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; 3414f1b2e48SStefano Zampini ierr = PetscCalloc1(1,&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); 4514f1b2e48SStefano Zampini ierr = ISDestroy(&zerodiag_subs[s]);CHKERRQ(ierr); 452339f8db1SStefano Zampini ierr = PetscFree(array);CHKERRQ(ierr); 4534f1b2e48SStefano Zampini } 454339f8db1SStefano Zampini ierr = MatAssemblyBegin(pcbddc->benign_change,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 455339f8db1SStefano Zampini ierr = MatAssemblyEnd(pcbddc->benign_change,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 456339f8db1SStefano Zampini /* TODO: need optimization? */ 45797d764eeSStefano Zampini ierr = MatPtAP(pcbddc->local_mat,pcbddc->benign_change,MAT_INITIAL_MATRIX,2.0,&pcbddc->benign_original_mat);CHKERRQ(ierr); 458339f8db1SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 45997d764eeSStefano Zampini ierr = MatSeqAIJCompress(pcbddc->benign_original_mat,&pcbddc->local_mat);CHKERRQ(ierr); 46097d764eeSStefano Zampini ierr = MatDestroy(&pcbddc->benign_original_mat);CHKERRQ(ierr); 4614f1b2e48SStefano Zampini /* store global idxs for p0 */ 4624f1b2e48SStefano Zampini ierr = ISLocalToGlobalMappingApply(pc->pmat->rmap->mapping,pcbddc->benign_n,pcbddc->benign_p0_lidx,pcbddc->benign_p0_gidx);CHKERRQ(ierr); 463339f8db1SStefano Zampini } 4644f1b2e48SStefano Zampini ierr = PetscFree(zerodiag_subs);CHKERRQ(ierr); 4654f1b2e48SStefano Zampini ierr = ISDestroy(&pressures);CHKERRQ(ierr); 466b0f5fe93SStefano Zampini 467b0f5fe93SStefano Zampini /* determines if the coarse solver will be singular or not */ 468b0f5fe93SStefano Zampini ierr = MPI_Allreduce(&have_null,&pcbddc->benign_null,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 469339f8db1SStefano Zampini *zerodiaglocal = zerodiag; 470339f8db1SStefano Zampini PetscFunctionReturn(0); 471339f8db1SStefano Zampini } 472339f8db1SStefano Zampini 473339f8db1SStefano Zampini #undef __FUNCT__ 474015636ebSStefano Zampini #define __FUNCT__ "PCBDDCBenignGetOrSetP0" 475015636ebSStefano Zampini PetscErrorCode PCBDDCBenignGetOrSetP0(PC pc, Vec v, PetscBool get) 476efc2fbd9SStefano Zampini { 477efc2fbd9SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 478efc2fbd9SStefano Zampini PetscErrorCode ierr; 479efc2fbd9SStefano Zampini 480efc2fbd9SStefano Zampini PetscFunctionBegin; 481efc2fbd9SStefano Zampini if (!pcbddc->benign_sf) { 482efc2fbd9SStefano Zampini ierr = PetscSFCreate(PetscObjectComm((PetscObject)pc),&pcbddc->benign_sf);CHKERRQ(ierr); 4834f1b2e48SStefano Zampini ierr = PetscSFSetGraphLayout(pcbddc->benign_sf,pc->pmat->rmap,pcbddc->benign_n,NULL,PETSC_OWN_POINTER,pcbddc->benign_p0_gidx);CHKERRQ(ierr); 484efc2fbd9SStefano Zampini } 485015636ebSStefano Zampini if (get) { /* use SF to get values */ 486efc2fbd9SStefano Zampini PetscScalar *array; 487efc2fbd9SStefano Zampini 488efc2fbd9SStefano Zampini ierr = VecGetArrayRead(v,(const PetscScalar**)&array);CHKERRQ(ierr); 4894f1b2e48SStefano Zampini ierr = PetscSFBcastBegin(pcbddc->benign_sf,MPIU_SCALAR,array,pcbddc->benign_p0);CHKERRQ(ierr); 4904f1b2e48SStefano Zampini ierr = PetscSFBcastEnd(pcbddc->benign_sf,MPIU_SCALAR,array,pcbddc->benign_p0);CHKERRQ(ierr); 491efc2fbd9SStefano Zampini ierr = VecRestoreArrayRead(v,(const PetscScalar**)&array);CHKERRQ(ierr); 4924f1b2e48SStefano Zampini } else { /* use VecSetValues (not scalable, I should try to find a better solution (defining a new MPI_OP for reduction) */ 4934f1b2e48SStefano Zampini ierr = VecSetValues(v,pcbddc->benign_n,pcbddc->benign_p0_gidx,pcbddc->benign_p0,INSERT_VALUES);CHKERRQ(ierr); 494efc2fbd9SStefano Zampini ierr = VecAssemblyBegin(v);CHKERRQ(ierr); 495efc2fbd9SStefano Zampini ierr = VecAssemblyEnd(v);CHKERRQ(ierr); 496efc2fbd9SStefano Zampini } 497efc2fbd9SStefano Zampini PetscFunctionReturn(0); 498efc2fbd9SStefano Zampini } 499efc2fbd9SStefano Zampini 500efc2fbd9SStefano Zampini #undef __FUNCT__ 501c263805aSStefano Zampini #define __FUNCT__ "PCBDDCBenignPopOrPushB0" 502c263805aSStefano Zampini PetscErrorCode PCBDDCBenignPopOrPushB0(PC pc, PetscBool pop) 503c263805aSStefano Zampini { 504c263805aSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 505c263805aSStefano Zampini PetscErrorCode ierr; 506c263805aSStefano Zampini 507c263805aSStefano Zampini PetscFunctionBegin; 508c263805aSStefano Zampini /* TODO: add error checking 509c263805aSStefano Zampini - avoid nested pop (or push) calls. 510c263805aSStefano Zampini - cannot push before pop. 5111c604dc7SStefano Zampini - cannot call this if pcbddc->local_mat is NULL 512c263805aSStefano Zampini */ 5134f1b2e48SStefano Zampini if (!pcbddc->benign_n) { 514efc2fbd9SStefano Zampini PetscFunctionReturn(0); 515efc2fbd9SStefano Zampini } 516c263805aSStefano Zampini if (pop) { 5174f1b2e48SStefano Zampini IS is_p0; 5184f1b2e48SStefano Zampini MatReuse reuse; 519c263805aSStefano Zampini 520c263805aSStefano Zampini /* extract B_0 */ 5214f1b2e48SStefano Zampini reuse = MAT_INITIAL_MATRIX; 5224f1b2e48SStefano Zampini if (pcbddc->benign_B0) { 5234f1b2e48SStefano Zampini reuse = MAT_REUSE_MATRIX; 5244f1b2e48SStefano Zampini } 5254f1b2e48SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,pcbddc->benign_n,pcbddc->benign_p0_lidx,PETSC_COPY_VALUES,&is_p0);CHKERRQ(ierr); 5264f1b2e48SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,is_p0,NULL,reuse,&pcbddc->benign_B0);CHKERRQ(ierr); 527c263805aSStefano Zampini /* remove rows and cols from local problem */ 528c263805aSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_KEEP_NONZERO_PATTERN,PETSC_TRUE);CHKERRQ(ierr); 52997d764eeSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_FALSE);CHKERRQ(ierr); 5304f1b2e48SStefano Zampini ierr = MatZeroRowsColumnsIS(pcbddc->local_mat,is_p0,1.0,NULL,NULL);CHKERRQ(ierr); 5314f1b2e48SStefano Zampini ierr = ISDestroy(&is_p0);CHKERRQ(ierr); 532c263805aSStefano Zampini } else { /* push */ 5334f1b2e48SStefano Zampini PetscInt i; 5344f1b2e48SStefano Zampini 5354f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 5364f1b2e48SStefano Zampini PetscScalar *B0_vals; 5374f1b2e48SStefano Zampini PetscInt *B0_cols,B0_ncol; 5384f1b2e48SStefano Zampini 5394f1b2e48SStefano Zampini ierr = MatGetRow(pcbddc->benign_B0,i,&B0_ncol,(const PetscInt**)&B0_cols,(const PetscScalar**)&B0_vals);CHKERRQ(ierr); 5404f1b2e48SStefano Zampini ierr = MatSetValues(pcbddc->local_mat,1,pcbddc->benign_p0_lidx+i,B0_ncol,B0_cols,B0_vals,INSERT_VALUES);CHKERRQ(ierr); 5417b034428SStefano Zampini ierr = MatSetValues(pcbddc->local_mat,B0_ncol,B0_cols,1,pcbddc->benign_p0_lidx+i,B0_vals,INSERT_VALUES);CHKERRQ(ierr); 5424f1b2e48SStefano Zampini ierr = MatSetValue(pcbddc->local_mat,pcbddc->benign_p0_lidx[i],pcbddc->benign_p0_lidx[i],0.0,INSERT_VALUES);CHKERRQ(ierr); 5434f1b2e48SStefano Zampini ierr = MatRestoreRow(pcbddc->benign_B0,i,&B0_ncol,(const PetscInt**)&B0_cols,(const PetscScalar**)&B0_vals);CHKERRQ(ierr); 5444f1b2e48SStefano Zampini } 545c263805aSStefano Zampini ierr = MatAssemblyBegin(pcbddc->local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 546c263805aSStefano Zampini ierr = MatAssemblyEnd(pcbddc->local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 547c263805aSStefano Zampini } 548c263805aSStefano Zampini PetscFunctionReturn(0); 549c263805aSStefano Zampini } 550c263805aSStefano Zampini 551c263805aSStefano Zampini #undef __FUNCT__ 552b1b3d7a2SStefano Zampini #define __FUNCT__ "PCBDDCAdaptiveSelection" 55308122e43SStefano Zampini PetscErrorCode PCBDDCAdaptiveSelection(PC pc) 554b1b3d7a2SStefano Zampini { 555b1b3d7a2SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 55608122e43SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 55708122e43SStefano Zampini PetscBLASInt B_dummyint,B_neigs,B_ierr,B_lwork; 55808122e43SStefano Zampini PetscBLASInt *B_iwork,*B_ifail; 55908122e43SStefano Zampini PetscScalar *work,lwork; 56008122e43SStefano Zampini PetscScalar *St,*S,*eigv; 56108122e43SStefano Zampini PetscScalar *Sarray,*Starray; 56208122e43SStefano Zampini PetscReal *eigs,thresh; 5631b968477SStefano Zampini PetscInt i,nmax,nmin,nv,cum,mss,cum2,cumarray,maxneigs; 564f6f667cfSStefano Zampini PetscBool allocated_S_St; 56508122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 56608122e43SStefano Zampini PetscReal *rwork; 56708122e43SStefano Zampini #endif 568b1b3d7a2SStefano Zampini PetscErrorCode ierr; 569b1b3d7a2SStefano Zampini 570b1b3d7a2SStefano Zampini PetscFunctionBegin; 57108122e43SStefano Zampini if (!sub_schurs->use_mumps) { 57208122e43SStefano Zampini SETERRQ(PetscObjectComm((PetscObject)pc),PETSC_ERR_SUP,"Adaptive selection of constraints requires MUMPS"); 57308122e43SStefano Zampini } 57408122e43SStefano