKgesvd (or LAPACKsyev if SVD is missing) */ 2881a773dcb8SStefano Zampini if (!skip_lapack) { 2882674ae819SStefano Zampini PetscScalar temp_work; 2883911cabfeSStefano Zampini 2884674ae819SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 2885984c4197SStefano Zampini /* Proper Orthogonal Decomposition (POD) using the snapshot method */ 2886785e854fSJed Brown ierr = PetscMalloc1(max_constraints*max_constraints,&correlation_mat);CHKERRQ(ierr); 2887785e854fSJed Brown ierr = PetscMalloc1(max_constraints,&singular_vals);CHKERRQ(ierr); 2888785e854fSJed Brown ierr = PetscMalloc1(max_size_of_constraint*max_constraints,&temp_basis);CHKERRQ(ierr); 2889674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX) 2890785e854fSJed Brown ierr = PetscMalloc1(3*max_constraints,&rwork);CHKERRQ(ierr); 2891674ae819SStefano Zampini #endif 2892674ae819SStefano Zampini /* now we evaluate the optimal workspace using query with lwork=-1 */ 2893c8244a33SStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_N);CHKERRQ(ierr); 2894c8244a33SStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_LDA);CHKERRQ(ierr); 2895674ae819SStefano Zampini lwork = -1; 2896674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 2897674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 2898c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,&temp_work,&lwork,&lierr)); 2899674ae819SStefano Zampini #else 2900c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,&temp_work,&lwork,rwork,&lierr)); 2901674ae819SStefano Zampini #endif 2902674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2903984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to SYEV Lapack routine %d",(int)lierr); 2904674ae819SStefano Zampini #else /* on missing GESVD */ 2905674ae819SStefano Zampini /* SVD */ 2906674ae819SStefano Zampini PetscInt max_n,min_n; 2907674ae819SStefano Zampini max_n = max_size_of_constraint; 2908984c4197SStefano Zampini min_n = max_constraints; 2909984c4197SStefano Zampini if (max_size_of_constraint < max_constraints) { 2910674ae819SStefano Zampini min_n = max_size_of_constraint; 2911984c4197SStefano Zampini max_n = max_constraints; 2912674ae819SStefano Zampini } 2913785e854fSJed Brown ierr = PetscMalloc1(min_n,&singular_vals);CHKERRQ(ierr); 2914674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX) 2915785e854fSJed Brown ierr = PetscMalloc1(5*min_n,&rwork);CHKERRQ(ierr); 2916674ae819SStefano Zampini #endif 2917674ae819SStefano Zampini /* now we evaluate the optimal workspace using query with lwork=-1 */ 2918674ae819SStefano Zampini lwork = -1; 2919e310c8b4SStefano Zampini ierr = PetscBLASIntCast(max_n,&Blas_M);CHKERRQ(ierr); 2920e310c8b4SStefano Zampini ierr = PetscBLASIntCast(min_n,&Blas_N);CHKERRQ(ierr); 2921b7d8b9f8SStefano Zampini ierr = PetscBLASIntCast(max_n,&Blas_LDA);CHKERRQ(ierr); 2922674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 2923674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 29249162d606SStefano 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)); 2925674ae819SStefano Zampini #else 29269162d606SStefano 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)); 2927674ae819SStefano Zampini #endif 2928674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2929984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to GESVD Lapack routine %d",(int)lierr); 2930984c4197SStefano Zampini #endif /* on missing GESVD */ 2931674ae819SStefano Zampini /* Allocate optimal workspace */ 2932674ae819SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(temp_work),&lwork);CHKERRQ(ierr); 2933854ce69bSBarry Smith ierr = PetscMalloc1(lwork,&work);CHKERRQ(ierr); 2934674ae819SStefano Zampini } 2935674ae819SStefano Zampini /* Now we can loop on constraining sets */ 2936674ae819SStefano Zampini total_counts = 0; 29379162d606SStefano Zampini constraints_idxs_ptr[0] = 0; 29389162d606SStefano Zampini constraints_data_ptr[0] = 0; 2939674ae819SStefano Zampini /* vertices */ 29409162d606SStefano Zampini if (n_vertices) { 2941674ae819SStefano Zampini ierr = ISGetIndices(ISForVertices,(const PetscInt**)&is_indices);CHKERRQ(ierr); 29429162d606SStefano Zampini ierr = PetscMemcpy(constraints_idxs,is_indices,n_vertices*sizeof(PetscInt));CHKERRQ(ierr); 2943674ae819SStefano Zampini for (i=0;i<n_vertices;i++) { 29449162d606SStefano Zampini constraints_n[total_counts] = 1; 29459162d606SStefano Zampini constraints_data[total_counts] = 1.0; 29469162d606SStefano Zampini constraints_idxs_ptr[total_counts+1] = constraints_idxs_ptr[total_counts]+1; 29479162d606SStefano Zampini constraints_data_ptr[total_counts+1] = constraints_data_ptr[total_counts]+1; 2948674ae819SStefano Zampini total_counts++; 2949674ae819SStefano Zampini } 2950674ae819SStefano Zampini ierr = ISRestoreIndices(ISForVertices,(const PetscInt**)&is_indices);CHKERRQ(ierr); 2951674ae819SStefano Zampini n_vertices = total_counts; 2952674ae819SStefano Zampini } 2953984c4197SStefano Zampini 2954674ae819SStefano Zampini /* edges and faces */ 29559162d606SStefano Zampini total_counts_cc = total_counts; 2956911cabfeSStefano Zampini for (ncc=0;ncc<n_ISForEdges+n_ISForFaces;ncc++) { 29579162d606SStefano Zampini IS used_is; 29589162d606SStefano Zampini PetscBool idxs_copied = PETSC_FALSE; 29599162d606SStefano Zampini 2960911cabfeSStefano Zampini if (ncc<n_ISForEdges) { 29619162d606SStefano Zampini used_is = ISForEdges[ncc]; 2962984c4197SStefano Zampini boolforchange = pcbddc->use_change_of_basis; /* change or not the basis on the edge */ 2963674ae819SStefano Zampini } else { 29649162d606SStefano Zampini used_is = ISForFaces[ncc-n_ISForEdges]; 2965984c4197SStefano Zampini boolforchange = (PetscBool)(pcbddc->use_change_of_basis && pcbddc->use_change_on_faces); /* change or not the basis on the face */ 2966674ae819SStefano Zampini } 2967674ae819SStefano Zampini temp_constraints = 0; /* zero the number of constraints I have on this conn comp */ 29689162d606SStefano Zampini 29699162d606SStefano Zampini ierr = ISGetSize(used_is,&size_of_constraint);CHKERRQ(ierr); 29709162d606SStefano Zampini ierr = ISGetIndices(used_is,(const PetscInt**)&is_indices);CHKERRQ(ierr); 2971984c4197SStefano Zampini /* change of basis should not be performed on local periodic nodes */ 2972984c4197SStefano Zampini if (pcbddc->mat_graph->mirrors && pcbddc->mat_graph->mirrors[is_indices[0]]) boolforchange = PETSC_FALSE; 2973674ae819SStefano Zampini if (nnsp_has_cnst) { 29745b08dc53SStefano Zampini PetscScalar quad_value; 29759162d606SStefano Zampini 29769162d606SStefano Zampini ierr = PetscMemcpy(constraints_idxs + constraints_idxs_ptr[total_counts_cc],is_indices,size_of_constraint*sizeof(PetscInt));CHKERRQ(ierr); 29779162d606SStefano Zampini idxs_copied = PETSC_TRUE; 29789162d606SStefano Zampini 2979a773dcb8SStefano Zampini if (!pcbddc->use_nnsp_true) { 2980674ae819SStefano Zampini quad_value = (PetscScalar)(1.0/PetscSqrtReal((PetscReal)size_of_constraint)); 2981a773dcb8SStefano Zampini } else { 2982a773dcb8SStefano Zampini quad_value = 1.0; 2983a773dcb8SStefano Zampini } 2984674ae819SStefano Zampini for (j=0;j<size_of_constraint;j++) { 29859162d606SStefano Zampini constraints_data[constraints_data_ptr[total_counts_cc]+j] = quad_value; 2986674ae819SStefano Zampini } 29879162d606SStefano Zampini temp_constraints++; 2988674ae819SStefano Zampini total_counts++; 2989674ae819SStefano Zampini } 2990674ae819SStefano Zampini for (k=0;k<nnsp_size;k++) { 2991984c4197SStefano Zampini PetscReal real_value; 29929162d606SStefano Zampini PetscScalar *ptr_to_data; 29939162d606SStefano Zampini 2994984c4197SStefano Zampini ierr = VecGetArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 29959162d606SStefano Zampini ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]+temp_constraints*size_of_constraint]; 2996674ae819SStefano Zampini for (j=0;j<size_of_constraint;j++) { 29979162d606SStefano Zampini ptr_to_data[j] = array[is_indices[j]]; 2998674ae819SStefano Zampini } 2999984c4197SStefano Zampini ierr = VecRestoreArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 3000984c4197SStefano Zampini /* check if array is null on the connected component */ 3001e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 30029162d606SStefano Zampini PetscStackCallBLAS("BLASasum",real_value = BLASasum_(&Blas_N,ptr_to_data,&Blas_one)); 30035b08dc53SStefano Zampini if (real_value > 0.0) { /* keep indices and values */ 3004674ae819SStefano Zampini temp_constraints++; 3005674ae819SStefano Zampini total_counts++; 30069162d606SStefano Zampini if (!idxs_copied) { 30079162d606SStefano Zampini ierr = PetscMemcpy(constraints_idxs + constraints_idxs_ptr[total_counts_cc],is_indices,size_of_constraint*sizeof(PetscInt));CHKERRQ(ierr); 30089162d606SStefano Zampini idxs_copied = PETSC_TRUE; 3009674ae819SStefano Zampini } 3010674ae819SStefano Zampini } 30119162d606SStefano Zampini } 30129162d606SStefano Zampini ierr = ISRestoreIndices(used_is,(const PetscInt**)&is_indices);CHKERRQ(ierr); 301345a1bb75SStefano Zampini valid_constraints = temp_constraints; 3014eb97c9d2SStefano Zampini if (!pcbddc->use_nnsp_true && temp_constraints) { 3015a773dcb8SStefano Zampini if (temp_constraints == 1) { /* just normalize the constraint */ 30169162d606SStefano Zampini PetscScalar norm,*ptr_to_data; 30179162d606SStefano Zampini 30189162d606SStefano Zampini ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]]; 3019a773dcb8SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 30209162d606SStefano Zampini PetscStackCallBLAS("BLASdot",norm = BLASdot_(&Blas_N,ptr_to_data,&Blas_one,ptr_to_data,&Blas_one)); 3021a773dcb8SStefano Zampini norm = 1.0/PetscSqrtReal(PetscRealPart(norm)); 30229162d606SStefano Zampini PetscStackCallBLAS("BLASscal",BLASscal_(&Blas_N,&norm,ptr_to_data,&Blas_one)); 3023a773dcb8SStefano Zampini } else { /* perform SVD */ 3024984c4197SStefano Zampini PetscReal tol = 1.0e-8; /* tolerance for retaining eigenmodes */ 30259162d606SStefano Zampini PetscScalar *ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]]; 3026674ae819SStefano Zampini 3027674ae819SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 3028984c4197SStefano Zampini /* SVD: Y = U*S*V^H -> U (eigenvectors of Y*Y^H) = Y*V*(S)^\dag 3029984c4197SStefano Zampini POD: Y^H*Y = V*D*V^H, D = S^H*S -> U = Y*V*D^(-1/2) 3030984c4197SStefano Zampini -> When PETSC_USE_COMPLEX and PETSC_MISSING_LAPACK_GESVD are defined 3031984c4197SStefano Zampini the constraints basis will differ (by a complex factor with absolute value equal to 1) 3032984c4197SStefano Zampini from that computed using LAPACKgesvd 3033984c4197SStefano Zampini -> This is due to a different computation of eigenvectors in LAPACKheev 3034984c4197SStefano Zampini -> The quality of the POD-computed basis will be the same */ 3035984c4197SStefano Zampini ierr = PetscMemzero(correlation_mat,temp_constraints*temp_constraints*sizeof(PetscScalar));CHKERRQ(ierr); 3036674ae819SStefano Zampini /* Store upper triangular part of correlation matrix */ 3037e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 3038984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 3039674ae819SStefano Zampini for (j=0;j<temp_constraints;j++) { 3040674ae819SStefano Zampini for (k=0;k<j+1;k++) { 30419162d606SStefano 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)); 3042674ae819SStefano Zampini } 3043674ae819SStefano Zampini } 3044e310c8b4SStefano Zampini /* compute eigenvalues and eigenvectors of correlation matrix */ 3045e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 3046e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_LDA);CHKERRQ(ierr); 3047674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 3048c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,work,&lwork,&lierr)); 3049674ae819SStefano Zampini #else 3050c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,work,&lwork,rwork,&lierr)); 3051674ae819SStefano Zampini #endif 3052674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 3053984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYEV Lapack routine %d",(int)lierr); 3054984c4197SStefano Zampini /* retain eigenvalues greater than tol: note that LAPACKsyev gives eigs in ascending order */ 3055674ae819SStefano Zampini j = 0; 3056984c4197SStefano Zampini while (j < temp_constraints && singular_vals[j] < tol) j++; 3057674ae819SStefano Zampini total_counts = total_counts-j; 305845a1bb75SStefano Zampini valid_constraints = temp_constraints-j; 3059e310c8b4SStefano Zampini /* scale and copy POD basis into used quadrature memory */ 3060c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 3061c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 3062c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_K);CHKERRQ(ierr); 3063c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 3064c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_LDB);CHKERRQ(ierr); 3065c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDC);CHKERRQ(ierr); 3066674ae819SStefano Zampini if (j<temp_constraints) { 3067984c4197SStefano Zampini PetscInt ii; 3068984c4197SStefano Zampini for (k=j;k<temp_constraints;k++) singular_vals[k] = 1.0/PetscSqrtReal(singular_vals[k]); 3069674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 30709162d606SStefano 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)); 3071674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 3072984c4197SStefano Zampini for (k=0;k<temp_constraints-j;k++) { 3073674ae819SStefano Zampini for (ii=0;ii<size_of_constraint;ii++) { 30749162d606SStefano 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]; 3075674ae819SStefano Zampini } 3076674ae819SStefano Zampini } 3077674ae819SStefano Zampini } 3078674ae819SStefano Zampini #else /* on missing GESVD */ 3079e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 3080e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 3081b7d8b9f8SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 3082674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 3083674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 30849162d606SStefano 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)); 3085674ae819SStefano Zampini #else 30869162d606SStefano 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)); 3087674ae819SStefano Zampini #endif 3088984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GESVD Lapack routine %d",(int)lierr); 3089674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 3090984c4197SStefano Zampini /* retain eigenvalues greater than tol: note that LAPACKgesvd gives eigs in descending order */ 3091e310c8b4SStefano Zampini k = temp_constraints; 3092e310c8b4SStefano Zampini if (k > size_of_constraint) k = size_of_constraint; 3093674ae819SStefano Zampini j = 0; 3094e310c8b4SStefano Zampini while (j < k && singular_vals[k-j-1] < tol) j++; 309545a1bb75SStefano Zampini valid_constraints = k-j; 3096911cabfeSStefano Zampini total_counts = total_counts-temp_constraints+valid_constraints; 3097984c4197SStefano Zampini #endif /* on missing GESVD */ 3098674ae819SStefano Zampini } 3099a773dcb8SStefano Zampini } 31009162d606SStefano Zampini /* update pointers information */ 31019162d606SStefano Zampini if (valid_constraints) { 31029162d606SStefano Zampini constraints_n[total_counts_cc] = valid_constraints; 31039162d606SStefano Zampini constraints_idxs_ptr[total_counts_cc+1] = constraints_idxs_ptr[total_counts_cc]+size_of_constraint; 31049162d606SStefano Zampini constraints_data_ptr[total_counts_cc+1] = constraints_data_ptr[total_counts_cc]+size_of_constraint*valid_constraints; 31059162d606SStefano Zampini /* set change_of_basis flag */ 310645a1bb75SStefano Zampini if (boolforchange) { 3107b3d85658SStefano Zampini PetscBTSet(change_basis,total_counts_cc); 31089162d606SStefano Zampini } 3109b3d85658SStefano Zampini total_counts_cc++; 311045a1bb75SStefano Zampini } 311145a1bb75SStefano Zampini } 3112984c4197SStefano Zampini /* free workspace */ 31138f1c130eSStefano Zampini if (!skip_lapack) { 3114984c4197SStefano Zampini ierr = PetscFree(work);CHKERRQ(ierr); 3115984c4197SStefano Zampini #if defined(PETSC_USE_COMPLEX) 3116984c4197SStefano Zampini ierr = PetscFree(rwork);CHKERRQ(ierr); 3117984c4197SStefano Zampini #endif 3118984c4197SStefano Zampini ierr = PetscFree(singular_vals);CHKERRQ(ierr); 3119984c4197SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 3120984c4197SStefano Zampini ierr = PetscFree(correlation_mat);CHKERRQ(ierr); 3121984c4197SStefano Zampini ierr = PetscFree(temp_basis);CHKERRQ(ierr); 3122984c4197SStefano Zampini #endif 3123984c4197SStefano Zampini } 3124984c4197SStefano Zampini for (k=0;k<nnsp_size;k++) { 3125984c4197SStefano Zampini ierr = VecDestroy(&localnearnullsp[k]);CHKERRQ(ierr); 3126984c4197SStefano Zampini } 3127984c4197SStefano Zampini ierr = PetscFree(localnearnullsp);CHKERRQ(ierr); 3128cf5a6209SStefano Zampini /* free index sets of faces, edges and vertices */ 3129cf5a6209SStefano Zampini for (i=0;i<n_ISForFaces;i++) { 3130cf5a6209SStefano Zampini ierr = ISDestroy(&ISForFaces[i]);CHKERRQ(ierr); 3131cf5a6209SStefano Zampini } 3132cf5a6209SStefano Zampini if (n_ISForFaces) { 3133cf5a6209SStefano Zampini ierr = PetscFree(ISForFaces);CHKERRQ(ierr); 3134cf5a6209SStefano Zampini } 3135cf5a6209SStefano Zampini for (i=0;i<n_ISForEdges;i++) { 3136cf5a6209SStefano Zampini ierr = ISDestroy(&ISForEdges[i]);CHKERRQ(ierr); 3137cf5a6209SStefano Zampini } 3138cf5a6209SStefano Zampini if (n_ISForEdges) { 3139cf5a6209SStefano Zampini ierr = PetscFree(ISForEdges);CHKERRQ(ierr); 3140cf5a6209SStefano Zampini } 3141cf5a6209SStefano Zampini ierr = ISDestroy(&ISForVertices);CHKERRQ(ierr); 314208122e43SStefano Zampini } else { 314308122e43SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 3144984c4197SStefano Zampini 314508122e43SStefano Zampini total_counts = 0; 314608122e43SStefano Zampini n_vertices = 0; 3147d62866d3SStefano Zampini if (sub_schurs->is_vertices && pcbddc->use_vertices) { 3148d62866d3SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_vertices,&n_vertices);CHKERRQ(ierr); 314908122e43SStefano Zampini } 315008122e43SStefano Zampini max_constraints = 0; 31519162d606SStefano Zampini total_counts_cc = 0; 315208122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs+n_vertices;i++) { 315308122e43SStefano Zampini total_counts += pcbddc->adaptive_constraints_n[i]; 31549162d606SStefano Zampini if (pcbddc->adaptive_constraints_n[i]) total_counts_cc++; 315508122e43SStefano Zampini max_constraints = PetscMax(max_constraints,pcbddc->adaptive_constraints_n[i]); 315608122e43SStefano Zampini } 31579162d606SStefano Zampini constraints_idxs_ptr = pcbddc->adaptive_constraints_idxs_ptr; 31589162d606SStefano Zampini constraints_data_ptr = pcbddc->adaptive_constraints_data_ptr; 31599162d606SStefano Zampini constraints_idxs = pcbddc->adaptive_constraints_idxs; 31609162d606SStefano Zampini constraints_data = pcbddc->adaptive_constraints_data; 316174d5cdf7SStefano Zampini /* constraints_n differs from pcbddc->adaptive_constraints_n */ 31629162d606SStefano Zampini ierr = PetscMalloc1(total_counts_cc,&constraints_n);CHKERRQ(ierr); 31639162d606SStefano Zampini total_counts_cc = 0; 31649162d606SStefano Zampini for (i=0;i<sub_schurs->n_subs+n_vertices;i++) { 31659162d606SStefano Zampini if (pcbddc->adaptive_constraints_n[i]) { 31669162d606SStefano Zampini constraints_n[total_counts_cc++] = pcbddc->adaptive_constraints_n[i]; 316708122e43SStefano Zampini } 316808122e43SStefano Zampini } 31699162d606SStefano Zampini #if 0 31709162d606SStefano Zampini printf("Found %d totals (%d)\n",total_counts_cc,total_counts); 31719162d606SStefano Zampini for (i=0;i<total_counts_cc;i++) { 31729162d606SStefano Zampini printf("const %d, start %d",i,constraints_idxs_ptr[i]); 31739162d606SStefano Zampini printf(" end %d:\n",constraints_idxs_ptr[i+1]); 31749162d606SStefano Zampini for (j=constraints_idxs_ptr[i];j<constraints_idxs_ptr[i+1];j++) { 31759162d606SStefano Zampini printf(" %d",constraints_idxs[j]); 31769162d606SStefano Zampini } 31779162d606SStefano Zampini printf("\n"); 31789162d606SStefano Zampini printf("number of cc: %d\n",constraints_n[i]); 31799162d606SStefano Zampini } 31801b968477SStefano Zampini for (i=0;i<n_vertices;i++) { 31818bec7fa6SStefano Zampini PetscPrintf(PETSC_COMM_SELF,"[%d] vertex %d, n %d\n",PetscGlobalRank,i,pcbddc->adaptive_constraints_n[i]); 31821b968477SStefano Zampini } 31831b968477SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 31848bec7fa6SStefano 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]); 31851b968477SStefano Zampini } 318608122e43SStefano Zampini #endif 318708122e43SStefano Zampini 31888bec7fa6SStefano Zampini max_size_of_constraint = 0; 31899162d606SStefano 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]); 31909162d606SStefano Zampini ierr = PetscMalloc1(constraints_idxs_ptr[total_counts_cc],&constraints_idxs_B);CHKERRQ(ierr); 319108122e43SStefano Zampini /* Change of basis */ 3192b3d85658SStefano Zampini ierr = PetscBTCreate(total_counts_cc,&change_basis);CHKERRQ(ierr); 319308122e43SStefano Zampini if (pcbddc->use_change_of_basis) { 319408122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 319508122e43SStefano Zampini if (PetscBTLookup(sub_schurs->is_edge,i) || pcbddc->use_change_on_faces) { 3196b3d85658SStefano Zampini ierr = PetscBTSet(change_basis,i+n_vertices);CHKERRQ(ierr); 319708122e43SStefano Zampini } 319808122e43SStefano Zampini } 319908122e43SStefano Zampini } 320008122e43SStefano Zampini } 3201984c4197SStefano Zampini pcbddc->local_primal_size = total_counts; 32024f1b2e48SStefano Zampini ierr = PetscMalloc1(pcbddc->local_primal_size+pcbddc->benign_n,&pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 320308122e43SStefano Zampini 32049162d606SStefano Zampini /* map constraints_idxs in boundary numbering */ 32059162d606SStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,constraints_idxs_ptr[total_counts_cc],constraints_idxs,&i,constraints_idxs_B);CHKERRQ(ierr); 32069162d606SStefano Zampini if (i != constraints_idxs_ptr[total_counts_cc]) { 32079162d606SStefano 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); 320808122e43SStefano Zampini } 3209674ae819SStefano Zampini 3210674ae819SStefano Zampini /* Create constraint matrix */ 3211674ae819SStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 321216f15bc4SStefano Zampini ierr = MatSetType(pcbddc->ConstraintMatrix,MATAIJ);CHKERRQ(ierr); 3213984c4197SStefano Zampini ierr = MatSetSizes(pcbddc->ConstraintMatrix,pcbddc->local_primal_size,pcis->n,pcbddc->local_primal_size,pcis->n);CHKERRQ(ierr); 3214984c4197SStefano Zampini 3215984c4197SStefano Zampini /* find primal_dofs: subdomain corners plus dofs selected as primal after change of basis */ 3216a717540cSStefano Zampini /* determine if a QR strategy is needed for change of basis */ 3217a717540cSStefano Zampini qr_needed = PETSC_FALSE; 321874d5cdf7SStefano Zampini ierr = PetscBTCreate(total_counts_cc,&qr_needed_idx);CHKERRQ(ierr); 3219984c4197SStefano Zampini total_primal_vertices=0; 3220b3d85658SStefano Zampini pcbddc->local_primal_size_cc = 0; 32219162d606SStefano Zampini for (i=0;i<total_counts_cc;i++) { 32229162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 3223984c4197SStefano Zampini if (size_of_constraint == 1) { 32249162d606SStefano Zampini pcbddc->primal_indices_local_idxs[total_primal_vertices++] = constraints_idxs[constraints_idxs_ptr[i]]; 3225b3d85658SStefano Zampini pcbddc->local_primal_size_cc += 1; 322664efe560SStefano Zampini } else if (PetscBTLookup(change_basis,i)) { 32279162d606SStefano Zampini for (k=0;k<constraints_n[i];k++) { 32289162d606SStefano Zampini pcbddc->primal_indices_local_idxs[total_primal_vertices++] = constraints_idxs[constraints_idxs_ptr[i]+k]; 3229a717540cSStefano Zampini } 3230b3d85658SStefano Zampini pcbddc->local_primal_size_cc += constraints_n[i]; 323174d5cdf7SStefano Zampini if (constraints_n[i] > 1 || pcbddc->use_qr_single || pcbddc->faster_deluxe) { 3232a717540cSStefano Zampini PetscBTSet(qr_needed_idx,i); 3233a717540cSStefano Zampini qr_needed = PETSC_TRUE; 3234a717540cSStefano Zampini } 3235fa434743SStefano Zampini } else { 3236b3d85658SStefano Zampini pcbddc->local_primal_size_cc += 1; 3237fa434743SStefano Zampini } 3238a717540cSStefano Zampini } 3239b371cd4fSStefano Zampini /* note that the local variable n_vertices used below stores the number of pointwise constraints */ 3240b371cd4fSStefano Zampini pcbddc->n_vertices = total_primal_vertices; 3241674ae819SStefano Zampini /* permute indices in order to have a sorted set of vertices */ 324270022509SStefano Zampini ierr = PetscSortInt(total_primal_vertices,pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 3243b3d85658SStefano Zampini 32444f1b2e48SStefano 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); 32450e6343abSStefano Zampini ierr = PetscMemcpy(pcbddc->local_primal_ref_node,pcbddc->primal_indices_local_idxs,total_primal_vertices*sizeof(PetscInt));CHKERRQ(ierr); 32460e6343abSStefano Zampini for (i=0;i<total_primal_vertices;i++) pcbddc->local_primal_ref_mult[i] = 1; 3247984c4197SStefano Zampini 3248984c4197SStefano Zampini /* nonzero structure of constraint matrix */ 324974d5cdf7SStefano Zampini /* and get reference dof for local constraints */ 3250785e854fSJed Brown ierr = PetscMalloc1(pcbddc->local_primal_size,&nnz);CHKERRQ(ierr); 3251984c4197SStefano Zampini for (i=0;i<total_primal_vertices;i++) nnz[i] = 1; 325274d5cdf7SStefano Zampini 3253984c4197SStefano Zampini j = total_primal_vertices; 325474d5cdf7SStefano Zampini total_counts = total_primal_vertices; 3255b3d85658SStefano Zampini cum = total_primal_vertices; 32569162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 32574641a718SStefano Zampini if (!PetscBTLookup(change_basis,i)) { 3258b3d85658SStefano Zampini pcbddc->local_primal_ref_node[cum] = constraints_idxs[constraints_idxs_ptr[i]]; 3259b3d85658SStefano Zampini pcbddc->local_primal_ref_mult[cum] = constraints_n[i]; 3260b3d85658SStefano Zampini cum++; 32619162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 326274d5cdf7SStefano Zampini for (k=0;k<constraints_n[i];k++) { 326374d5cdf7SStefano Zampini pcbddc->primal_indices_local_idxs[total_counts++] = constraints_idxs[constraints_idxs_ptr[i]+k]; 326474d5cdf7SStefano Zampini nnz[j+k] = size_of_constraint; 326574d5cdf7SStefano Zampini } 32669162d606SStefano Zampini j += constraints_n[i]; 3267674ae819SStefano Zampini } 3268674ae819SStefano Zampini } 3269674ae819SStefano Zampini ierr = MatSeqAIJSetPreallocation(pcbddc->ConstraintMatrix,0,nnz);CHKERRQ(ierr); 3270674ae819SStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 3271088faed8SStefano