Zampini 57506a4e24aSStefano Zampini if (sub_schurs->n_subs && (!sub_schurs->is_hermitian || !sub_schurs->is_posdef)) { 57606a4e24aSStefano 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); 57706a4e24aSStefano Zampini } 57806a4e24aSStefano Zampini 579fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 580fd14bc51SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 581fd14bc51SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 582fd14bc51SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Check adaptive selection of constraints\n");CHKERRQ(ierr); 5831575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 584fd14bc51SStefano Zampini } 585fd14bc51SStefano Zampini 586e496cd5dSStefano Zampini if (pcbddc->dbg_flag) { 587e496cd5dSStefano 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); 588e496cd5dSStefano Zampini } 589e496cd5dSStefano Zampini 59008122e43SStefano Zampini /* max size of subsets */ 59108122e43SStefano Zampini mss = 0; 59208122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 59308122e43SStefano Zampini PetscInt subset_size; 594862806e4SStefano Zampini 59508122e43SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 59608122e43SStefano Zampini mss = PetscMax(mss,subset_size); 59708122e43SStefano Zampini } 59808122e43SStefano Zampini 59908122e43SStefano Zampini /* min/max and threshold */ 60008122e43SStefano Zampini nmax = pcbddc->adaptive_nmax > 0 ? pcbddc->adaptive_nmax : mss; 601f6f667cfSStefano Zampini nmin = pcbddc->adaptive_nmin > 0 ? pcbddc->adaptive_nmin : 0; 60208122e43SStefano Zampini nmax = PetscMax(nmin,nmax); 603f6f667cfSStefano Zampini allocated_S_St = PETSC_FALSE; 604f6f667cfSStefano Zampini if (nmin) { 605f6f667cfSStefano Zampini allocated_S_St = PETSC_TRUE; 606f6f667cfSStefano Zampini } 60708122e43SStefano Zampini 60808122e43SStefano Zampini /* allocate lapack workspace */ 60908122e43SStefano Zampini cum = cum2 = 0; 61008122e43SStefano Zampini maxneigs = 0; 61108122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 61208122e43SStefano Zampini PetscInt n,subset_size; 613f6f667cfSStefano Zampini 61408122e43SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 61508122e43SStefano Zampini n = PetscMin(subset_size,nmax); 6169162d606SStefano Zampini cum += subset_size; 6179162d606SStefano Zampini cum2 += subset_size*n; 61808122e43SStefano Zampini maxneigs = PetscMax(maxneigs,n); 61908122e43SStefano Zampini } 62008122e43SStefano Zampini if (mss) { 6219ab7bb16SStefano Zampini if (sub_schurs->is_hermitian && sub_schurs->is_posdef) { 62208122e43SStefano Zampini PetscBLASInt B_itype = 1; 62308122e43SStefano Zampini PetscBLASInt B_N = mss; 6244c6709b3SStefano Zampini PetscReal zero = 0.0; 6254c6709b3SStefano Zampini PetscReal eps = 0.0; /* dlamch? */ 62608122e43SStefano Zampini 62708122e43SStefano Zampini B_lwork = -1; 62808122e43SStefano Zampini S = NULL; 62908122e43SStefano Zampini St = NULL; 630a58a30b4SStefano Zampini eigs = NULL; 631a58a30b4SStefano Zampini eigv = NULL; 632a58a30b4SStefano Zampini B_iwork = NULL; 633a58a30b4SStefano Zampini B_ifail = NULL; 634d1710679SStefano Zampini #if defined(PETSC_USE_COMPLEX) 635d1710679SStefano Zampini rwork = NULL; 636d1710679SStefano Zampini #endif 6378bec7fa6SStefano Zampini thresh = 1.0; 63808122e43SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 63908122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 64008122e43SStefano 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)); 64108122e43SStefano Zampini #else 64208122e43SStefano 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)); 64308122e43SStefano Zampini #endif 64408122e43SStefano Zampini if (B_ierr != 0) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to SYGVX Lapack routine %d",(int)B_ierr); 64508122e43SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 64608122e43SStefano Zampini } else { 64708122e43SStefano Zampini /* TODO */ 64808122e43SStefano Zampini } 64908122e43SStefano Zampini } else { 65008122e43SStefano Zampini lwork = 0; 65108122e43SStefano Zampini } 65208122e43SStefano Zampini 65308122e43SStefano Zampini nv = 0; 654d62866d3SStefano 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) */ 655d62866d3SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_vertices,&nv);CHKERRQ(ierr); 65608122e43SStefano Zampini } 6574c6709b3SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lwork),&B_lwork);CHKERRQ(ierr); 658f6f667cfSStefano Zampini if (allocated_S_St) { 659f6f667cfSStefano Zampini ierr = PetscMalloc2(mss*mss,&S,mss*mss,&St);CHKERRQ(ierr); 660f6f667cfSStefano Zampini } 661f6f667cfSStefano Zampini ierr = PetscMalloc5(mss*mss,&eigv,mss,&eigs,B_lwork,&work,5*mss,&B_iwork,mss,&B_ifail);CHKERRQ(ierr); 66208122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 66308122e43SStefano Zampini ierr = PetscMalloc1(7*mss,&rwork);CHKERRQ(ierr); 66408122e43SStefano Zampini #endif 6659162d606SStefano Zampini ierr = PetscMalloc5(nv+sub_schurs->n_subs,&pcbddc->adaptive_constraints_n, 6669162d606SStefano Zampini nv+sub_schurs->n_subs+1,&pcbddc->adaptive_constraints_idxs_ptr, 6679162d606SStefano Zampini nv+sub_schurs->n_subs+1,&pcbddc->adaptive_constraints_data_ptr, 66808122e43SStefano Zampini nv+cum,&pcbddc->adaptive_constraints_idxs, 6699162d606SStefano Zampini nv+cum2,&pcbddc->adaptive_constraints_data);CHKERRQ(ierr); 67008122e43SStefano Zampini ierr = PetscMemzero(pcbddc->adaptive_constraints_n,(nv+sub_schurs->n_subs)*sizeof(PetscInt));CHKERRQ(ierr); 67108122e43SStefano Zampini 67208122e43SStefano Zampini maxneigs = 0; 67308122e43SStefano Zampini cum = cum2 = cumarray = 0; 6749162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr[0] = 0; 6759162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr[0] = 0; 676d62866d3SStefano Zampini if (sub_schurs->is_vertices && pcbddc->use_vertices) { 67708122e43SStefano Zampini const PetscInt *idxs; 67808122e43SStefano Zampini 679d62866d3SStefano Zampini ierr = ISGetIndices(sub_schurs->is_vertices,&idxs);CHKERRQ(ierr); 68008122e43SStefano Zampini for (cum=0;cum<nv;cum++) { 68108122e43SStefano Zampini pcbddc->adaptive_constraints_n[cum] = 1; 68208122e43SStefano Zampini pcbddc->adaptive_constraints_idxs[cum] = idxs[cum]; 68308122e43SStefano Zampini pcbddc->adaptive_constraints_data[cum] = 1.0; 6849162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr[cum+1] = pcbddc->adaptive_constraints_idxs_ptr[cum]+1; 6859162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr[cum+1] = pcbddc->adaptive_constraints_data_ptr[cum]+1; 68608122e43SStefano Zampini } 68708122e43SStefano Zampini cum2 = cum; 688d62866d3SStefano Zampini ierr = ISRestoreIndices(sub_schurs->is_vertices,&idxs);CHKERRQ(ierr); 68908122e43SStefano Zampini } 69008122e43SStefano Zampini 69108122e43SStefano Zampini if (mss) { /* multilevel */ 69208122e43SStefano Zampini ierr = MatSeqAIJGetArray(sub_schurs->sum_S_Ej_inv_all,&Sarray);CHKERRQ(ierr); 69308122e43SStefano Zampini ierr = MatSeqAIJGetArray(sub_schurs->sum_S_Ej_tilda_all,&Starray);CHKERRQ(ierr); 69408122e43SStefano Zampini } 69508122e43SStefano Zampini 696ffd830a3SStefano Zampini thresh = pcbddc->adaptive_threshold; 69708122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 69808122e43SStefano Zampini const PetscInt *idxs; 699f6f667cfSStefano Zampini PetscReal infty = PETSC_MAX_REAL; 700862806e4SStefano Zampini PetscInt j,subset_size,eigs_start = 0; 70108122e43SStefano Zampini PetscBLASInt B_N; 702aff50787SStefano Zampini PetscBool same_data = PETSC_FALSE; 70308122e43SStefano Zampini 704862806e4SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 705ffd830a3SStefano Zampini ierr = ISGetIndices(sub_schurs->is_subs[i],&idxs);CHKERRQ(ierr); 706f6f667cfSStefano Zampini ierr = PetscBLASIntCast(subset_size,&B_N);CHKERRQ(ierr); 707f6f667cfSStefano Zampini if (allocated_S_St) { /* S and S_t should be copied since we could need them later */ 7089ab7bb16SStefano Zampini if (sub_schurs->is_hermitian) { 709aff50787SStefano Zampini PetscInt j,k; 710aff50787SStefano Zampini if (sub_schurs->n_subs == 1) { /* zeroing memory to use PetscMemcmp later */ 711aff50787SStefano Zampini ierr = PetscMemzero(S,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 712aff50787SStefano Zampini ierr = PetscMemzero(St,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 71308122e43SStefano Zampini } 71408122e43SStefano Zampini for (j=0;j<subset_size;j++) { 715aff50787SStefano Zampini for (k=j;k<subset_size;k++) { 716aff50787SStefano Zampini S [j*subset_size+k] = Sarray [cumarray+j*subset_size+k]; 717aff50787SStefano Zampini St[j*subset_size+k] = Starray[cumarray+j*subset_size+k]; 718aff50787SStefano Zampini } 71908122e43SStefano Zampini } 