Zampini 3272674ae819SStefano Zampini /* set values in constraint matrix */ 3273984c4197SStefano Zampini for (i=0;i<total_primal_vertices;i++) { 32740e6343abSStefano Zampini ierr = MatSetValue(pcbddc->ConstraintMatrix,i,pcbddc->local_primal_ref_node[i],1.0,INSERT_VALUES);CHKERRQ(ierr); 3275674ae819SStefano Zampini } 3276984c4197SStefano Zampini total_counts = total_primal_vertices; 32779162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 32784641a718SStefano Zampini if (!PetscBTLookup(change_basis,i)) { 32799162d606SStefano Zampini PetscInt *cols; 32809162d606SStefano Zampini 32819162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 32829162d606SStefano Zampini cols = constraints_idxs+constraints_idxs_ptr[i]; 32839162d606SStefano Zampini for (k=0;k<constraints_n[i];k++) { 32849162d606SStefano Zampini PetscInt row = total_counts+k; 32859162d606SStefano Zampini PetscScalar *vals; 32869162d606SStefano Zampini 32879162d606SStefano Zampini vals = constraints_data+constraints_data_ptr[i]+k*size_of_constraint; 32889162d606SStefano Zampini ierr = MatSetValues(pcbddc->ConstraintMatrix,1,&row,size_of_constraint,cols,vals,INSERT_VALUES);CHKERRQ(ierr); 32899162d606SStefano Zampini } 32909162d606SStefano Zampini total_counts += constraints_n[i]; 3291674ae819SStefano Zampini } 3292674ae819SStefano Zampini } 3293674ae819SStefano Zampini /* assembling */ 3294674ae819SStefano Zampini ierr = MatAssemblyBegin(pcbddc->ConstraintMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3295674ae819SStefano Zampini ierr = MatAssemblyEnd(pcbddc->ConstraintMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3296088faed8SStefano Zampini 3297984c4197SStefano Zampini /* 329845a1bb75SStefano Zampini ierr = PetscViewerSetFormat(PETSC_VIEWER_STDOUT_SELF,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 3299984c4197SStefano Zampini ierr = MatView(pcbddc->ConstraintMatrix,(PetscViewer)0);CHKERRQ(ierr); 3300984c4197SStefano Zampini */ 3301674ae819SStefano Zampini /* Create matrix for change of basis. We don't need it in case pcbddc->use_change_of_basis is FALSE */ 3302674ae819SStefano Zampini if (pcbddc->use_change_of_basis) { 3303026de310SStefano Zampini /* dual and primal dofs on a single cc */ 3304984c4197SStefano Zampini PetscInt dual_dofs,primal_dofs; 3305984c4197SStefano Zampini /* working stuff for GEQRF */ 330681d9aea3SBarry Smith PetscScalar *qr_basis,*qr_tau = NULL,*qr_work,lqr_work_t; 3307984c4197SStefano Zampini PetscBLASInt lqr_work; 3308984c4197SStefano Zampini /* working stuff for UNGQR */ 3309984c4197SStefano Zampini PetscScalar *gqr_work,lgqr_work_t; 3310984c4197SStefano Zampini PetscBLASInt lgqr_work; 3311984c4197SStefano Zampini /* working stuff for TRTRS */ 3312984c4197SStefano Zampini PetscScalar *trs_rhs; 33133f08241aSStefano Zampini PetscBLASInt Blas_NRHS; 3314984c4197SStefano Zampini /* pointers for values insertion into change of basis matrix */ 3315984c4197SStefano Zampini PetscInt *start_rows,*start_cols; 3316984c4197SStefano Zampini PetscScalar *start_vals; 3317984c4197SStefano Zampini /* working stuff for values insertion */ 33184641a718SStefano Zampini PetscBT is_primal; 331964efe560SStefano Zampini PetscInt *aux_primal_numbering_B; 3320906d46d4SStefano Zampini /* matrix sizes */ 3321906d46d4SStefano Zampini PetscInt global_size,local_size; 3322906d46d4SStefano Zampini /* temporary change of basis */ 3323906d46d4SStefano Zampini Mat localChangeOfBasisMatrix; 3324cf5a6209SStefano Zampini /* extra space for debugging */ 3325cf5a6209SStefano Zampini PetscScalar *dbg_work; 3326984c4197SStefano Zampini 3327906d46d4SStefano Zampini /* local temporary change of basis acts on local interfaces -> dimension is n_B x n_B */ 3328906d46d4SStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&localChangeOfBasisMatrix);CHKERRQ(ierr); 332916f15bc4SStefano Zampini ierr = MatSetType(localChangeOfBasisMatrix,MATAIJ);CHKERRQ(ierr); 3330bbb9e6c6SStefano Zampini ierr = MatSetSizes(localChangeOfBasisMatrix,pcis->n,pcis->n,pcis->n,pcis->n);CHKERRQ(ierr); 3331906d46d4SStefano Zampini /* nonzeros for local mat */ 3332bbb9e6c6SStefano Zampini ierr = PetscMalloc1(pcis->n,&nnz);CHKERRQ(ierr); 3333bbb9e6c6SStefano Zampini for (i=0;i<pcis->n;i++) nnz[i]=1; 33349162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 3335a717540cSStefano Zampini if (PetscBTLookup(change_basis,i)) { 33369162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 3337a717540cSStefano Zampini if (PetscBTLookup(qr_needed_idx,i)) { 33389162d606SStefano Zampini for (j=0;j<size_of_constraint;j++) nnz[constraints_idxs[constraints_idxs_ptr[i]+j]] = size_of_constraint; 3339a717540cSStefano Zampini } else { 33409162d606SStefano Zampini nnz[constraints_idxs[constraints_idxs_ptr[i]]] = size_of_constraint; 33419162d606SStefano Zampini for (j=1;j<size_of_constraint;j++) nnz[constraints_idxs[constraints_idxs_ptr[i]+j]] = 2; 3342a717540cSStefano Zampini } 3343a717540cSStefano Zampini } 3344a717540cSStefano Zampini } 3345906d46d4SStefano Zampini ierr = MatSeqAIJSetPreallocation(localChangeOfBasisMatrix,0,nnz);CHKERRQ(ierr); 3346bbb9e6c6SStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 3347a717540cSStefano Zampini /* Set initial identity in the matrix */ 3348bbb9e6c6SStefano Zampini for (i=0;i<pcis->n;i++) { 3349906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,i,i,1.0,INSERT_VALUES);CHKERRQ(ierr); 3350a717540cSStefano Zampini } 3351a717540cSStefano Zampini 3352a717540cSStefano Zampini if (pcbddc->dbg_flag) { 3353a717540cSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"--------------------------------------------------------------\n");CHKERRQ(ierr); 3354a717540cSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Checking change of basis computation for subdomain %04d\n",PetscGlobalRank);CHKERRQ(ierr); 3355a717540cSStefano Zampini } 3356a717540cSStefano Zampini 3357a717540cSStefano Zampini 3358a717540cSStefano Zampini /* Now we loop on the constraints which need a change of basis */ 3359a717540cSStefano Zampini /* 3360a717540cSStefano Zampini Change of basis matrix is evaluated similarly to the FIRST APPROACH in 3361a717540cSStefano Zampini Klawonn and Widlund, Dual-primal FETI-DP methods for linear elasticity, (see Sect 6.2.1) 3362a717540cSStefano Zampini 3363a6b551f4SStefano Zampini Basic blocks of change of basis matrix T computed by 3364a717540cSStefano Zampini 3365a6b551f4SStefano 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) 3366a6b551f4SStefano Zampini 3367a6b551f4SStefano Zampini | 1 0 ... 0 s_1/S | 3368a6b551f4SStefano Zampini | 0 1 ... 0 s_2/S | 3369a717540cSStefano Zampini | ... | 3370a6b551f4SStefano Zampini | 0 ... 1 s_{n-1}/S | 3371a6b551f4SStefano Zampini | -s_1/s_n ... -s_{n-1}/s_n s_n/S | 3372a717540cSStefano Zampini 3373a6b551f4SStefano Zampini with S = \sum_{i=1}^n s_i^2 3374a6b551f4SStefano Zampini NOTE: in the above example, the primal dof is the last one of the edge in LOCAL ordering 3375a6b551f4SStefano Zampini in the current implementation, the primal dof is the first one of the edge in GLOBAL ordering 3376a6b551f4SStefano Zampini 3377a6b551f4SStefano Zampini - QR decomposition of constraints otherwise 3378a717540cSStefano Zampini */ 3379a717540cSStefano Zampini if (qr_needed) { 3380984c4197SStefano Zampini /* space to store Q */ 3381854ce69bSBarry Smith ierr = PetscMalloc1(max_size_of_constraint*max_size_of_constraint,&qr_basis);CHKERRQ(ierr); 3382984c4197SStefano Zampini /* first we issue queries for optimal work */ 33833f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_M);CHKERRQ(ierr); 33843f08241aSStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_N);CHKERRQ(ierr); 33853f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 3386984c4197SStefano Zampini lqr_work = -1; 33873f08241aSStefano Zampini PetscStackCallBLAS("LAPACKgeqrf",LAPACKgeqrf_(&Blas_M,&Blas_N,qr_basis,&Blas_LDA,qr_tau,&lqr_work_t,&lqr_work,&lierr)); 3388984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to GEQRF Lapack routine %d",(int)lierr); 3389984c4197SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lqr_work_t),&lqr_work);CHKERRQ(ierr); 3390785e854fSJed Brown ierr = PetscMalloc1((PetscInt)PetscRealPart(lqr_work_t),&qr_work);CHKERRQ(ierr); 3391984c4197SStefano Zampini lgqr_work = -1; 33923f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_M);CHKERRQ(ierr); 33933f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_N);CHKERRQ(ierr); 33943f08241aSStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_K);CHKERRQ(ierr); 33953f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 33963f08241aSStefano Zampini if (Blas_K>Blas_M) Blas_K=Blas_M; /* adjust just for computing optimal work */ 33973f08241aSStefano Zampini PetscStackCallBLAS("LAPACKungqr",LAPACKungqr_(&Blas_M,&Blas_N,&Blas_K,qr_basis,&Blas_LDA,qr_tau,&lgqr_work_t,&lgqr_work,&lierr)); 3398984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to UNGQR Lapack routine %d",(int)lierr); 3399984c4197SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lgqr_work_t),&lgqr_work);CHKERRQ(ierr); 3400785e854fSJed Brown ierr = PetscMalloc1((PetscInt)PetscRealPart(lgqr_work_t),&gqr_work);CHKERRQ(ierr); 3401984c4197SStefano Zampini /* array to store scaling factors for reflectors */ 3402785e854fSJed Brown ierr = PetscMalloc1(max_constraints,&qr_tau);CHKERRQ(ierr); 3403984c4197SStefano Zampini /* array to store rhs and solution of triangular solver */ 3404785e854fSJed Brown ierr = PetscMalloc1(max_constraints*max_constraints,&trs_rhs);CHKERRQ(ierr); 3405a717540cSStefano Zampini /* allocating workspace for check */ 3406a717540cSStefano Zampini if (pcbddc->dbg_flag) { 3407cf5a6209SStefano Zampini ierr = PetscMalloc1(max_size_of_constraint*(max_constraints+max_size_of_constraint),&dbg_work);CHKERRQ(ierr); 3408a717540cSStefano Zampini } 3409a717540cSStefano Zampini } 3410984c4197SStefano Zampini /* array to store whether a node is primal or not */ 34114641a718SStefano Zampini ierr = PetscBTCreate(pcis->n_B,&is_primal);CHKERRQ(ierr); 3412473ba861SJed Brown ierr = PetscMalloc1(total_primal_vertices,&aux_primal_numbering_B);CHKERRQ(ierr); 34130e6343abSStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,total_primal_vertices,pcbddc->local_primal_ref_node,&i,aux_primal_numbering_B);CHKERRQ(ierr); 341439e2fb2aSStefano Zampini if (i != total_primal_vertices) { 341539e2fb2aSStefano Zampini SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_SUP,"Error in boundary numbering for BDDC vertices! %d != %d\n",total_primal_vertices,i); 34164641a718SStefano Zampini } 341739e2fb2aSStefano Zampini for (i=0;i<total_primal_vertices;i++) { 341839e2fb2aSStefano Zampini ierr = PetscBTSet(is_primal,aux_primal_numbering_B[i]);CHKERRQ(ierr); 341939e2fb2aSStefano Zampini } 342039e2fb2aSStefano Zampini ierr = PetscFree(aux_primal_numbering_B);CHKERRQ(ierr); 3421984c4197SStefano Zampini 3422a717540cSStefano Zampini /* loop on constraints and see whether or not they need a change of basis and compute it */ 34239162d606SStefano Zampini for (total_counts=n_vertices;total_counts<total_counts_cc;total_counts++) { 34249162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[total_counts+1]-constraints_idxs_ptr[total_counts]; 34254641a718SStefano Zampini if (PetscBTLookup(change_basis,total_counts)) { 3426984c4197SStefano Zampini /* get constraint info */ 34279162d606SStefano Zampini primal_dofs = constraints_n[total_counts]; 3428984c4197SStefano Zampini dual_dofs = size_of_constraint-primal_dofs; 3429984c4197SStefano Zampini 3430984c4197SStefano Zampini if (pcbddc->dbg_flag) { 34319162d606SStefano 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); 3432674ae819SStefano Zampini } 3433984c4197SStefano Zampini 3434fa434743SStefano Zampini if (PetscBTLookup(qr_needed_idx,total_counts)) { /* QR */ 3435a717540cSStefano Zampini 3436a717540cSStefano Zampini /* copy quadrature constraints for change of basis check */ 3437a717540cSStefano Zampini if (pcbddc->dbg_flag) { 34389162d606SStefano Zampini ierr = PetscMemcpy(dbg_work,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 3439a717540cSStefano Zampini } 3440984c4197SStefano Zampini /* copy temporary constraints into larger work vector (in order to store all columns of Q) */ 34419162d606SStefano Zampini ierr = PetscMemcpy(qr_basis,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 3442984c4197SStefano Zampini 3443984c4197SStefano Zampini /* compute QR decomposition of constraints */ 34443f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 34453f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 34463f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 3447674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 34483f08241aSStefano Zampini PetscStackCallBLAS("LAPACKgeqrf",LAPACKgeqrf_(&Blas_M,&Blas_N,qr_basis,&Blas_LDA,qr_tau,qr_work,&lqr_work,&lierr)); 3449984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GEQRF Lapack routine %d",(int)lierr); 3450674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 3451984c4197SStefano Zampini 3452984c4197SStefano Zampini /* explictly compute R^-T */ 3453984c4197SStefano Zampini ierr = PetscMemzero(trs_rhs,primal_dofs*primal_dofs*sizeof(*trs_rhs));CHKERRQ(ierr); 3454984c4197SStefano Zampini for (j=0;j<primal_dofs;j++) trs_rhs[j*(primal_dofs+1)] = 1.0; 34553f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 34563f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_NRHS);CHKERRQ(ierr); 34573f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 34583f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDB);CHKERRQ(ierr); 3459984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 34603f08241aSStefano Zampini PetscStackCallBLAS("LAPACKtrtrs",LAPACKtrtrs_("U","T","N",&Blas_N,&Blas_NRHS,qr_basis,&Blas_LDA,trs_rhs,&Blas_LDB,&lierr)); 3461984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in TRTRS Lapack routine %d",(int)lierr); 3462984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 3463984c4197SStefano Zampini 3464a717540cSStefano Zampini /* explicitly compute all columns of Q (Q = [Q1 | Q2] ) overwriting QR factorization in qr_basis */ 34653f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 34663f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 34673f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_K);CHKERRQ(ierr); 34683f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 3469984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 34703f08241aSStefano Zampini PetscStackCallBLAS("LAPACKungqr",LAPACKungqr_(&Blas_M,&Blas_N,&Blas_K,qr_basis,&Blas_LDA,qr_tau,gqr_work,&lgqr_work,&lierr)); 3471984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in UNGQR Lapack routine %d",(int)lierr); 3472984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 3473984c4197SStefano Zampini 3474984c4197SStefano Zampini /* first primal_dofs columns of Q need to be re-scaled in order to be unitary w.r.t constraints 3475984c4197SStefano Zampini i.e. C_{pxn}*Q_{nxn} should be equal to [I_pxp | 0_pxd] (see check below) 3476984c4197SStefano Zampini where n=size_of_constraint, p=primal_dofs, d=dual_dofs (n=p+d), I and 0 identity and null matrix resp. */ 34773f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 34783f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 34793f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_K);CHKERRQ(ierr); 34803f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 34813f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDB);CHKERRQ(ierr); 34823f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDC);CHKERRQ(ierr); 3483984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 34849162d606SStefano 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)); 3485984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 34869162d606SStefano Zampini ierr = PetscMemcpy(qr_basis,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 3487984c4197SStefano Zampini 3488984c4197SStefano Zampini /* insert values in change of basis matrix respecting global ordering of new primal dofs */ 34899162d606SStefano Zampini start_rows = &constraints_idxs[constraints_idxs_ptr[total_counts]]; 3490984c4197SStefano Zampini /* insert cols for primal dofs */ 3491984c4197SStefano Zampini for (j=0;j<primal_dofs;j++) { 3492984c4197SStefano Zampini start_vals = &qr_basis[j*size_of_constraint]; 34939162d606SStefano Zampini start_cols = &constraints_idxs[constraints_idxs_ptr[total_counts]+j]; 3494906d46d4SStefano Zampini ierr = MatSetValues(localChangeOfBasisMatrix,size_of_constraint,start_rows,1,start_cols,start_vals,INSERT_VALUES);CHKERRQ(ierr); 3495984c4197SStefano Zampini } 3496984c4197SStefano Zampini /* insert cols for dual dofs */ 3497984c4197SStefano Zampini for (j=0,k=0;j<dual_dofs;k++) { 34989162d606SStefano Zampini if (!PetscBTLookup(is_primal,constraints_idxs_B[constraints_idxs_ptr[total_counts]+k])) { 3499984c4197SStefano Zampini start_vals = &qr_basis[(primal_dofs+j)*size_of_constraint]; 35009162d606SStefano Zampini start_cols = &constraints_idxs[constraints_idxs_ptr[total_counts]+k]; 3501906d46d4SStefano Zampini ierr = MatSetValues(localChangeOfBasisMatrix,size_of_constraint,start_rows,1,start_cols,start_vals,INSERT_VALUES);CHKERRQ(ierr); 3502984c4197SStefano Zampini j++; 3503674ae819SStefano Zampini } 3504674ae819SStefano Zampini } 3505984c4197SStefano Zampini 3506984c4197SStefano Zampini /* check change of basis */ 3507984c4197SStefano Zampini if (pcbddc->dbg_flag) { 3508984c4197SStefano Zampini PetscInt ii,jj; 3509984c4197SStefano Zampini PetscBool valid_qr=PETSC_TRUE; 3510c4303822SStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_M);CHKERRQ(ierr); 3511c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 3512c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_K);CHKERRQ(ierr); 3513c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 3514c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDB);CHKERRQ(ierr); 3515c4303822SStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDC);CHKERRQ(ierr); 3516984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 3517cf5a6209SStefano 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)); 3518984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 3519984c4197SStefano Zampini for (jj=0;jj<size_of_constraint;jj++) { 3520984c4197SStefano Zampini for (ii=0;ii<primal_dofs;ii++) { 3521cf5a6209SStefano Zampini if (ii != jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]) > 1.e-12) valid_qr = PETSC_FALSE; 3522cf5a6209SStefano 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; 3523674ae819SStefano Zampini } 3524674ae819SStefano Zampini } 3525984c4197SStefano Zampini if (!valid_qr) { 352622d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> wrong change of basis!\n");CHKERRQ(ierr); 3527984c4197SStefano Zampini for (jj=0;jj<size_of_constraint;jj++) { 3528984c4197SStefano Zampini for (ii=0;ii<primal_dofs;ii++) { 3529cf5a6209SStefano Zampini if (ii != jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]) > 1.e-12) { 3530cf5a6209SStefano 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])); 3531674ae819SStefano Zampini } 3532cf5a6209SStefano Zampini if (ii == jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]-1.0) > 1.e-12) { 3533cf5a6209SStefano 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])); 3534984c4197SStefano Zampini } 3535984c4197SStefano Zampini } 3536984c4197SStefano Zampini } 3537674ae819SStefano Zampini } else { 353822d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> right change of basis!\n");CHKERRQ(ierr); 3539674ae819SStefano Zampini } 3540674ae819SStefano Zampini } 3541a717540cSStefano Zampini } else { /* simple transformation block */ 3542a717540cSStefano Zampini PetscInt row,col; 3543a6b551f4SStefano Zampini PetscScalar val,norm; 3544a6b551f4SStefano Zampini 3545a6b551f4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 35469162d606SStefano 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)); 3547a717540cSStefano Zampini for (j=0;j<size_of_constraint;j++) { 35489162d606SStefano Zampini PetscInt row_B = constraints_idxs_B[constraints_idxs_ptr[total_counts]+j]; 35499162d606SStefano Zampini row = constraints_idxs[constraints_idxs_ptr[total_counts]+j]; 3550bbb9e6c6SStefano Zampini if (!PetscBTLookup(is_primal,row_B)) { 35519162d606SStefano Zampini col = constraints_idxs[constraints_idxs_ptr[total_counts]]; 3552906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,row,1.0,INSERT_VALUES);CHKERRQ(ierr); 35539162d606SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,col,constraints_data[constraints_data_ptr[total_counts]+j]/norm,INSERT_VALUES);CHKERRQ(ierr); 3554a717540cSStefano Zampini } else { 3555a717540cSStefano Zampini for (k=0;k<size_of_constraint;k++) { 35569162d606SStefano Zampini col = constraints_idxs[constraints_idxs_ptr[total_counts]+k]; 3557a717540cSStefano Zampini if (row != col) { 35589162d606SStefano Zampini val = -constraints_data[constraints_data_ptr[total_counts]+k]/constraints_data[constraints_data_ptr[total_counts]]; 3559a717540cSStefano Zampini } else { 35609162d606SStefano Zampini val = constraints_data[constraints_data_ptr[total_counts]]/norm; 3561a717540cSStefano Zampini } 3562906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,col,val,INSERT_VALUES);CHKERRQ(ierr); 3563a717540cSStefano Zampini } 3564a717540cSStefano Zampini } 3565a717540cSStefano Zampini } 356698a51de6SStefano Zampini if (pcbddc->dbg_flag) { 356722d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> using standard change of basis\n");CHKERRQ(ierr); 3568a717540cSStefano Zampini } 3569674ae819SStefano Zampini } 3570984c4197SStefano Zampini } else { 3571984c4197SStefano Zampini if (pcbddc->dbg_flag) { 35729162d606SStefano 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); 3573674ae819SStefano Zampini } 3574674ae819SStefano Zampini } 3575674ae819SStefano Zampini } 3576a717540cSStefano Zampini 3577a717540cSStefano Zampini /* free workspace */ 3578a717540cSStefano Zampini if (qr_needed) { 3579984c4197SStefano Zampini if (pcbddc->dbg_flag) { 3580cf5a6209SStefano Zampini ierr = PetscFree(dbg_work);CHKERRQ(ierr); 3581984c4197SStefano Zampini } 3582984c4197SStefano Zampini ierr = PetscFree(trs_rhs);CHKERRQ(ierr); 3583984c4197SStefano Zampini ierr = PetscFree(qr_tau);CHKERRQ(ierr); 3584984c4197SStefano Zampini ierr = PetscFree(qr_work);CHKERRQ(ierr); 3585984c4197SStefano Zampini ierr = PetscFree(gqr_work);CHKERRQ(ierr); 3586984c4197SStefano Zampini ierr = PetscFree(qr_basis);CHKERRQ(ierr); 3587674ae819SStefano Zampini } 3588a717540cSStefano Zampini ierr = PetscBTDestroy(&is_primal);CHKERRQ(ierr); 3589906d46d4SStefano Zampini ierr = MatAssemblyBegin(localChangeOfBasisMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3590906d46d4SStefano Zampini ierr = MatAssemblyEnd(localChangeOfBasisMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3591906d46d4SStefano Zampini 3592906d46d4SStefano Zampini /* assembling of global change of variable */ 3593bbb9e6c6SStefano Zampini { 3594bbb9e6c6SStefano Zampini Mat tmat; 359516f15bc4SStefano Zampini PetscInt bs; 359616f15bc4SStefano Zampini 3597906d46d4SStefano Zampini ierr = VecGetSize(pcis->vec1_global,&global_size);CHKERRQ(ierr); 3598906d46d4SStefano Zampini ierr = VecGetLocalSize(pcis->vec1_global,&local_size);CHKERRQ(ierr); 3599bbb9e6c6SStefano Zampini ierr = MatDuplicate(pc->pmat,MAT_DO_NOT_COPY_VALUES,&tmat);CHKERRQ(ierr); 3600bbb9e6c6SStefano Zampini ierr = MatISSetLocalMat(tmat,localChangeOfBasisMatrix);CHKERRQ(ierr); 3601bbb9e6c6SStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)pc),&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 3602bbb9e6c6SStefano Zampini ierr = MatSetType(pcbddc->ChangeOfBasisMatrix,MATAIJ);CHKERRQ(ierr); 360316f15bc4SStefano Zampini ierr = MatGetBlockSize(pc->pmat,&bs);CHKERRQ(ierr); 360416f15bc4SStefano Zampini ierr = MatSetBlockSize(pcbddc->ChangeOfBasisMatrix,bs);CHKERRQ(ierr); 3605906d46d4SStefano Zampini ierr = MatSetSizes(pcbddc->ChangeOfBasisMatrix,local_size,local_size,global_size,global_size);CHKERRQ(ierr); 3606bbb9e6c6SStefano Zampini ierr = MatISSetMPIXAIJPreallocation_Private(tmat,pcbddc->ChangeOfBasisMatrix,PETSC_TRUE);CHKERRQ(ierr); 3607bbb9e6c6SStefano Zampini ierr = MatISGetMPIXAIJ(tmat,MAT_REUSE_MATRIX,&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 3608bbb9e6c6SStefano Zampini ierr = MatDestroy(&tmat);CHKERRQ(ierr); 3609bbb9e6c6SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 3610bbb9e6c6SStefano Zampini ierr = VecSet(pcis->vec1_N,1.0);CHKERRQ(ierr); 3611e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 3612e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 3613bbb9e6c6SStefano Zampini ierr = VecReciprocal(pcis->vec1_global);CHKERRQ(ierr); 3614bbb9e6c6SStefano Zampini ierr = MatDiagonalScale(pcbddc->ChangeOfBasisMatrix,pcis->vec1_global,NULL);CHKERRQ(ierr); 3615906d46d4SStefano Zampini } 3616906d46d4SStefano Zampini /* check */ 3617906d46d4SStefano Zampini if (pcbddc->dbg_flag) { 3618906d46d4SStefano Zampini PetscReal error; 3619906d46d4SStefano Zampini Vec x,x_change; 3620906d46d4SStefano Zampini 3621906d46d4SStefano Zampini ierr = VecDuplicate(pcis->vec1_global,&x);CHKERRQ(ierr); 3622906d46d4SStefano Zampini ierr = VecDuplicate(pcis->vec1_global,&x_change);CHKERRQ(ierr); 3623906d46d4SStefano Zampini ierr = VecSetRandom(x,NULL);CHKERRQ(ierr); 3624906d46d4SStefano Zampini ierr = VecCopy(x,pcis->vec1_global);CHKERRQ(ierr); 3625e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,x,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 3626e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,x,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 3627bbb9e6c6SStefano Zampini ierr = MatMult(localChangeOfBasisMatrix,pcis->vec1_N,pcis->vec2_N);CHKERRQ(ierr); 3628e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec2_N,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 3629e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec2_N,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 3630906d46d4SStefano Zampini ierr = MatMult(pcbddc->ChangeOfBasisMatrix,pcis->vec1_global,x_change);CHKERRQ(ierr); 3631906d46d4SStefano Zampini ierr = VecAXPY(x,-1.0,x_change);CHKERRQ(ierr); 3632906d46d4SStefano Zampini ierr = VecNorm(x,NORM_INFINITY,&error);CHKERRQ(ierr); 3633906d46d4SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 3634bbb9e6c6SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Error global vs local change: %1.6e\n",error);CHKERRQ(ierr); 3635906d46d4SStefano Zampini ierr = VecDestroy(&x);CHKERRQ(ierr); 3636906d46d4SStefano Zampini ierr = VecDestroy(&x_change);CHKERRQ(ierr); 3637906d46d4SStefano Zampini } 3638b96c3477SStefano Zampini 3639b96c3477SStefano Zampini /* adapt sub_schurs computed (if any) */ 3640b96c3477SStefano Zampini if (pcbddc->use_deluxe_scaling) { 3641b96c3477SStefano Zampini PCBDDCSubSchurs sub_schurs=pcbddc->sub_schurs; 3642b96c3477SStefano Zampini if (sub_schurs->S_Ej_all) { 3643ac632422SStefano Zampini Mat S_new,tmat; 3644b087196eSStefano Zampini ISLocalToGlobalMapping NtoSall; 3645b087196eSStefano Zampini IS is_all_N,is_V,is_V_Sall; 3646b087196eSStefano Zampini const PetscScalar *array; 3647b087196eSStefano Zampini const PetscInt *idxs_V,*idxs_all; 3648b087196eSStefano Zampini PetscInt i,n_V; 3649bbb9e6c6SStefano Zampini 3650bbb9e6c6SStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(pcis->BtoNmap,sub_schurs->is_Ej_all,&is_all_N);CHKERRQ(ierr); 36516816873aSStefano Zampini ierr = MatGetSubMatrix(localChangeOfBasisMatrix,is_all_N,is_all_N,MAT_INITIAL_MATRIX,&tmat);CHKERRQ(ierr); 3652b087196eSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,pcbddc->n_vertices,pcbddc->local_primal_ref_node,PETSC_COPY_VALUES,&is_V);CHKERRQ(ierr); 3653b087196eSStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(is_all_N,&NtoSall);CHKERRQ(ierr); 3654b087196eSStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(NtoSall,IS_GTOLM_DROP,is_V,&is_V_Sall);CHKERRQ(ierr); 3655b087196eSStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&NtoSall);CHKERRQ(ierr); 3656bbb9e6c6SStefano Zampini ierr = ISDestroy(&is_all_N);CHKERRQ(ierr); 3657b087196eSStefano Zampini ierr = ISDestroy(&is_V);CHKERRQ(ierr); 3658ac632422SStefano Zampini ierr = MatPtAP(sub_schurs->S_Ej_all,tmat,MAT_INITIAL_MATRIX,1.0,&S_new);CHKERRQ(ierr); 3659b96c3477SStefano Zampini ierr = MatDestroy(&sub_schurs->S_Ej_all);CHKERRQ(ierr); 3660ac632422SStefano Zampini ierr = PetscObjectReference((PetscObject)S_new);CHKERRQ(ierr); 3661b087196eSStefano Zampini ierr = MatZeroRowsColumnsIS(S_new,is_V_Sall,1.,NULL,NULL);CHKERRQ(ierr); 3662b087196eSStefano Zampini ierr = ISGetLocalSize(is_V_Sall,&n_V);CHKERRQ(ierr); 3663b087196eSStefano Zampini ierr = ISGetIndices(is_V_Sall,&idxs_V);CHKERRQ(ierr); 3664b087196eSStefano Zampini ierr = ISGetIndices(sub_schurs->is_Ej_all,&idxs_all);CHKERRQ(ierr); 3665b087196eSStefano Zampini ierr = VecGetArrayRead(pcis->D,&array);CHKERRQ(ierr); 3666b087196eSStefano Zampini for (i=0;i<n_V;i++) { 3667b087196eSStefano Zampini PetscScalar val; 3668b087196eSStefano Zampini PetscInt idx; 3669b087196eSStefano Zampini 3670b087196eSStefano Zampini idx = idxs_V[i]; 3671b087196eSStefano Zampini val = array[idxs_all[idxs_V[i]]]; 3672b087196eSStefano Zampini ierr = MatSetValue(S_new,idx,idx,val,INSERT_VALUES);CHKERRQ(ierr); 3673b087196eSStefano Zampini } 3674b087196eSStefano Zampini ierr = MatAssemblyBegin(S_new,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3675b087196eSStefano Zampini ierr = MatAssemblyEnd(S_new,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3676ac632422SStefano Zampini sub_schurs->S_Ej_all = S_new; 3677ac632422SStefano Zampini ierr = MatDestroy(&S_new);CHKERRQ(ierr); 3678ac632422SStefano Zampini if (sub_schurs->sum_S_Ej_all) { 3679ac632422SStefano Zampini ierr = MatPtAP(sub_schurs->sum_S_Ej_all,tmat,MAT_INITIAL_MATRIX,1.0,&S_new);CHKERRQ(ierr); 3680b96c3477SStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_all);CHKERRQ(ierr); 3681ac632422SStefano Zampini ierr = PetscObjectReference((PetscObject)S_new);CHKERRQ(ierr); 3682b087196eSStefano Zampini ierr = MatZeroRowsColumnsIS(S_new,is_V_Sall,1.,NULL,NULL);CHKERRQ(ierr); 3683ac632422SStefano Zampini sub_schurs->sum_S_Ej_all = S_new; 3684ac632422SStefano Zampini ierr = MatDestroy(&S_new);CHKERRQ(ierr); 3685ac632422SStefano Zampini } 3686b087196eSStefano Zampini ierr = VecRestoreArrayRead(pcis->D,&array);CHKERRQ(ierr); 3687b087196eSStefano Zampini ierr = ISRestoreIndices(sub_schurs->is_Ej_all,&idxs_all);CHKERRQ(ierr); 3688b087196eSStefano Zampini ierr = ISRestoreIndices(is_V_Sall,&idxs_V);CHKERRQ(ierr); 3689b96c3477SStefano Zampini ierr = MatDestroy(&tmat);CHKERRQ(ierr); 3690b087196eSStefano Zampini ierr = ISDestroy(&is_V_Sall);CHKERRQ(ierr); 3691b96c3477SStefano Zampini } 3692b96c3477SStefano Zampini } 3693906d46d4SStefano Zampini ierr = MatDestroy(&localChangeOfBasisMatrix);CHKERRQ(ierr); 3694906d46d4SStefano Zampini } else if (pcbddc->user_ChangeOfBasisMatrix) { 3695b9b85e73SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->user_ChangeOfBasisMatrix);CHKERRQ(ierr); 3696b9b85e73SStefano Zampini pcbddc->ChangeOfBasisMatrix = pcbddc->user_ChangeOfBasisMatrix; 3697b9b85e73SStefano Zampini } 3698906d46d4SStefano Zampini 3699906d46d4SStefano Zampini /* set up change of basis context */ 3700906d46d4SStefano Zampini if (pcbddc->ChangeOfBasisMatrix) { 3701906d46d4SStefano Zampini PCBDDCChange_ctx change_ctx; 3702906d46d4SStefano Zampini 3703906d46d4SStefano Zampini if (!pcbddc->new_global_mat) { 3704906d46d4SStefano Zampini PetscInt global_size,local_size; 3705906d46d4SStefano Zampini 3706906d46d4SStefano Zampini ierr = VecGetSize(pcis->vec1_global,&global_size);CHKERRQ(ierr); 3707906d46d4SStefano Zampini ierr = VecGetLocalSize(pcis->vec1_global,&local_size);CHKERRQ(ierr); 3708906d46d4SStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)pc),&pcbddc->new_global_mat);CHKERRQ(ierr); 3709906d46d4SStefano Zampini ierr = MatSetSizes(pcbddc->new_global_mat,local_size,local_size,global_size,global_size);CHKERRQ(ierr); 3710906d46d4SStefano Zampini ierr = MatSetType(pcbddc->new_global_mat,MATSHELL);CHKERRQ(ierr); 3711906d46d4SStefano Zampini ierr = MatShellSetOperation(pcbddc->new_global_mat,MATOP_MULT,(void (*)(void))PCBDDCMatMult_Private);CHKERRQ(ierr); 3712906d46d4SStefano Zampini ierr = MatShellSetOperation(pcbddc->new_global_mat,MATOP_MULT_TRANSPOSE,(void (*)(void))PCBDDCMatMultTranspose_Private);CHKERRQ(ierr); 3713906d46d4SStefano Zampini ierr = PetscNew(&change_ctx);CHKERRQ(ierr); 3714906d46d4SStefano Zampini ierr = MatShellSetContext(pcbddc->new_global_mat,change_ctx);CHKERRQ(ierr); 3715906d46d4SStefano Zampini } else { 3716906d46d4SStefano Zampini ierr = MatShellGetContext(pcbddc->new_global_mat,&change_ctx);CHKERRQ(ierr); 3717906d46d4SStefano Zampini ierr = MatDestroy(&change_ctx->global_change);CHKERRQ(ierr); 3718906d46d4SStefano Zampini ierr = VecDestroyVecs(2,&change_ctx->work);CHKERRQ(ierr); 3719906d46d4SStefano Zampini } 3720906d46d4SStefano Zampini if (!pcbddc->user_ChangeOfBasisMatrix) { 3721906d46d4SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 3722906d46d4SStefano Zampini change_ctx->global_change = pcbddc->ChangeOfBasisMatrix; 3723906d46d4SStefano Zampini } else { 3724906d46d4SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->user_ChangeOfBasisMatrix);CHKERRQ(ierr); 3725906d46d4SStefano Zampini change_ctx->global_change = pcbddc->user_ChangeOfBasisMatrix; 3726906d46d4SStefano Zampini } 3727906d46d4SStefano Zampini ierr = VecDuplicateVecs(pcis->vec1_global,2,&change_ctx->work);CHKERRQ(ierr); 3728906d46d4SStefano Zampini ierr = MatSetUp(pcbddc->new_global_mat);CHKERRQ(ierr); 3729906d46d4SStefano Zampini ierr = MatAssemblyBegin(pcbddc->new_global_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3730906d46d4SStefano Zampini ierr = MatAssemblyEnd(pcbddc->new_global_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3731a8661815SStefano Zampini } else { 3732a8661815SStefano Zampini ierr = MatDestroy(&pcbddc->new_global_mat);CHKERRQ(ierr); 3733b9b85e73SStefano Zampini } 3734a717540cSStefano Zampini 37354f1b2e48SStefano Zampini /* add pressure dofs to set of primal nodes for numbering purposes */ 37364f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 37374f1b2e48SStefano Zampini pcbddc->local_primal_ref_node[pcbddc->local_primal_size_cc] = pcbddc->benign_p0_lidx[i]; 37384f1b2e48SStefano Zampini pcbddc->primal_indices_local_idxs[pcbddc->local_primal_size] = pcbddc->benign_p0_lidx[i]; 3739019a44ceSStefano Zampini pcbddc->local_primal_ref_mult[pcbddc->local_primal_size_cc] = 1; 3740019a44ceSStefano Zampini pcbddc->local_primal_size_cc++; 3741019a44ceSStefano Zampini pcbddc->local_primal_size++; 3742019a44ceSStefano Zampini } 3743019a44ceSStefano Zampini 3744019a44ceSStefano Zampini /* check if a new primal space has been introduced (also take into account benign trick) */ 3745727cdba6SStefano Zampini pcbddc->new_primal_space_local = PETSC_TRUE; 3746727cdba6SStefano Zampini if (olocal_primal_size == pcbddc->local_primal_size) { 37479f47a83aSStefano Zampini ierr = PetscMemcmp(pcbddc->local_primal_ref_node,olocal_primal_ref_node,olocal_primal_size_cc*sizeof(PetscInt),&pcbddc->new_primal_space_local);CHKERRQ(ierr); 3748c1c8e736SStefano Zampini pcbddc->new_primal_space_local = (PetscBool)(!pcbddc->new_primal_space_local); 37490e6343abSStefano Zampini if (!pcbddc->new_primal_space_local) { 37509f47a83aSStefano Zampini ierr = PetscMemcmp(pcbddc->local_primal_ref_mult,olocal_primal_ref_mult,olocal_primal_size_cc*sizeof(PetscInt),&pcbddc->new_primal_space_local);CHKERRQ(ierr); 3751727cdba6SStefano Zampini pcbddc->new_primal_space_local = (PetscBool)(!pcbddc->new_primal_space_local); 3752727cdba6SStefano Zampini } 37530e6343abSStefano Zampini } 37540e6343abSStefano Zampini ierr = PetscFree2(olocal_primal_ref_node,olocal_primal_ref_mult);CHKERRQ(ierr); 3755727cdba6SStefano Zampini /* new_primal_space will be used for numbering of coarse dofs, so it should be the same across all subdomains */ 3756727cdba6SStefano Zampini ierr = MPI_Allreduce(&pcbddc->new_primal_space_local,&pcbddc->new_primal_space,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 3757727cdba6SStefano Zampini 3758a717540cSStefano Zampini /* flush dbg viewer */ 3759b8ffe317SStefano Zampini if (pcbddc->dbg_flag) { 3760b8ffe317SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 3761b8ffe317SStefano Zampini } 3762a717540cSStefano Zampini 3763e310c8b4SStefano Zampini /* free workspace */ 3764a717540cSStefano Zampini ierr = PetscBTDestroy(&qr_needed_idx);CHKERRQ(ierr); 37654641a718SStefano Zampini ierr = PetscBTDestroy(&change_basis);CHKERRQ(ierr); 376608122e43SStefano Zampini if (!pcbddc->adaptive_selection) { 37679162d606SStefano Zampini ierr = PetscFree3(constraints_idxs_ptr,constraints_data_ptr,constraints_n);CHKERRQ(ierr); 37689162d606SStefano Zampini ierr = PetscFree3(constraints_data,constraints_idxs,constraints_idxs_B);CHKERRQ(ierr); 376908122e43SStefano Zampini } else { 37709162d606SStefano Zampini ierr = PetscFree5(pcbddc->adaptive_constraints_n, 37719162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr, 37729162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr, 377308122e43SStefano Zampini pcbddc->adaptive_constraints_idxs, 377408122e43SStefano Zampini pcbddc->adaptive_constraints_data);CHKERRQ(ierr); 37759162d606SStefano Zampini ierr = PetscFree(constraints_n);CHKERRQ(ierr); 37769162d606SStefano Zampini ierr = PetscFree(constraints_idxs_B);CHKERRQ(ierr); 377708122e43SStefano Zampini } 3778674ae819SStefano Zampini PetscFunctionReturn(0); 3779674ae819SStefano Zampini } 3780674ae819SStefano Zampini 3781674ae819SStefano Zampini #undef __FUNCT__ 3782674ae819SStefano Zampini #define __FUNCT__ "PCBDDCAnalyzeInterface" 3783674ae819SStefano Zampini PetscErrorCode PCBDDCAnalyzeInterface(PC pc) 3784674ae819SStefano Zampini { 3785674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 3786674ae819SStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 3787674ae819SStefano Zampini Mat_IS *matis = (Mat_IS*)pc->pmat->data; 37887fb0e2dbSStefano Zampini PetscInt ierr,i,vertex_size,N; 3789674ae819SStefano Zampini PetscViewer viewer=pcbddc->dbg_viewer; 3790674ae819SStefano Zampini 3791674ae819SStefano Zampini PetscFunctionBegin; 37928e61c736SStefano Zampini /* Reset previously computed graph */ 37938e61c736SStefano Zampini ierr = PCBDDCGraphReset(pcbddc->mat_graph);CHKERRQ(ierr); 3794674ae819SStefano Zampini /* Init local Graph struct */ 37957fb0e2dbSStefano Zampini ierr = MatGetSize(pc->pmat,&N,NULL);CHKERRQ(ierr); 37963bbff08aSStefano Zampini ierr = PCBDDCGraphInit(pcbddc->mat_graph,pcis->mapping,N);CHKERRQ(ierr); 3797674ae819SStefano Zampini 3798575ad6abSStefano Zampini /* Check validity of the csr graph passed in by the user */ 37995099eff2SStefano Zampini if (pcbddc->mat_graph->nvtxs_csr && pcbddc->mat_graph->nvtxs_csr != pcbddc->mat_graph->nvtxs) { 38005099eff2SStefano 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); 3801575ad6abSStefano Zampini } 38029577ea80SStefano Zampini 3803674ae819SStefano Zampini /* Set default CSR adjacency of local dofs if not provided by the user with PCBDDCSetLocalAdjacencyGraph */ 38044d379d7bSStefano Zampini if (!pcbddc->mat_graph->xadj || !pcbddc->mat_graph->adjncy) { 38054d379d7bSStefano Zampini PetscInt *xadj,*adjncy; 38064d379d7bSStefano Zampini PetscInt nvtxs; 3807e496cd5dSStefano Zampini PetscBool flg_row=PETSC_FALSE; 3808674ae819SStefano Zampini 38094d379d7bSStefano Zampini if (pcbddc->use_local_adj) { 38102fffb893SStefano Zampini 38112fffb893SStefano Zampini ierr = MatGetRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 38122fffb893SStefano Zampini if (flg_row) { 38134d379d7bSStefano Zampini ierr = PCBDDCSetLocalAdjacencyGraph(pc,nvtxs,xadj,adjncy,PETSC_COPY_VALUES);CHKERRQ(ierr); 3814b96c3477SStefano Zampini pcbddc->computed_rowadj = PETSC_TRUE; 38152fffb893SStefano Zampini } 38162fffb893SStefano Zampini ierr = MatRestoreRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 38179b28b941SStefano 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 */ 38184d379d7bSStefano Zampini IS is_dummy; 38194d379d7bSStefano Zampini ISLocalToGlobalMapping l2gmap_dummy; 38204d379d7bSStefano Zampini PetscInt j,sum; 38214d379d7bSStefano Zampini PetscInt *cxadj,*cadjncy; 38224d379d7bSStefano Zampini const PetscInt *idxs; 38234d379d7bSStefano Zampini PCBDDCGraph graph; 38244d379d7bSStefano Zampini PetscBT is_on_boundary; 38254d379d7bSStefano Zampini 38264d379d7bSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcis->n,0,1,&is_dummy);CHKERRQ(ierr); 38274d379d7bSStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(is_dummy,&l2gmap_dummy);CHKERRQ(ierr); 38284d379d7bSStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 38294d379d7bSStefano Zampini ierr = PCBDDCGraphCreate(&graph);CHKERRQ(ierr); 38307fb0e2dbSStefano Zampini ierr = PCBDDCGraphInit(graph,l2gmap_dummy,pcis->n);CHKERRQ(ierr); 38314d379d7bSStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&l2gmap_dummy);CHKERRQ(ierr); 3832e496cd5dSStefano Zampini ierr = MatGetRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 3833e496cd5dSStefano Zampini if (flg_row) { 38344d379d7bSStefano Zampini graph->xadj = xadj; 38354d379d7bSStefano Zampini graph->adjncy = adjncy; 3836e496cd5dSStefano Zampini } 38374d379d7bSStefano Zampini ierr = PCBDDCGraphSetUp(graph,1,NULL,NULL,0,NULL,NULL);CHKERRQ(ierr); 38384d379d7bSStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(graph);CHKERRQ(ierr); 3839e496cd5dSStefano Zampini ierr = MatRestoreRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 38404d379d7bSStefano Zampini 38414d379d7bSStefano Zampini if (pcbddc->dbg_flag) { 38429b28b941SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"[%d] Found %d subdomains (local size %d)\n",PetscGlobalRank,graph->ncc,pcis->n);CHKERRQ(ierr); 38434d379d7bSStefano Zampini for (i=0;i<graph->ncc;i++) { 38444d379d7bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"[%d] %d cc size %d\n",PetscGlobalRank,i,graph->cptr[i+1]-graph->cptr[i]);CHKERRQ(ierr); 38454d379d7bSStefano Zampini } 38464d379d7bSStefano Zampini } 38474d379d7bSStefano Zampini 3848e496cd5dSStefano Zampini ierr = PetscBTCreate(pcis->n,&is_on_boundary);CHKERRQ(ierr); 38494d379d7bSStefano Zampini ierr = ISGetIndices(pcis->is_B_local,&idxs);CHKERRQ(ierr); 38504d379d7bSStefano Zampini for (i=0;i<pcis->n_B;i++) { 38514d379d7bSStefano Zampini ierr = PetscBTSet(is_on_boundary,idxs[i]);CHKERRQ(ierr); 38524d379d7bSStefano Zampini } 38534d379d7bSStefano Zampini ierr = ISRestoreIndices(pcis->is_B_local,&idxs);CHKERRQ(ierr); 38544d379d7bSStefano Zampini 3855e496cd5dSStefano Zampini ierr = PetscCalloc1(pcis->n+1,&cxadj);CHKERRQ(ierr); 38564d379d7bSStefano Zampini sum = 0; 38574d379d7bSStefano Zampini for (i=0;i<graph->ncc;i++) { 38584d379d7bSStefano Zampini PetscInt sizecc = 0; 38594d379d7bSStefano Zampini for (j=graph->cptr[i];j<graph->cptr[i+1];j++) { 38604d379d7bSStefano Zampini if (PetscBTLookup(is_on_boundary,graph->queue[j])) { 38614d379d7bSStefano Zampini sizecc++; 38624d379d7bSStefano Zampini } 38634d379d7bSStefano Zampini } 38644d379d7bSStefano Zampini for (j=graph->cptr[i];j<graph->cptr[i+1];j++) { 38654d379d7bSStefano Zampini if (PetscBTLookup(is_on_boundary,graph->queue[j])) { 38664d379d7bSStefano Zampini cxadj[graph->queue[j]] = sizecc; 38674d379d7bSStefano Zampini } 38684d379d7bSStefano Zampini } 38694d379d7bSStefano Zampini sum += sizecc*sizecc; 38704d379d7bSStefano Zampini } 38714d379d7bSStefano Zampini ierr = PetscMalloc1(sum,&cadjncy);CHKERRQ(ierr); 38724d379d7bSStefano Zampini sum = 0; 3873e496cd5dSStefano Zampini for (i=0;i<pcis->n;i++) { 38744d379d7bSStefano Zampini PetscInt temp = cxadj[i]; 38754d379d7bSStefano Zampini cxadj[i] = sum; 38764d379d7bSStefano Zampini sum += temp; 38774d379d7bSStefano Zampini } 3878e496cd5dSStefano Zampini cxadj[pcis->n] = sum; 38794d379d7bSStefano Zampini for (i=0;i<graph->ncc;i++) { 38804d379d7bSStefano Zampini for (j=graph->cptr[i];j<graph->cptr[i+1];j++) { 38814d379d7bSStefano Zampini if (PetscBTLookup(is_on_boundary,graph->queue[j])) { 38824d379d7bSStefano Zampini PetscInt k,sizecc = 0; 38834d379d7bSStefano Zampini for (k=graph->cptr[i];k<graph->cptr[i+1];k++) { 38844d379d7bSStefano Zampini if (PetscBTLookup(is_on_boundary,graph->queue[k])) { 38854d379d7bSStefano Zampini cadjncy[cxadj[graph->queue[j]]+sizecc]=graph->queue[k]; 38864d379d7bSStefano Zampini sizecc++; 38874d379d7bSStefano Zampini } 38884d379d7bSStefano Zampini } 38894d379d7bSStefano Zampini } 38904d379d7bSStefano Zampini } 38914d379d7bSStefano Zampini } 38929b28b941SStefano Zampini if (sum) { 3893e496cd5dSStefano Zampini ierr = PCBDDCSetLocalAdjacencyGraph(pc,pcis->n,cxadj,cadjncy,PETSC_OWN_POINTER);CHKERRQ(ierr); 38944d379d7bSStefano Zampini } else { 38954d379d7bSStefano Zampini ierr = PetscFree(cxadj);CHKERRQ(ierr); 38964d379d7bSStefano Zampini ierr = PetscFree(cadjncy);CHKERRQ(ierr); 38974d379d7bSStefano Zampini } 38984d379d7bSStefano Zampini graph->xadj = 0; 38994d379d7bSStefano Zampini graph->adjncy = 0; 39004d379d7bSStefano Zampini ierr = PCBDDCGraphDestroy(&graph);CHKERRQ(ierr); 39014d379d7bSStefano Zampini ierr = PetscBTDestroy(&is_on_boundary);CHKERRQ(ierr); 39024d379d7bSStefano Zampini } 3903674ae819SStefano Zampini } 39049b28b941SStefano Zampini if (pcbddc->dbg_flag) { 39059b28b941SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 3906674ae819SStefano Zampini } 3907674ae819SStefano Zampini 390863602bcaSStefano Zampini /* Set default dofs' splitting if no information has been provided by the user with PCBDDCSetDofsSplitting or PCBDDCSetDofsSplittingLocal */ 3909674ae819SStefano Zampini vertex_size = 1; 391063602bcaSStefano Zampini if (pcbddc->user_provided_isfordofs) { 391163602bcaSStefano Zampini if (pcbddc->n_ISForDofs) { /* need to convert from global to local and remove references to global dofs splitting */ 391295ecbf38SStefano Zampini ierr = PetscMalloc1(pcbddc->n_ISForDofs,&pcbddc->ISForDofsLocal);CHKERRQ(ierr); 391363602bcaSStefano Zampini for (i=0;i<pcbddc->n_ISForDofs;i++) { 3914e176bc59SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->rctx,pcis->vec1_global,pcis->vec1_N,pcbddc->ISForDofs[i],&pcbddc->ISForDofsLocal[i]);CHKERRQ(ierr); 391563602bcaSStefano Zampini ierr = ISDestroy(&pcbddc->ISForDofs[i]);CHKERRQ(ierr); 3916674ae819SStefano Zampini } 391763602bcaSStefano Zampini pcbddc->n_ISForDofsLocal = pcbddc->n_ISForDofs; 391863602bcaSStefano Zampini pcbddc->n_ISForDofs = 0; 391963602bcaSStefano Zampini ierr = PetscFree(pcbddc->ISForDofs);CHKERRQ(ierr); 3920674ae819SStefano Zampini } 392163602bcaSStefano Zampini /* mat block size as vertex size (used for elasticity with rigid body modes as nearnullspace) */ 3922674ae819SStefano Zampini ierr = MatGetBlockSize(matis->A,&vertex_size);CHKERRQ(ierr); 392363602bcaSStefano Zampini } else { 392463602bcaSStefano Zampini if (!pcbddc->n_ISForDofsLocal) { /* field split not present, create it in local ordering */ 392563602bcaSStefano Zampini ierr = MatGetBlockSize(pc->pmat,&pcbddc->n_ISForDofsLocal);CHKERRQ(ierr); 3926854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->n_ISForDofsLocal,&pcbddc->ISForDofsLocal);CHKERRQ(ierr); 392763602bcaSStefano Zampini for (i=0;i<pcbddc->n_ISForDofsLocal;i++) { 392863602bcaSStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)pc),pcis->n/pcbddc->n_ISForDofsLocal,i,pcbddc->n_ISForDofsLocal,&pcbddc->ISForDofsLocal[i]);CHKERRQ(ierr); 392963602bcaSStefano Zampini } 393063602bcaSStefano Zampini } 3931674ae819SStefano Zampini } 3932674ae819SStefano Zampini 3933674ae819SStefano Zampini /* Setup of Graph */ 3934785d1243SStefano Zampini if (!pcbddc->DirichletBoundariesLocal && pcbddc->DirichletBoundaries) { /* need to convert from global to local */ 3935e176bc59SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->rctx,pcis->vec1_global,pcis->vec1_N,pcbddc->DirichletBoundaries,&pcbddc->DirichletBoundariesLocal);CHKERRQ(ierr); 3936785d1243SStefano Zampini } 3937785d1243SStefano Zampini if (!pcbddc->NeumannBoundariesLocal && pcbddc->NeumannBoundaries) { /* need to convert from global to local */ 3938e176bc59SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->rctx,pcis->vec1_global,pcis->vec1_N,pcbddc->NeumannBoundaries,&pcbddc->NeumannBoundariesLocal);CHKERRQ(ierr); 3939785d1243SStefano Zampini } 394030368db7SStefano Zampini if (!pcbddc->user_primal_vertices_local && pcbddc->user_primal_vertices) { /* need to convert from global to local */ 394130368db7SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->rctx,pcis->vec1_global,pcis->vec1_N,pcbddc->user_primal_vertices,&pcbddc->user_primal_vertices_local);CHKERRQ(ierr); 394230368db7SStefano Zampini } 394330368db7SStefano Zampini ierr = PCBDDCGraphSetUp(pcbddc->mat_graph,vertex_size,pcbddc->NeumannBoundariesLocal,pcbddc->DirichletBoundariesLocal,pcbddc->n_ISForDofsLocal,pcbddc->ISForDofsLocal,pcbddc->user_primal_vertices_local);CHKERRQ(ierr); 3944674ae819SStefano Zampini 39454f1b2e48SStefano Zampini /* attach info on disconnected subdomains if present */ 39464f1b2e48SStefano Zampini if (pcbddc->n_local_subs) { 39474f1b2e48SStefano Zampini PetscInt *local_subs; 39484f1b2e48SStefano Zampini 39494f1b2e48SStefano Zampini ierr = PetscMalloc1(N,&local_subs);CHKERRQ(ierr); 39504f1b2e48SStefano Zampini for (i=0;i<pcbddc->n_local_subs;i++) { 39514f1b2e48SStefano Zampini const PetscInt *idxs; 39524f1b2e48SStefano Zampini PetscInt nl,j; 39534f1b2e48SStefano Zampini 39544f1b2e48SStefano Zampini ierr = ISGetLocalSize(pcbddc->local_subs[i],&nl);CHKERRQ(ierr); 39554f1b2e48SStefano Zampini ierr = ISGetIndices(pcbddc->local_subs[i],&idxs);CHKERRQ(ierr); 39564f1b2e48SStefano Zampini for (j=0;j<nl;j++) { 39574f1b2e48SStefano Zampini local_subs[idxs[j]] = i; 39584f1b2e48SStefano Zampini } 39594f1b2e48SStefano Zampini ierr = ISRestoreIndices(pcbddc->local_subs[i],&idxs);CHKERRQ(ierr); 39604f1b2e48SStefano Zampini } 39614f1b2e48SStefano Zampini pcbddc->mat_graph->n_local_subs = pcbddc->n_local_subs; 39624f1b2e48SStefano Zampini pcbddc->mat_graph->local_subs = local_subs; 39634f1b2e48SStefano Zampini } 39644f1b2e48SStefano Zampini 3965674ae819SStefano Zampini /* Graph's connected components analysis */ 3966674ae819SStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(pcbddc->mat_graph);CHKERRQ(ierr); 3967674ae819SStefano Zampini 3968674ae819SStefano Zampini /* print some info to stdout */ 3969674ae819SStefano Zampini if (pcbddc->dbg_flag) { 3970302440fdSBarry Smith ierr = PCBDDCGraphASCIIView(pcbddc->mat_graph,pcbddc->dbg_flag,viewer);CHKERRQ(ierr); 3971674ae819SStefano Zampini } 3972fb180af4SStefano Zampini 3973fb180af4SStefano Zampini /* mark topography has done */ 3974fb180af4SStefano Zampini pcbddc->recompute_topography = PETSC_FALSE; 3975674ae819SStefano Zampini PetscFunctionReturn(0); 3976674ae819SStefano Zampini } 3977674ae819SStefano Zampini 3978dc456d91SStefano Zampini /* given an index sets possibly with holes, renumbers the indexes removing the holes */ 3979674ae819SStefano Zampini #undef __FUNCT__ 3980674ae819SStefano Zampini #define __FUNCT__ "PCBDDCSubsetNumbering" 3981dc456d91SStefano Zampini PetscErrorCode PCBDDCSubsetNumbering(IS subset, IS subset_mult, PetscInt *N_n, IS *subset_n) 3982674ae819SStefano Zampini { 3983dc456d91SStefano Zampini PetscSF sf; 3984dc456d91SStefano Zampini PetscLayout map; 3985dc456d91SStefano Zampini const PetscInt *idxs; 3986dc456d91SStefano Zampini PetscInt *leaf_data,*root_data,*gidxs; 3987dc456d91SStefano Zampini PetscInt N,n,i,lbounds[2],gbounds[2],Nl; 3988dc456d91SStefano Zampini PetscInt n_n,nlocals,start,first_index; 3989dc456d91SStefano Zampini PetscMPIInt commsize; 3990674ae819SStefano Zampini PetscBool first_found; 3991674ae819SStefano Zampini PetscErrorCode ierr; 3992674ae819SStefano Zampini 3993674ae819SStefano Zampini PetscFunctionBegin; 3994dc456d91SStefano Zampini ierr = ISGetLocalSize(subset,&n);CHKERRQ(ierr); 3995dc456d91SStefano Zampini if (subset_mult) { 3996dc456d91SStefano Zampini PetscCheckSameComm(subset,1,subset_mult,2); 3997dc456d91SStefano Zampini ierr = ISGetLocalSize(subset,&i);CHKERRQ(ierr); 3998dc456d91SStefano Zampini if (i != n) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Local subset and multiplicity sizes don't match! %d != %d",n,i); 3999674ae819SStefano Zampini } 4000dc456d91SStefano Zampini /* create workspace layout for computing global indices of subset */ 4001dc456d91SStefano Zampini ierr = ISGetIndices(subset,&idxs);CHKERRQ(ierr); 4002dc456d91SStefano Zampini lbounds[0] = lbounds[1] = 0; 4003dc456d91SStefano Zampini for (i=0;i<n;i++) { 4004dc456d91SStefano Zampini if (idxs[i] < lbounds[0]) lbounds[0] = idxs[i]; 4005dc456d91SStefano Zampini else if (idxs[i] > lbounds[1]) lbounds[1] = idxs[i]; 4006674ae819SStefano Zampini } 4007dc456d91SStefano Zampini lbounds[0] = -lbounds[0]; 4008dc456d91SStefano Zampini ierr = MPI_Allreduce(lbounds,gbounds,2,MPIU_INT,MPI_MAX,PetscObjectComm((PetscObject)subset));CHKERRQ(ierr); 4009dc456d91SStefano Zampini gbounds[0] = -gbounds[0]; 4010dc456d91SStefano Zampini N = gbounds[1] - gbounds[0] + 1; 4011dc456d91SStefano Zampini ierr = PetscLayoutCreate(PetscObjectComm((PetscObject)subset),&map);CHKERRQ(ierr); 