72008122e43SStefano Zampini } else { 72108122e43SStefano Zampini ierr = PetscMemcpy(S,Sarray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 72208122e43SStefano Zampini ierr = PetscMemcpy(St,Starray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 72308122e43SStefano Zampini } 7248bec7fa6SStefano Zampini } else { 725f6f667cfSStefano Zampini S = Sarray + cumarray; 726f6f667cfSStefano Zampini St = Starray + cumarray; 7278bec7fa6SStefano Zampini } 72808122e43SStefano Zampini 729aff50787SStefano Zampini /* see if we can save some work */ 730aff50787SStefano Zampini if (sub_schurs->n_subs == 1) { 731aff50787SStefano Zampini ierr = PetscMemcmp(S,St,subset_size*subset_size*sizeof(PetscScalar),&same_data);CHKERRQ(ierr); 732aff50787SStefano Zampini } 733aff50787SStefano Zampini 734aff50787SStefano Zampini if (same_data) { /* there's no need of constraints here, deluxe scaling is enough */ 735aff50787SStefano Zampini B_neigs = 0; 736aff50787SStefano Zampini } else { 7379ab7bb16SStefano Zampini if (sub_schurs->is_hermitian && sub_schurs->is_posdef) { 73808122e43SStefano Zampini PetscBLASInt B_itype = 1; 739f6f667cfSStefano Zampini PetscBLASInt B_IL, B_IU; 7404c6709b3SStefano Zampini PetscReal eps = -1.0; /* dlamch? */ 7419552c7c7SStefano Zampini PetscInt nmin_s; 74208122e43SStefano Zampini 743fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 7448bec7fa6SStefano 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]]); 745fd14bc51SStefano Zampini } 746d16cbb6bSStefano Zampini 74708122e43SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 748d16cbb6bSStefano Zampini if (thresh > 1.+PETSC_SMALL) { 749d16cbb6bSStefano Zampini 750d16cbb6bSStefano Zampini /* ask for eigenvalues larger than thresh */ 75108122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 752f6f667cfSStefano 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)); 75308122e43SStefano Zampini #else 754f6f667cfSStefano 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)); 75508122e43SStefano Zampini #endif 756d16cbb6bSStefano Zampini } else { 757d16cbb6bSStefano Zampini B_IU = PetscMax(1,PetscMin(B_N,nmax)); 758d16cbb6bSStefano Zampini B_IL = 1; 759d16cbb6bSStefano Zampini #if defined(PETSC_USE_COMPLEX) 760d16cbb6bSStefano 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)); 761d16cbb6bSStefano Zampini #else 762d16cbb6bSStefano 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)); 763d16cbb6bSStefano Zampini #endif 764d16cbb6bSStefano Zampini } 76508122e43SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 76608122e43SStefano Zampini if (B_ierr) { 76708122e43SStefano Zampini if (B_ierr < 0 ) { 76808122e43SStefano Zampini SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYGVX Lapack routine: illegal value for argument %d",-(int)B_ierr); 76908122e43SStefano Zampini } else if (B_ierr <= B_N) { 77008122e43SStefano Zampini SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYGVX Lapack routine: %d eigenvalues failed to converge",(int)B_ierr); 77108122e43SStefano Zampini } else { 7729552c7c7SStefano 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); 77308122e43SStefano Zampini } 77408122e43SStefano Zampini } 77508122e43SStefano Zampini 77608122e43SStefano Zampini if (B_neigs > nmax) { 777fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 778fd14bc51SStefano Zampini PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," found %d eigs, more than maximum required %d.\n",B_neigs,nmax); 779fd14bc51SStefano Zampini } 780f6f667cfSStefano Zampini eigs_start = B_neigs -nmax; 78108122e43SStefano Zampini B_neigs = nmax; 78208122e43SStefano Zampini } 78308122e43SStefano Zampini 7849552c7c7SStefano Zampini nmin_s = PetscMin(nmin,B_N); 7859552c7c7SStefano Zampini if (B_neigs < nmin_s) { 78608122e43SStefano Zampini PetscBLASInt B_neigs2; 78708122e43SStefano Zampini 788f6f667cfSStefano Zampini B_IU = B_N - B_neigs; 789f6f667cfSStefano Zampini B_IL = B_N - nmin_s + 1; 790fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 791fd14bc51SStefano 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); 792fd14bc51SStefano Zampini } 7939ab7bb16SStefano Zampini if (sub_schurs->is_hermitian) { 79408122e43SStefano Zampini PetscInt j; 79508122e43SStefano Zampini for (j=0;j<subset_size;j++) { 79608122e43SStefano Zampini ierr = PetscMemcpy(S+j*(subset_size+1),Sarray+cumarray+j*(subset_size+1),(subset_size-j)*sizeof(PetscScalar));CHKERRQ(ierr); 79708122e43SStefano Zampini } 79808122e43SStefano Zampini for (j=0;j<subset_size;j++) { 79908122e43SStefano Zampini ierr = PetscMemcpy(St+j*(subset_size+1),Starray+cumarray+j*(subset_size+1),(subset_size-j)*sizeof(PetscScalar));CHKERRQ(ierr); 80008122e43SStefano Zampini } 80108122e43SStefano Zampini } else { 80208122e43SStefano Zampini ierr = PetscMemcpy(S,Sarray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 80308122e43SStefano Zampini ierr = PetscMemcpy(St,Starray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 80408122e43SStefano Zampini } 80508122e43SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 80608122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 807f6f667cfSStefano 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)); 80808122e43SStefano Zampini #else 809f6f667cfSStefano 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)); 81008122e43SStefano Zampini #endif 81108122e43SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 81208122e43SStefano Zampini B_neigs += B_neigs2; 81308122e43SStefano Zampini } 81408122e43SStefano Zampini if (B_ierr) { 81508122e43SStefano Zampini if (B_ierr < 0 ) { 81608122e43SStefano Zampini SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYGVX Lapack routine: illegal value for argument %d",-(int)B_ierr); 81708122e43SStefano Zampini } else if (B_ierr <= B_N) { 81808122e43SStefano Zampini SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYGVX Lapack routine: %d eigenvalues failed to converge",(int)B_ierr); 81908122e43SStefano Zampini } else { 8209552c7c7SStefano 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); 82108122e43SStefano Zampini } 82208122e43SStefano Zampini } 823fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 824ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," -> Got %d eigs\n",B_neigs);CHKERRQ(ierr); 82508122e43SStefano Zampini for (j=0;j<B_neigs;j++) { 82608122e43SStefano Zampini if (eigs[j] == 0.0) { 827ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," Inf\n");CHKERRQ(ierr); 82808122e43SStefano Zampini } else { 829ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," %1.6e\n",eigs[j+eigs_start]);CHKERRQ(ierr); 830fd14bc51SStefano Zampini } 83108122e43SStefano Zampini } 83208122e43SStefano Zampini } 83308122e43SStefano Zampini } else { 83408122e43SStefano Zampini /* TODO */ 83508122e43SStefano Zampini } 836aff50787SStefano Zampini } 8378bec7fa6SStefano Zampini maxneigs = PetscMax(B_neigs,maxneigs); 8388bec7fa6SStefano Zampini pcbddc->adaptive_constraints_n[i+nv] = B_neigs; 8399162d606SStefano Zampini if (B_neigs) { 8409162d606SStefano 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); 841fd14bc51SStefano Zampini 842fd14bc51SStefano Zampini if (pcbddc->dbg_flag > 1) { 8439552c7c7SStefano Zampini PetscInt ii; 8449552c7c7SStefano Zampini for (ii=0;ii<B_neigs;ii++) { 845ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," -> Eigenvector %d/%d (%d)\n",ii,B_neigs,B_N);CHKERRQ(ierr); 8469552c7c7SStefano Zampini for (j=0;j<B_N;j++) { 847ac47001eSStefano Zampini #if defined(PETSC_USE_COMPLEX) 848ac47001eSStefano Zampini PetscReal r = PetscRealPart(pcbddc->adaptive_constraints_data[ii*subset_size+j+pcbddc->adaptive_constraints_data_ptr[cum]]); 849ac47001eSStefano Zampini PetscReal c = PetscImaginaryPart(pcbddc->adaptive_constraints_data[ii*subset_size+j+pcbddc->adaptive_constraints_data_ptr[cum]]); 850ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," %1.4e + %1.4e i\n",r,c);CHKERRQ(ierr); 851ac47001eSStefano Zampini #else 852ac47001eSStefano 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); 853ac47001eSStefano Zampini #endif 8549552c7c7SStefano Zampini } 8559552c7c7SStefano Zampini } 856fd14bc51SStefano Zampini } 85708122e43SStefano Zampini #if 0 8589162d606SStefano Zampini for (j=0;j<B_neigs;j++) { 85908122e43SStefano Zampini PetscBLASInt Blas_N,Blas_one = 1.0; 86008122e43SStefano Zampini PetscScalar norm; 86108122e43SStefano Zampini ierr = PetscBLASIntCast(subset_size,&Blas_N);CHKERRQ(ierr); 8629162d606SStefano Zampini PetscStackCallBLAS("BLASdot",norm = BLASdot_(&Blas_N,pcbddc->adaptive_constraints_data+pcbddc->adaptive_constraints_data_ptr[cum]+j*subset_size, 8639162d606SStefano Zampini &Blas_one,pcbddc->adaptive_constraints_data+pcbddc->adaptive_constraints_data_ptr[cum]+j*subset_size,&Blas_one)); 86408122e43SStefano Zampini if (pcbddc->adaptive_constraints_data[cum2] > 0.0) { 86508122e43SStefano Zampini norm = 1.0/PetscSqrtReal(PetscRealPart(norm)); 86608122e43SStefano Zampini } else { 86708122e43SStefano Zampini norm = -1.0/PetscSqrtReal(PetscRealPart(norm)); 86808122e43SStefano Zampini } 8699162d606SStefano Zampini PetscStackCallBLAS("BLASscal",BLASscal_(&Blas_N,&norm,pcbddc->adaptive_constraints_data+pcbddc->adaptive_constraints_data_ptr[cum]+j*subset_size,&Blas_one)); 870b1b3d7a2SStefano Zampini } 871b1b3d7a2SStefano Zampini #endif 8729162d606SStefano Zampini ierr = PetscMemcpy(pcbddc->adaptive_constraints_idxs+pcbddc->adaptive_constraints_idxs_ptr[cum],idxs,subset_size*sizeof(PetscInt));CHKERRQ(ierr); 8739162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr[cum+1] = pcbddc->adaptive_constraints_idxs_ptr[cum] + subset_size; 8749162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr[cum+1] = pcbddc->adaptive_constraints_data_ptr[cum] + subset_size*B_neigs; 8759162d606SStefano Zampini cum++; 87608122e43SStefano Zampini } 87708122e43SStefano Zampini ierr = ISRestoreIndices(sub_schurs->is_subs[i],&idxs);CHKERRQ(ierr); 87808122e43SStefano Zampini /* shift for next computation */ 87908122e43SStefano Zampini cumarray += subset_size*subset_size; 88008122e43SStefano Zampini } 881fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 882fd14bc51SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 883fd14bc51SStefano Zampini } 88408122e43SStefano Zampini 88508122e43SStefano Zampini if (mss) { 88608122e43SStefano Zampini ierr = MatSeqAIJRestoreArray(sub_schurs->sum_S_Ej_inv_all,&Sarray);CHKERRQ(ierr); 88708122e43SStefano Zampini ierr = MatSeqAIJRestoreArray(sub_schurs->sum_S_Ej_tilda_all,&Starray);CHKERRQ(ierr); 888f6f667cfSStefano Zampini /* destroy matrices (junk) */ 889f6f667cfSStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_inv_all);CHKERRQ(ierr); 890f6f667cfSStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_tilda_all);CHKERRQ(ierr); 89108122e43SStefano Zampini } 892f6f667cfSStefano Zampini if (allocated_S_St) { 893f6f667cfSStefano Zampini ierr = PetscFree2(S,St);CHKERRQ(ierr); 894f6f667cfSStefano Zampini } 895f6f667cfSStefano Zampini ierr = PetscFree5(eigv,eigs,work,B_iwork,B_ifail);CHKERRQ(ierr); 89608122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 89708122e43SStefano Zampini ierr = PetscFree(rwork);CHKERRQ(ierr); 89808122e43SStefano Zampini #endif 89908122e43SStefano Zampini if (pcbddc->dbg_flag) { 9001b968477SStefano Zampini PetscInt maxneigs_r; 90108122e43SStefano Zampini ierr = MPI_Allreduce(&maxneigs,&maxneigs_r,1,MPIU_INT,MPI_MAX,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 9029b28b941SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Maximum number of constraints per cc %d\n",maxneigs_r);CHKERRQ(ierr); 90308122e43SStefano Zampini } 90408122e43SStefano Zampini PetscFunctionReturn(0); 90508122e43SStefano Zampini } 906b1b3d7a2SStefano Zampini 907674ae819SStefano Zampini #undef __FUNCT__ 908c8587f34SStefano Zampini #define __FUNCT__ "PCBDDCSetUpSolvers" 909c8587f34SStefano Zampini PetscErrorCode PCBDDCSetUpSolvers(PC pc) 910c8587f34SStefano Zampini { 911c8587f34SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 9128629588bSStefano Zampini PetscScalar *coarse_submat_vals; 913c8587f34SStefano Zampini PetscErrorCode ierr; 914c8587f34SStefano Zampini 915c8587f34SStefano Zampini PetscFunctionBegin; 916f4ddd8eeSStefano Zampini /* Setup local scatters R_to_B and (optionally) R_to_D */ 9175e8657edSStefano Zampini /* PCBDDCSetUpLocalWorkVectors should be called first! */ 918c8587f34SStefano Zampini ierr = PCBDDCSetUpLocalScatters(pc);CHKERRQ(ierr); 919c8587f34SStefano Zampini 920684f6988SStefano Zampini /* Setup local neumann solver ksp_R */ 9210fccc4e9SStefano Zampini /* PCBDDCSetUpLocalScatters should be called first! */ 922684f6988SStefano Zampini ierr = PCBDDCSetUpLocalSolvers(pc,PETSC_FALSE,PETSC_TRUE);CHKERRQ(ierr); 923c8587f34SStefano Zampini 924c8587f34SStefano Zampini /* Change global null space passed in by the user if change of basis has been requested */ 925b9b85e73SStefano Zampini if (pcbddc->NullSpace && pcbddc->ChangeOfBasisMatrix) { 926c8587f34SStefano Zampini ierr = PCBDDCNullSpaceAdaptGlobal(pc);CHKERRQ(ierr); 927c8587f34SStefano Zampini } 928c8587f34SStefano Zampini 9298629588bSStefano Zampini /* 9308629588bSStefano Zampini Setup local correction and local part of coarse basis. 9318629588bSStefano Zampini Gives back the dense local part of the coarse matrix in column major ordering 9328629588bSStefano Zampini */ 93347f4ddc3SStefano Zampini ierr = PCBDDCSetUpCorrection(pc,&coarse_submat_vals);CHKERRQ(ierr); 9348629588bSStefano Zampini 9358629588bSStefano Zampini /* Compute total number of coarse nodes and setup coarse solver */ 9368629588bSStefano Zampini ierr = PCBDDCSetUpCoarseSolver(pc,coarse_submat_vals);CHKERRQ(ierr); 9378629588bSStefano Zampini 9388629588bSStefano Zampini /* free */ 9398629588bSStefano Zampini ierr = PetscFree(coarse_submat_vals);CHKERRQ(ierr); 940c8587f34SStefano Zampini PetscFunctionReturn(0); 941c8587f34SStefano Zampini } 942c8587f34SStefano Zampini 943c8587f34SStefano Zampini #undef __FUNCT__ 944674ae819SStefano Zampini #define __FUNCT__ "PCBDDCResetCustomization" 945674ae819SStefano Zampini PetscErrorCode PCBDDCResetCustomization(PC pc) 946674ae819SStefano Zampini { 947674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 948674ae819SStefano Zampini PetscErrorCode ierr; 949674ae819SStefano Zampini 950674ae819SStefano Zampini PetscFunctionBegin; 951674ae819SStefano Zampini ierr = PCBDDCGraphResetCSR(pcbddc->mat_graph);CHKERRQ(ierr); 952674ae819SStefano Zampini ierr = ISDestroy(&pcbddc->user_primal_vertices);CHKERRQ(ierr); 95330368db7SStefano Zampini ierr = ISDestroy(&pcbddc->user_primal_vertices_local);CHKERRQ(ierr); 954674ae819SStefano Zampini ierr = MatNullSpaceDestroy(&pcbddc->NullSpace);CHKERRQ(ierr); 955674ae819SStefano Zampini ierr = ISDestroy(&pcbddc->NeumannBoundaries);CHKERRQ(ierr); 956785d1243SStefano Zampini ierr = ISDestroy(&pcbddc->NeumannBoundariesLocal);CHKERRQ(ierr); 957674ae819SStefano Zampini ierr = ISDestroy(&pcbddc->DirichletBoundaries);CHKERRQ(ierr); 958f4ddd8eeSStefano Zampini ierr = MatNullSpaceDestroy(&pcbddc->onearnullspace);CHKERRQ(ierr); 959f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->onearnullvecs_state);CHKERRQ(ierr); 960785d1243SStefano Zampini ierr = ISDestroy(&pcbddc->DirichletBoundariesLocal);CHKERRQ(ierr); 96163602bcaSStefano Zampini ierr = PCBDDCSetDofsSplitting(pc,0,NULL);CHKERRQ(ierr); 96263602bcaSStefano Zampini ierr = PCBDDCSetDofsSplittingLocal(pc,0,NULL);CHKERRQ(ierr); 963674ae819SStefano Zampini PetscFunctionReturn(0); 964674ae819SStefano Zampini } 965674ae819SStefano Zampini 966674ae819SStefano Zampini #undef __FUNCT__ 967674ae819SStefano Zampini #define __FUNCT__ "PCBDDCResetTopography" 968674ae819SStefano Zampini PetscErrorCode PCBDDCResetTopography(PC pc) 969674ae819SStefano Zampini { 970674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 9714f1b2e48SStefano Zampini PetscInt i; 972674ae819SStefano Zampini PetscErrorCode ierr; 973674ae819SStefano Zampini 974674ae819SStefano Zampini PetscFunctionBegin; 975b9b85e73SStefano Zampini ierr = MatDestroy(&pcbddc->user_ChangeOfBasisMatrix);CHKERRQ(ierr); 976674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 977674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 978674ae819SStefano Zampini ierr = PCBDDCGraphReset(pcbddc->mat_graph);CHKERRQ(ierr); 9794f1b2e48SStefano Zampini for (i=0;i<pcbddc->n_local_subs;i++) { 9804f1b2e48SStefano Zampini ierr = ISDestroy(&pcbddc->local_subs[i]);CHKERRQ(ierr); 9814f1b2e48SStefano Zampini } 9824f1b2e48SStefano Zampini ierr = PetscFree(pcbddc->local_subs);CHKERRQ(ierr); 983b96c3477SStefano Zampini ierr = PCBDDCSubSchursReset(pcbddc->sub_schurs);CHKERRQ(ierr); 984674ae819SStefano Zampini PetscFunctionReturn(0); 985674ae819SStefano Zampini } 986674ae819SStefano Zampini 987674ae819SStefano Zampini #undef __FUNCT__ 988674ae819SStefano Zampini #define __FUNCT__ "PCBDDCResetSolvers" 989674ae819SStefano Zampini PetscErrorCode PCBDDCResetSolvers(PC pc) 990674ae819SStefano Zampini { 991674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 99206656605SStefano Zampini PetscScalar *array; 993674ae819SStefano Zampini PetscErrorCode ierr; 994674ae819SStefano Zampini 995674ae819SStefano Zampini PetscFunctionBegin; 996674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->coarse_vec);CHKERRQ(ierr); 99758da7f69SStefano Zampini if (pcbddc->coarse_phi_B) { 99806656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&array);CHKERRQ(ierr); 99906656605SStefano Zampini ierr = PetscFree(array);CHKERRQ(ierr); 100058da7f69SStefano Zampini } 1001674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_B);CHKERRQ(ierr); 1002674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_D);CHKERRQ(ierr); 100315aaf578SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_B);CHKERRQ(ierr); 100415aaf578SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_D);CHKERRQ(ierr); 1005674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec1_P);CHKERRQ(ierr); 1006674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec1_C);CHKERRQ(ierr); 1007674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat2);CHKERRQ(ierr); 100806656605SStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat1);CHKERRQ(ierr); 1009674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec1_R);CHKERRQ(ierr); 1010674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec2_R);CHKERRQ(ierr); 10118ce42a96SStefano Zampini ierr = ISDestroy(&pcbddc->is_R_local);CHKERRQ(ierr); 1012674ae819SStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_B);CHKERRQ(ierr); 1013674ae819SStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_D);CHKERRQ(ierr); 1014674ae819SStefano Zampini ierr = VecScatterDestroy(&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 1015f4ddd8eeSStefano Zampini ierr = KSPDestroy(&pcbddc->ksp_D);CHKERRQ(ierr); 1016f4ddd8eeSStefano Zampini ierr = KSPDestroy(&pcbddc->ksp_R);CHKERRQ(ierr); 1017f4ddd8eeSStefano Zampini ierr = KSPDestroy(&pcbddc->coarse_ksp);CHKERRQ(ierr); 1018f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 1019727cdba6SStefano Zampini ierr = PetscFree(pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 10200e6343abSStefano Zampini ierr = PetscFree2(pcbddc->local_primal_ref_node,pcbddc->local_primal_ref_mult);CHKERRQ(ierr); 1021f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->global_primal_indices);CHKERRQ(ierr); 102270cf5478SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling);CHKERRQ(ierr); 10236e683305SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 102481d14e9dSStefano Zampini ierr = MatDestroy(&pcbddc->benign_change);CHKERRQ(ierr); 10250369aaf7SStefano Zampini ierr = VecDestroy(&pcbddc->benign_vec);CHKERRQ(ierr); 10268b9f24d4SStefano Zampini ierr = MatDestroy(&pcbddc->benign_original_mat);CHKERRQ(ierr); 10274f1b2e48SStefano Zampini ierr = MatDestroy(&pcbddc->benign_B0);CHKERRQ(ierr); 10288b9f24d4SStefano Zampini ierr = PetscSFDestroy(&pcbddc->benign_sf);CHKERRQ(ierr); 10294f1b2e48SStefano Zampini ierr = PetscFree3(pcbddc->benign_p0_lidx,pcbddc->benign_p0_gidx,pcbddc->benign_p0);CHKERRQ(ierr); 1030674ae819SStefano Zampini PetscFunctionReturn(0); 1031674ae819SStefano Zampini } 1032674ae819SStefano Zampini 1033674ae819SStefano Zampini #undef __FUNCT__ 1034f4ddd8eeSStefano Zampini #define __FUNCT__ "PCBDDCSetUpLocalWorkVectors" 1035f4ddd8eeSStefano Zampini PetscErrorCode PCBDDCSetUpLocalWorkVectors(PC pc) 10366bfb1811SStefano Zampini { 10376bfb1811SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 10386bfb1811SStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 10396bfb1811SStefano Zampini VecType impVecType; 10404f1b2e48SStefano Zampini PetscInt n_constraints,n_R,old_size; 10416bfb1811SStefano Zampini PetscErrorCode ierr; 10426bfb1811SStefano Zampini 10436bfb1811SStefano Zampini PetscFunctionBegin; 1044f4ddd8eeSStefano Zampini if (!pcbddc->ConstraintMatrix) { 1045019a44ceSStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"BDDC Constraint matrix has not been created"); 1046f4ddd8eeSStefano Zampini } 1047e7b262bdSStefano Zampini /* get sizes */ 10484f1b2e48SStefano Zampini n_constraints = pcbddc->local_primal_size - pcbddc->benign_n - pcbddc->n_vertices; 1049b371cd4fSStefano Zampini n_R = pcis->n - pcbddc->n_vertices; 10506bfb1811SStefano Zampini ierr = VecGetType(pcis->vec1_N,&impVecType);CHKERRQ(ierr); 1051e7b262bdSStefano Zampini /* local work vectors (try to avoid unneeded work)*/ 1052e7b262bdSStefano Zampini /* R nodes */ 1053e7b262bdSStefano Zampini old_size = -1; 1054e7b262bdSStefano Zampini if (pcbddc->vec1_R) { 1055e7b262bdSStefano Zampini ierr = VecGetSize(pcbddc->vec1_R,&old_size);CHKERRQ(ierr); 1056e7b262bdSStefano Zampini } 1057e7b262bdSStefano Zampini if (n_R != old_size) { 1058e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec1_R);CHKERRQ(ierr); 1059e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec2_R);CHKERRQ(ierr); 10606bfb1811SStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_R);CHKERRQ(ierr); 10616bfb1811SStefano Zampini ierr = VecSetSizes(pcbddc->vec1_R,PETSC_DECIDE,n_R);CHKERRQ(ierr); 10626bfb1811SStefano Zampini ierr = VecSetType(pcbddc->vec1_R,impVecType);CHKERRQ(ierr); 10636bfb1811SStefano Zampini ierr = VecDuplicate(pcbddc->vec1_R,&pcbddc->vec2_R);CHKERRQ(ierr); 1064e7b262bdSStefano Zampini } 1065e7b262bdSStefano Zampini /* local primal dofs */ 1066e7b262bdSStefano Zampini old_size = -1; 1067e7b262bdSStefano Zampini if (pcbddc->vec1_P) { 1068e7b262bdSStefano Zampini ierr = VecGetSize(pcbddc->vec1_P,&old_size);CHKERRQ(ierr); 1069e7b262bdSStefano Zampini } 1070e9189074SStefano Zampini if (pcbddc->local_primal_size != old_size) { 1071e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec1_P);CHKERRQ(ierr); 107283b7ccabSStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_P);CHKERRQ(ierr); 1073e9189074SStefano Zampini ierr = VecSetSizes(pcbddc->vec1_P,PETSC_DECIDE,pcbddc->local_primal_size);CHKERRQ(ierr); 10746bfb1811SStefano Zampini ierr = VecSetType(pcbddc->vec1_P,impVecType);CHKERRQ(ierr); 1075e7b262bdSStefano Zampini } 1076e7b262bdSStefano Zampini /* local explicit constraints */ 1077e7b262bdSStefano Zampini old_size = -1; 1078e7b262bdSStefano Zampini if (pcbddc->vec1_C) { 1079e7b262bdSStefano Zampini ierr = VecGetSize(pcbddc->vec1_C,&old_size);CHKERRQ(ierr); 1080e7b262bdSStefano Zampini } 1081e7b262bdSStefano Zampini if (n_constraints && n_constraints != old_size) { 1082e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec1_C);CHKERRQ(ierr); 108383b7ccabSStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_C);CHKERRQ(ierr); 108483b7ccabSStefano Zampini ierr = VecSetSizes(pcbddc->vec1_C,PETSC_DECIDE,n_constraints);CHKERRQ(ierr); 108583b7ccabSStefano Zampini ierr = VecSetType(pcbddc->vec1_C,impVecType);CHKERRQ(ierr); 108683b7ccabSStefano Zampini } 10876bfb1811SStefano Zampini PetscFunctionReturn(0); 10886bfb1811SStefano Zampini } 10896bfb1811SStefano Zampini 10906bfb1811SStefano Zampini #undef __FUNCT__ 109147f4ddc3SStefano Zampini #define __FUNCT__ "PCBDDCSetUpCorrection" 109247f4ddc3SStefano Zampini PetscErrorCode PCBDDCSetUpCorrection(PC pc, PetscScalar **coarse_submat_vals_n) 109388ebb749SStefano Zampini { 109425084f0cSStefano Zampini PetscErrorCode ierr; 109525084f0cSStefano Zampini /* pointers to pcis and pcbddc */ 109688ebb749SStefano Zampini PC_IS* pcis = (PC_IS*)pc->data; 109788ebb749SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 1098d62866d3SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 109925084f0cSStefano Zampini /* submatrices of local problem */ 110080677318SStefano Zampini Mat A_RV,A_VR,A_VV,local_auxmat2_R; 1101019a44ceSStefano Zampini /* submatrices of benign trick */ 1102d16cbb6bSStefano Zampini Mat B0_V = NULL; 110306656605SStefano Zampini /* submatrices of local coarse problem */ 110406656605SStefano Zampini Mat S_VV,S_CV,S_VC,S_CC; 110525084f0cSStefano Zampini /* working matrices */ 110606656605SStefano Zampini Mat C_CR; 110725084f0cSStefano Zampini /* additional working stuff */ 110806656605SStefano Zampini PC pc_R; 11094f1b2e48SStefano Zampini Mat F; 111006656605SStefano Zampini PetscBool isLU,isCHOL,isILU; 111106656605SStefano Zampini 111225084f0cSStefano Zampini PetscScalar *coarse_submat_vals; /* TODO: use a PETSc matrix */ 111306656605SStefano Zampini PetscScalar *work; 111406656605SStefano Zampini PetscInt *idx_V_B; 1115ffd830a3SStefano Zampini PetscInt lda_rhs,n,n_vertices,n_constraints,*p0_lidx_I; 111606656605SStefano Zampini PetscInt i,n_R,n_D,n_B; 1117ffd830a3SStefano Zampini 111825084f0cSStefano Zampini /* some shortcuts to scalars */ 111906656605SStefano Zampini PetscScalar one=1.0,m_one=-1.0; 112088ebb749SStefano Zampini 112188ebb749SStefano Zampini PetscFunctionBegin; 1122ffd830a3SStefano Zampini if (!pcbddc->symmetric_primal && pcbddc->benign_n) { 1123ffd830a3SStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Non-symmetric primal basis computation with benign trick not yet implemented"); 1124ffd830a3SStefano Zampini } 1125ffd830a3SStefano Zampini 