4012dc456d91SStefano Zampini ierr = PetscLayoutSetBlockSize(map,1);CHKERRQ(ierr); 4013dc456d91SStefano Zampini ierr = PetscLayoutSetSize(map,N);CHKERRQ(ierr); 4014dc456d91SStefano Zampini ierr = PetscLayoutSetUp(map);CHKERRQ(ierr); 4015dc456d91SStefano Zampini ierr = PetscLayoutGetLocalSize(map,&Nl);CHKERRQ(ierr); 4016dc456d91SStefano Zampini 4017dc456d91SStefano Zampini /* create sf : leaf_data == multiplicity of indexes, root data == global index in layout */ 4018dc456d91SStefano Zampini ierr = PetscMalloc2(n,&leaf_data,Nl,&root_data);CHKERRQ(ierr); 4019dc456d91SStefano Zampini if (subset_mult) { 4020dc456d91SStefano Zampini const PetscInt* idxs_mult; 4021dc456d91SStefano Zampini 4022dc456d91SStefano Zampini ierr = ISGetIndices(subset_mult,&idxs_mult);CHKERRQ(ierr); 4023dc456d91SStefano Zampini ierr = PetscMemcpy(leaf_data,idxs_mult,n*sizeof(PetscInt));CHKERRQ(ierr); 4024dc456d91SStefano Zampini ierr = ISRestoreIndices(subset_mult,&idxs_mult);CHKERRQ(ierr); 4025674ae819SStefano Zampini } else { 4026dc456d91SStefano Zampini for (i=0;i<n;i++) leaf_data[i] = 1; 4027674ae819SStefano Zampini } 4028dc456d91SStefano Zampini /* local size of new subset */ 4029dc456d91SStefano Zampini n_n = 0; 4030dc456d91SStefano Zampini for (i=0;i<n;i++) n_n += leaf_data[i]; 4031dc456d91SStefano Zampini 4032dc456d91SStefano Zampini /* global indexes in layout */ 4033dc456d91SStefano Zampini ierr = PetscMalloc1(n_n,&gidxs);CHKERRQ(ierr); /* allocating possibly extra space in gidxs which will be used later */ 4034dc456d91SStefano Zampini for (i=0;i<n;i++) gidxs[i] = idxs[i] - gbounds[0]; 4035dc456d91SStefano Zampini ierr = ISRestoreIndices(subset,&idxs);CHKERRQ(ierr); 4036dc456d91SStefano Zampini ierr = PetscSFCreate(PetscObjectComm((PetscObject)subset),&sf);CHKERRQ(ierr); 4037dc456d91SStefano Zampini ierr = PetscSFSetGraphLayout(sf,map,n,NULL,PETSC_COPY_VALUES,gidxs);CHKERRQ(ierr); 4038dc456d91SStefano Zampini ierr = PetscLayoutDestroy(&map);CHKERRQ(ierr); 4039dc456d91SStefano Zampini 4040dc456d91SStefano Zampini /* reduce from leaves to roots */ 4041dc456d91SStefano Zampini ierr = PetscMemzero(root_data,Nl*sizeof(PetscInt));CHKERRQ(ierr); 404264a8e5bcSStefano Zampini ierr = PetscSFReduceBegin(sf,MPIU_INT,leaf_data,root_data,MPI_MAX);CHKERRQ(ierr); 404364a8e5bcSStefano Zampini ierr = PetscSFReduceEnd(sf,MPIU_INT,leaf_data,root_data,MPI_MAX);CHKERRQ(ierr); 4044dc456d91SStefano Zampini 4045dc456d91SStefano Zampini /* count indexes in local part of layout */ 4046674ae819SStefano Zampini nlocals = 0; 4047674ae819SStefano Zampini first_index = -1; 4048674ae819SStefano Zampini first_found = PETSC_FALSE; 4049dc456d91SStefano Zampini for (i=0;i<Nl;i++) { 4050dc456d91SStefano Zampini if (!first_found && root_data[i]) { 4051674ae819SStefano Zampini first_found = PETSC_TRUE; 4052674ae819SStefano Zampini first_index = i; 4053674ae819SStefano Zampini } 4054dc456d91SStefano Zampini nlocals += root_data[i]; 4055674ae819SStefano Zampini } 4056dc456d91SStefano Zampini 4057dc456d91SStefano Zampini /* cumulative of number of indexes and size of subset without holes */ 40585fa240b1SStefano Zampini #if defined(PETSC_HAVE_MPI_EXSCAN) 4059dc456d91SStefano Zampini start = 0; 406064a8e5bcSStefano Zampini ierr = MPI_Exscan(&nlocals,&start,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)subset));CHKERRQ(ierr); 40615fa240b1SStefano Zampini #else 406264a8e5bcSStefano Zampini ierr = MPI_Scan(&nlocals,&start,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)subset));CHKERRQ(ierr); 40635fa240b1SStefano Zampini start = start-nlocals; 40645fa240b1SStefano Zampini #endif 40655fa240b1SStefano Zampini 4066dc456d91SStefano Zampini if (N_n) { /* compute total size of new subset if requested */ 4067dc456d91SStefano Zampini *N_n = start + nlocals; 4068dc456d91SStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)subset),&commsize);CHKERRQ(ierr); 4069dc456d91SStefano Zampini ierr = MPI_Bcast(N_n,1,MPIU_INT,commsize-1,PetscObjectComm((PetscObject)subset));CHKERRQ(ierr); 4070674ae819SStefano Zampini } 40715fa240b1SStefano Zampini 40725fa240b1SStefano Zampini /* adapt root data with cumulative */ 4073674ae819SStefano Zampini if (first_found) { 4074dc456d91SStefano Zampini PetscInt old_index; 4075dc456d91SStefano Zampini 4076dc456d91SStefano Zampini root_data[first_index] += start; 4077674ae819SStefano Zampini old_index = first_index; 4078dc456d91SStefano Zampini for (i=first_index+1;i<Nl;i++) { 4079dc456d91SStefano Zampini if (root_data[i]) { 4080dc456d91SStefano Zampini root_data[i] += root_data[old_index]; 4081674ae819SStefano Zampini old_index = i; 4082674ae819SStefano Zampini } 4083674ae819SStefano Zampini } 4084674ae819SStefano Zampini } 4085dc456d91SStefano Zampini 4086dc456d91SStefano Zampini /* from roots to leaves */ 4087dc456d91SStefano Zampini ierr = PetscSFBcastBegin(sf,MPIU_INT,root_data,leaf_data);CHKERRQ(ierr); 4088dc456d91SStefano Zampini ierr = PetscSFBcastEnd(sf,MPIU_INT,root_data,leaf_data);CHKERRQ(ierr); 4089dc456d91SStefano Zampini ierr = PetscSFDestroy(&sf);CHKERRQ(ierr); 4090dc456d91SStefano Zampini 4091dc456d91SStefano Zampini /* create new IS with global indexes without holes */ 4092dc456d91SStefano Zampini if (subset_mult) { 4093dc456d91SStefano Zampini const PetscInt* idxs_mult; 4094dc456d91SStefano Zampini PetscInt cum; 4095dc456d91SStefano Zampini 4096dc456d91SStefano Zampini cum = 0; 4097dc456d91SStefano Zampini ierr = ISGetIndices(subset_mult,&idxs_mult);CHKERRQ(ierr); 4098dc456d91SStefano Zampini for (i=0;i<n;i++) { 4099dc456d91SStefano Zampini PetscInt j; 4100dc456d91SStefano Zampini for (j=0;j<idxs_mult[i];j++) gidxs[cum++] = leaf_data[i] - idxs_mult[i] + j; 4101674ae819SStefano Zampini } 4102dc456d91SStefano Zampini ierr = ISRestoreIndices(subset_mult,&idxs_mult);CHKERRQ(ierr); 4103674ae819SStefano Zampini } else { 4104dc456d91SStefano Zampini for (i=0;i<n;i++) { 4105dc456d91SStefano Zampini gidxs[i] = leaf_data[i]-1; 4106674ae819SStefano Zampini } 4107674ae819SStefano Zampini } 4108dc456d91SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)subset),n_n,gidxs,PETSC_OWN_POINTER,subset_n);CHKERRQ(ierr); 4109dc456d91SStefano Zampini ierr = PetscFree2(leaf_data,root_data);CHKERRQ(ierr); 4110674ae819SStefano Zampini PetscFunctionReturn(0); 4111674ae819SStefano Zampini } 41129a7d3425SStefano Zampini 41139a7d3425SStefano Zampini #undef __FUNCT__ 41149a7d3425SStefano Zampini #define __FUNCT__ "PCBDDCOrthonormalizeVecs" 41159a7d3425SStefano Zampini PetscErrorCode PCBDDCOrthonormalizeVecs(PetscInt n, Vec vecs[]) 41169a7d3425SStefano Zampini { 41179a7d3425SStefano Zampini PetscInt i,j; 41189a7d3425SStefano Zampini PetscScalar *alphas; 41199a7d3425SStefano Zampini PetscErrorCode ierr; 41209a7d3425SStefano Zampini 41219a7d3425SStefano Zampini PetscFunctionBegin; 41229a7d3425SStefano Zampini /* this implements stabilized Gram-Schmidt */ 4123785e854fSJed Brown ierr = PetscMalloc1(n,&alphas);CHKERRQ(ierr); 41249a7d3425SStefano Zampini for (i=0;i<n;i++) { 41259a7d3425SStefano Zampini ierr = VecNormalize(vecs[i],NULL);CHKERRQ(ierr); 41269a7d3425SStefano Zampini if (i<n) { ierr = VecMDot(vecs[i],n-i-1,&vecs[i+1],&alphas[i+1]);CHKERRQ(ierr); } 41279a7d3425SStefano Zampini for (j=i+1;j<n;j++) { ierr = VecAXPY(vecs[j],PetscConj(-alphas[j]),vecs[i]);CHKERRQ(ierr); } 41289a7d3425SStefano Zampini } 41299a7d3425SStefano Zampini ierr = PetscFree(alphas);CHKERRQ(ierr); 41309a7d3425SStefano Zampini PetscFunctionReturn(0); 41319a7d3425SStefano Zampini } 41329a7d3425SStefano Zampini 4133e7931f94SStefano Zampini #undef __FUNCT__ 413470cf5478SStefano Zampini #define __FUNCT__ "MatISGetSubassemblingPattern" 4135b0c7d250SStefano Zampini PetscErrorCode MatISGetSubassemblingPattern(Mat mat, PetscInt n_subdomains, PetscInt redprocs, IS* is_sends) 4136e7931f94SStefano Zampini { 413752e5ac9dSStefano Zampini IS ranks_send_to; 4138e7931f94SStefano Zampini PetscInt n_neighs,*neighs,*n_shared,**shared; 4139e7931f94SStefano Zampini PetscMPIInt size,rank,color; 414052e5ac9dSStefano Zampini PetscInt *xadj,*adjncy; 414152e5ac9dSStefano Zampini PetscInt *adjncy_wgt,*v_wgt,*ranks_send_to_idx; 41423837a79fSStefano Zampini PetscInt i,local_size,threshold=0; 41432b510759SStefano Zampini PetscBool use_vwgt=PETSC_FALSE,use_square=PETSC_FALSE; 4144e7931f94SStefano Zampini PetscSubcomm subcomm; 414552e5ac9dSStefano Zampini PetscErrorCode ierr; 4146a57a6d2fSStefano Zampini 4147e7931f94SStefano Zampini PetscFunctionBegin; 41482b510759SStefano Zampini ierr = PetscOptionsGetBool(NULL,"-matis_partitioning_use_square",&use_square,NULL);CHKERRQ(ierr); 41492b510759SStefano Zampini ierr = PetscOptionsGetBool(NULL,"-matis_partitioning_use_vwgt",&use_vwgt,NULL);CHKERRQ(ierr); 41502b510759SStefano Zampini ierr = PetscOptionsGetInt(NULL,"-matis_partitioning_threshold",&threshold,NULL);CHKERRQ(ierr); 4151e7931f94SStefano Zampini 4152e7931f94SStefano Zampini /* Get info on mapping */ 41533bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetSize(mat->rmap->mapping,&local_size);CHKERRQ(ierr); 41543bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetInfo(mat->rmap->mapping,&n_neighs,&neighs,&n_shared,&shared);CHKERRQ(ierr); 4155e7931f94SStefano Zampini 4156e7931f94SStefano Zampini /* build local CSR graph of subdomains' connectivity */ 4157785e854fSJed Brown ierr = PetscMalloc1(2,&xadj);CHKERRQ(ierr); 4158e7931f94SStefano Zampini xadj[0] = 0; 4159e7931f94SStefano Zampini xadj[1] = PetscMax(n_neighs-1,0); 4160785e854fSJed Brown ierr = PetscMalloc1(xadj[1],&adjncy);CHKERRQ(ierr); 4161785e854fSJed Brown ierr = PetscMalloc1(xadj[1],&adjncy_wgt);CHKERRQ(ierr); 4162e7931f94SStefano Zampini 41632b510759SStefano Zampini if (threshold) { 4164d023bfaeSStefano Zampini PetscInt xadj_count = 0; 41652b510759SStefano Zampini for (i=1;i<n_neighs;i++) { 4166d023bfaeSStefano Zampini if (n_shared[i] > threshold) { 4167d023bfaeSStefano Zampini adjncy[xadj_count] = neighs[i]; 4168d023bfaeSStefano Zampini adjncy_wgt[xadj_count] = n_shared[i]; 4169d023bfaeSStefano Zampini xadj_count++; 4170e7931f94SStefano Zampini } 4171e7931f94SStefano Zampini } 4172d023bfaeSStefano Zampini xadj[1] = xadj_count; 4173c8587f34SStefano Zampini } else { 4174e7931f94SStefano Zampini if (xadj[1]) { 4175e7931f94SStefano Zampini ierr = PetscMemcpy(adjncy,&neighs[1],xadj[1]*sizeof(*adjncy));CHKERRQ(ierr); 4176e7931f94SStefano Zampini ierr = PetscMemcpy(adjncy_wgt,&n_shared[1],xadj[1]*sizeof(*adjncy_wgt));CHKERRQ(ierr); 4177c8587f34SStefano Zampini } 4178e7931f94SStefano Zampini } 41793bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingRestoreInfo(mat->rmap->mapping,&n_neighs,&neighs,&n_shared,&shared);CHKERRQ(ierr); 4180e7931f94SStefano Zampini if (use_square) { 4181e7931f94SStefano Zampini for (i=0;i<xadj[1];i++) { 4182e7931f94SStefano Zampini adjncy_wgt[i] = adjncy_wgt[i]*adjncy_wgt[i]; 4183e7931f94SStefano Zampini } 4184e7931f94SStefano Zampini } 4185e7931f94SStefano Zampini ierr = PetscSortIntWithArray(xadj[1],adjncy,adjncy_wgt);CHKERRQ(ierr); 4186e7931f94SStefano Zampini 41873837a79fSStefano Zampini ierr = PetscMalloc1(1,&ranks_send_to_idx);CHKERRQ(ierr); 4188e7931f94SStefano Zampini 4189e7931f94SStefano Zampini /* 4190e7931f94SStefano Zampini Restrict work on active processes only. 4191e7931f94SStefano Zampini */ 4192e7931f94SStefano Zampini ierr = PetscSubcommCreate(PetscObjectComm((PetscObject)mat),&subcomm);CHKERRQ(ierr); 4193e7931f94SStefano Zampini ierr = PetscSubcommSetNumber(subcomm,2);CHKERRQ(ierr); /* 2 groups, active process and not active processes */ 4194e7931f94SStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)mat),&rank);CHKERRQ(ierr); 41952b510759SStefano Zampini ierr = PetscMPIIntCast(!local_size,&color);CHKERRQ(ierr); 4196d3531aaaSJed Brown ierr = PetscSubcommSetTypeGeneral(subcomm,color,rank);CHKERRQ(ierr); 4197e7931f94SStefano Zampini if (color) { 4198e7931f94SStefano Zampini ierr = PetscFree(xadj);CHKERRQ(ierr); 4199e7931f94SStefano Zampini ierr = PetscFree(adjncy);CHKERRQ(ierr); 4200e7931f94SStefano Zampini ierr = PetscFree(adjncy_wgt);CHKERRQ(ierr); 4201c8587f34SStefano Zampini } else { 420252e5ac9dSStefano Zampini Mat subdomain_adj; 420352e5ac9dSStefano Zampini IS new_ranks,new_ranks_contig; 420452e5ac9dSStefano Zampini MatPartitioning partitioner; 420552e5ac9dSStefano Zampini PetscInt prank,rstart=0,rend=0; 420652e5ac9dSStefano Zampini PetscInt *is_indices,*oldranks; 4207b0c7d250SStefano Zampini PetscBool aggregate; 4208b0c7d250SStefano Zampini 4209306c2d5bSBarry Smith ierr = MPI_Comm_size(PetscSubcommChild(subcomm),&size);CHKERRQ(ierr); 4210785e854fSJed Brown ierr = PetscMalloc1(size,&oldranks);CHKERRQ(ierr); 4211e7931f94SStefano Zampini prank = rank; 4212306c2d5bSBarry Smith ierr = MPI_Allgather(&prank,1,MPIU_INT,oldranks,1,MPIU_INT,PetscSubcommChild(subcomm));CHKERRQ(ierr); 42138002ef2cSStefano Zampini /* 4214e7931f94SStefano Zampini for (i=0;i<size;i++) { 4215e7931f94SStefano Zampini PetscPrintf(subcomm->comm,"oldranks[%d] = %d\n",i,oldranks[i]); 4216c8587f34SStefano Zampini } 42178002ef2cSStefano Zampini */ 4218e7931f94SStefano Zampini for (i=0;i<xadj[1];i++) { 4219e7931f94SStefano Zampini ierr = PetscFindInt(adjncy[i],size,oldranks,&adjncy[i]);CHKERRQ(ierr); 4220c8587f34SStefano Zampini } 4221e7931f94SStefano Zampini ierr = PetscSortIntWithArray(xadj[1],adjncy,adjncy_wgt);CHKERRQ(ierr); 4222b0c7d250SStefano Zampini aggregate = ((redprocs > 0 && redprocs < size) ? PETSC_TRUE : PETSC_FALSE); 4223b0c7d250SStefano Zampini if (aggregate) { 4224b0c7d250SStefano Zampini PetscInt lrows,row,ncols,*cols; 4225b0c7d250SStefano Zampini PetscMPIInt nrank; 4226b0c7d250SStefano Zampini PetscScalar *vals; 4227b0c7d250SStefano Zampini 4228b0c7d250SStefano Zampini ierr = MPI_Comm_rank(PetscSubcommChild(subcomm),&nrank);CHKERRQ(ierr); 4229b0c7d250SStefano Zampini lrows = 0; 4230b0c7d250SStefano Zampini if (nrank<redprocs) { 4231b0c7d250SStefano Zampini lrows = size/redprocs; 4232b0c7d250SStefano Zampini if (nrank<size%redprocs) lrows++; 4233b0c7d250SStefano Zampini } 42345fa240b1SStefano Zampini ierr = MatCreateAIJ(PetscSubcommChild(subcomm),lrows,lrows,size,size,50,NULL,50,NULL,&subdomain_adj);CHKERRQ(ierr); 4235b0c7d250SStefano Zampini ierr = MatGetOwnershipRange(subdomain_adj,&rstart,&rend);CHKERRQ(ierr); 4236b0c7d250SStefano Zampini ierr = MatSetOption(subdomain_adj,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_FALSE);CHKERRQ(ierr); 4237b0c7d250SStefano Zampini ierr = MatSetOption(subdomain_adj,MAT_NEW_NONZERO_ALLOCATION_ERR,PETSC_FALSE);CHKERRQ(ierr); 4238b0c7d250SStefano Zampini row = nrank; 4239b0c7d250SStefano Zampini ncols = xadj[1]-xadj[0]; 4240b0c7d250SStefano Zampini cols = adjncy; 4241b0c7d250SStefano Zampini ierr = PetscMalloc1(ncols,&vals);CHKERRQ(ierr); 4242b0c7d250SStefano Zampini for (i=0;i<ncols;i++) vals[i] = adjncy_wgt[i]; 4243b0c7d250SStefano Zampini ierr = MatSetValues(subdomain_adj,1,&row,ncols,cols,vals,INSERT_VALUES);CHKERRQ(ierr); 4244b0c7d250SStefano Zampini ierr = MatAssemblyBegin(subdomain_adj,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4245b0c7d250SStefano Zampini ierr = MatAssemblyEnd(subdomain_adj,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 424652e5ac9dSStefano Zampini ierr = PetscFree(xadj);CHKERRQ(ierr); 424752e5ac9dSStefano Zampini ierr = PetscFree(adjncy);CHKERRQ(ierr); 424852e5ac9dSStefano Zampini ierr = PetscFree(adjncy_wgt);CHKERRQ(ierr); 4249b0c7d250SStefano Zampini ierr = PetscFree(vals);CHKERRQ(ierr); 4250b0c7d250SStefano Zampini } else { 4251306c2d5bSBarry Smith ierr = MatCreateMPIAdj(PetscSubcommChild(subcomm),1,(PetscInt)size,xadj,adjncy,adjncy_wgt,&subdomain_adj);CHKERRQ(ierr); 4252b0c7d250SStefano Zampini } 425322b6e8a2SStefano Zampini /* ierr = MatView(subdomain_adj,0);CHKERRQ(ierr); */ 4254e7931f94SStefano Zampini 4255e7931f94SStefano Zampini /* Partition */ 4256306c2d5bSBarry Smith ierr = MatPartitioningCreate(PetscSubcommChild(subcomm),&partitioner);CHKERRQ(ierr); 4257e7931f94SStefano Zampini ierr = MatPartitioningSetAdjacency(partitioner,subdomain_adj);CHKERRQ(ierr); 4258e7931f94SStefano Zampini if (use_vwgt) { 42593837a79fSStefano Zampini ierr = PetscMalloc1(1,&v_wgt);CHKERRQ(ierr); 4260e7931f94SStefano Zampini v_wgt[0] = local_size; 4261e7931f94SStefano Zampini ierr = MatPartitioningSetVertexWeights(partitioner,v_wgt);CHKERRQ(ierr); 4262c8587f34SStefano Zampini } 426328143c3dSStefano Zampini n_subdomains = PetscMin((PetscInt)size,n_subdomains); 426428143c3dSStefano Zampini ierr = MatPartitioningSetNParts(partitioner,n_subdomains);CHKERRQ(ierr); 4265e7931f94SStefano Zampini ierr = MatPartitioningSetFromOptions(partitioner);CHKERRQ(ierr); 4266e7931f94SStefano Zampini ierr = MatPartitioningApply(partitioner,&new_ranks);CHKERRQ(ierr); 426722b6e8a2SStefano Zampini /* ierr = MatPartitioningView(partitioner,0);CHKERRQ(ierr); */ 4268e7931f94SStefano Zampini 426952e5ac9dSStefano Zampini /* renumber new_ranks to avoid "holes" in new set of processors */ 427052e5ac9dSStefano Zampini ierr = PCBDDCSubsetNumbering(new_ranks,NULL,NULL,&new_ranks_contig);CHKERRQ(ierr); 427152e5ac9dSStefano Zampini ierr = ISDestroy(&new_ranks);CHKERRQ(ierr); 427252e5ac9dSStefano Zampini ierr = ISGetIndices(new_ranks_contig,(const PetscInt**)&is_indices);CHKERRQ(ierr); 4273b0c7d250SStefano Zampini if (!redprocs) { 4274b0c7d250SStefano Zampini ranks_send_to_idx[0] = oldranks[is_indices[0]]; 427528143c3dSStefano Zampini } else { 4276b0c7d250SStefano Zampini PetscInt idxs[1]; 4277b0c7d250SStefano Zampini PetscMPIInt tag; 4278b0c7d250SStefano Zampini MPI_Request *reqs; 4279b0c7d250SStefano Zampini 4280b0c7d250SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)subdomain_adj,&tag);CHKERRQ(ierr); 4281b0c7d250SStefano Zampini ierr = PetscMalloc1(rend-rstart,&reqs);CHKERRQ(ierr); 4282b0c7d250SStefano Zampini for (i=rstart;i<rend;i++) { 4283b0c7d250SStefano Zampini ierr = MPI_Isend(is_indices+i-rstart,1,MPIU_INT,i,tag,PetscSubcommChild(subcomm),&reqs[i-rstart]);CHKERRQ(ierr); 428428143c3dSStefano Zampini } 4285b0c7d250SStefano Zampini ierr = MPI_Recv(idxs,1,MPIU_INT,MPI_ANY_SOURCE,tag,PetscSubcommChild(subcomm),MPI_STATUS_IGNORE);CHKERRQ(ierr); 4286b0c7d250SStefano Zampini ierr = MPI_Waitall(rend-rstart,reqs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 4287b0c7d250SStefano Zampini ierr = PetscFree(reqs);CHKERRQ(ierr); 4288b0c7d250SStefano Zampini ranks_send_to_idx[0] = oldranks[idxs[0]]; 4289e7931f94SStefano Zampini } 429052e5ac9dSStefano Zampini ierr = ISRestoreIndices(new_ranks_contig,(const PetscInt**)&is_indices);CHKERRQ(ierr); 4291e7931f94SStefano Zampini /* clean up */ 4292e7931f94SStefano Zampini ierr = PetscFree(oldranks);CHKERRQ(ierr); 429352e5ac9dSStefano Zampini ierr = ISDestroy(&new_ranks_contig);CHKERRQ(ierr); 4294e7931f94SStefano Zampini ierr = MatDestroy(&subdomain_adj);CHKERRQ(ierr); 4295e7931f94SStefano Zampini ierr = MatPartitioningDestroy(&partitioner);CHKERRQ(ierr); 4296e7931f94SStefano Zampini } 4297e7931f94SStefano Zampini ierr = PetscSubcommDestroy(&subcomm);CHKERRQ(ierr); 4298e7931f94SStefano Zampini 4299e7931f94SStefano Zampini /* assemble parallel IS for sends */ 4300e7931f94SStefano Zampini i = 1; 4301e7931f94SStefano Zampini if (color) i=0; 4302e7931f94SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)mat),i,ranks_send_to_idx,PETSC_OWN_POINTER,&ranks_send_to);CHKERRQ(ierr); 4303e7931f94SStefano Zampini /* get back IS */ 4304e7931f94SStefano Zampini *is_sends = ranks_send_to; 4305e7931f94SStefano Zampini PetscFunctionReturn(0); 4306e7931f94SStefano Zampini } 4307e7931f94SStefano Zampini 4308e7931f94SStefano Zampini typedef enum {MATDENSE_PRIVATE=0,MATAIJ_PRIVATE,MATBAIJ_PRIVATE,MATSBAIJ_PRIVATE}MatTypePrivate; 4309e7931f94SStefano Zampini 4310e7931f94SStefano Zampini #undef __FUNCT__ 4311e7931f94SStefano Zampini #define __FUNCT__ "MatISSubassemble" 431253a05cb3SStefano 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[]) 4313e7931f94SStefano Zampini { 431470cf5478SStefano Zampini Mat local_mat; 4315e7931f94SStefano Zampini IS is_sends_internal; 43169d30be91SStefano Zampini PetscInt rows,cols,new_local_rows; 431728143c3dSStefano Zampini PetscInt i,bs,buf_size_idxs,buf_size_idxs_is,buf_size_vals; 43189d30be91SStefano Zampini PetscBool ismatis,isdense,newisdense,destroy_mat; 4319e7931f94SStefano Zampini ISLocalToGlobalMapping l2gmap; 4320e7931f94SStefano Zampini PetscInt* l2gmap_indices; 4321e7931f94SStefano Zampini const PetscInt* is_indices; 4322e7931f94SStefano Zampini MatType new_local_type; 4323e7931f94SStefano Zampini /* buffers */ 4324e7931f94SStefano Zampini PetscInt *ptr_idxs,*send_buffer_idxs,*recv_buffer_idxs; 432528143c3dSStefano Zampini PetscInt *ptr_idxs_is,*send_buffer_idxs_is,*recv_buffer_idxs_is; 43269d30be91SStefano Zampini PetscInt *recv_buffer_idxs_local; 4327e7931f94SStefano Zampini PetscScalar *ptr_vals,*send_buffer_vals,*recv_buffer_vals; 4328e7931f94SStefano Zampini /* MPI */ 432928143c3dSStefano Zampini MPI_Comm comm,comm_n; 433028143c3dSStefano Zampini PetscSubcomm subcomm; 4331e7931f94SStefano Zampini PetscMPIInt n_sends,n_recvs,commsize; 433228143c3dSStefano Zampini PetscMPIInt *iflags,*ilengths_idxs,*ilengths_vals,*ilengths_idxs_is; 433328143c3dSStefano Zampini PetscMPIInt *onodes,*onodes_is,*olengths_idxs,*olengths_idxs_is,*olengths_vals; 433428143c3dSStefano Zampini PetscMPIInt len,tag_idxs,tag_idxs_is,tag_vals,source_dest; 433528143c3dSStefano Zampini MPI_Request *send_req_idxs,*send_req_idxs_is,*send_req_vals; 433628143c3dSStefano Zampini MPI_Request *recv_req_idxs,*recv_req_idxs_is,*recv_req_vals; 4337e7931f94SStefano Zampini PetscErrorCode ierr; 4338e7931f94SStefano Zampini 4339e7931f94SStefano Zampini PetscFunctionBegin; 434028143c3dSStefano Zampini /* TODO: add missing checks */ 434128143c3dSStefano Zampini PetscValidLogicalCollectiveInt(mat,n_subdomains,3); 434228143c3dSStefano Zampini PetscValidLogicalCollectiveBool(mat,restrict_comm,4); 434328143c3dSStefano Zampini PetscValidLogicalCollectiveEnum(mat,reuse,5); 434428143c3dSStefano Zampini PetscValidLogicalCollectiveInt(mat,nis,7); 4345e7931f94SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)mat,MATIS,&ismatis);CHKERRQ(ierr); 434628143c3dSStefano Zampini if (!ismatis) SETERRQ1(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot use %s on a matrix object which is not of type MATIS",__FUNCT__); 4347e7931f94SStefano Zampini ierr = MatISGetLocalMat(mat,&local_mat);CHKERRQ(ierr); 4348e7931f94SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)local_mat,MATSEQDENSE,&isdense);CHKERRQ(ierr); 4349e7931f94SStefano Zampini if (!isdense) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Currently cannot subassemble MATIS when local matrix type is not of type SEQDENSE"); 4350e7931f94SStefano Zampini ierr = MatGetSize(local_mat,&rows,&cols);CHKERRQ(ierr); 4351e7931f94SStefano Zampini if (rows != cols) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Local MATIS matrices should be square"); 435228143c3dSStefano Zampini if (reuse == MAT_REUSE_MATRIX && *mat_n) { 435370cf5478SStefano Zampini PetscInt mrows,mcols,mnrows,mncols; 435470cf5478SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)*mat_n,MATIS,&ismatis);CHKERRQ(ierr); 435528143c3dSStefano Zampini if (!ismatis) SETERRQ(PetscObjectComm((PetscObject)*mat_n),PETSC_ERR_SUP,"Cannot reuse a matrix which is not of type MATIS"); 435670cf5478SStefano Zampini ierr = MatGetSize(mat,&mrows,&mcols);CHKERRQ(ierr); 435770cf5478SStefano Zampini ierr = MatGetSize(*mat_n,&mnrows,&mncols);CHKERRQ(ierr); 435870cf5478SStefano Zampini if (mrows != mnrows) SETERRQ2(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot reuse matrix! Wrong number of rows %D != %D",mrows,mnrows); 435970cf5478SStefano Zampini if (mcols != mncols) SETERRQ2(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot reuse matrix! Wrong number of cols %D != %D",mcols,mncols); 436070cf5478SStefano Zampini } 4361e7931f94SStefano Zampini ierr = MatGetBlockSize(local_mat,&bs);CHKERRQ(ierr); 4362e7931f94SStefano Zampini PetscValidLogicalCollectiveInt(mat,bs,0); 4363e7931f94SStefano Zampini /* prepare IS for sending if not provided */ 4364e7931f94SStefano Zampini if (!is_sends) { 436528143c3dSStefano Zampini if (!n_subdomains) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"You should specify either an IS or a target number of subdomains"); 4366b0c7d250SStefano Zampini ierr = MatISGetSubassemblingPattern(mat,n_subdomains,0,&is_sends_internal);CHKERRQ(ierr); 4367c8587f34SStefano Zampini } else { 4368e7931f94SStefano Zampini ierr = PetscObjectReference((PetscObject)is_sends);CHKERRQ(ierr); 4369e7931f94SStefano Zampini is_sends_internal = is_sends; 4370c8587f34SStefano Zampini } 4371e7931f94SStefano Zampini 4372e7931f94SStefano Zampini /* get comm */ 4373a316fed8SStefano Zampini ierr = PetscObjectGetComm((PetscObject)mat,&comm);CHKERRQ(ierr); 4374e7931f94SStefano Zampini 4375e7931f94SStefano Zampini /* compute number of sends */ 4376e7931f94SStefano Zampini ierr = ISGetLocalSize(is_sends_internal,&i);CHKERRQ(ierr); 4377e7931f94SStefano Zampini ierr = PetscMPIIntCast(i,&n_sends);CHKERRQ(ierr); 4378e7931f94SStefano Zampini 4379e7931f94SStefano Zampini /* compute number of receives */ 4380e7931f94SStefano Zampini ierr = MPI_Comm_size(comm,&commsize);CHKERRQ(ierr); 4381785e854fSJed Brown ierr = PetscMalloc1(commsize,&iflags);CHKERRQ(ierr); 4382e7931f94SStefano Zampini ierr = PetscMemzero(iflags,commsize*sizeof(*iflags));CHKERRQ(ierr); 4383e7931f94SStefano Zampini ierr = ISGetIndices(is_sends_internal,&is_indices);CHKERRQ(ierr); 4384e7931f94SStefano Zampini for (i=0;i<n_sends;i++) iflags[is_indices[i]] = 1; 4385e7931f94SStefano Zampini ierr = PetscGatherNumberOfMessages(comm,iflags,NULL,&n_recvs);CHKERRQ(ierr); 4386e7931f94SStefano Zampini ierr = PetscFree(iflags);CHKERRQ(ierr); 4387e7931f94SStefano Zampini 438828143c3dSStefano Zampini /* restrict comm if requested */ 438928143c3dSStefano Zampini subcomm = 0; 439028143c3dSStefano Zampini destroy_mat = PETSC_FALSE; 439128143c3dSStefano Zampini if (restrict_comm) { 4392779c1cceSStefano Zampini PetscMPIInt color,subcommsize; 4393779c1cceSStefano Zampini 439428143c3dSStefano Zampini color = 0; 439553a05cb3SStefano Zampini if (restrict_full) { 439653a05cb3SStefano Zampini if (!n_recvs) color = 1; /* processes not receiving anything will not partecipate in new comm (full restriction) */ 439753a05cb3SStefano Zampini } else { 439853a05cb3SStefano Zampini if (!n_recvs && n_sends) color = 1; /* just those processes that are sending but not receiving anything will not partecipate in new comm */ 439953a05cb3SStefano Zampini } 440028143c3dSStefano Zampini ierr = MPI_Allreduce(&color,&subcommsize,1,MPI_INT,MPI_SUM,comm);CHKERRQ(ierr); 440128143c3dSStefano Zampini subcommsize = commsize - subcommsize; 440228143c3dSStefano Zampini /* check if reuse has been requested */ 440328143c3dSStefano Zampini if (reuse == MAT_REUSE_MATRIX) { 440428143c3dSStefano Zampini if (*mat_n) { 440528143c3dSStefano Zampini PetscMPIInt subcommsize2; 440628143c3dSStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)*mat_n),&subcommsize2);CHKERRQ(ierr); 440728143c3dSStefano Zampini if (subcommsize != subcommsize2) SETERRQ2(PetscObjectComm((PetscObject)*mat_n),PETSC_ERR_PLIB,"Cannot reuse matrix! wrong subcomm size %d != %d",subcommsize,subcommsize2); 440828143c3dSStefano Zampini comm_n = PetscObjectComm((PetscObject)*mat_n); 440928143c3dSStefano Zampini } else { 441028143c3dSStefano Zampini comm_n = PETSC_COMM_SELF; 441128143c3dSStefano Zampini } 441228143c3dSStefano Zampini } else { /* MAT_INITIAL_MATRIX */ 4413779c1cceSStefano Zampini PetscMPIInt rank; 4414779c1cceSStefano Zampini 4415779c1cceSStefano Zampini ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 441628143c3dSStefano Zampini ierr = PetscSubcommCreate(comm,&subcomm);CHKERRQ(ierr); 441728143c3dSStefano Zampini ierr = PetscSubcommSetNumber(subcomm,2);CHKERRQ(ierr); 441828143c3dSStefano Zampini ierr = PetscSubcommSetTypeGeneral(subcomm,color,rank);CHKERRQ(ierr); 4419306c2d5bSBarry Smith comm_n = PetscSubcommChild(subcomm); 442028143c3dSStefano Zampini } 442128143c3dSStefano Zampini /* flag to destroy *mat_n if not significative */ 442228143c3dSStefano Zampini if (color) destroy_mat = PETSC_TRUE; 442328143c3dSStefano Zampini } else { 442428143c3dSStefano Zampini comm_n = comm; 442528143c3dSStefano Zampini } 442628143c3dSStefano Zampini 4427e7931f94SStefano Zampini /* prepare send/receive buffers */ 4428785e854fSJed Brown ierr = PetscMalloc1(commsize,&ilengths_idxs);CHKERRQ(ierr); 4429e7931f94SStefano Zampini ierr = PetscMemzero(ilengths_idxs,commsize*sizeof(*ilengths_idxs));CHKERRQ(ierr); 4430785e854fSJed Brown ierr = PetscMalloc1(commsize,&ilengths_vals);CHKERRQ(ierr); 4431e7931f94SStefano Zampini ierr = PetscMemzero(ilengths_vals,commsize*sizeof(*ilengths_vals));CHKERRQ(ierr); 443228143c3dSStefano Zampini if (nis) { 4433854ce69bSBarry Smith ierr = PetscCalloc1(commsize,&ilengths_idxs_is);CHKERRQ(ierr); 443428143c3dSStefano Zampini } 4435e7931f94SStefano Zampini 443628143c3dSStefano Zampini /* Get data from local matrices */ 4437e7931f94SStefano Zampini if (!isdense) { 4438a26c9d0eSStefano Zampini SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Subassembling of AIJ local matrices not yet implemented"); 4439e7931f94SStefano Zampini /* TODO: See below some guidelines on how to prepare the local buffers */ 4440e7931f94SStefano Zampini /* 4441e7931f94SStefano Zampini send_buffer_vals should contain the raw values of the local matrix 4442e7931f94SStefano Zampini send_buffer_idxs should contain: 4443e7931f94SStefano Zampini - MatType_PRIVATE type 4444e7931f94SStefano Zampini - PetscInt size_of_l2gmap 4445e7931f94SStefano Zampini - PetscInt global_row_indices[size_of_l2gmap] 4446e7931f94SStefano Zampini - PetscInt all_other_info_which_is_needed_to_compute_preallocation_and_set_values 4447e7931f94SStefano Zampini */ 4448e7931f94SStefano Zampini } else { 4449e7931f94SStefano Zampini ierr = MatDenseGetArray(local_mat,&send_buffer_vals);CHKERRQ(ierr); 44503bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetSize(mat->rmap->mapping,&i);CHKERRQ(ierr); 4451854ce69bSBarry Smith ierr = PetscMalloc1(i+2,&send_buffer_idxs);CHKERRQ(ierr); 4452e7931f94SStefano Zampini send_buffer_idxs[0] = (PetscInt)MATDENSE_PRIVATE; 4453e7931f94SStefano Zampini send_buffer_idxs[1] = i; 44543bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetIndices(mat->rmap->mapping,(const PetscInt**)&ptr_idxs);CHKERRQ(ierr); 4455e7931f94SStefano Zampini ierr = PetscMemcpy(&send_buffer_idxs[2],ptr_idxs,i*sizeof(PetscInt));CHKERRQ(ierr); 44563bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingRestoreIndices(mat->rmap->mapping,(const PetscInt**)&ptr_idxs);CHKERRQ(ierr); 4457e7931f94SStefano Zampini ierr = PetscMPIIntCast(i,&len);CHKERRQ(ierr); 4458e7931f94SStefano Zampini for (i=0;i<n_sends;i++) { 4459e7931f94SStefano Zampini ilengths_vals[is_indices[i]] = len*len; 4460e7931f94SStefano Zampini ilengths_idxs[is_indices[i]] = len+2; 4461c8587f34SStefano Zampini } 4462c8587f34SStefano Zampini } 4463e7931f94SStefano Zampini ierr = PetscGatherMessageLengths2(comm,n_sends,n_recvs,ilengths_idxs,ilengths_vals,&onodes,&olengths_idxs,&olengths_vals);CHKERRQ(ierr); 446428143c3dSStefano Zampini /* additional is (if any) */ 446528143c3dSStefano Zampini if (nis) { 446628143c3dSStefano Zampini PetscMPIInt psum; 446728143c3dSStefano Zampini PetscInt j; 446828143c3dSStefano Zampini for (j=0,psum=0;j<nis;j++) { 446928143c3dSStefano Zampini PetscInt plen; 447028143c3dSStefano Zampini ierr = ISGetLocalSize(isarray[j],&plen);CHKERRQ(ierr); 447128143c3dSStefano Zampini ierr = PetscMPIIntCast(plen,&len);CHKERRQ(ierr); 447228143c3dSStefano Zampini psum += len+1; /* indices + lenght */ 447328143c3dSStefano Zampini } 4474854ce69bSBarry Smith ierr = PetscMalloc1(psum,&send_buffer_idxs_is);CHKERRQ(ierr); 447528143c3dSStefano Zampini for (j=0,psum=0;j<nis;j++) { 447628143c3dSStefano Zampini PetscInt plen; 447728143c3dSStefano Zampini const PetscInt *is_array_idxs; 447828143c3dSStefano Zampini ierr = ISGetLocalSize(isarray[j],&plen);CHKERRQ(ierr); 447928143c3dSStefano Zampini send_buffer_idxs_is[psum] = plen; 448028143c3dSStefano Zampini ierr = ISGetIndices(isarray[j],&is_array_idxs);CHKERRQ(ierr); 448128143c3dSStefano Zampini ierr = PetscMemcpy(&send_buffer_idxs_is[psum+1],is_array_idxs,plen*sizeof(PetscInt));CHKERRQ(ierr); 448228143c3dSStefano Zampini ierr = ISRestoreIndices(isarray[j],&is_array_idxs);CHKERRQ(ierr); 448328143c3dSStefano Zampini psum += plen+1; /* indices + lenght */ 448428143c3dSStefano Zampini } 448528143c3dSStefano Zampini for (i=0;i<n_sends;i++) { 448628143c3dSStefano Zampini ilengths_idxs_is[is_indices[i]] = psum; 448728143c3dSStefano Zampini } 448828143c3dSStefano Zampini ierr = PetscGatherMessageLengths(comm,n_sends,n_recvs,ilengths_idxs_is,&onodes_is,&olengths_idxs_is);CHKERRQ(ierr); 448928143c3dSStefano Zampini } 449028143c3dSStefano Zampini 4491e7931f94SStefano Zampini buf_size_idxs = 0; 4492e7931f94SStefano Zampini buf_size_vals = 0; 449328143c3dSStefano Zampini buf_size_idxs_is = 0; 4494e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 4495e7931f94SStefano Zampini buf_size_idxs += (PetscInt)olengths_idxs[i]; 4496e7931f94SStefano Zampini buf_size_vals += (PetscInt)olengths_vals[i]; 449728143c3dSStefano Zampini if (nis) buf_size_idxs_is += (PetscInt)olengths_idxs_is[i]; 4498e7931f94SStefano Zampini } 4499785e854fSJed Brown ierr = PetscMalloc1(buf_size_idxs,&recv_buffer_idxs);CHKERRQ(ierr); 4500785e854fSJed Brown ierr = PetscMalloc1(buf_size_vals,&recv_buffer_vals);CHKERRQ(ierr); 450195ecbf38SStefano Zampini ierr = PetscMalloc1(buf_size_idxs_is,&recv_buffer_idxs_is);CHKERRQ(ierr); 4502e7931f94SStefano Zampini 4503e7931f94SStefano Zampini /* get new tags for clean communications */ 4504e7931f94SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_idxs);CHKERRQ(ierr); 4505e7931f94SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_vals);CHKERRQ(ierr); 450628143c3dSStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_idxs_is);CHKERRQ(ierr); 4507e7931f94SStefano Zampini 4508e7931f94SStefano Zampini /* allocate for requests */ 4509785e854fSJed Brown ierr = PetscMalloc1(n_sends,&send_req_idxs);CHKERRQ(ierr); 4510785e854fSJed Brown ierr = PetscMalloc1(n_sends,&send_req_vals);CHKERRQ(ierr); 451195ecbf38SStefano Zampini ierr = PetscMalloc1(n_sends,&send_req_idxs_is);CHKERRQ(ierr); 4512785e854fSJed Brown ierr = PetscMalloc1(n_recvs,&recv_req_idxs);CHKERRQ(ierr); 4513785e854fSJed Brown ierr = PetscMalloc1(n_recvs,&recv_req_vals);CHKERRQ(ierr); 451495ecbf38SStefano Zampini ierr = PetscMalloc1(n_recvs,&recv_req_idxs_is);CHKERRQ(ierr); 4515e7931f94SStefano Zampini 4516e7931f94SStefano Zampini /* communications */ 4517e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 4518e7931f94SStefano Zampini ptr_vals = recv_buffer_vals; 451928143c3dSStefano Zampini ptr_idxs_is = recv_buffer_idxs_is; 4520e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 4521e7931f94SStefano Zampini source_dest = onodes[i]; 4522e7931f94SStefano Zampini ierr = MPI_Irecv(ptr_idxs,olengths_idxs[i],MPIU_INT,source_dest,tag_idxs,comm,&recv_req_idxs[i]);CHKERRQ(ierr); 4523e7931f94SStefano Zampini ierr = MPI_Irecv(ptr_vals,olengths_vals[i],MPIU_SCALAR,source_dest,tag_vals,comm,&recv_req_vals[i]);CHKERRQ(ierr); 4524e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 4525e7931f94SStefano Zampini ptr_vals += olengths_vals[i]; 452628143c3dSStefano Zampini if (nis) { 452728143c3dSStefano 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); 452828143c3dSStefano Zampini ptr_idxs_is += olengths_idxs_is[i]; 452928143c3dSStefano Zampini } 4530e7931f94SStefano Zampini } 4531e7931f94SStefano Zampini for (i=0;i<n_sends;i++) { 4532e7931f94SStefano Zampini ierr = PetscMPIIntCast(is_indices[i],&source_dest);CHKERRQ(ierr); 4533e7931f94SStefano Zampini ierr = MPI_Isend(send_buffer_idxs,ilengths_idxs[source_dest],MPIU_INT,source_dest,tag_idxs,comm,&send_req_idxs[i]);CHKERRQ(ierr); 4534e7931f94SStefano Zampini ierr = MPI_Isend(send_buffer_vals,ilengths_vals[source_dest],MPIU_SCALAR,source_dest,tag_vals,comm,&send_req_vals[i]);CHKERRQ(ierr); 453528143c3dSStefano Zampini if (nis) { 453628143c3dSStefano 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); 453728143c3dSStefano Zampini } 4538e7931f94SStefano Zampini } 4539e7931f94SStefano Zampini ierr = ISRestoreIndices(is_sends_internal,&is_indices);CHKERRQ(ierr); 4540e7931f94SStefano Zampini ierr = ISDestroy(&is_sends_internal);CHKERRQ(ierr); 4541e7931f94SStefano Zampini 4542e7931f94SStefano Zampini /* assemble new l2g map */ 4543e7931f94SStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_idxs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 4544e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 45459d30be91SStefano Zampini new_local_rows = 0; 4546e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 45479d30be91SStefano Zampini new_local_rows += *(ptr_idxs+1); /* second element is the local size of the l2gmap */ 4548e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 4549e7931f94SStefano Zampini } 45509d30be91SStefano Zampini ierr = PetscMalloc1(new_local_rows,&l2gmap_indices);CHKERRQ(ierr); 4551e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 45529d30be91SStefano Zampini new_local_rows = 0; 4553e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 45549d30be91SStefano Zampini ierr = PetscMemcpy(&l2gmap_indices[new_local_rows],ptr_idxs+2,(*(ptr_idxs+1))*sizeof(PetscInt));CHKERRQ(ierr); 45559d30be91SStefano Zampini new_local_rows += *(ptr_idxs+1); /* second element is the local size of the l2gmap */ 4556e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 4557e7931f94SStefano Zampini } 45589d30be91SStefano Zampini ierr = PetscSortRemoveDupsInt(&new_local_rows,l2gmap_indices);CHKERRQ(ierr); 45599d30be91SStefano Zampini ierr = ISLocalToGlobalMappingCreate(comm_n,1,new_local_rows,l2gmap_indices,PETSC_COPY_VALUES,&l2gmap);CHKERRQ(ierr); 4560e7931f94SStefano Zampini ierr = PetscFree(l2gmap_indices);CHKERRQ(ierr); 4561e7931f94SStefano Zampini 4562e7931f94SStefano Zampini /* infer new local matrix type from received local matrices type */ 4563e7931f94SStefano Zampini /* currently if all local matrices are of type X, then the resulting matrix will be of type X, except for the dense case */ 4564e7931f94SStefano 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) */ 4565e7931f94SStefano Zampini if (n_recvs) { 456628143c3dSStefano Zampini MatTypePrivate new_local_type_private = (MatTypePrivate)send_buffer_idxs[0]; 4567e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 4568e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 4569e7931f94SStefano Zampini if ((PetscInt)new_local_type_private != *ptr_idxs) { 4570e7931f94SStefano Zampini new_local_type_private = MATAIJ_PRIVATE; 4571e7931f94SStefano Zampini break; 4572e7931f94SStefano Zampini } 4573e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 4574e7931f94SStefano Zampini } 4575e7931f94SStefano Zampini switch (new_local_type_private) { 457628143c3dSStefano Zampini case MATDENSE_PRIVATE: 457728143c3dSStefano Zampini if (n_recvs>1) { /* subassembling of dense matrices does not give a dense matrix! */ 4578e7931f94SStefano Zampini new_local_type = MATSEQAIJ; 4579e7931f94SStefano Zampini bs = 1; 458028143c3dSStefano Zampini } else { /* if I receive only 1 dense matrix */ 458128143c3dSStefano Zampini new_local_type = MATSEQDENSE; 458228143c3dSStefano Zampini bs = 1; 458328143c3dSStefano Zampini } 4584e7931f94SStefano Zampini break; 4585e7931f94SStefano Zampini case MATAIJ_PRIVATE: 4586e7931f94SStefano Zampini new_local_type = MATSEQAIJ; 4587e7931f94SStefano Zampini bs = 1; 4588e7931f94SStefano Zampini break; 4589e7931f94SStefano Zampini case MATBAIJ_PRIVATE: 4590e7931f94SStefano Zampini new_local_type = MATSEQBAIJ; 4591e7931f94SStefano Zampini break; 4592e7931f94SStefano Zampini case MATSBAIJ_PRIVATE: 4593e7931f94SStefano Zampini new_local_type = MATSEQSBAIJ; 4594e7931f94SStefano Zampini break; 4595e7931f94SStefano Zampini default: 45969d30be91SStefano Zampini SETERRQ2(comm,PETSC_ERR_SUP,"Unsupported private type %d in %s",new_local_type_private,__FUNCT__); 4597e7931f94SStefano Zampini break; 4598e7931f94SStefano Zampini } 459928143c3dSStefano Zampini } else { /* by default, new_local_type is seqdense */ 460028143c3dSStefano Zampini new_local_type = MATSEQDENSE; 460128143c3dSStefano Zampini bs = 1; 4602e7931f94SStefano Zampini } 4603e7931f94SStefano Zampini 460470cf5478SStefano Zampini /* create MATIS object if needed */ 460570cf5478SStefano Zampini if (reuse == MAT_INITIAL_MATRIX) { 4606e7931f94SStefano Zampini ierr = MatGetSize(mat,&rows,&cols);CHKERRQ(ierr); 4607e176bc59SStefano Zampini ierr = MatCreateIS(comm_n,bs,PETSC_DECIDE,PETSC_DECIDE,rows,cols,l2gmap,NULL,mat_n);CHKERRQ(ierr); 460870cf5478SStefano Zampini } else { 460970cf5478SStefano Zampini /* it also destroys the local matrices */ 461070cf5478SStefano Zampini ierr = MatSetLocalToGlobalMapping(*mat_n,l2gmap,l2gmap);CHKERRQ(ierr); 461170cf5478SStefano Zampini } 461270cf5478SStefano Zampini ierr = MatISGetLocalMat(*mat_n,&local_mat);CHKERRQ(ierr); 4613e7931f94SStefano Zampini ierr = MatSetType(local_mat,new_local_type);CHKERRQ(ierr); 46149d30be91SStefano Zampini 46159d30be91SStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_vals,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 46169d30be91SStefano Zampini 46179d30be91SStefano Zampini /* Global to local map of received indices */ 46189d30be91SStefano Zampini ierr = PetscMalloc1(buf_size_idxs,&recv_buffer_idxs_local);CHKERRQ(ierr); /* needed for values insertion */ 46199d30be91SStefano Zampini ierr = ISGlobalToLocalMappingApply(l2gmap,IS_GTOLM_MASK,buf_size_idxs,recv_buffer_idxs,&i,recv_buffer_idxs_local);CHKERRQ(ierr); 46209d30be91SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&l2gmap);CHKERRQ(ierr); 46219d30be91SStefano Zampini 46229d30be91SStefano Zampini /* restore attributes -> type of incoming data and its size */ 46239d30be91SStefano Zampini buf_size_idxs = 0; 46249d30be91SStefano Zampini for (i=0;i<n_recvs;i++) { 46259d30be91SStefano Zampini recv_buffer_idxs_local[buf_size_idxs] = recv_buffer_idxs[buf_size_idxs]; 46269d30be91SStefano Zampini recv_buffer_idxs_local[buf_size_idxs+1] = recv_buffer_idxs[buf_size_idxs+1]; 46279d30be91SStefano Zampini buf_size_idxs += (PetscInt)olengths_idxs[i]; 46289d30be91SStefano Zampini } 46299d30be91SStefano Zampini ierr = PetscFree(recv_buffer_idxs);CHKERRQ(ierr); 46309d30be91SStefano Zampini 46319d30be91SStefano Zampini /* set preallocation */ 46329d30be91SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)local_mat,MATSEQDENSE,&newisdense);CHKERRQ(ierr); 46339d30be91SStefano Zampini if (!newisdense) { 46349d30be91SStefano Zampini PetscInt *new_local_nnz=0; 46359d30be91SStefano Zampini 46369d30be91SStefano Zampini ptr_vals = recv_buffer_vals; 46379d30be91SStefano Zampini ptr_idxs = recv_buffer_idxs_local; 46389d30be91SStefano Zampini if (n_recvs) { 46399d30be91SStefano Zampini ierr = PetscCalloc1(new_local_rows,&new_local_nnz);CHKERRQ(ierr); 46409d30be91SStefano Zampini } 46419d30be91SStefano Zampini for (i=0;i<n_recvs;i++) { 46429d30be91SStefano Zampini PetscInt j; 46439d30be91SStefano Zampini if (*ptr_idxs == (PetscInt)MATDENSE_PRIVATE) { /* preallocation provided for dense case only */ 46449d30be91SStefano Zampini for (j=0;j<*(ptr_idxs+1);j++) { 46459d30be91SStefano Zampini new_local_nnz[*(ptr_idxs+2+j)] += *(ptr_idxs+1); 46469d30be91SStefano Zampini } 46479d30be91SStefano Zampini } else { 46489d30be91SStefano Zampini /* TODO */ 46499d30be91SStefano Zampini } 46509d30be91SStefano Zampini ptr_idxs += olengths_idxs[i]; 46519d30be91SStefano Zampini } 46529d30be91SStefano Zampini if (new_local_nnz) { 46539d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] = PetscMin(new_local_nnz[i],new_local_rows); 46549d30be91SStefano Zampini ierr = MatSeqAIJSetPreallocation(local_mat,0,new_local_nnz);CHKERRQ(ierr); 46559d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] /= bs; 46569d30be91SStefano Zampini ierr = MatSeqBAIJSetPreallocation(local_mat,bs,0,new_local_nnz);CHKERRQ(ierr); 46579d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] = PetscMax(new_local_nnz[i]-i,0); 46589d30be91SStefano Zampini ierr = MatSeqSBAIJSetPreallocation(local_mat,bs,0,new_local_nnz);CHKERRQ(ierr); 46599d30be91SStefano Zampini } else { 46609d30be91SStefano Zampini ierr = MatSetUp(local_mat);CHKERRQ(ierr); 46619d30be91SStefano Zampini } 46629d30be91SStefano Zampini ierr = PetscFree(new_local_nnz);CHKERRQ(ierr); 46639d30be91SStefano Zampini } else { 46649d30be91SStefano Zampini ierr = MatSetUp(local_mat);CHKERRQ(ierr); 46659d30be91SStefano Zampini } 4666e7931f94SStefano Zampini 4667e7931f94SStefano Zampini /* set values */ 4668e7931f94SStefano Zampini ptr_vals = recv_buffer_vals; 46699d30be91SStefano Zampini ptr_idxs = recv_buffer_idxs_local; 4670e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 4671e7931f94SStefano Zampini if (*ptr_idxs == (PetscInt)MATDENSE_PRIVATE) { /* values insertion provided for dense case only */ 4672e7931f94SStefano Zampini ierr = MatSetOption(local_mat,MAT_ROW_ORIENTED,PETSC_FALSE);CHKERRQ(ierr); 46739d30be91SStefano Zampini ierr = MatSetValues(local_mat,*(ptr_idxs+1),ptr_idxs+2,*(ptr_idxs+1),ptr_idxs+2,ptr_vals,ADD_VALUES);CHKERRQ(ierr); 4674e7931f94SStefano Zampini ierr = MatAssemblyBegin(local_mat,MAT_FLUSH_ASSEMBLY);CHKERRQ(ierr); 4675e7931f94SStefano Zampini ierr = MatAssemblyEnd(local_mat,MAT_FLUSH_ASSEMBLY);CHKERRQ(ierr); 4676e7931f94SStefano Zampini ierr = MatSetOption(local_mat,MAT_ROW_ORIENTED,PETSC_TRUE);CHKERRQ(ierr); 467728143c3dSStefano Zampini } else { 467828143c3dSStefano Zampini /* TODO */ 4679e7931f94SStefano Zampini } 4680e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 4681e7931f94SStefano Zampini ptr_vals += olengths_vals[i]; 4682e7931f94SStefano Zampini } 4683e7931f94SStefano Zampini ierr = MatAssemblyBegin(local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4684e7931f94SStefano Zampini ierr = MatAssemblyEnd(local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 468570cf5478SStefano Zampini ierr = MatAssemblyBegin(*mat_n,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 468670cf5478SStefano Zampini ierr = MatAssemblyEnd(*mat_n,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 46879d30be91SStefano Zampini ierr = PetscFree(recv_buffer_vals);CHKERRQ(ierr); 46889d30be91SStefano Zampini ierr = PetscFree(recv_buffer_idxs_local);CHKERRQ(ierr); 4689e7931f94SStefano Zampini 4690dfd14d43SStefano Zampini #if 0 469128143c3dSStefano Zampini if (!restrict_comm) { /* check */ 4692e7931f94SStefano Zampini Vec lvec,rvec; 4693e7931f94SStefano Zampini PetscReal infty_error; 4694e7931f94SStefano Zampini 46952a7a6963SBarry Smith ierr = MatCreateVecs(mat,&rvec,&lvec);CHKERRQ(ierr); 4696e7931f94SStefano Zampini ierr = VecSetRandom(rvec,NULL);CHKERRQ(ierr); 4697e7931f94SStefano Zampini ierr = MatMult(mat,rvec,lvec);CHKERRQ(ierr); 4698e7931f94SStefano Zampini ierr = VecScale(lvec,-1.0);CHKERRQ(ierr); 469970cf5478SStefano Zampini ierr = MatMultAdd(*mat_n,rvec,lvec,lvec);CHKERRQ(ierr); 4700e7931f94SStefano Zampini ierr = VecNorm(lvec,NORM_INFINITY,&infty_error);CHKERRQ(ierr); 4701e7931f94SStefano Zampini ierr = PetscPrintf(PetscObjectComm((PetscObject)mat),"Infinity error subassembling %1.6e\n",infty_error); 4702e7931f94SStefano Zampini ierr = VecDestroy(&rvec);CHKERRQ(ierr); 4703e7931f94SStefano Zampini ierr = VecDestroy(&lvec);CHKERRQ(ierr); 4704e7931f94SStefano Zampini } 470528143c3dSStefano Zampini #endif 4706e7931f94SStefano Zampini 470728143c3dSStefano Zampini /* assemble new additional is (if any) */ 470828143c3dSStefano Zampini if (nis) { 470928143c3dSStefano Zampini PetscInt **temp_idxs,*count_is,j,psum; 471028143c3dSStefano Zampini 471128143c3dSStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_idxs_is,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 4712854ce69bSBarry Smith ierr = PetscCalloc1(nis,&count_is);CHKERRQ(ierr); 471328143c3dSStefano Zampini ptr_idxs = recv_buffer_idxs_is; 471428143c3dSStefano Zampini psum = 0; 471528143c3dSStefano Zampini for (i=0;i<n_recvs;i++) { 471628143c3dSStefano Zampini for (j=0;j<nis;j++) { 471728143c3dSStefano Zampini PetscInt plen = *(ptr_idxs); /* first element is the local size of IS's indices */ 471828143c3dSStefano Zampini count_is[j] += plen; /* increment counting of buffer for j-th IS */ 471928143c3dSStefano Zampini psum += plen; 472028143c3dSStefano Zampini ptr_idxs += plen+1; /* shift pointer to received data */ 472128143c3dSStefano Zampini } 472228143c3dSStefano Zampini } 4723854ce69bSBarry Smith ierr = PetscMalloc1(nis,&temp_idxs);CHKERRQ(ierr); 4724854ce69bSBarry Smith ierr = PetscMalloc1(psum,&temp_idxs[0]);CHKERRQ(ierr); 472528143c3dSStefano Zampini for (i=1;i<nis;i++) { 472628143c3dSStefano Zampini temp_idxs[i] = temp_idxs[i-1]+count_is[i-1]; 472728143c3dSStefano Zampini } 472828143c3dSStefano Zampini ierr = PetscMemzero(count_is,nis*sizeof(PetscInt));CHKERRQ(ierr); 472928143c3dSStefano Zampini ptr_idxs = recv_buffer_idxs_is; 473028143c3dSStefano Zampini for (i=0;i<n_recvs;i++) { 473128143c3dSStefano Zampini for (j=0;j<nis;j++) { 473228143c3dSStefano Zampini PetscInt plen = *(ptr_idxs); /* first element is the local size of IS's indices */ 473328143c3dSStefano Zampini ierr = PetscMemcpy(&temp_idxs[j][count_is[j]],ptr_idxs+1,plen*sizeof(PetscInt));CHKERRQ(ierr); 473428143c3dSStefano Zampini count_is[j] += plen; /* increment starting point of buffer for j-th IS */ 473528143c3dSStefano Zampini ptr_idxs += plen+1; /* shift pointer to received data */ 473628143c3dSStefano Zampini } 473728143c3dSStefano Zampini } 473828143c3dSStefano Zampini for (i=0;i<nis;i++) { 473928143c3dSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 474028143c3dSStefano Zampini ierr = PetscSortRemoveDupsInt(&count_is[i],temp_idxs[i]);CHKERRQ(ierr);CHKERRQ(ierr); 474128143c3dSStefano Zampini ierr = ISCreateGeneral(comm_n,count_is[i],temp_idxs[i],PETSC_COPY_VALUES,&isarray[i]);CHKERRQ(ierr); 474228143c3dSStefano Zampini } 474328143c3dSStefano Zampini ierr = PetscFree(count_is);CHKERRQ(ierr); 474428143c3dSStefano Zampini ierr = PetscFree(temp_idxs[0]);CHKERRQ(ierr); 474528143c3dSStefano Zampini ierr = PetscFree(temp_idxs);CHKERRQ(ierr); 474628143c3dSStefano Zampini } 4747e7931f94SStefano Zampini /* free workspace */ 474828143c3dSStefano Zampini ierr = PetscFree(recv_buffer_idxs_is);CHKERRQ(ierr); 4749e7931f94SStefano Zampini ierr = MPI_Waitall(n_sends,send_req_idxs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 4750e7931f94SStefano Zampini ierr = PetscFree(send_buffer_idxs);CHKERRQ(ierr); 4751e7931f94SStefano Zampini ierr = MPI_Waitall(n_sends,send_req_vals,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 4752e7931f94SStefano Zampini if (isdense) { 4753e7931f94SStefano Zampini ierr = MatISGetLocalMat(mat,&local_mat);CHKERRQ(ierr); 4754e7931f94SStefano Zampini ierr = MatDenseRestoreArray(local_mat,&send_buffer_vals);CHKERRQ(ierr); 4755e7931f94SStefano Zampini } else { 4756e7931f94SStefano Zampini /* ierr = PetscFree(send_buffer_vals);CHKERRQ(ierr); */ 4757e7931f94SStefano Zampini } 475828143c3dSStefano Zampini if (nis) { 475928143c3dSStefano Zampini ierr = MPI_Waitall(n_sends,send_req_idxs_is,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 476028143c3dSStefano Zampini ierr = PetscFree(send_buffer_idxs_is);CHKERRQ(ierr); 476128143c3dSStefano Zampini } 4762e7931f94SStefano Zampini ierr = PetscFree(recv_req_idxs);CHKERRQ(ierr); 4763e7931f94SStefano Zampini ierr = PetscFree(recv_req_vals);CHKERRQ(ierr); 476428143c3dSStefano Zampini ierr = PetscFree(recv_req_idxs_is);CHKERRQ(ierr); 4765e7931f94SStefano Zampini ierr = PetscFree(send_req_idxs);CHKERRQ(ierr); 4766e7931f94SStefano Zampini ierr = PetscFree(send_req_vals);CHKERRQ(ierr); 476728143c3dSStefano Zampini ierr = PetscFree(send_req_idxs_is);CHKERRQ(ierr); 4768e7931f94SStefano Zampini ierr = PetscFree(ilengths_vals);CHKERRQ(ierr); 4769e7931f94SStefano Zampini ierr = PetscFree(ilengths_idxs);CHKERRQ(ierr); 4770e7931f94SStefano Zampini ierr = PetscFree(olengths_vals);CHKERRQ(ierr); 4771e7931f94SStefano Zampini ierr = PetscFree(olengths_idxs);CHKERRQ(ierr); 4772e7931f94SStefano Zampini ierr = PetscFree(onodes);CHKERRQ(ierr); 477328143c3dSStefano Zampini if (nis) { 477428143c3dSStefano Zampini ierr = PetscFree(ilengths_idxs_is);CHKERRQ(ierr); 477528143c3dSStefano Zampini ierr = PetscFree(olengths_idxs_is);CHKERRQ(ierr); 477628143c3dSStefano Zampini ierr = PetscFree(onodes_is);CHKERRQ(ierr); 477728143c3dSStefano Zampini } 477828143c3dSStefano Zampini ierr = PetscSubcommDestroy(&subcomm);CHKERRQ(ierr); 477928143c3dSStefano Zampini if (destroy_mat) { /* destroy mat is true only if restrict comm is true and process will not partecipate */ 478028143c3dSStefano Zampini ierr = MatDestroy(mat_n);CHKERRQ(ierr); 478128143c3dSStefano Zampini for (i=0;i<nis;i++) { 478228143c3dSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 478328143c3dSStefano Zampini } 478453a05cb3SStefano Zampini *mat_n = NULL; 478528143c3dSStefano Zampini } 4786e7931f94SStefano Zampini PetscFunctionReturn(0); 4787e7931f94SStefano Zampini } 4788a57a6d2fSStefano Zampini 478912edc857SStefano Zampini /* temporary hack into ksp private data structure */ 4790af0996ceSBarry Smith #include <petsc/private/kspimpl.h> 479112edc857SStefano Zampini 4792c8587f34SStefano Zampini #undef __FUNCT__ 4793c8587f34SStefano Zampini #define __FUNCT__ "PCBDDCSetUpCoarseSolver" 4794c8587f34SStefano Zampini PetscErrorCode PCBDDCSetUpCoarseSolver(PC pc,PetscScalar* coarse_submat_vals) 4795c8587f34SStefano Zampini { 4796c8587f34SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 4797c8587f34SStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 479820a2ab83SStefano Zampini Mat coarse_mat,coarse_mat_is,coarse_submat_dense; 47999881197aSStefano Zampini MatNullSpace CoarseNullSpace=NULL; 480020a2ab83SStefano Zampini ISLocalToGlobalMapping coarse_islg; 48016e683305SStefano Zampini IS coarse_is,*isarray; 48026e683305SStefano Zampini PetscInt i,im_active=-1,active_procs=-1; 480330368db7SStefano Zampini PetscInt nis,nisdofs,nisneu,nisvert; 4804f9eb5b7dSStefano Zampini PC pc_temp; 4805c8587f34SStefano Zampini PCType coarse_pc_type; 4806c8587f34SStefano Zampini KSPType coarse_ksp_type; 4807f9eb5b7dSStefano Zampini PetscBool multilevel_requested,multilevel_allowed; 48084f3a063dSStefano Zampini PetscBool isredundant,isbddc,isnn,coarse_reuse; 48096e683305SStefano Zampini Mat t_coarse_mat_is; 48106e683305SStefano Zampini PetscInt void_procs,ncoarse_ml,ncoarse_ds,ncoarse; 48116e683305SStefano Zampini PetscMPIInt all_procs; 481274e2c79eSStefano Zampini PetscBool csin_ml,csin_ds,csin,csin_type_simple,redist; 481368457ee5SStefano Zampini PetscBool compute_vecs = PETSC_FALSE; 481422bc73bbSStefano Zampini PetscScalar *array; 48159881197aSStefano Zampini PetscErrorCode ierr; 4816fdc09c96SStefano Zampini 4817c8587f34SStefano Zampini PetscFunctionBegin; 4818c8587f34SStefano Zampini /* Assign global numbering to coarse dofs */ 481968457ee5SStefano 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 */ 4820fa7f1dd8SStefano Zampini PetscInt ocoarse_size; 48215a75c04eSSatish Balay compute_vecs = PETSC_TRUE; 4822fa7f1dd8SStefano Zampini ocoarse_size = pcbddc->coarse_size; 4823f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->global_primal_indices);CHKERRQ(ierr); 4824f4ddd8eeSStefano Zampini ierr = PCBDDCComputePrimalNumbering(pc,&pcbddc->coarse_size,&pcbddc->global_primal_indices);CHKERRQ(ierr); 4825f4ddd8eeSStefano Zampini /* see if we can avoid some work */ 4826fa7f1dd8SStefano Zampini if (pcbddc->coarse_ksp) { /* coarse ksp has already been created */ 482751bea450SStefano Zampini /* if the coarse size is different or we are using adaptive selection, better to not reuse the coarse matrix */ 482851bea450SStefano Zampini if (ocoarse_size != pcbddc->coarse_size || pcbddc->adaptive_selection) { 4829dc4bcba2SStefano Zampini PC pc; 4830dc4bcba2SStefano Zampini PetscBool isbddc; 4831dc4bcba2SStefano Zampini 4832dc4bcba2SStefano Zampini /* temporary workaround since PCBDDC does not have a reset method so far */ 4833dc4bcba2SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&pc);CHKERRQ(ierr); 4834dc4bcba2SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc,PCBDDC,&isbddc);CHKERRQ(ierr); 4835dc4bcba2SStefano Zampini if (isbddc) { 4836dc4bcba2SStefano Zampini ierr = PCDestroy(&pc);CHKERRQ(ierr); 4837dc4bcba2SStefano Zampini } 4838727cdba6SStefano Zampini ierr = KSPReset(pcbddc->coarse_ksp);CHKERRQ(ierr); 4839fa7f1dd8SStefano Zampini coarse_reuse = PETSC_FALSE; 4840fa7f1dd8SStefano Zampini } else { /* we can safely reuse already computed coarse matrix */ 4841fa7f1dd8SStefano Zampini coarse_reuse = PETSC_TRUE; 4842f4ddd8eeSStefano Zampini } 4843fa7f1dd8SStefano Zampini } else { /* there's no coarse ksp, so we need to create the coarse matrix too */ 4844fa7f1dd8SStefano Zampini coarse_reuse = PETSC_FALSE; 4845f4ddd8eeSStefano Zampini } 484670cf5478SStefano Zampini /* reset any subassembling information */ 484770cf5478SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling);CHKERRQ(ierr); 48486e683305SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 48496e683305SStefano Zampini } else { /* primal space is unchanged, so we can reuse coarse matrix */ 4850fa7f1dd8SStefano Zampini coarse_reuse = PETSC_TRUE; 4851f4ddd8eeSStefano Zampini } 4852c8587f34SStefano Zampini 48536e683305SStefano Zampini /* count "active" (i.e. with positive local size) and "void" processes */ 48542b510759SStefano Zampini im_active = !!(pcis->n); 48552b510759SStefano Zampini ierr = MPI_Allreduce(&im_active,&active_procs,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 48566e683305SStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)pc),&all_procs);CHKERRQ(ierr); 48576e683305SStefano Zampini void_procs = all_procs-active_procs; 48586e683305SStefano Zampini csin_type_simple = PETSC_TRUE; 485974e2c79eSStefano Zampini redist = PETSC_FALSE; 486022bc73bbSStefano Zampini if (pcbddc->current_level && void_procs) { 48616e683305SStefano Zampini csin_ml = PETSC_TRUE; 48626e683305SStefano Zampini ncoarse_ml = void_procs; 4863779c1cceSStefano Zampini /* it has no sense to redistribute on a set of processors larger than the number of active processes */ 4864779c1cceSStefano Zampini if (pcbddc->redistribute_coarse > 0 && pcbddc->redistribute_coarse < active_procs) { 48656e683305SStefano Zampini csin_ds = PETSC_TRUE; 486618a45a71SStefano Zampini ncoarse_ds = pcbddc->redistribute_coarse; 486718a45a71SStefano Zampini redist = PETSC_TRUE; 486818a45a71SStefano Zampini } else { 48696e683305SStefano Zampini csin_ds = PETSC_TRUE; 4870779c1cceSStefano Zampini ncoarse_ds = active_procs; 4871779c1cceSStefano Zampini redist = PETSC_TRUE; 487218a45a71SStefano Zampini } 48736e683305SStefano Zampini } else { 48746e683305SStefano Zampini csin_ml = PETSC_FALSE; 48756e683305SStefano Zampini ncoarse_ml = all_procs; 48766e683305SStefano Zampini if (void_procs) { 48776e683305SStefano Zampini csin_ds = PETSC_TRUE; 48786e683305SStefano Zampini ncoarse_ds = void_procs; 48796e683305SStefano Zampini csin_type_simple = PETSC_FALSE; 48806e683305SStefano Zampini } else { 4881779c1cceSStefano Zampini if (pcbddc->redistribute_coarse > 0 && pcbddc->redistribute_coarse < all_procs) { 488274e2c79eSStefano Zampini csin_ds = PETSC_TRUE; 488374e2c79eSStefano Zampini ncoarse_ds = pcbddc->redistribute_coarse; 488474e2c79eSStefano Zampini redist = PETSC_TRUE; 488574e2c79eSStefano Zampini } else { 48866e683305SStefano Zampini csin_ds = PETSC_FALSE; 48876e683305SStefano Zampini ncoarse_ds = all_procs; 48886e683305SStefano Zampini } 48896e683305SStefano Zampini } 489074e2c79eSStefano Zampini } 48916e683305SStefano Zampini 48926e683305SStefano Zampini /* 48936e683305SStefano Zampini test if we can go multilevel: three conditions must be satisfied: 48946e683305SStefano Zampini - we have not exceeded the number of levels requested 48956e683305SStefano Zampini - we can actually subassemble the active processes 48966e683305SStefano Zampini - we can find a suitable number of MPI processes where we can place the subassembled problem 48976e683305SStefano Zampini */ 48986e683305SStefano Zampini multilevel_allowed = PETSC_FALSE; 48996e683305SStefano Zampini multilevel_requested = PETSC_FALSE; 49006e683305SStefano Zampini if (pcbddc->current_level < pcbddc->max_levels) { 49016e683305SStefano Zampini multilevel_requested = PETSC_TRUE; 49026e683305SStefano Zampini if (active_procs/pcbddc->coarsening_ratio < 2 || ncoarse_ml/pcbddc->coarsening_ratio < 2) { 4903f9eb5b7dSStefano Zampini multilevel_allowed = PETSC_FALSE; 49042b510759SStefano Zampini } else { 4905f9eb5b7dSStefano Zampini multilevel_allowed = PETSC_TRUE; 4906c8587f34SStefano Zampini } 4907c8587f34SStefano Zampini } 49086e683305SStefano Zampini /* determine number of process partecipating to coarse solver */ 49096e683305SStefano Zampini if (multilevel_allowed) { 49106e683305SStefano Zampini ncoarse = ncoarse_ml; 49116e683305SStefano Zampini csin = csin_ml; 491258da7f69SStefano Zampini redist = PETSC_FALSE; 49136e683305SStefano Zampini } else { 49146e683305SStefano Zampini ncoarse = ncoarse_ds; 49156e683305SStefano Zampini csin = csin_ds; 49166e683305SStefano Zampini } 4917e7931f94SStefano Zampini 4918abbbba34SStefano Zampini /* creates temporary l2gmap and IS for coarse indexes */ 4919abbbba34SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),pcbddc->local_primal_size,pcbddc->global_primal_indices,PETSC_COPY_VALUES,&coarse_is);CHKERRQ(ierr); 4920abbbba34SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(coarse_is,&coarse_islg);CHKERRQ(ierr); 4921abbbba34SStefano Zampini 4922abbbba34SStefano Zampini /* creates temporary MATIS object for coarse matrix */ 492322bc73bbSStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,pcbddc->local_primal_size,pcbddc->local_primal_size,NULL,&coarse_submat_dense);CHKERRQ(ierr); 492422bc73bbSStefano Zampini ierr = MatDenseGetArray(coarse_submat_dense,&array);CHKERRQ(ierr); 492522bc73bbSStefano Zampini ierr = PetscMemcpy(array,coarse_submat_vals,sizeof(*coarse_submat_vals)*pcbddc->local_primal_size*pcbddc->local_primal_size);CHKERRQ(ierr); 492622bc73bbSStefano Zampini ierr = MatDenseRestoreArray(coarse_submat_dense,&array);CHKERRQ(ierr); 4927e176bc59SStefano 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); 49286e683305SStefano Zampini ierr = MatISSetLocalMat(t_coarse_mat_is,coarse_submat_dense);CHKERRQ(ierr); 49296e683305SStefano Zampini ierr = MatAssemblyBegin(t_coarse_mat_is,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 49306e683305SStefano Zampini ierr = MatAssemblyEnd(t_coarse_mat_is,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4931abbbba34SStefano Zampini ierr = MatDestroy(&coarse_submat_dense);CHKERRQ(ierr); 4932abbbba34SStefano Zampini 49336e683305SStefano Zampini /* compute dofs splitting and neumann boundaries for coarse dofs */ 493430368db7SStefano Zampini if (multilevel_allowed && (pcbddc->n_ISForDofsLocal || pcbddc->NeumannBoundariesLocal || (pcbddc->benign_saddle_point && pcbddc->user_primal_vertices_local))) { /* protects from unneded computations */ 49356e683305SStefano Zampini PetscInt *tidxs,*tidxs2,nout,tsize,i; 49366e683305SStefano Zampini const PetscInt *idxs; 49376e683305SStefano Zampini ISLocalToGlobalMapping tmap; 49386e683305SStefano Zampini 49396e683305SStefano Zampini /* create map between primal indices (in local representative ordering) and local primal numbering */ 49400be93d54SStefano Zampini ierr = ISLocalToGlobalMappingCreate(PETSC_COMM_SELF,1,pcbddc->local_primal_size,pcbddc->primal_indices_local_idxs,PETSC_COPY_VALUES,&tmap);CHKERRQ(ierr); 49416e683305SStefano Zampini /* allocate space for temporary storage */ 4942854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->local_primal_size,&tidxs);CHKERRQ(ierr); 4943854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->local_primal_size,&tidxs2);CHKERRQ(ierr); 49446e683305SStefano Zampini /* allocate for IS array */ 49456e683305SStefano Zampini nisdofs = pcbddc->n_ISForDofsLocal; 49466e683305SStefano Zampini nisneu = !!pcbddc->NeumannBoundariesLocal; 494730368db7SStefano Zampini nisvert = 0; 494830368db7SStefano Zampini if (pcbddc->benign_saddle_point && pcbddc->user_primal_vertices_local) { 494930368db7SStefano Zampini nisvert = 1; 495030368db7SStefano Zampini } 495130368db7SStefano Zampini nis = nisdofs + nisneu + nisvert; 4952854ce69bSBarry Smith ierr = PetscMalloc1(nis,&isarray);CHKERRQ(ierr); 49536e683305SStefano Zampini /* dofs splitting */ 49546e683305SStefano Zampini for (i=0;i<nisdofs;i++) { 49556e683305SStefano Zampini /* ierr = ISView(pcbddc->ISForDofsLocal[i],0);CHKERRQ(ierr); */ 49566e683305SStefano Zampini ierr = ISGetLocalSize(pcbddc->ISForDofsLocal[i],&tsize);CHKERRQ(ierr); 49576e683305SStefano Zampini ierr = ISGetIndices(pcbddc->ISForDofsLocal[i],&idxs);CHKERRQ(ierr); 49586e683305SStefano Zampini ierr = ISGlobalToLocalMappingApply(tmap,IS_GTOLM_DROP,tsize,idxs,&nout,tidxs);CHKERRQ(ierr); 49596e683305SStefano Zampini ierr = ISRestoreIndices(pcbddc->ISForDofsLocal[i],&idxs);CHKERRQ(ierr); 49606e683305SStefano Zampini ierr = ISLocalToGlobalMappingApply(coarse_islg,nout,tidxs,tidxs2);CHKERRQ(ierr); 496130368db7SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),nout,tidxs2,PETSC_COPY_VALUES,&isarray[i]);CHKERRQ(ierr); 49626e683305SStefano Zampini /* ierr = ISView(isarray[i],0);CHKERRQ(ierr); */ 49636e683305SStefano Zampini } 49646e683305SStefano Zampini /* neumann boundaries */ 49656e683305SStefano Zampini if (pcbddc->NeumannBoundariesLocal) { 49666e683305SStefano Zampini /* ierr = ISView(pcbddc->NeumannBoundariesLocal,0);CHKERRQ(ierr); */ 49676e683305SStefano Zampini ierr = ISGetLocalSize(pcbddc->NeumannBoundariesLocal,&tsize);CHKERRQ(ierr); 49686e683305SStefano Zampini ierr = ISGetIndices(pcbddc->NeumannBoundariesLocal,&idxs);CHKERRQ(ierr); 49696e683305SStefano Zampini ierr = ISGlobalToLocalMappingApply(tmap,IS_GTOLM_DROP,tsize,idxs,&nout,tidxs);CHKERRQ(ierr); 49706e683305SStefano Zampini ierr = ISRestoreIndices(pcbddc->NeumannBoundariesLocal,&idxs);CHKERRQ(ierr); 49716e683305SStefano Zampini ierr = ISLocalToGlobalMappingApply(coarse_islg,nout,tidxs,tidxs2);CHKERRQ(ierr); 497230368db7SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),nout,tidxs2,PETSC_COPY_VALUES,&isarray[nisdofs]);CHKERRQ(ierr); 49736e683305SStefano Zampini /* ierr = ISView(isarray[nisdofs],0);CHKERRQ(ierr); */ 49746e683305SStefano Zampini } 497530368db7SStefano Zampini /* primal vertices (benign) */ 497630368db7SStefano Zampini if (pcbddc->benign_saddle_point && pcbddc->user_primal_vertices_local) { 497730368db7SStefano Zampini ierr = ISGetLocalSize(pcbddc->user_primal_vertices_local,&tsize);CHKERRQ(ierr); 497830368db7SStefano Zampini ierr = ISGetIndices(pcbddc->user_primal_vertices_local,&idxs);CHKERRQ(ierr); 497930368db7SStefano Zampini ierr = ISGlobalToLocalMappingApply(tmap,IS_GTOLM_DROP,tsize,idxs,&nout,tidxs);CHKERRQ(ierr); 498030368db7SStefano Zampini ierr = ISRestoreIndices(pcbddc->user_primal_vertices_local,&idxs);CHKERRQ(ierr); 498130368db7SStefano Zampini ierr = ISLocalToGlobalMappingApply(coarse_islg,nout,tidxs,tidxs2);CHKERRQ(ierr); 498230368db7SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),nout,tidxs2,PETSC_COPY_VALUES,&isarray[nis-1]);CHKERRQ(ierr); 498330368db7SStefano Zampini } 49846e683305SStefano Zampini /* free memory */ 49856e683305SStefano Zampini ierr = PetscFree(tidxs);CHKERRQ(ierr); 49866e683305SStefano Zampini ierr = PetscFree(tidxs2);CHKERRQ(ierr); 49876e683305SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&tmap);CHKERRQ(ierr); 49886e683305SStefano Zampini } else { 49896e683305SStefano Zampini nis = 0; 49906e683305SStefano Zampini nisdofs = 0; 49916e683305SStefano Zampini nisneu = 0; 499230368db7SStefano Zampini nisvert = 0; 49936e683305SStefano Zampini isarray = NULL; 49946e683305SStefano Zampini } 49956e683305SStefano Zampini /* destroy no longer needed map */ 49966e683305SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&coarse_islg);CHKERRQ(ierr); 49976e683305SStefano Zampini 49986e683305SStefano Zampini /* restrict on coarse candidates (if needed) */ 49996e683305SStefano Zampini coarse_mat_is = NULL; 50006e683305SStefano Zampini if (csin) { 50016e683305SStefano Zampini if (!pcbddc->coarse_subassembling_init ) { /* creates subassembling init pattern if not present */ 500274e2c79eSStefano Zampini if (redist) { 500374e2c79eSStefano Zampini PetscMPIInt rank; 5004779c1cceSStefano Zampini PetscInt spc,n_spc_p1,dest[1],destsize; 500574e2c79eSStefano Zampini 500674e2c79eSStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)pc),&rank);CHKERRQ(ierr); 500758da7f69SStefano Zampini spc = active_procs/ncoarse; 500858da7f69SStefano Zampini n_spc_p1 = active_procs%ncoarse; 5009779c1cceSStefano Zampini if (im_active) { 5010779c1cceSStefano Zampini destsize = 1; 501174e2c79eSStefano Zampini if (rank > n_spc_p1*(spc+1)-1) { 501274e2c79eSStefano Zampini dest[0] = n_spc_p1+(rank-(n_spc_p1*(spc+1)))/spc; 501374e2c79eSStefano Zampini } else { 501474e2c79eSStefano Zampini dest[0] = rank/(spc+1); 501574e2c79eSStefano Zampini } 501674e2c79eSStefano Zampini } else { 5017779c1cceSStefano Zampini destsize = 0; 50186e683305SStefano Zampini } 5019779c1cceSStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),destsize,dest,PETSC_COPY_VALUES,&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 5020779c1cceSStefano Zampini } else if (csin_type_simple) { 50216e683305SStefano Zampini PetscMPIInt rank; 50226e683305SStefano Zampini PetscInt issize,isidx; 5023779c1cceSStefano Zampini 50246e683305SStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)pc),&rank);CHKERRQ(ierr); 50256e683305SStefano Zampini if (im_active) { 50266e683305SStefano Zampini issize = 1; 50276e683305SStefano Zampini isidx = (PetscInt)rank; 50286e683305SStefano Zampini } else { 50296e683305SStefano Zampini issize = 0; 50306e683305SStefano Zampini isidx = -1; 50316e683305SStefano Zampini } 50326e683305SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),issize,&isidx,PETSC_COPY_VALUES,&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 5033779c1cceSStefano Zampini } else { /* get a suitable subassembling pattern from MATIS code */ 5034b0c7d250SStefano Zampini ierr = MatISGetSubassemblingPattern(t_coarse_mat_is,ncoarse,pcbddc->coarse_adj_red,&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 50356e683305SStefano Zampini } 5036779c1cceSStefano Zampini 5037779c1cceSStefano Zampini /* we need to shift on coarse candidates either if we are not redistributing or we are redistributing and we have enough void processes */ 5038779c1cceSStefano Zampini if (!redist || ncoarse <= void_procs) { 5039779c1cceSStefano Zampini PetscInt ncoarse_cand,tissize,*nisindices; 5040779c1cceSStefano Zampini PetscInt *coarse_candidates; 5041779c1cceSStefano Zampini const PetscInt* tisindices; 5042779c1cceSStefano Zampini 5043779c1cceSStefano Zampini /* get coarse candidates' ranks in pc communicator */ 5044779c1cceSStefano Zampini ierr = PetscMalloc1(all_procs,&coarse_candidates);CHKERRQ(ierr); 5045779c1cceSStefano Zampini ierr = MPI_Allgather(&im_active,1,MPIU_INT,coarse_candidates,1,MPIU_INT,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 5046779c1cceSStefano Zampini for (i=0,ncoarse_cand=0;i<all_procs;i++) { 5047779c1cceSStefano Zampini if (!coarse_candidates[i]) { 5048779c1cceSStefano Zampini coarse_candidates[ncoarse_cand++]=i; 5049779c1cceSStefano Zampini } 5050779c1cceSStefano Zampini } 5051779c1cceSStefano Zampini if (ncoarse_cand < ncoarse) SETERRQ2(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"This should not happen! %d < %d",ncoarse_cand,ncoarse); 5052779c1cceSStefano Zampini 5053779c1cceSStefano Zampini 50546e683305SStefano Zampini if (pcbddc->dbg_flag) { 50556e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 50566e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Subassembling pattern init (before shift)\n");CHKERRQ(ierr); 50576e683305SStefano Zampini ierr = ISView(pcbddc->coarse_subassembling_init,pcbddc->dbg_viewer);CHKERRQ(ierr); 50586e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Coarse candidates\n");CHKERRQ(ierr); 5059779c1cceSStefano Zampini for (i=0;i<ncoarse_cand;i++) { 50606e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"%d ",coarse_candidates[i]);CHKERRQ(ierr); 50616e683305SStefano Zampini } 50626e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"\n");CHKERRQ(ierr); 50636e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 50646e683305SStefano Zampini } 50656e683305SStefano Zampini /* shift the pattern on coarse candidates */ 50666e683305SStefano Zampini ierr = ISGetLocalSize(pcbddc->coarse_subassembling_init,&tissize);CHKERRQ(ierr); 50676e683305SStefano Zampini ierr = ISGetIndices(pcbddc->coarse_subassembling_init,&tisindices);CHKERRQ(ierr); 5068854ce69bSBarry Smith ierr = PetscMalloc1(tissize,&nisindices);CHKERRQ(ierr); 50696e683305SStefano Zampini for (i=0;i<tissize;i++) nisindices[i] = coarse_candidates[tisindices[i]]; 50706e683305SStefano Zampini ierr = ISRestoreIndices(pcbddc->coarse_subassembling_init,&tisindices);CHKERRQ(ierr); 50716e683305SStefano Zampini ierr = ISGeneralSetIndices(pcbddc->coarse_subassembling_init,tissize,nisindices,PETSC_OWN_POINTER);CHKERRQ(ierr); 50726e683305SStefano Zampini ierr = PetscFree(coarse_candidates);CHKERRQ(ierr); 50736e683305SStefano Zampini } 50746e683305SStefano Zampini if (pcbddc->dbg_flag) { 50756e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 50766e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Subassembling pattern init\n");CHKERRQ(ierr); 50776e683305SStefano Zampini ierr = ISView(pcbddc->coarse_subassembling_init,pcbddc->dbg_viewer);CHKERRQ(ierr); 50786e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 50796e683305SStefano Zampini } 5080779c1cceSStefano Zampini } 50816e683305SStefano Zampini /* get temporary coarse mat in IS format restricted on coarse procs (plus additional index sets of isarray) */ 508253a05cb3SStefano Zampini if (multilevel_allowed) { /* we need to keep tracking of void processes for future placements */ 508353a05cb3SStefano 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); 508453a05cb3SStefano Zampini } else { /* this is the last level, so use just receiving processes in subcomm */ 508553a05cb3SStefano 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); 508653a05cb3SStefano Zampini } 50876e683305SStefano Zampini } else { 50886e683305SStefano Zampini if (pcbddc->dbg_flag) { 50896e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 50906e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Subassembling pattern init not needed\n");CHKERRQ(ierr); 50916e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 50926e683305SStefano Zampini } 50936e683305SStefano Zampini ierr = PetscObjectReference((PetscObject)t_coarse_mat_is);CHKERRQ(ierr); 50946e683305SStefano Zampini coarse_mat_is = t_coarse_mat_is; 50956e683305SStefano Zampini } 50966e683305SStefano Zampini 50976e683305SStefano Zampini /* create local to global scatters for coarse problem */ 509868457ee5SStefano Zampini if (compute_vecs) { 50996e683305SStefano Zampini PetscInt lrows; 51006e683305SStefano Zampini ierr = VecDestroy(&pcbddc->coarse_vec);CHKERRQ(ierr); 51016e683305SStefano Zampini if (coarse_mat_is) { 51026e683305SStefano Zampini ierr = MatGetLocalSize(coarse_mat_is,&lrows,NULL);CHKERRQ(ierr); 51036e683305SStefano Zampini } else { 51046e683305SStefano Zampini lrows = 0; 51056e683305SStefano Zampini } 51066e683305SStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pc),&pcbddc->coarse_vec);CHKERRQ(ierr); 51076e683305SStefano Zampini ierr = VecSetSizes(pcbddc->coarse_vec,lrows,PETSC_DECIDE);CHKERRQ(ierr); 51086e683305SStefano Zampini ierr = VecSetType(pcbddc->coarse_vec,VECSTANDARD);CHKERRQ(ierr); 51096e683305SStefano Zampini ierr = VecScatterDestroy(&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 51106e683305SStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_P,NULL,pcbddc->coarse_vec,coarse_is,&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 51116e683305SStefano Zampini } 51126e683305SStefano Zampini ierr = ISDestroy(&coarse_is);CHKERRQ(ierr); 51136e683305SStefano Zampini ierr = MatDestroy(&t_coarse_mat_is);CHKERRQ(ierr); 5114c8587f34SStefano Zampini 5115f9eb5b7dSStefano Zampini /* set defaults for coarse KSP and PC */ 5116f9eb5b7dSStefano Zampini if (multilevel_allowed) { 5117f9eb5b7dSStefano Zampini coarse_ksp_type = KSPRICHARDSON; 5118f9eb5b7dSStefano Zampini coarse_pc_type = PCBDDC; 5119f9eb5b7dSStefano Zampini } else { 5120f9eb5b7dSStefano Zampini coarse_ksp_type = KSPPREONLY; 5121f9eb5b7dSStefano Zampini coarse_pc_type = PCREDUNDANT; 5122c8587f34SStefano Zampini } 5123c8587f34SStefano Zampini 51246e683305SStefano Zampini /* print some info if requested */ 51256e683305SStefano Zampini if (pcbddc->dbg_flag) { 51266e683305SStefano Zampini if (!multilevel_allowed) { 51276e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 51286e683305SStefano Zampini if (multilevel_requested) { 51296e683305SStefano 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); 51306e683305SStefano Zampini } else if (pcbddc->max_levels) { 51316e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Maximum number of requested levels reached (%d)\n",pcbddc->max_levels);CHKERRQ(ierr); 51326e683305SStefano Zampini } 51336e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 51346e683305SStefano Zampini } 51356e683305SStefano Zampini } 51366e683305SStefano Zampini 5137f9eb5b7dSStefano Zampini /* create the coarse KSP object only once with defaults */ 51386e683305SStefano Zampini if (coarse_mat_is) { 51396e683305SStefano Zampini MatReuse coarse_mat_reuse; 51406a1308c2SStefano Zampini PetscViewer dbg_viewer = NULL; 51416e683305SStefano Zampini if (pcbddc->dbg_flag) { 51426e683305SStefano Zampini dbg_viewer = PETSC_VIEWER_STDOUT_(PetscObjectComm((PetscObject)coarse_mat_is)); 51436e683305SStefano Zampini ierr = PetscViewerASCIIAddTab(dbg_viewer,2*pcbddc->current_level);CHKERRQ(ierr); 51446e683305SStefano Zampini } 5145f9eb5b7dSStefano Zampini if (!pcbddc->coarse_ksp) { 5146312be037SStefano Zampini char prefix[256],str_level[16]; 5147e604994aSStefano Zampini size_t len; 51486e683305SStefano Zampini ierr = KSPCreate(PetscObjectComm((PetscObject)coarse_mat_is),&pcbddc->coarse_ksp);CHKERRQ(ierr); 5149422a814eSBarry Smith ierr = KSPSetErrorIfNotConverged(pcbddc->coarse_ksp,pc->erroriffailure);CHKERRQ(ierr); 5150c8587f34SStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->coarse_ksp,(PetscObject)pc,1);CHKERRQ(ierr); 5151f9eb5b7dSStefano Zampini ierr = KSPSetTolerances(pcbddc->coarse_ksp,PETSC_DEFAULT,PETSC_DEFAULT,PETSC_DEFAULT,1);CHKERRQ(ierr); 51525f76c7aeSStefano Zampini ierr = KSPSetOperators(pcbddc->coarse_ksp,coarse_mat_is,coarse_mat_is);CHKERRQ(ierr); 5153c8587f34SStefano Zampini ierr = KSPSetType(pcbddc->coarse_ksp,coarse_ksp_type);CHKERRQ(ierr); 51546e683305SStefano Zampini ierr = KSPSetNormType(pcbddc->coarse_ksp,KSP_NORM_NONE);CHKERRQ(ierr); 5155c8587f34SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&pc_temp);CHKERRQ(ierr); 5156c8587f34SStefano Zampini ierr = PCSetType(pc_temp,coarse_pc_type);CHKERRQ(ierr); 5157e604994aSStefano Zampini /* prefix */ 5158e604994aSStefano Zampini ierr = PetscStrcpy(prefix,"");CHKERRQ(ierr); 5159e604994aSStefano Zampini ierr = PetscStrcpy(str_level,"");CHKERRQ(ierr); 5160e604994aSStefano Zampini if (!pcbddc->current_level) { 5161e604994aSStefano Zampini ierr = PetscStrcpy(prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 