1126ffd830a3SStefano Zampini /* Set Non-overlapping dimensions */ 1127b371cd4fSStefano Zampini n_vertices = pcbddc->n_vertices; 11284f1b2e48SStefano Zampini n_constraints = pcbddc->local_primal_size - pcbddc->benign_n - n_vertices; 1129b371cd4fSStefano Zampini n_B = pcis->n_B; 1130b371cd4fSStefano Zampini n_D = pcis->n - n_B; 113188ebb749SStefano Zampini n_R = pcis->n - n_vertices; 113288ebb749SStefano Zampini 113388ebb749SStefano Zampini /* vertices in boundary numbering */ 1134785e854fSJed Brown ierr = PetscMalloc1(n_vertices,&idx_V_B);CHKERRQ(ierr); 11350e6343abSStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,n_vertices,pcbddc->local_primal_ref_node,&i,idx_V_B);CHKERRQ(ierr); 113688ebb749SStefano Zampini if (i != n_vertices) { 113722d5777bSStefano Zampini SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Error in boundary numbering for BDDC vertices! %d != %d\n",n_vertices,i); 113888ebb749SStefano Zampini } 113988ebb749SStefano Zampini 114006656605SStefano Zampini /* Subdomain contribution (Non-overlapping) to coarse matrix */ 1141019a44ceSStefano Zampini ierr = PetscCalloc1(pcbddc->local_primal_size*pcbddc->local_primal_size,&coarse_submat_vals);CHKERRQ(ierr); 114206656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_vertices,n_vertices,coarse_submat_vals,&S_VV);CHKERRQ(ierr); 114306656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_VV,pcbddc->local_primal_size);CHKERRQ(ierr); 114406656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_constraints,n_vertices,coarse_submat_vals+n_vertices,&S_CV);CHKERRQ(ierr); 114506656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_CV,pcbddc->local_primal_size);CHKERRQ(ierr); 114606656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_vertices,n_constraints,coarse_submat_vals+pcbddc->local_primal_size*n_vertices,&S_VC);CHKERRQ(ierr); 114706656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_VC,pcbddc->local_primal_size);CHKERRQ(ierr); 114806656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_constraints,n_constraints,coarse_submat_vals+(pcbddc->local_primal_size+1)*n_vertices,&S_CC);CHKERRQ(ierr); 114906656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_CC,pcbddc->local_primal_size);CHKERRQ(ierr); 115006656605SStefano Zampini 115106656605SStefano Zampini /* determine if can use MatSolve routines instead of calling KSPSolve on ksp_R */ 115206656605SStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_R);CHKERRQ(ierr); 115306656605SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_R,PCLU,&isLU);CHKERRQ(ierr); 115406656605SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_R,PCILU,&isILU);CHKERRQ(ierr); 115506656605SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_R,PCCHOLESKY,&isCHOL);CHKERRQ(ierr); 1156ffd830a3SStefano Zampini lda_rhs = n_R; 115706656605SStefano Zampini if (isLU || isILU || isCHOL) { 115806656605SStefano Zampini ierr = PCFactorGetMatrix(pc_R,&F);CHKERRQ(ierr); 1159d62866d3SStefano Zampini } else if (sub_schurs->reuse_mumps) { 1160d62866d3SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 1161d62866d3SStefano Zampini MatFactorType type; 1162d62866d3SStefano Zampini 11636816873aSStefano Zampini F = reuse_mumps->F; 11646816873aSStefano Zampini ierr = MatGetFactorType(F,&type);CHKERRQ(ierr); 1165d62866d3SStefano Zampini if (type == MAT_FACTOR_CHOLESKY) isCHOL = PETSC_TRUE; 1166ffd830a3SStefano Zampini ierr = MatGetSize(F,&lda_rhs,NULL);CHKERRQ(ierr); 116706656605SStefano Zampini } else { 116806656605SStefano Zampini F = NULL; 116906656605SStefano Zampini } 117006656605SStefano Zampini 1171ffd830a3SStefano Zampini /* allocate workspace */ 1172ffd830a3SStefano Zampini n = 0; 1173ffd830a3SStefano Zampini if (n_constraints) { 1174ffd830a3SStefano Zampini n += lda_rhs*n_constraints; 1175ffd830a3SStefano Zampini } 1176ffd830a3SStefano Zampini if (n_vertices) { 1177ffd830a3SStefano Zampini n = PetscMax(2*lda_rhs*n_vertices,n); 1178ffd830a3SStefano Zampini n = PetscMax((lda_rhs+n_B)*n_vertices,n); 1179ffd830a3SStefano Zampini } 1180ffd830a3SStefano Zampini ierr = PetscMalloc1(n,&work);CHKERRQ(ierr); 1181ffd830a3SStefano Zampini 118288ebb749SStefano Zampini /* Precompute stuffs needed for preprocessing and application of BDDC*/ 118388ebb749SStefano Zampini if (n_constraints) { 118406656605SStefano Zampini Mat M1,M2,M3; 118580677318SStefano Zampini Mat auxmat; 118606656605SStefano Zampini IS is_aux; 118780677318SStefano Zampini PetscScalar *array,*array2; 118806656605SStefano Zampini 1189f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat1);CHKERRQ(ierr); 119080677318SStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat2);CHKERRQ(ierr); 119188ebb749SStefano Zampini 119225084f0cSStefano Zampini /* Extract constraints on R nodes: C_{CR} */ 119325084f0cSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,n_constraints,n_vertices,1,&is_aux);CHKERRQ(ierr); 11948ce42a96SStefano Zampini ierr = MatGetSubMatrix(pcbddc->ConstraintMatrix,is_aux,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&C_CR);CHKERRQ(ierr); 119580677318SStefano Zampini ierr = MatGetSubMatrix(pcbddc->ConstraintMatrix,is_aux,pcis->is_B_local,MAT_INITIAL_MATRIX,&auxmat);CHKERRQ(ierr); 119688ebb749SStefano Zampini 119780677318SStefano Zampini /* Assemble local_auxmat2_R = (- A_{RR}^{-1} C^T_{CR}) needed by BDDC setup */ 119880677318SStefano Zampini /* Assemble pcbddc->local_auxmat2 = R_to_B (- A_{RR}^{-1} C^T_{CR}) needed by BDDC application */ 1199ffd830a3SStefano Zampini ierr = PetscMemzero(work,lda_rhs*n_constraints*sizeof(PetscScalar));CHKERRQ(ierr); 120088ebb749SStefano Zampini for (i=0;i<n_constraints;i++) { 120106656605SStefano Zampini const PetscScalar *row_cmat_values; 120206656605SStefano Zampini const PetscInt *row_cmat_indices; 120306656605SStefano Zampini PetscInt size_of_constraint,j; 120488ebb749SStefano Zampini 120506656605SStefano Zampini ierr = MatGetRow(C_CR,i,&size_of_constraint,&row_cmat_indices,&row_cmat_values);CHKERRQ(ierr); 120606656605SStefano Zampini for (j=0;j<size_of_constraint;j++) { 1207ffd830a3SStefano Zampini work[row_cmat_indices[j]+i*lda_rhs] = -row_cmat_values[j]; 120806656605SStefano Zampini } 120906656605SStefano Zampini ierr = MatRestoreRow(C_CR,i,&size_of_constraint,&row_cmat_indices,&row_cmat_values);CHKERRQ(ierr); 121006656605SStefano Zampini } 1211ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_constraints,NULL,&local_auxmat2_R);CHKERRQ(ierr); 121206656605SStefano Zampini if (F) { 121306656605SStefano Zampini Mat B; 121406656605SStefano Zampini 1215ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_constraints,work,&B);CHKERRQ(ierr); 121680677318SStefano Zampini ierr = MatMatSolve(F,B,local_auxmat2_R);CHKERRQ(ierr); 121706656605SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 121806656605SStefano Zampini } else { 121980677318SStefano Zampini PetscScalar *marr; 122080677318SStefano Zampini 122180677318SStefano Zampini ierr = MatDenseGetArray(local_auxmat2_R,&marr);CHKERRQ(ierr); 122206656605SStefano Zampini for (i=0;i<n_constraints;i++) { 1223ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*lda_rhs);CHKERRQ(ierr); 1224ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,marr+i*lda_rhs);CHKERRQ(ierr); 122506656605SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 122606656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 122706656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 122806656605SStefano Zampini } 122980677318SStefano Zampini ierr = MatDenseRestoreArray(local_auxmat2_R,&marr);CHKERRQ(ierr); 123006656605SStefano Zampini } 123180677318SStefano Zampini if (!pcbddc->switch_static) { 123280677318SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,n_constraints,NULL,&pcbddc->local_auxmat2);CHKERRQ(ierr); 123380677318SStefano Zampini ierr = MatDenseGetArray(pcbddc->local_auxmat2,&array);CHKERRQ(ierr); 123480677318SStefano Zampini ierr = MatDenseGetArray(local_auxmat2_R,&array2);CHKERRQ(ierr); 123580677318SStefano Zampini for (i=0;i<n_constraints;i++) { 1236ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,array2+i*lda_rhs);CHKERRQ(ierr); 123780677318SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,array+i*n_B);CHKERRQ(ierr); 123880677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 123980677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 124080677318SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 124180677318SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 124280677318SStefano Zampini } 124380677318SStefano Zampini ierr = MatDenseRestoreArray(local_auxmat2_R,&array2);CHKERRQ(ierr); 124480677318SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->local_auxmat2,&array);CHKERRQ(ierr); 