5162e604994aSStefano Zampini ierr = PetscStrcat(prefix,"pc_bddc_coarse_");CHKERRQ(ierr); 5163c8587f34SStefano Zampini } else { 5164e604994aSStefano Zampini ierr = PetscStrlen(((PetscObject)pc)->prefix,&len);CHKERRQ(ierr); 5165312be037SStefano Zampini if (pcbddc->current_level>1) len -= 3; /* remove "lX_" with X level number */ 5166312be037SStefano Zampini if (pcbddc->current_level>10) len -= 1; /* remove another char from level number */ 516734d6797cSStefano Zampini ierr = PetscStrncpy(prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 5168312be037SStefano Zampini sprintf(str_level,"l%d_",(int)(pcbddc->current_level)); 5169e604994aSStefano Zampini ierr = PetscStrcat(prefix,str_level);CHKERRQ(ierr); 5170e604994aSStefano Zampini } 5171e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->coarse_ksp,prefix);CHKERRQ(ierr); 51723e3c6dadSStefano Zampini /* propagate BDDC info to the next level (these are dummy calls if pc_temp is not of type PCBDDC) */ 51733e3c6dadSStefano Zampini ierr = PCBDDCSetLevel(pc_temp,pcbddc->current_level+1);CHKERRQ(ierr); 51743e3c6dadSStefano Zampini ierr = PCBDDCSetCoarseningRatio(pc_temp,pcbddc->coarsening_ratio);CHKERRQ(ierr); 51753e3c6dadSStefano Zampini ierr = PCBDDCSetLevels(pc_temp,pcbddc->max_levels);CHKERRQ(ierr); 5176f9eb5b7dSStefano Zampini /* allow user customization */ 5177f9eb5b7dSStefano Zampini ierr = KSPSetFromOptions(pcbddc->coarse_ksp);CHKERRQ(ierr); 51783e3c6dadSStefano Zampini } 51793e3c6dadSStefano Zampini /* propagate BDDC info to the next level (these are dummy calls if pc_temp is not of type PCBDDC) */ 518051bea450SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&pc_temp);CHKERRQ(ierr); 51813e3c6dadSStefano Zampini if (nisdofs) { 51823e3c6dadSStefano Zampini ierr = PCBDDCSetDofsSplitting(pc_temp,nisdofs,isarray);CHKERRQ(ierr); 51833e3c6dadSStefano Zampini for (i=0;i<nisdofs;i++) { 51843e3c6dadSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 51853e3c6dadSStefano Zampini } 51863e3c6dadSStefano Zampini } 51873e3c6dadSStefano Zampini if (nisneu) { 51883e3c6dadSStefano Zampini ierr = PCBDDCSetNeumannBoundaries(pc_temp,isarray[nisdofs]);CHKERRQ(ierr); 51893e3c6dadSStefano Zampini ierr = ISDestroy(&isarray[nisdofs]);CHKERRQ(ierr); 5190312be037SStefano Zampini } 519130368db7SStefano Zampini if (nisvert) { 519230368db7SStefano Zampini ierr = PCBDDCSetPrimalVerticesIS(pc_temp,isarray[nis-1]);CHKERRQ(ierr); 519330368db7SStefano Zampini ierr = ISDestroy(&isarray[nis-1]);CHKERRQ(ierr); 519430368db7SStefano Zampini } 5195f9eb5b7dSStefano Zampini 5196f9eb5b7dSStefano Zampini /* get some info after set from options */ 5197f9eb5b7dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCNN,&isnn);CHKERRQ(ierr); 5198f9eb5b7dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCBDDC,&isbddc);CHKERRQ(ierr); 51994f3a063dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCREDUNDANT,&isredundant);CHKERRQ(ierr); 52006e683305SStefano Zampini if (isbddc && !multilevel_allowed) { /* multilevel can only be requested via pc_bddc_set_levels */ 5201f9eb5b7dSStefano Zampini ierr = PCSetType(pc_temp,coarse_pc_type);CHKERRQ(ierr); 5202f9eb5b7dSStefano Zampini isbddc = PETSC_FALSE; 5203f9eb5b7dSStefano Zampini } 520439f0f02cSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 52054f3a063dSStefano Zampini if (isredundant) { 52064f3a063dSStefano Zampini KSP inner_ksp; 52074f3a063dSStefano Zampini PC inner_pc; 52084f3a063dSStefano Zampini ierr = PCRedundantGetKSP(pc_temp,&inner_ksp);CHKERRQ(ierr); 52094f3a063dSStefano Zampini ierr = KSPGetPC(inner_ksp,&inner_pc);CHKERRQ(ierr); 52104f3a063dSStefano Zampini ierr = PCFactorSetReuseFill(inner_pc,PETSC_TRUE);CHKERRQ(ierr); 52114f3a063dSStefano Zampini } 5212f9eb5b7dSStefano Zampini 5213f9eb5b7dSStefano Zampini /* assemble coarse matrix */ 5214fa7f1dd8SStefano Zampini if (coarse_reuse) { 521581d9aea3SBarry Smith ierr = KSPGetOperators(pcbddc->coarse_ksp,&coarse_mat,NULL);CHKERRQ(ierr); 5216fa7f1dd8SStefano Zampini ierr = PetscObjectReference((PetscObject)coarse_mat);CHKERRQ(ierr); 52176e683305SStefano Zampini coarse_mat_reuse = MAT_REUSE_MATRIX; 5218fa7f1dd8SStefano Zampini } else { 52196e683305SStefano Zampini coarse_mat_reuse = MAT_INITIAL_MATRIX; 5220fa7f1dd8SStefano Zampini } 5221c8587f34SStefano Zampini if (isbddc || isnn) { 522222bc73bbSStefano Zampini if (pcbddc->coarsening_ratio > 1) { 522370cf5478SStefano Zampini if (!pcbddc->coarse_subassembling) { /* subassembling info is not present */ 5224b0c7d250SStefano Zampini ierr = MatISGetSubassemblingPattern(coarse_mat_is,active_procs/pcbddc->coarsening_ratio,pcbddc->coarse_adj_red,&pcbddc->coarse_subassembling);CHKERRQ(ierr); 522522b6e8a2SStefano Zampini if (pcbddc->dbg_flag) { 52266e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 52276e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Subassembling pattern\n");CHKERRQ(ierr); 52286e683305SStefano Zampini ierr = ISView(pcbddc->coarse_subassembling,dbg_viewer);CHKERRQ(ierr); 52296e683305SStefano Zampini ierr = PetscViewerFlush(dbg_viewer);CHKERRQ(ierr); 523022b6e8a2SStefano Zampini } 523170cf5478SStefano Zampini } 523253a05cb3SStefano Zampini ierr = MatISSubassemble(coarse_mat_is,pcbddc->coarse_subassembling,0,PETSC_FALSE,PETSC_FALSE,coarse_mat_reuse,&coarse_mat,0,NULL);CHKERRQ(ierr); 523370cf5478SStefano Zampini } else { 523422bc73bbSStefano Zampini ierr = PetscObjectReference((PetscObject)coarse_mat_is);CHKERRQ(ierr); 523522bc73bbSStefano Zampini coarse_mat = coarse_mat_is; 523622bc73bbSStefano Zampini } 523722bc73bbSStefano Zampini } else { 52382e1e5fa4SStefano Zampini ierr = MatISGetMPIXAIJ(coarse_mat_is,coarse_mat_reuse,&coarse_mat);CHKERRQ(ierr); 5239c8587f34SStefano Zampini } 5240c8587f34SStefano Zampini ierr = MatDestroy(&coarse_mat_is);CHKERRQ(ierr); 5241c8587f34SStefano Zampini 52423301b35fSStefano Zampini /* propagate symmetry info of coarse matrix */ 52435a16e3a0SStefano Zampini ierr = MatSetOption(coarse_mat,MAT_STRUCTURALLY_SYMMETRIC,PETSC_TRUE);CHKERRQ(ierr); 52443301b35fSStefano Zampini if (pc->pmat->symmetric_set) { 52453301b35fSStefano Zampini ierr = MatSetOption(coarse_mat,MAT_SYMMETRIC,pc->pmat->symmetric);CHKERRQ(ierr); 52463301b35fSStefano Zampini } 52473301b35fSStefano Zampini if (pc->pmat->hermitian_set) { 52483301b35fSStefano Zampini ierr = MatSetOption(coarse_mat,MAT_HERMITIAN,pc->pmat->hermitian);CHKERRQ(ierr); 52493301b35fSStefano Zampini } 52503301b35fSStefano Zampini if (pc->pmat->spd_set) { 52513301b35fSStefano Zampini ierr = MatSetOption(coarse_mat,MAT_SPD,pc->pmat->spd);CHKERRQ(ierr); 52523301b35fSStefano Zampini } 52536e683305SStefano Zampini /* set operators */ 52545f76c7aeSStefano Zampini ierr = KSPSetOperators(pcbddc->coarse_ksp,coarse_mat,coarse_mat);CHKERRQ(ierr); 52556e683305SStefano Zampini if (pcbddc->dbg_flag) { 52566e683305SStefano Zampini ierr = PetscViewerASCIISubtractTab(dbg_viewer,2*pcbddc->current_level);CHKERRQ(ierr); 52576e683305SStefano Zampini } 52586e683305SStefano Zampini } else { /* processes non partecipating to coarse solver (if any) */ 52596e683305SStefano Zampini coarse_mat = 0; 52606e683305SStefano Zampini } 52616e683305SStefano Zampini ierr = PetscFree(isarray);CHKERRQ(ierr); 5262b1ecc7b1SStefano Zampini #if 0 5263b9b85e73SStefano Zampini { 5264b9b85e73SStefano Zampini PetscViewer viewer; 5265b9b85e73SStefano Zampini char filename[256]; 5266b1ecc7b1SStefano Zampini sprintf(filename,"coarse_mat_level%d.m",pcbddc->current_level); 5267b1ecc7b1SStefano Zampini ierr = PetscViewerASCIIOpen(PetscObjectComm((PetscObject)coarse_mat),filename,&viewer);CHKERRQ(ierr); 5268b9b85e73SStefano Zampini ierr = PetscViewerSetFormat(viewer,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 5269b9b85e73SStefano Zampini ierr = MatView(coarse_mat,viewer);CHKERRQ(ierr); 5270b9b85e73SStefano Zampini ierr = PetscViewerDestroy(&viewer);CHKERRQ(ierr); 5271b9b85e73SStefano Zampini } 5272b9b85e73SStefano Zampini #endif 5273c8587f34SStefano Zampini 5274c8587f34SStefano Zampini /* Compute coarse null space (special handling by BDDC only) */ 5275298c0119SStefano Zampini #if 0 5276c8587f34SStefano Zampini if (pcbddc->NullSpace) { 5277c8587f34SStefano Zampini ierr = PCBDDCNullSpaceAssembleCoarse(pc,coarse_mat,&CoarseNullSpace);CHKERRQ(ierr); 527898a51de6SStefano Zampini } 5279298c0119SStefano Zampini #endif 5280b0f5fe93SStefano Zampini /* hack */ 528198a51de6SStefano Zampini if (pcbddc->coarse_ksp) { 528298a51de6SStefano Zampini Vec crhs,csol; 528304708bb6SStefano Zampini 5284f347579bSStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&csol);CHKERRQ(ierr); 5285f347579bSStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&crhs);CHKERRQ(ierr); 5286f347579bSStefano Zampini if (!csol) { 52872a7a6963SBarry Smith ierr = MatCreateVecs(coarse_mat,&((pcbddc->coarse_ksp)->vec_sol),NULL);CHKERRQ(ierr); 5288f9eb5b7dSStefano Zampini } 5289f347579bSStefano Zampini if (!crhs) { 52902a7a6963SBarry Smith ierr = MatCreateVecs(coarse_mat,NULL,&((pcbddc->coarse_ksp)->vec_rhs));CHKERRQ(ierr); 5291f347579bSStefano Zampini } 5292b0f5fe93SStefano Zampini } 5293b0f5fe93SStefano Zampini 5294b0f5fe93SStefano Zampini /* compute null space for coarse solver if the benign trick has been requested */ 5295b0f5fe93SStefano Zampini if (pcbddc->benign_null) { 5296b0f5fe93SStefano Zampini 5297b0f5fe93SStefano Zampini ierr = VecSet(pcbddc->vec1_P,0.);CHKERRQ(ierr); 52984f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 52994f1b2e48SStefano Zampini ierr = VecSetValue(pcbddc->vec1_P,pcbddc->local_primal_size-pcbddc->benign_n+i,1.0,INSERT_VALUES);CHKERRQ(ierr); 53004f1b2e48SStefano Zampini } 5301b0f5fe93SStefano Zampini ierr = VecAssemblyBegin(pcbddc->vec1_P);CHKERRQ(ierr); 5302b0f5fe93SStefano Zampini ierr = VecAssemblyEnd(pcbddc->vec1_P);CHKERRQ(ierr); 5303b0f5fe93SStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 5304b0f5fe93SStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 5305b0f5fe93SStefano Zampini if (coarse_mat) { 5306b0f5fe93SStefano Zampini Vec nullv; 5307b0f5fe93SStefano Zampini PetscScalar *array,*array2; 5308b0f5fe93SStefano Zampini PetscInt nl; 5309b0f5fe93SStefano Zampini 5310b0f5fe93SStefano Zampini ierr = MatCreateVecs(coarse_mat,&nullv,NULL);CHKERRQ(ierr); 5311b0f5fe93SStefano Zampini ierr = VecGetLocalSize(nullv,&nl);CHKERRQ(ierr); 5312b0f5fe93SStefano Zampini ierr = VecGetArrayRead(pcbddc->coarse_vec,(const PetscScalar**)&array);CHKERRQ(ierr); 5313b0f5fe93SStefano Zampini ierr = VecGetArray(nullv,&array2);CHKERRQ(ierr); 5314b0f5fe93SStefano Zampini ierr = PetscMemcpy(array2,array,nl*sizeof(*array));CHKERRQ(ierr); 5315b0f5fe93SStefano Zampini ierr = VecRestoreArray(nullv,&array2);CHKERRQ(ierr); 5316b0f5fe93SStefano Zampini ierr = VecRestoreArrayRead(pcbddc->coarse_vec,(const PetscScalar**)&array);CHKERRQ(ierr); 5317b0f5fe93SStefano Zampini ierr = VecNormalize(nullv,NULL);CHKERRQ(ierr); 5318b0f5fe93SStefano Zampini ierr = MatNullSpaceCreate(PetscObjectComm((PetscObject)coarse_mat),PETSC_FALSE,1,&nullv,&CoarseNullSpace);CHKERRQ(ierr); 5319b0f5fe93SStefano Zampini ierr = VecDestroy(&nullv);CHKERRQ(ierr); 5320b0f5fe93SStefano Zampini } 5321b0f5fe93SStefano Zampini } 5322b0f5fe93SStefano Zampini 5323b0f5fe93SStefano Zampini if (pcbddc->coarse_ksp) { 5324b0f5fe93SStefano Zampini PetscBool ispreonly; 5325b0f5fe93SStefano Zampini 5326b0f5fe93SStefano Zampini if (CoarseNullSpace) { 5327b0f5fe93SStefano Zampini PetscBool isnull; 5328b0f5fe93SStefano Zampini ierr = MatNullSpaceTest(CoarseNullSpace,coarse_mat,&isnull);CHKERRQ(ierr); 5329b0f5fe93SStefano Zampini if (isnull) { 533030368db7SStefano Zampini if (isbddc && !pcbddc->benign_saddle_point) { 5331b0f5fe93SStefano Zampini ierr = PCBDDCSetNullSpace(pc_temp,CoarseNullSpace);CHKERRQ(ierr); 5332b0f5fe93SStefano Zampini } else { 5333b0f5fe93SStefano Zampini ierr = MatSetNullSpace(coarse_mat,CoarseNullSpace);CHKERRQ(ierr); 5334b0f5fe93SStefano Zampini } 5335b0f5fe93SStefano Zampini } else { 5336b0f5fe93SStefano Zampini ierr = MatNullSpaceDestroy(&CoarseNullSpace);CHKERRQ(ierr); 5337b0f5fe93SStefano Zampini } 5338b0f5fe93SStefano Zampini } 5339b0f5fe93SStefano Zampini /* setup coarse ksp */ 5340b0f5fe93SStefano Zampini ierr = KSPSetUp(pcbddc->coarse_ksp);CHKERRQ(ierr); 5341cbcc2c2aSStefano Zampini /* Check coarse problem if in debug mode or if solving with an iterative method */ 5342cbcc2c2aSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcbddc->coarse_ksp,KSPPREONLY,&ispreonly);CHKERRQ(ierr); 53436e683305SStefano Zampini if (pcbddc->dbg_flag || (!ispreonly && pcbddc->use_coarse_estimates) ) { 5344c8587f34SStefano Zampini KSP check_ksp; 53452b510759SStefano Zampini KSPType check_ksp_type; 5346c8587f34SStefano Zampini PC check_pc; 53476e683305SStefano Zampini Vec check_vec,coarse_vec; 53486a1308c2SStefano Zampini PetscReal abs_infty_error,infty_error,lambda_min=1.0,lambda_max=1.0; 53492b510759SStefano Zampini PetscInt its; 53506e683305SStefano Zampini PetscBool compute_eigs; 53516e683305SStefano Zampini PetscReal *eigs_r,*eigs_c; 53526e683305SStefano Zampini PetscInt neigs; 53538e185a42SStefano Zampini const char *prefix; 5354c8587f34SStefano Zampini 53552b510759SStefano Zampini /* Create ksp object suitable for estimation of extreme eigenvalues */ 53566e683305SStefano Zampini ierr = KSPCreate(PetscObjectComm((PetscObject)pcbddc->coarse_ksp),&check_ksp);CHKERRQ(ierr); 5357422a814eSBarry Smith ierr = KSPSetErrorIfNotConverged(pcbddc->coarse_ksp,pc->erroriffailure);CHKERRQ(ierr); 535823ee1639SBarry Smith ierr = KSPSetOperators(check_ksp,coarse_mat,coarse_mat);CHKERRQ(ierr); 5359f4ddd8eeSStefano Zampini ierr = KSPSetTolerances(check_ksp,1.e-12,1.e-12,PETSC_DEFAULT,pcbddc->coarse_size);CHKERRQ(ierr); 53602b510759SStefano Zampini if (ispreonly) { 53612b510759SStefano Zampini check_ksp_type = KSPPREONLY; 53626e683305SStefano Zampini compute_eigs = PETSC_FALSE; 53632b510759SStefano Zampini } else { 5364cbcc2c2aSStefano Zampini check_ksp_type = KSPGMRES; 53656e683305SStefano Zampini compute_eigs = PETSC_TRUE; 5366c8587f34SStefano Zampini } 5367c8587f34SStefano Zampini ierr = KSPSetType(check_ksp,check_ksp_type);CHKERRQ(ierr); 53686e683305SStefano Zampini ierr = KSPSetComputeSingularValues(check_ksp,compute_eigs);CHKERRQ(ierr); 53696e683305SStefano Zampini ierr = KSPSetComputeEigenvalues(check_ksp,compute_eigs);CHKERRQ(ierr); 53706e683305SStefano Zampini ierr = KSPGMRESSetRestart(check_ksp,pcbddc->coarse_size+1);CHKERRQ(ierr); 5371a7dc3881SStefano Zampini ierr = KSPGetOptionsPrefix(pcbddc->coarse_ksp,&prefix);CHKERRQ(ierr); 5372a7dc3881SStefano Zampini ierr = KSPSetOptionsPrefix(check_ksp,prefix);CHKERRQ(ierr); 5373a7dc3881SStefano Zampini ierr = KSPAppendOptionsPrefix(check_ksp,"check_");CHKERRQ(ierr); 5374a7dc3881SStefano Zampini ierr = KSPSetFromOptions(check_ksp);CHKERRQ(ierr); 5375c8587f34SStefano Zampini ierr = KSPSetUp(check_ksp);CHKERRQ(ierr); 5376c8587f34SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&check_pc);CHKERRQ(ierr); 5377c8587f34SStefano Zampini ierr = KSPSetPC(check_ksp,check_pc);CHKERRQ(ierr); 5378c8587f34SStefano Zampini /* create random vec */ 53796e683305SStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&coarse_vec);CHKERRQ(ierr); 53806e683305SStefano Zampini ierr = VecDuplicate(coarse_vec,&check_vec);CHKERRQ(ierr); 5381c8587f34SStefano Zampini ierr = VecSetRandom(check_vec,NULL);CHKERRQ(ierr); 5382c8587f34SStefano Zampini if (CoarseNullSpace) { 5383c8587f34SStefano Zampini ierr = MatNullSpaceRemove(CoarseNullSpace,check_vec);CHKERRQ(ierr); 5384c8587f34SStefano Zampini } 53856e683305SStefano Zampini ierr = MatMult(coarse_mat,check_vec,coarse_vec);CHKERRQ(ierr); 5386c8587f34SStefano Zampini /* solve coarse problem */ 53876e683305SStefano Zampini ierr = KSPSolve(check_ksp,coarse_vec,coarse_vec);CHKERRQ(ierr); 5388c8587f34SStefano Zampini if (CoarseNullSpace) { 53896e683305SStefano Zampini ierr = MatNullSpaceRemove(CoarseNullSpace,coarse_vec);CHKERRQ(ierr); 5390c8587f34SStefano Zampini } 5391cbcc2c2aSStefano Zampini /* set eigenvalue estimation if preonly has not been requested */ 53926e683305SStefano Zampini if (compute_eigs) { 5393854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->coarse_size+1,&eigs_r);CHKERRQ(ierr); 5394854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->coarse_size+1,&eigs_c);CHKERRQ(ierr); 53956e683305SStefano Zampini ierr = KSPComputeEigenvalues(check_ksp,pcbddc->coarse_size+1,eigs_r,eigs_c,&neigs);CHKERRQ(ierr); 53966e683305SStefano Zampini lambda_max = eigs_r[neigs-1]; 53976e683305SStefano Zampini lambda_min = eigs_r[0]; 53986e683305SStefano Zampini if (pcbddc->use_coarse_estimates) { 53996e683305SStefano Zampini if (lambda_max>lambda_min) { 5400cbcc2c2aSStefano Zampini ierr = KSPChebyshevSetEigenvalues(pcbddc->coarse_ksp,lambda_max,lambda_min);CHKERRQ(ierr); 5401cbcc2c2aSStefano Zampini ierr = KSPRichardsonSetScale(pcbddc->coarse_ksp,2.0/(lambda_max+lambda_min));CHKERRQ(ierr); 5402cbcc2c2aSStefano Zampini } 5403c8587f34SStefano Zampini } 5404c8587f34SStefano Zampini } 5405cbcc2c2aSStefano Zampini 5406c8587f34SStefano Zampini /* check coarse problem residual error */ 54076e683305SStefano Zampini if (pcbddc->dbg_flag) { 54086e683305SStefano Zampini PetscViewer dbg_viewer = PETSC_VIEWER_STDOUT_(PetscObjectComm((PetscObject)pcbddc->coarse_ksp)); 54096e683305SStefano Zampini ierr = PetscViewerASCIIAddTab(dbg_viewer,2*(pcbddc->current_level+1));CHKERRQ(ierr); 54106e683305SStefano Zampini ierr = VecAXPY(check_vec,-1.0,coarse_vec);CHKERRQ(ierr); 5411c8587f34SStefano Zampini ierr = VecNorm(check_vec,NORM_INFINITY,&infty_error);CHKERRQ(ierr); 54126e683305SStefano Zampini ierr = MatMult(coarse_mat,check_vec,coarse_vec);CHKERRQ(ierr); 54136e683305SStefano Zampini ierr = VecNorm(coarse_vec,NORM_INFINITY,&abs_infty_error);CHKERRQ(ierr); 5414c8587f34SStefano Zampini ierr = VecDestroy(&check_vec);CHKERRQ(ierr); 5415779c1cceSStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem details (use estimates %d)\n",pcbddc->use_coarse_estimates);CHKERRQ(ierr); 54166e683305SStefano Zampini ierr = PetscObjectPrintClassNamePrefixType((PetscObject)(pcbddc->coarse_ksp),dbg_viewer);CHKERRQ(ierr); 54176e683305SStefano Zampini ierr = PetscObjectPrintClassNamePrefixType((PetscObject)(check_pc),dbg_viewer);CHKERRQ(ierr); 54186e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem exact infty_error : %1.6e\n",infty_error);CHKERRQ(ierr); 54196e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem residual infty_error: %1.6e\n",abs_infty_error);CHKERRQ(ierr); 5420b0f5fe93SStefano Zampini if (CoarseNullSpace) { 5421b0f5fe93SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem is singular\n");CHKERRQ(ierr); 5422b0f5fe93SStefano Zampini } 54236e683305SStefano Zampini if (compute_eigs) { 54246e683305SStefano Zampini PetscReal lambda_max_s,lambda_min_s; 5425deec49d1SStefano Zampini ierr = KSPGetType(check_ksp,&check_ksp_type);CHKERRQ(ierr); 5426c8587f34SStefano Zampini ierr = KSPGetIterationNumber(check_ksp,&its);CHKERRQ(ierr); 54276e683305SStefano Zampini ierr = KSPComputeExtremeSingularValues(check_ksp,&lambda_max_s,&lambda_min_s);CHKERRQ(ierr); 54286e683305SStefano 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); 54296e683305SStefano Zampini for (i=0;i<neigs;i++) { 54306e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"%1.6e %1.6ei\n",eigs_r[i],eigs_c[i]);CHKERRQ(ierr); 5431c8587f34SStefano Zampini } 54326e683305SStefano Zampini } 54336e683305SStefano Zampini ierr = PetscViewerFlush(dbg_viewer);CHKERRQ(ierr); 54346e683305SStefano Zampini ierr = PetscViewerASCIISubtractTab(dbg_viewer,2*(pcbddc->current_level+1));CHKERRQ(ierr); 54356e683305SStefano Zampini } 5436c8587f34SStefano Zampini ierr = KSPDestroy(&check_ksp);CHKERRQ(ierr); 54376e683305SStefano Zampini if (compute_eigs) { 54386e683305SStefano Zampini ierr = PetscFree(eigs_r);CHKERRQ(ierr); 54396e683305SStefano Zampini ierr = PetscFree(eigs_c);CHKERRQ(ierr); 5440c8587f34SStefano Zampini } 54416e683305SStefano Zampini } 54426e683305SStefano Zampini } 5443cbcc2c2aSStefano Zampini /* print additional info */ 5444cbcc2c2aSStefano Zampini if (pcbddc->dbg_flag) { 54456e683305SStefano Zampini /* waits until all processes reaches this point */ 54466e683305SStefano Zampini ierr = PetscBarrier((PetscObject)pc);CHKERRQ(ierr); 5447cbcc2c2aSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Coarse solver setup completed at level %d\n",pcbddc->current_level);CHKERRQ(ierr); 5448cbcc2c2aSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 5449cbcc2c2aSStefano Zampini } 5450cbcc2c2aSStefano Zampini 54512b510759SStefano Zampini /* free memory */ 5452c8587f34SStefano Zampini ierr = MatNullSpaceDestroy(&CoarseNullSpace);CHKERRQ(ierr); 5453fdc635d7SStefano Zampini ierr = MatDestroy(&coarse_mat);CHKERRQ(ierr); 5454c8587f34SStefano Zampini PetscFunctionReturn(0); 5455c8587f34SStefano Zampini } 5456674ae819SStefano Zampini 5457f34684f1SStefano Zampini #undef __FUNCT__ 5458f34684f1SStefano Zampini #define __FUNCT__ "PCBDDCComputePrimalNumbering" 5459f34684f1SStefano Zampini PetscErrorCode PCBDDCComputePrimalNumbering(PC pc,PetscInt* coarse_size_n,PetscInt** local_primal_indices_n) 5460f34684f1SStefano Zampini { 5461f34684f1SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 5462f34684f1SStefano Zampini PC_IS* pcis = (PC_IS*)pc->data; 5463f34684f1SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 5464dc456d91SStefano Zampini IS subset,subset_mult,subset_n; 5465dc456d91SStefano Zampini PetscInt local_size,coarse_size=0; 546673be2a3aSStefano Zampini PetscInt *local_primal_indices=NULL; 5467dc456d91SStefano Zampini const PetscInt *t_local_primal_indices; 5468f34684f1SStefano Zampini PetscErrorCode ierr; 5469f34684f1SStefano Zampini 5470f34684f1SStefano Zampini PetscFunctionBegin; 5471f34684f1SStefano Zampini /* Compute global number of coarse dofs */ 54720e6343abSStefano Zampini if (pcbddc->local_primal_size && !pcbddc->local_primal_ref_node) { 54730e6343abSStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"BDDC ConstraintsSetUp should be called first"); 5474727cdba6SStefano Zampini } 5475dc456d91SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)(pc->pmat)),pcbddc->local_primal_size_cc,pcbddc->local_primal_ref_node,PETSC_COPY_VALUES,&subset_n);CHKERRQ(ierr); 54763bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(pcis->mapping,subset_n,&subset);CHKERRQ(ierr); 5477dc456d91SStefano Zampini ierr = ISDestroy(&subset_n);CHKERRQ(ierr); 5478dc456d91SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)(pc->pmat)),pcbddc->local_primal_size_cc,pcbddc->local_primal_ref_mult,PETSC_COPY_VALUES,&subset_mult);CHKERRQ(ierr); 5479dc456d91SStefano Zampini ierr = PCBDDCSubsetNumbering(subset,subset_mult,&coarse_size,&subset_n);CHKERRQ(ierr); 5480dc456d91SStefano Zampini ierr = ISDestroy(&subset);CHKERRQ(ierr); 5481dc456d91SStefano Zampini ierr = ISDestroy(&subset_mult);CHKERRQ(ierr); 5482dc456d91SStefano Zampini ierr = ISGetLocalSize(subset_n,&local_size);CHKERRQ(ierr); 54830e6343abSStefano Zampini if (local_size != pcbddc->local_primal_size) { 54840e6343abSStefano Zampini SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Invalid number of local primal indices computed %d != %d",local_size,pcbddc->local_primal_size); 54850e6343abSStefano Zampini } 5486dc456d91SStefano Zampini ierr = PetscMalloc1(local_size,&local_primal_indices);CHKERRQ(ierr); 5487dc456d91SStefano Zampini ierr = ISGetIndices(subset_n,&t_local_primal_indices);CHKERRQ(ierr); 5488dc456d91SStefano Zampini ierr = PetscMemcpy(local_primal_indices,t_local_primal_indices,local_size*sizeof(PetscInt));CHKERRQ(ierr); 5489dc456d91SStefano Zampini ierr = ISRestoreIndices(subset_n,&t_local_primal_indices);CHKERRQ(ierr); 5490dc456d91SStefano Zampini ierr = ISDestroy(&subset_n);CHKERRQ(ierr); 5491f34684f1SStefano Zampini 5492f34684f1SStefano Zampini /* check numbering */ 5493f34684f1SStefano Zampini if (pcbddc->dbg_flag) { 5494019a44ceSStefano Zampini PetscScalar coarsesum,*array,*array2; 5495dc456d91SStefano Zampini PetscInt i; 5496b9b85e73SStefano Zampini PetscBool set_error = PETSC_FALSE,set_error_reduced = PETSC_FALSE; 5497f34684f1SStefano Zampini 5498f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 5499f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 5500f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Check coarse indices\n");CHKERRQ(ierr); 55011575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 5502019a44ceSStefano Zampini /* counter */ 5503019a44ceSStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 5504019a44ceSStefano Zampini ierr = VecSet(pcis->vec1_N,1.0);CHKERRQ(ierr); 5505019a44ceSStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 5506019a44ceSStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 5507019a44ceSStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_global,pcis->vec2_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 5508019a44ceSStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_global,pcis->vec2_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 5509f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_N,0.0);CHKERRQ(ierr); 5510f34684f1SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 5511727cdba6SStefano Zampini ierr = VecSetValue(pcis->vec1_N,pcbddc->primal_indices_local_idxs[i],1.0,INSERT_VALUES);CHKERRQ(ierr); 5512f34684f1SStefano Zampini } 5513f34684f1SStefano Zampini ierr = VecAssemblyBegin(pcis->vec1_N);CHKERRQ(ierr); 5514f34684f1SStefano Zampini ierr = VecAssemblyEnd(pcis->vec1_N);CHKERRQ(ierr); 5515f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 5516e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 5517e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 5518e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_global,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 5519e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_global,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 5520f34684f1SStefano Zampini ierr = VecGetArray(pcis->vec1_N,&array);CHKERRQ(ierr); 5521019a44ceSStefano Zampini ierr = VecGetArray(pcis->vec2_N,&array2);CHKERRQ(ierr); 5522f34684f1SStefano Zampini for (i=0;i<pcis->n;i++) { 5523019a44ceSStefano Zampini if (array[i] != 0.0 && array[i] != array2[i]) { 552475c01103SStefano Zampini PetscInt owned = (PetscInt)PetscRealPart(array[i]); 552575c01103SStefano Zampini PetscInt neigh = (PetscInt)PetscRealPart(array2[i]); 5526b9b85e73SStefano Zampini set_error = PETSC_TRUE; 5527019a44ceSStefano 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); 5528f34684f1SStefano Zampini } 5529f34684f1SStefano Zampini } 5530019a44ceSStefano Zampini