124580677318SStefano Zampini ierr = MatMatMult(auxmat,pcbddc->local_auxmat2,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&M3);CHKERRQ(ierr); 124680677318SStefano Zampini } else { 1247ffd830a3SStefano Zampini if (lda_rhs != n_R) { 1248ffd830a3SStefano Zampini IS dummy; 1249ffd830a3SStefano Zampini 1250ffd830a3SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,n_R,0,1,&dummy);CHKERRQ(ierr); 1251ffd830a3SStefano Zampini ierr = MatGetSubMatrix(local_auxmat2_R,dummy,dummy,MAT_INITIAL_MATRIX,&pcbddc->local_auxmat2);CHKERRQ(ierr); 1252ffd830a3SStefano Zampini ierr = ISDestroy(&dummy);CHKERRQ(ierr); 1253ffd830a3SStefano Zampini } else { 125480677318SStefano Zampini ierr = PetscObjectReference((PetscObject)local_auxmat2_R);CHKERRQ(ierr); 125580677318SStefano Zampini pcbddc->local_auxmat2 = local_auxmat2_R; 1256ffd830a3SStefano Zampini } 125725084f0cSStefano Zampini ierr = MatMatMult(C_CR,pcbddc->local_auxmat2,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&M3);CHKERRQ(ierr); 125880677318SStefano Zampini } 125980677318SStefano Zampini ierr = ISDestroy(&is_aux);CHKERRQ(ierr); 126080677318SStefano Zampini /* Assemble explicitly S_CC = ( C_{CR} A_{RR}^{-1} C^T_{CR} )^{-1} */ 126180677318SStefano Zampini ierr = MatScale(M3,m_one);CHKERRQ(ierr); 126206656605SStefano Zampini ierr = MatDuplicate(M3,MAT_DO_NOT_COPY_VALUES,&M1);CHKERRQ(ierr); 126306656605SStefano Zampini ierr = MatDuplicate(M3,MAT_DO_NOT_COPY_VALUES,&M2);CHKERRQ(ierr); 126480677318SStefano Zampini if (isCHOL) { 126580677318SStefano Zampini ierr = MatCholeskyFactor(M3,NULL,NULL);CHKERRQ(ierr); 126680677318SStefano Zampini } else { 126725084f0cSStefano Zampini ierr = MatLUFactor(M3,NULL,NULL,NULL);CHKERRQ(ierr); 126880677318SStefano Zampini } 126980677318SStefano Zampini ierr = VecSet(pcbddc->vec1_C,one);CHKERRQ(ierr); 127006656605SStefano Zampini ierr = MatDiagonalSet(M2,pcbddc->vec1_C,INSERT_VALUES);CHKERRQ(ierr); 127125084f0cSStefano Zampini ierr = MatMatSolve(M3,M2,M1);CHKERRQ(ierr); 127225084f0cSStefano Zampini ierr = MatDestroy(&M2);CHKERRQ(ierr); 127325084f0cSStefano Zampini ierr = MatDestroy(&M3);CHKERRQ(ierr); 127480677318SStefano Zampini /* Assemble local_auxmat1 = S_CC*C_{CB} needed by BDDC application in KSP and in preproc */ 127580677318SStefano Zampini ierr = MatMatMult(M1,auxmat,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&pcbddc->local_auxmat1);CHKERRQ(ierr); 127680677318SStefano Zampini ierr = MatDestroy(&auxmat);CHKERRQ(ierr); 127706656605SStefano Zampini ierr = MatCopy(M1,S_CC,SAME_NONZERO_PATTERN);CHKERRQ(ierr); /* S_CC can have a different LDA, MatMatSolve doesn't support it */ 127806656605SStefano Zampini ierr = MatDestroy(&M1);CHKERRQ(ierr); 1279f4ddd8eeSStefano Zampini } 128088ebb749SStefano Zampini /* Get submatrices from subdomain matrix */ 12814f1b2e48SStefano Zampini if (pcbddc->benign_n) { 1282d16cbb6bSStefano Zampini IS dummy; 1283d16cbb6bSStefano Zampini Mat B0_R; 1284d16cbb6bSStefano Zampini PetscReal norm; 1285d16cbb6bSStefano Zampini 12864f1b2e48SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcbddc->benign_n,0,1,&dummy);CHKERRQ(ierr); 12874f1b2e48SStefano Zampini ierr = MatGetSubMatrix(pcbddc->benign_B0,dummy,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&B0_R);CHKERRQ(ierr); 1288d16cbb6bSStefano Zampini ierr = MatNorm(B0_R,NORM_INFINITY,&norm);CHKERRQ(ierr); 1289d16cbb6bSStefano Zampini if (norm > PETSC_SMALL) { 1290d16cbb6bSStefano Zampini SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_SUP,"Benign trick can not be applied! ||B0_R|| = %f (should be numerically 0.)",norm); 1291d16cbb6bSStefano Zampini } 1292d16cbb6bSStefano Zampini ierr = MatDestroy(&B0_R);CHKERRQ(ierr); 1293d16cbb6bSStefano Zampini ierr = ISDestroy(&dummy);CHKERRQ(ierr); 1294d16cbb6bSStefano Zampini } 1295d16cbb6bSStefano Zampini 129688ebb749SStefano Zampini if (n_vertices) { 129706656605SStefano Zampini IS is_aux; 12983a50541eSStefano Zampini 12996816873aSStefano Zampini if (sub_schurs->reuse_mumps) { /* is_R_local is not sorted, ISComplement doesn't like it */ 13006816873aSStefano Zampini IS tis; 13016816873aSStefano Zampini 13026816873aSStefano Zampini ierr = ISDuplicate(pcbddc->is_R_local,&tis);CHKERRQ(ierr); 13036816873aSStefano Zampini ierr = ISSort(tis);CHKERRQ(ierr); 13046816873aSStefano Zampini ierr = ISComplement(tis,0,pcis->n,&is_aux);CHKERRQ(ierr); 13056816873aSStefano Zampini ierr = ISDestroy(&tis);CHKERRQ(ierr); 13066816873aSStefano Zampini } else { 13073a50541eSStefano Zampini ierr = ISComplement(pcbddc->is_R_local,0,pcis->n,&is_aux);CHKERRQ(ierr); 13086816873aSStefano Zampini } 13099577ea80SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,pcbddc->is_R_local,is_aux,MAT_INITIAL_MATRIX,&A_RV);CHKERRQ(ierr); 13109577ea80SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,is_aux,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&A_VR);CHKERRQ(ierr); 131104708bb6SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,is_aux,is_aux,MAT_INITIAL_MATRIX,&A_VV);CHKERRQ(ierr); 13124f1b2e48SStefano Zampini if (pcbddc->benign_n) { 1313019a44ceSStefano Zampini IS dummy; 1314019a44ceSStefano Zampini 13154f1b2e48SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcbddc->benign_n,0,1,&dummy);CHKERRQ(ierr); 13164f1b2e48SStefano Zampini ierr = MatGetSubMatrix(pcbddc->benign_B0,dummy,is_aux,MAT_INITIAL_MATRIX,&B0_V);CHKERRQ(ierr); 1317019a44ceSStefano Zampini ierr = ISDestroy(&dummy);CHKERRQ(ierr); 1318019a44ceSStefano Zampini } 131925084f0cSStefano Zampini ierr = ISDestroy(&is_aux);CHKERRQ(ierr); 132088ebb749SStefano Zampini } 132188ebb749SStefano Zampini 132288ebb749SStefano Zampini /* Matrix of coarse basis functions (local) */ 1323f4ddd8eeSStefano Zampini if (pcbddc->coarse_phi_B) { 132406656605SStefano Zampini PetscInt on_B,on_primal,on_D=n_D; 132506656605SStefano Zampini if (pcbddc->coarse_phi_D) { 132606656605SStefano Zampini ierr = MatGetSize(pcbddc->coarse_phi_D,&on_D,NULL);CHKERRQ(ierr); 132706656605SStefano Zampini } 1328f4ddd8eeSStefano Zampini ierr = MatGetSize(pcbddc->coarse_phi_B,&on_B,&on_primal);CHKERRQ(ierr); 132906656605SStefano Zampini if (on_B != n_B || on_primal != pcbddc->local_primal_size || on_D != n_D) { 133006656605SStefano Zampini PetscScalar *marray; 133106656605SStefano Zampini 133206656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&marray);CHKERRQ(ierr); 133306656605SStefano Zampini ierr = PetscFree(marray);CHKERRQ(ierr); 1334f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_B);CHKERRQ(ierr); 1335f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_B);CHKERRQ(ierr); 1336f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_D);CHKERRQ(ierr); 1337f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_D);CHKERRQ(ierr); 1338f4ddd8eeSStefano Zampini } 1339f4ddd8eeSStefano Zampini } 134006656605SStefano Zampini 1341f4ddd8eeSStefano Zampini if (!pcbddc->coarse_phi_B) { 134206656605SStefano Zampini PetscScalar *marray; 134388ebb749SStefano Zampini 134406656605SStefano Zampini n = n_B*pcbddc->local_primal_size; 13458eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 134606656605SStefano Zampini n += n_D*pcbddc->local_primal_size; 134788ebb749SStefano Zampini } 13483301b35fSStefano Zampini if (!pcbddc->symmetric_primal) { 134906656605SStefano Zampini n *= 2; 135088ebb749SStefano Zampini } 135106656605SStefano Zampini ierr = PetscCalloc1(n,&marray);CHKERRQ(ierr); 135206656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,pcbddc->local_primal_size,marray,&pcbddc->coarse_phi_B);CHKERRQ(ierr); 135306656605SStefano Zampini n = n_B*pcbddc->local_primal_size; 13548eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 135506656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_D,pcbddc->local_primal_size,marray+n,&pcbddc->coarse_phi_D);CHKERRQ(ierr); 135606656605SStefano Zampini n += n_D*pcbddc->local_primal_size; 135788ebb749SStefano Zampini } 13583301b35fSStefano Zampini if (!pcbddc->symmetric_primal) { 135906656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,pcbddc->local_primal_size,marray+n,&pcbddc->coarse_psi_B);CHKERRQ(ierr); 13608eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 136106656605SStefano Zampini n = n_B*pcbddc->local_primal_size; 136206656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_D,pcbddc->local_primal_size,marray+n,&pcbddc->coarse_psi_D);CHKERRQ(ierr); 136388ebb749SStefano Zampini } 136488ebb749SStefano Zampini } else { 1365c0553b1fSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->coarse_phi_B);CHKERRQ(ierr); 1366c0553b1fSStefano Zampini pcbddc->coarse_psi_B = pcbddc->coarse_phi_B; 13671b968477SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 