ierr = VecRestoreArray(pcis->vec2_N,&array2);CHKERRQ(ierr); 5531b9b85e73SStefano Zampini ierr = MPI_Allreduce(&set_error,&set_error_reduced,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 5532f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 5533f34684f1SStefano Zampini for (i=0;i<pcis->n;i++) { 5534f34684f1SStefano Zampini if (PetscRealPart(array[i]) > 0.0) array[i] = 1.0/PetscRealPart(array[i]); 5535f34684f1SStefano Zampini } 5536f34684f1SStefano Zampini ierr = VecRestoreArray(pcis->vec1_N,&array);CHKERRQ(ierr); 5537f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 5538e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 5539e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 5540f34684f1SStefano Zampini ierr = VecSum(pcis->vec1_global,&coarsesum);CHKERRQ(ierr); 5541f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Size of coarse problem is %d (%lf)\n",coarse_size,PetscRealPart(coarsesum));CHKERRQ(ierr); 5542b9b85e73SStefano Zampini if (pcbddc->dbg_flag > 1 || set_error_reduced) { 5543ca8b9ea9SStefano Zampini PetscInt *gidxs; 5544ca8b9ea9SStefano Zampini 5545ca8b9ea9SStefano Zampini ierr = PetscMalloc1(pcbddc->local_primal_size,&gidxs);CHKERRQ(ierr); 55463bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingApply(pcis->mapping,pcbddc->local_primal_size,pcbddc->primal_indices_local_idxs,gidxs);CHKERRQ(ierr); 5547f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Distribution of local primal indices\n");CHKERRQ(ierr); 5548f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 5549f34684f1SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d\n",PetscGlobalRank);CHKERRQ(ierr); 5550f34684f1SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 55514bc2dc4bSStefano 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); 5552f34684f1SStefano Zampini } 5553f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 5554ca8b9ea9SStefano Zampini ierr = PetscFree(gidxs);CHKERRQ(ierr); 5555f34684f1SStefano Zampini } 5556f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 55571575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 5558302440fdSBarry Smith if (set_error_reduced) SETERRQ(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"BDDC Numbering of coarse dofs failed"); 5559f34684f1SStefano Zampini } 55608bec7fa6SStefano Zampini /* ierr = PetscPrintf(PetscObjectComm((PetscObject)pc),"Size of coarse problem is %d\n",coarse_size);CHKERRQ(ierr); */ 5561f34684f1SStefano Zampini /* get back data */ 5562f34684f1SStefano Zampini *coarse_size_n = coarse_size; 5563f34684f1SStefano Zampini *local_primal_indices_n = local_primal_indices; 5564674ae819SStefano Zampini PetscFunctionReturn(0); 5565674ae819SStefano Zampini } 5566674ae819SStefano Zampini 5567e456f2a8SStefano Zampini #undef __FUNCT__ 5568e456f2a8SStefano Zampini #define __FUNCT__ "PCBDDCGlobalToLocal" 5569a7dc3881SStefano Zampini PetscErrorCode PCBDDCGlobalToLocal(VecScatter g2l_ctx,Vec gwork, Vec lwork, IS globalis, IS* localis) 5570e456f2a8SStefano Zampini { 5571e456f2a8SStefano Zampini IS localis_t; 5572a7dc3881SStefano Zampini PetscInt i,lsize,*idxs,n; 5573e456f2a8SStefano Zampini PetscScalar *vals; 5574e456f2a8SStefano Zampini PetscErrorCode ierr; 5575e456f2a8SStefano Zampini 5576e456f2a8SStefano Zampini PetscFunctionBegin; 5577a7dc3881SStefano Zampini /* get indices in local ordering exploiting local to global map */ 5578e456f2a8SStefano Zampini ierr = ISGetLocalSize(globalis,&lsize);CHKERRQ(ierr); 5579854ce69bSBarry Smith ierr = PetscMalloc1(lsize,&vals);CHKERRQ(ierr); 5580e456f2a8SStefano Zampini for (i=0;i<lsize;i++) vals[i] = 1.0; 5581e456f2a8SStefano Zampini ierr = ISGetIndices(globalis,(const PetscInt**)&idxs);CHKERRQ(ierr); 5582a7dc3881SStefano Zampini ierr = VecSet(gwork,0.0);CHKERRQ(ierr); 5583a7dc3881SStefano Zampini ierr = VecSet(lwork,0.0);CHKERRQ(ierr); 55841035eff8SStefano Zampini if (idxs) { /* multilevel guard */ 5585a7dc3881SStefano Zampini ierr = VecSetValues(gwork,lsize,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 55861035eff8SStefano Zampini } 5587a7dc3881SStefano Zampini ierr = VecAssemblyBegin(gwork);CHKERRQ(ierr); 5588e456f2a8SStefano Zampini ierr = ISRestoreIndices(globalis,(const PetscInt**)&idxs);CHKERRQ(ierr); 5589e456f2a8SStefano Zampini ierr = PetscFree(vals);CHKERRQ(ierr); 5590a7dc3881SStefano Zampini ierr = VecAssemblyEnd(gwork);CHKERRQ(ierr); 5591a7dc3881SStefano Zampini /* now compute set in local ordering */ 5592a7dc3881SStefano Zampini ierr = VecScatterBegin(g2l_ctx,gwork,lwork,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 5593a7dc3881SStefano Zampini ierr = VecScatterEnd(g2l_ctx,gwork,lwork,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 5594a7dc3881SStefano Zampini ierr = VecGetArrayRead(lwork,(const PetscScalar**)&vals);CHKERRQ(ierr); 5595a7dc3881SStefano Zampini ierr = VecGetSize(lwork,&n);CHKERRQ(ierr); 5596a7dc3881SStefano Zampini for (i=0,lsize=0;i<n;i++) { 5597ff92baa0SMatthew G. Knepley if (PetscRealPart(vals[i]) > 0.5) { 5598e456f2a8SStefano Zampini lsize++; 5599e456f2a8SStefano Zampini } 5600e456f2a8SStefano Zampini } 5601854ce69bSBarry Smith ierr = PetscMalloc1(lsize,&idxs);CHKERRQ(ierr); 5602a7dc3881SStefano Zampini for (i=0,lsize=0;i<n;i++) { 5603ff92baa0SMatthew G. Knepley if (PetscRealPart(vals[i]) > 0.5) { 5604e456f2a8SStefano Zampini idxs[lsize++] = i; 5605e456f2a8SStefano Zampini } 5606e456f2a8SStefano Zampini } 5607a7dc3881SStefano Zampini ierr = VecRestoreArrayRead(lwork,(const PetscScalar**)&vals);CHKERRQ(ierr); 5608a7dc3881SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)gwork),lsize,idxs,PETSC_OWN_POINTER,&localis_t);CHKERRQ(ierr); 5609e456f2a8SStefano Zampini *localis = localis_t; 5610e456f2a8SStefano Zampini PetscFunctionReturn(0); 5611e456f2a8SStefano Zampini } 5612906d46d4SStefano Zampini 5613906d46d4SStefano Zampini /* the next two functions will be called in KSPMatMult if a change of basis has been requested */ 5614906d46d4SStefano Zampini #undef __FUNCT__ 5615906d46d4SStefano Zampini #define __FUNCT__ "PCBDDCMatMult_Private" 5616906d46d4SStefano Zampini static PetscErrorCode PCBDDCMatMult_Private(Mat A, Vec x, Vec y) 5617906d46d4SStefano Zampini { 5618906d46d4SStefano Zampini PCBDDCChange_ctx change_ctx; 5619906d46d4SStefano Zampini PetscErrorCode ierr; 5620906d46d4SStefano Zampini 5621906d46d4SStefano Zampini PetscFunctionBegin; 5622906d46d4SStefano Zampini ierr = MatShellGetContext(A,&change_ctx);CHKERRQ(ierr); 5623906d46d4SStefano Zampini ierr = MatMult(change_ctx->global_change,x,change_ctx->work[0]);CHKERRQ(ierr); 5624906d46d4SStefano Zampini ierr = MatMult(change_ctx->original_mat,change_ctx->work[0],change_ctx->work[1]);CHKERRQ(ierr); 5625906d46d4SStefano Zampini ierr = MatMultTranspose(change_ctx->global_change,change_ctx->work[1],y);CHKERRQ(ierr); 5626906d46d4SStefano Zampini PetscFunctionReturn(0); 5627906d46d4SStefano Zampini } 5628906d46d4SStefano Zampini 5629906d46d4SStefano Zampini #undef __FUNCT__ 5630906d46d4SStefano Zampini #define __FUNCT__ "PCBDDCMatMultTranspose_Private" 5631906d46d4SStefano Zampini static PetscErrorCode PCBDDCMatMultTranspose_Private(Mat A, Vec x, Vec y) 5632906d46d4SStefano Zampini { 5633906d46d4SStefano Zampini PCBDDCChange_ctx change_ctx; 5634906d46d4SStefano Zampini PetscErrorCode ierr; 5635906d46d4SStefano Zampini 5636906d46d4SStefano Zampini PetscFunctionBegin; 5637906d46d4SStefano Zampini ierr = MatShellGetContext(A,&change_ctx);CHKERRQ(ierr); 5638906d46d4SStefano Zampini ierr = MatMult(change_ctx->global_change,x,change_ctx->work[0]);CHKERRQ(ierr); 5639906d46d4SStefano Zampini ierr = MatMultTranspose(change_ctx->original_mat,change_ctx->work[0],change_ctx->work[1]);CHKERRQ(ierr); 5640906d46d4SStefano Zampini ierr = MatMultTranspose(change_ctx->global_change,change_ctx->work[1],y);CHKERRQ(ierr); 5641906d46d4SStefano Zampini PetscFunctionReturn(0); 5642906d46d4SStefano Zampini } 5643b96c3477SStefano Zampini 5644b96c3477SStefano Zampini #undef __FUNCT__ 5645b96c3477SStefano Zampini #define __FUNCT__ "PCBDDCSetUpSubSchurs" 564608122e43SStefano Zampini PetscErrorCode PCBDDCSetUpSubSchurs(PC pc) 5647b96c3477SStefano Zampini { 5648a64f4aa4SStefano Zampini PC_IS *pcis=(PC_IS*)pc->data; 5649b96c3477SStefano Zampini PC_BDDC *pcbddc=(PC_BDDC*)pc->data; 5650b96c3477SStefano Zampini PCBDDCSubSchurs sub_schurs=pcbddc->sub_schurs; 5651a64f4aa4SStefano Zampini Mat S_j; 5652b96c3477SStefano Zampini PetscInt *used_xadj,*used_adjncy; 5653b96c3477SStefano Zampini PetscBool free_used_adj; 5654b96c3477SStefano Zampini PetscErrorCode ierr; 5655b96c3477SStefano Zampini 5656b96c3477SStefano Zampini PetscFunctionBegin; 5657b96c3477SStefano Zampini /* decide the adjacency to be used for determining internal problems for local schur on subsets */ 5658b96c3477SStefano Zampini free_used_adj = PETSC_FALSE; 565908122e43SStefano Zampini if (pcbddc->sub_schurs_layers == -1) { 5660b96c3477SStefano Zampini used_xadj = NULL; 5661b96c3477SStefano Zampini used_adjncy = NULL; 5662b96c3477SStefano Zampini } else { 566308122e43SStefano Zampini if (pcbddc->sub_schurs_use_useradj && pcbddc->mat_graph->xadj) { 566408122e43SStefano Zampini used_xadj = pcbddc->mat_graph->xadj; 566508122e43SStefano Zampini used_adjncy = pcbddc->mat_graph->adjncy; 566608122e43SStefano Zampini } else if (pcbddc->computed_rowadj) { 5667b96c3477SStefano Zampini used_xadj = pcbddc->mat_graph->xadj; 5668b96c3477SStefano Zampini used_adjncy = pcbddc->mat_graph->adjncy; 5669b96c3477SStefano Zampini } else { 56702fffb893SStefano Zampini PetscBool flg_row=PETSC_FALSE; 5671b96c3477SStefano Zampini const PetscInt *xadj,*adjncy; 5672b96c3477SStefano Zampini PetscInt nvtxs; 5673b96c3477SStefano Zampini 56742fffb893SStefano Zampini ierr = MatGetRowIJ(pcbddc->local_mat,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 56752fffb893SStefano Zampini if (flg_row) { 5676b96c3477SStefano Zampini ierr = PetscMalloc2(nvtxs+1,&used_xadj,xadj[nvtxs],&used_adjncy);CHKERRQ(ierr); 5677b96c3477SStefano Zampini ierr = PetscMemcpy(used_xadj,xadj,(nvtxs+1)*sizeof(*xadj));CHKERRQ(ierr); 5678b96c3477SStefano Zampini ierr = PetscMemcpy(used_adjncy,adjncy,(xadj[nvtxs])*sizeof(*adjncy));CHKERRQ(ierr); 5679b96c3477SStefano Zampini free_used_adj = PETSC_TRUE; 56802fffb893SStefano Zampini } else { 56812fffb893SStefano Zampini pcbddc->sub_schurs_layers = -1; 56822fffb893SStefano Zampini used_xadj = NULL; 56832fffb893SStefano Zampini used_adjncy = NULL; 56842fffb893SStefano Zampini } 56852fffb893SStefano Zampini ierr = MatRestoreRowIJ(pcbddc->local_mat,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 5686b96c3477SStefano Zampini } 5687b96c3477SStefano Zampini } 5688d5574798SStefano Zampini 5689d5574798SStefano Zampini /* setup sub_schurs data */ 5690a64f4aa4SStefano Zampini ierr = MatCreateSchurComplement(pcis->A_II,pcis->A_II,pcis->A_IB,pcis->A_BI,pcis->A_BB,&S_j);CHKERRQ(ierr); 5691a64f4aa4SStefano Zampini if (!sub_schurs->use_mumps) { 5692a64f4aa4SStefano Zampini /* pcbddc->ksp_D up to date only if not using MUMPS */ 5693a64f4aa4SStefano Zampini ierr = MatSchurComplementSetKSP(S_j,pcbddc->ksp_D);CHKERRQ(ierr); 5694*ca92afb2SStefano Zampini ierr = PCBDDCSubSchursSetUp(sub_schurs,NULL,S_j,PETSC_FALSE,used_xadj,used_adjncy,pcbddc->sub_schurs_layers,pcbddc->faster_deluxe,pcbddc->adaptive_selection,PETSC_FALSE,PETSC_FALSE,0,NULL,NULL);CHKERRQ(ierr); 5695a64f4aa4SStefano Zampini } else { 56966816873aSStefano Zampini PetscBool reuse_solvers = (PetscBool)!pcbddc->use_change_of_basis; 569704708bb6SStefano Zampini PetscBool isseqaij; 5698*ca92afb2SStefano Zampini PetscInt i; 56995feab87aSStefano Zampini if (!pcbddc->use_vertices && reuse_solvers) { 57005feab87aSStefano Zampini PetscInt n_vertices; 57015feab87aSStefano Zampini 57025feab87aSStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_vertices,&n_vertices);CHKERRQ(ierr); 57032034aafcSStefano Zampini reuse_solvers = (PetscBool)!n_vertices; 57045feab87aSStefano Zampini } 570504708bb6SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcbddc->local_mat,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 570604708bb6SStefano Zampini if (!isseqaij) { 570704708bb6SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 570804708bb6SStefano Zampini if (matis->A == pcbddc->local_mat) { 570904708bb6SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 571004708bb6SStefano Zampini ierr = MatConvert(matis->A,MATSEQAIJ,MAT_INITIAL_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 571104708bb6SStefano Zampini } else { 571204708bb6SStefano Zampini ierr = MatConvert(pcbddc->local_mat,MATSEQAIJ,MAT_REUSE_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 571304708bb6SStefano Zampini } 571404708bb6SStefano Zampini } 5715*ca92afb2SStefano Zampini if (!pcbddc->benign_n) { 5716*ca92afb2SStefano Zampini ierr = PCBDDCSubSchursSetUp(sub_schurs,pcbddc->local_mat,S_j,pcbddc->sub_schurs_exact_schur,used_xadj,used_adjncy,pcbddc->sub_schurs_layers,pcbddc->faster_deluxe,pcbddc->adaptive_selection,reuse_solvers,pcbddc->benign_saddle_point,0,NULL,NULL);CHKERRQ(ierr); 5717*ca92afb2SStefano Zampini } else { 5718*ca92afb2SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 5719*ca92afb2SStefano Zampini ierr = PCBDDCSubSchursSetUp(sub_schurs,matis->A,S_j,pcbddc->sub_schurs_exact_schur,used_xadj,used_adjncy,pcbddc->sub_schurs_layers,pcbddc->faster_deluxe,pcbddc->adaptive_selection,reuse_solvers,pcbddc->benign_saddle_point,pcbddc->benign_n,pcbddc->benign_p0_lidx,pcbddc->benign_zerodiag_subs);CHKERRQ(ierr); 5720*ca92afb2SStefano Zampini } 5721*ca92afb2SStefano Zampini /* we no longer need these index sets */ 5722*ca92afb2SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 5723*ca92afb2SStefano Zampini ierr = ISDestroy(&pcbddc->benign_zerodiag_subs[i]);CHKERRQ(ierr); 5724*ca92afb2SStefano Zampini } 5725*ca92afb2SStefano Zampini ierr = PetscFree(pcbddc->benign_zerodiag_subs);CHKERRQ(ierr); 5726a64f4aa4SStefano Zampini } 5727a64f4aa4SStefano Zampini ierr = MatDestroy(&S_j);CHKERRQ(ierr); 5728b96c3477SStefano Zampini 5729b96c3477SStefano Zampini /* free adjacency */ 5730b96c3477SStefano Zampini if (free_used_adj) { 5731b96c3477SStefano Zampini ierr = PetscFree2(used_xadj,used_adjncy);CHKERRQ(ierr); 5732b96c3477SStefano Zampini } 5733b96c3477SStefano Zampini PetscFunctionReturn(0); 5734b96c3477SStefano Zampini } 5735b96c3477SStefano Zampini 5736b96c3477SStefano Zampini #undef __FUNCT__ 5737b96c3477SStefano Zampini #define __FUNCT__ "PCBDDCInitSubSchurs" 573808122e43SStefano Zampini PetscErrorCode PCBDDCInitSubSchurs(PC pc) 5739b96c3477SStefano Zampini { 5740b96c3477SStefano Zampini PC_IS *pcis=(PC_IS*)pc->data; 5741b96c3477SStefano Zampini PC_BDDC *pcbddc=(PC_BDDC*)pc->data; 5742b96c3477SStefano Zampini PCBDDCSubSchurs sub_schurs=pcbddc->sub_schurs; 5743b96c3477SStefano Zampini PCBDDCGraph graph; 5744b96c3477SStefano Zampini PetscErrorCode ierr; 5745b96c3477SStefano Zampini 5746b96c3477SStefano Zampini PetscFunctionBegin; 5747b96c3477SStefano Zampini /* attach interface graph for determining subsets */ 574808122e43SStefano Zampini if (pcbddc->sub_schurs_rebuild) { /* in case rebuild has been requested, it uses a graph generated only by the neighbouring information */ 57493301b35fSStefano Zampini IS verticesIS,verticescomm; 57503301b35fSStefano Zampini PetscInt vsize,*idxs; 5751b96c3477SStefano Zampini 5752b96c3477SStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,NULL,NULL,NULL,NULL,&verticesIS);CHKERRQ(ierr); 57533301b35fSStefano Zampini ierr = ISGetSize(verticesIS,&vsize);CHKERRQ(ierr); 57543301b35fSStefano Zampini ierr = ISGetIndices(verticesIS,(const PetscInt**)&idxs);CHKERRQ(ierr); 57553301b35fSStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),vsize,idxs,PETSC_COPY_VALUES,&verticescomm);CHKERRQ(ierr); 57563301b35fSStefano Zampini ierr = ISRestoreIndices(verticesIS,(const PetscInt**)&idxs);CHKERRQ(ierr); 57573301b35fSStefano Zampini ierr = ISDestroy(&verticesIS);CHKERRQ(ierr); 5758b96c3477SStefano Zampini ierr = PCBDDCGraphCreate(&graph);CHKERRQ(ierr); 57597fb0e2dbSStefano Zampini ierr = PCBDDCGraphInit(graph,pcbddc->mat_graph->l2gmap,pcbddc->mat_graph->nvtxs_global);CHKERRQ(ierr); 57603301b35fSStefano Zampini ierr = PCBDDCGraphSetUp(graph,0,NULL,pcbddc->DirichletBoundariesLocal,0,NULL,verticescomm);CHKERRQ(ierr); 57613301b35fSStefano Zampini ierr = ISDestroy(&verticescomm);CHKERRQ(ierr); 5762b96c3477SStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(graph);CHKERRQ(ierr); 5763b96c3477SStefano Zampini /* 5764b96c3477SStefano Zampini if (pcbddc->dbg_flag) { 5765b96c3477SStefano Zampini ierr = PCBDDCGraphASCIIView(graph,pcbddc->dbg_flag,pcbddc->dbg_viewer);CHKERRQ(ierr); 5766b96c3477SStefano Zampini } 5767b96c3477SStefano Zampini */ 5768b96c3477SStefano Zampini } else { 5769b96c3477SStefano Zampini graph = pcbddc->mat_graph; 5770b96c3477SStefano Zampini } 5771b96c3477SStefano Zampini 5772b96c3477SStefano Zampini /* sub_schurs init */ 5773a64f4aa4SStefano Zampini ierr = PCBDDCSubSchursInit(sub_schurs,pcis->is_I_local,pcis->is_B_local,graph,pcis->BtoNmap);CHKERRQ(ierr); 5774a64f4aa4SStefano Zampini 5775b96c3477SStefano Zampini /* free graph struct */ 577608122e43SStefano Zampini if (pcbddc->sub_schurs_rebuild) { 5777b96c3477SStefano Zampini ierr = PCBDDCGraphDestroy(&graph);CHKERRQ(ierr); 5778b96c3477SStefano Zampini } 5779b96c3477SStefano Zampini PetscFunctionReturn(0); 5780b96c3477SStefano Zampini } 5781fa34dd3eSStefano Zampini 5782fa34dd3eSStefano Zampini #undef __FUNCT__ 5783fa34dd3eSStefano Zampini #define __FUNCT__ "PCBDDCCheckOperator" 5784fa34dd3eSStefano Zampini PetscErrorCode PCBDDCCheckOperator(PC pc) 5785fa34dd3eSStefano Zampini { 5786fa34dd3eSStefano Zampini PC_IS *pcis=(PC_IS*)pc->data; 5787fa34dd3eSStefano Zampini PC_BDDC *pcbddc=(PC_BDDC*)pc->data; 5788fa34dd3eSStefano Zampini PetscErrorCode ierr; 5789fa34dd3eSStefano Zampini 5790fa34dd3eSStefano Zampini PetscFunctionBegin; 5791fa34dd3eSStefano Zampini if (pcbddc->n_vertices == pcbddc->local_primal_size) { 5792fa34dd3eSStefano Zampini IS zerodiag = NULL; 57934f1b2e48SStefano Zampini Mat S_j,B0_B=NULL; 5794fa34dd3eSStefano Zampini Vec dummy_vec=NULL,vec_check_B,vec_scale_P; 57954f1b2e48SStefano Zampini PetscScalar *p0_check,*array,*array2; 579675c01103SStefano Zampini PetscReal norm; 5797fa34dd3eSStefano Zampini PetscInt i; 5798fa34dd3eSStefano Zampini 5799fa34dd3eSStefano Zampini /* B0 and B0_B */ 5800fa34dd3eSStefano Zampini if (zerodiag) { 5801fa34dd3eSStefano Zampini IS dummy; 5802fa34dd3eSStefano Zampini 58034f1b2e48SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcbddc->benign_n,0,1,&dummy);CHKERRQ(ierr); 58044f1b2e48SStefano Zampini ierr = MatGetSubMatrix(pcbddc->benign_B0,dummy,pcis->is_B_local,MAT_INITIAL_MATRIX,&B0_B);CHKERRQ(ierr); 5805fa34dd3eSStefano Zampini ierr = MatCreateVecs(B0_B,NULL,&dummy_vec);CHKERRQ(ierr); 5806fa34dd3eSStefano Zampini ierr = ISDestroy(&dummy);CHKERRQ(ierr); 5807fa34dd3eSStefano Zampini } 5808fa34dd3eSStefano Zampini /* I need a primal vector to scale primal nodes since BDDC sums contibutions */ 5809fa34dd3eSStefano Zampini ierr = VecDuplicate(pcbddc->vec1_P,&vec_scale_P);CHKERRQ(ierr); 5810fa34dd3eSStefano Zampini ierr = VecSet(pcbddc->vec1_P,1.0);CHKERRQ(ierr); 5811fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 5812fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 5813fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,vec_scale_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 5814fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,vec_scale_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 5815fa34dd3eSStefano Zampini ierr = VecReciprocal(vec_scale_P);CHKERRQ(ierr); 5816fa34dd3eSStefano Zampini /* S_j */ 5817fa34dd3eSStefano Zampini ierr = MatCreateSchurComplement(pcis->A_II,pcis->A_II,pcis->A_IB,pcis->A_BI,pcis->A_BB,&S_j);CHKERRQ(ierr); 5818fa34dd3eSStefano Zampini ierr = MatSchurComplementSetKSP(S_j,pcbddc->ksp_D);CHKERRQ(ierr); 5819fa34dd3eSStefano Zampini 5820fa34dd3eSStefano Zampini /* mimic vector in \widetilde{W}_\Gamma */ 5821fa34dd3eSStefano Zampini ierr = VecSetRandom(pcis->vec1_N,NULL);CHKERRQ(ierr); 5822fa34dd3eSStefano Zampini /* continuous in primal space */ 5823fa34dd3eSStefano Zampini ierr = VecSetRandom(pcbddc->coarse_vec,NULL);CHKERRQ(ierr); 5824fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,pcbddc->vec1_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 5825fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,pcbddc->vec1_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 5826fa34dd3eSStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 58274f1b2e48SStefano Zampini ierr = PetscCalloc1(pcbddc->benign_n,&p0_check);CHKERRQ(ierr); 58284f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) p0_check[i] = array[pcbddc->local_primal_size-pcbddc->benign_n+i]; 5829fa34dd3eSStefano Zampini ierr = VecSetValues(pcis->vec1_N,pcbddc->local_primal_size,pcbddc->local_primal_ref_node,array,INSERT_VALUES);CHKERRQ(ierr); 5830fa34dd3eSStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 5831fa34dd3eSStefano Zampini ierr = VecAssemblyBegin(pcis->vec1_N);CHKERRQ(ierr); 5832fa34dd3eSStefano Zampini ierr = VecAssemblyEnd(pcis->vec1_N);CHKERRQ(ierr); 5833fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcis->N_to_B,pcis->vec1_N,pcis->vec2_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 5834fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcis->N_to_B,pcis->vec1_N,pcis->vec2_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 5835fa34dd3eSStefano Zampini ierr = VecDuplicate(pcis->vec2_B,&vec_check_B);CHKERRQ(ierr); 5836fa34dd3eSStefano Zampini ierr = VecCopy(pcis->vec2_B,vec_check_B);CHKERRQ(ierr); 5837fa34dd3eSStefano Zampini 5838fa34dd3eSStefano Zampini /* assemble rhs for coarse problem */ 5839fa34dd3eSStefano Zampini /* widetilde{S}_\Gamma w_\Gamma + \widetilde{B0}^T_B p0 */ 5840fa34dd3eSStefano Zampini /* local with Schur */ 5841fa34dd3eSStefano Zampini ierr = MatMult(S_j,pcis->vec2_B,pcis->vec1_B);CHKERRQ(ierr); 5842fa34dd3eSStefano Zampini if (zerodiag) { 5843fa34dd3eSStefano Zampini ierr = VecGetArray(dummy_vec,&array);CHKERRQ(ierr); 58444f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) array[i] = p0_check[i]; 5845fa34dd3eSStefano Zampini ierr = VecRestoreArray(dummy_vec,&array);CHKERRQ(ierr); 5846fa34dd3eSStefano Zampini ierr = MatMultTransposeAdd(B0_B,dummy_vec,pcis->vec1_B,pcis->vec1_B);CHKERRQ(ierr); 5847fa34dd3eSStefano Zampini } 5848fa34dd3eSStefano Zampini /* sum on primal nodes the local contributions */ 5849fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcis->N_to_B,pcis->vec1_B,pcis->vec1_N,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 5850fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcis->N_to_B,pcis->vec1_B,pcis->vec1_N,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 5851fa34dd3eSStefano Zampini ierr = VecGetArray(pcis->vec1_N,&array);CHKERRQ(ierr); 5852fa34dd3eSStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array2);CHKERRQ(ierr); 5853fa34dd3eSStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) array2[i] = array[pcbddc->local_primal_ref_node[i]]; 5854fa34dd3eSStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array2);CHKERRQ(ierr); 5855fa34dd3eSStefano Zampini ierr = VecRestoreArray(pcis->vec1_N,&array);CHKERRQ(ierr); 5856fa34dd3eSStefano Zampini ierr = VecSet(pcbddc->coarse_vec,0.);CHKERRQ(ierr); 5857fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 5858fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 5859fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,pcbddc->vec1_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 5860fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,pcbddc->vec1_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 5861fa34dd3eSStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 5862fa34dd3eSStefano Zampini /* scale primal nodes (BDDC sums contibutions) */ 5863fa34dd3eSStefano Zampini ierr = VecPointwiseMult(pcbddc->vec1_P,vec_scale_P,pcbddc->vec1_P);CHKERRQ(ierr); 5864fa34dd3eSStefano Zampini ierr = VecSetValues(pcis->vec1_N,pcbddc->local_primal_size,pcbddc->local_primal_ref_node,array,INSERT_VALUES);CHKERRQ(ierr); 5865fa34dd3eSStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 5866fa34dd3eSStefano Zampini ierr = VecAssemblyBegin(pcis->vec1_N);CHKERRQ(ierr); 5867fa34dd3eSStefano Zampini ierr = VecAssemblyEnd(pcis->vec1_N);CHKERRQ(ierr); 5868fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcis->N_to_B,pcis->vec1_N,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 5869fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcis->N_to_B,pcis->vec1_N,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 5870fa34dd3eSStefano Zampini /* global: \widetilde{B0}_B w_\Gamma */ 5871fa34dd3eSStefano Zampini if (zerodiag) { 5872fa34dd3eSStefano Zampini ierr = MatMult(B0_B,pcis->vec2_B,dummy_vec);CHKERRQ(ierr); 5873fa34dd3eSStefano Zampini ierr = VecGetArray(dummy_vec,&array);CHKERRQ(ierr); 58744f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) pcbddc->benign_p0[i] = array[i]; 5875fa34dd3eSStefano Zampini ierr = VecRestoreArray(dummy_vec,&array);CHKERRQ(ierr); 5876fa34dd3eSStefano Zampini } 5877fa34dd3eSStefano Zampini /* BDDC */ 5878fa34dd3eSStefano Zampini ierr = VecSet(pcis->vec1_D,0.);CHKERRQ(ierr); 5879fa34dd3eSStefano Zampini ierr = PCBDDCApplyInterfacePreconditioner(pc,PETSC_FALSE);CHKERRQ(ierr); 5880fa34dd3eSStefano Zampini 5881fa34dd3eSStefano Zampini ierr = VecCopy(pcis->vec1_B,pcis->vec2_B);CHKERRQ(ierr); 5882fa34dd3eSStefano Zampini ierr = VecAXPY(pcis->vec1_B,-1.0,vec_check_B);CHKERRQ(ierr); 5883fa34dd3eSStefano Zampini ierr = VecNorm(pcis->vec1_B,NORM_INFINITY,&norm);CHKERRQ(ierr); 5884fa34dd3eSStefano Zampini PetscPrintf(PETSC_COMM_SELF,"[%d] BDDC local error is %1.4e\n",PetscGlobalRank,norm); 58854f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 58864f1b2e48SStefano Zampini PetscPrintf(PETSC_COMM_SELF,"[%d] BDDC p0[%d] error is %1.4e\n",PetscGlobalRank,i,PetscAbsScalar(pcbddc->benign_p0[i]-p0_check[i])); 5887fa34dd3eSStefano Zampini } 58884f1b2e48SStefano Zampini ierr = PetscFree(p0_check);CHKERRQ(ierr); 5889fa34dd3eSStefano Zampini ierr = VecDestroy(&vec_scale_P);CHKERRQ(ierr); 5890fa34dd3eSStefano Zampini ierr = VecDestroy(&vec_check_B);CHKERRQ(ierr); 5891fa34dd3eSStefano Zampini ierr = VecDestroy(&dummy_vec);CHKERRQ(ierr); 5892fa34dd3eSStefano Zampini ierr = MatDestroy(&S_j);CHKERRQ(ierr); 5893fa34dd3eSStefano Zampini ierr = MatDestroy(&B0_B);CHKERRQ(ierr); 5894fa34dd3eSStefano Zampini } 5895fa34dd3eSStefano Zampini PetscFunctionReturn(0); 5896fa34dd3eSStefano Zampini } 5897