1368c0553b1fSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->coarse_phi_D);CHKERRQ(ierr); 1369c0553b1fSStefano Zampini pcbddc->coarse_psi_D = pcbddc->coarse_phi_D; 1370c0553b1fSStefano Zampini } 137188ebb749SStefano Zampini } 137206656605SStefano Zampini } 1373019a44ceSStefano Zampini 137406656605SStefano Zampini /* We are now ready to evaluate coarse basis functions and subdomain contribution to coarse problem */ 13754f1b2e48SStefano Zampini p0_lidx_I = NULL; 13764f1b2e48SStefano Zampini if (pcbddc->benign_n && (pcbddc->switch_static || pcbddc->dbg_flag)) { 1377d12edf2fSStefano Zampini const PetscInt *idxs; 1378d12edf2fSStefano Zampini 1379d12edf2fSStefano Zampini ierr = ISGetIndices(pcis->is_I_local,&idxs);CHKERRQ(ierr); 13804f1b2e48SStefano Zampini ierr = PetscMalloc1(pcbddc->benign_n,&p0_lidx_I);CHKERRQ(ierr); 13814f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 13824f1b2e48SStefano Zampini ierr = PetscFindInt(pcbddc->benign_p0_lidx[i],pcis->n-pcis->n_B,idxs,&p0_lidx_I[i]);CHKERRQ(ierr); 13834f1b2e48SStefano Zampini } 1384d12edf2fSStefano Zampini ierr = ISRestoreIndices(pcis->is_I_local,&idxs);CHKERRQ(ierr); 1385d12edf2fSStefano Zampini } 1386d16cbb6bSStefano Zampini 138706656605SStefano Zampini /* vertices */ 138806656605SStefano Zampini if (n_vertices) { 138916f15bc4SStefano Zampini 1390ffd830a3SStefano Zampini ierr = MatConvert(A_VV,MATDENSE,MAT_REUSE_MATRIX,&A_VV);CHKERRQ(ierr); 139104708bb6SStefano Zampini 139216f15bc4SStefano Zampini if (n_R) { 139306656605SStefano Zampini Mat A_RRmA_RV,S_VVt; /* S_VVt with LDA=N */ 139406656605SStefano Zampini PetscBLASInt B_N,B_one = 1; 139516f15bc4SStefano Zampini PetscScalar *x,*y; 139604708bb6SStefano Zampini PetscBool isseqaij; 139706656605SStefano Zampini 139821eccb56SStefano Zampini ierr = MatScale(A_RV,m_one);CHKERRQ(ierr); 1399ffd830a3SStefano Zampini if (lda_rhs == n_R) { 1400ffd830a3SStefano Zampini ierr = MatConvert(A_RV,MATDENSE,MAT_REUSE_MATRIX,&A_RV);CHKERRQ(ierr); 1401ffd830a3SStefano Zampini } else { 1402ffd830a3SStefano Zampini Mat A_RVt; 1403ffd830a3SStefano Zampini PetscScalar *array; 1404ffd830a3SStefano Zampini 1405ffd830a3SStefano Zampini /* TODO optimize */ 1406ffd830a3SStefano Zampini ierr = MatConvert(A_RV,MATDENSE,MAT_INITIAL_MATRIX,&A_RVt);CHKERRQ(ierr); 1407ffd830a3SStefano Zampini ierr = MatDestroy(&A_RV);CHKERRQ(ierr); 1408ffd830a3SStefano Zampini ierr = PetscMemzero(work+lda_rhs*n_vertices,lda_rhs*n_vertices*sizeof(PetscScalar));CHKERRQ(ierr); 1409ffd830a3SStefano Zampini ierr = MatDenseGetArray(A_RVt,&array);CHKERRQ(ierr); 1410ffd830a3SStefano Zampini for (i=0;i<n_vertices;i++) { 1411ffd830a3SStefano Zampini ierr = PetscMemcpy(work+(n_vertices+i)*lda_rhs,array+i*n_R,n_R*sizeof(PetscScalar));CHKERRQ(ierr); 1412ffd830a3SStefano Zampini } 1413ffd830a3SStefano Zampini ierr = MatDenseRestoreArray(A_RVt,&array);CHKERRQ(ierr); 1414ffd830a3SStefano Zampini ierr = MatDestroy(&A_RVt);CHKERRQ(ierr); 1415ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_vertices,work+lda_rhs*n_vertices,&A_RV);CHKERRQ(ierr); 1416ffd830a3SStefano Zampini } 1417ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_vertices,work,&A_RRmA_RV);CHKERRQ(ierr); 14186816873aSStefano Zampini if (F) { /* TODO could be optimized for symmetric problems */ 141906656605SStefano Zampini ierr = MatMatSolve(F,A_RV,A_RRmA_RV);CHKERRQ(ierr); 142006656605SStefano Zampini } else { 142106656605SStefano Zampini ierr = MatDenseGetArray(A_RV,&y);CHKERRQ(ierr); 142206656605SStefano Zampini for (i=0;i<n_vertices;i++) { 1423ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,y+i*lda_rhs);CHKERRQ(ierr); 1424ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,work+i*lda_rhs);CHKERRQ(ierr); 142506656605SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 142606656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 142706656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 142806656605SStefano Zampini } 142906656605SStefano Zampini ierr = MatDenseRestoreArray(A_RV,&y);CHKERRQ(ierr); 143006656605SStefano Zampini } 143180677318SStefano Zampini ierr = MatDestroy(&A_RV);CHKERRQ(ierr); 1432ffd830a3SStefano Zampini /* S_VV and S_CV */ 143306656605SStefano Zampini if (n_constraints) { 143406656605SStefano Zampini Mat B; 143580677318SStefano Zampini 1436ffd830a3SStefano Zampini ierr = PetscMemzero(work+lda_rhs*n_vertices,n_B*n_vertices*sizeof(PetscScalar));CHKERRQ(ierr); 143780677318SStefano Zampini for (i=0;i<n_vertices;i++) { 1438ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*lda_rhs);CHKERRQ(ierr); 1439ffd830a3SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,work+lda_rhs*n_vertices+i*n_B);CHKERRQ(ierr); 144080677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 144180677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 144280677318SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 144380677318SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 144480677318SStefano Zampini } 1445ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,n_vertices,work+lda_rhs*n_vertices,&B);CHKERRQ(ierr); 144680677318SStefano Zampini ierr = MatMatMult(pcbddc->local_auxmat1,B,MAT_REUSE_MATRIX,PETSC_DEFAULT,&S_CV);CHKERRQ(ierr); 144780677318SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 1448ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_vertices,work+lda_rhs*n_vertices,&B);CHKERRQ(ierr); 144980677318SStefano Zampini ierr = MatMatMult(local_auxmat2_R,S_CV,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B);CHKERRQ(ierr); 145006656605SStefano Zampini ierr = MatScale(S_CV,m_one);CHKERRQ(ierr); 1451ffd830a3SStefano Zampini ierr = PetscBLASIntCast(lda_rhs*n_vertices,&B_N);CHKERRQ(ierr); 1452ffd830a3SStefano Zampini PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&B_N,&one,work+lda_rhs*n_vertices,&B_one,work,&B_one)); 145306656605SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 145406656605SStefano Zampini } 145504708bb6SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)A_VR,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 145604708bb6SStefano Zampini if (!isseqaij) { /* MatMatMult with SEQ(S)BAIJ below will raise an error */ 145704708bb6SStefano Zampini ierr = MatConvert(A_VR,MATSEQAIJ,MAT_REUSE_MATRIX,&A_VR);CHKERRQ(ierr); 145804708bb6SStefano Zampini } 1459ffd830a3SStefano Zampini if (lda_rhs != n_R) { 1460ffd830a3SStefano Zampini ierr = MatDestroy(&A_RRmA_RV);CHKERRQ(ierr); 1461ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_vertices,work,&A_RRmA_RV);CHKERRQ(ierr); 1462ffd830a3SStefano Zampini ierr = MatSeqDenseSetLDA(A_RRmA_RV,lda_rhs);CHKERRQ(ierr); 1463ffd830a3SStefano Zampini } 146406656605SStefano Zampini ierr = MatMatMult(A_VR,A_RRmA_RV,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&S_VVt);CHKERRQ(ierr); 146580677318SStefano Zampini ierr = MatDestroy(&A_RRmA_RV);CHKERRQ(ierr); 146606656605SStefano Zampini ierr = PetscBLASIntCast(n_vertices*n_vertices,&B_N);CHKERRQ(ierr); 146706656605SStefano Zampini ierr = MatDenseGetArray(A_VV,&x);CHKERRQ(ierr); 146806656605SStefano Zampini ierr = MatDenseGetArray(S_VVt,&y);CHKERRQ(ierr); 146906656605SStefano Zampini PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&B_N,&one,x,&B_one,y,&B_one)); 147006656605SStefano Zampini ierr = MatDenseRestoreArray(A_VV,&x);CHKERRQ(ierr); 147106656605SStefano Zampini ierr = MatDenseRestoreArray(S_VVt,&y);CHKERRQ(ierr); 147206656605SStefano Zampini ierr = MatCopy(S_VVt,S_VV,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 1473d16cbb6bSStefano Zampini ierr = MatDestroy(&S_VVt);CHKERRQ(ierr); 1474019a44ceSStefano Zampini } else { 1475d16cbb6bSStefano Zampini ierr = MatCopy(A_VV,S_VV,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 1476d16cbb6bSStefano Zampini } 14774f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 1478019a44ceSStefano Zampini const PetscScalar *vals; 1479019a44ceSStefano Zampini const PetscInt *idxs; 14804f1b2e48SStefano Zampini PetscInt n,j,primal_idx; 1481019a44ceSStefano Zampini 14824f1b2e48SStefano Zampini ierr = MatGetRow(B0_V,i,&n,&idxs,&vals);CHKERRQ(ierr); 14834f1b2e48SStefano Zampini primal_idx = pcbddc->local_primal_size - pcbddc->benign_n + i; 1484d16cbb6bSStefano Zampini for (j=0;j<n;j++) { 14854f1b2e48SStefano Zampini coarse_submat_vals[primal_idx*pcbddc->local_primal_size+idxs[j]] = vals[j]; 14864f1b2e48SStefano Zampini coarse_submat_vals[idxs[j]*pcbddc->local_primal_size+primal_idx] = vals[j]; 1487019a44ceSStefano Zampini } 14884f1b2e48SStefano Zampini ierr = MatRestoreRow(B0_V,i,&n,&idxs,&vals);CHKERRQ(ierr); 148916f15bc4SStefano Zampini } 149021eccb56SStefano Zampini ierr = MatDestroy(&A_VV);CHKERRQ(ierr); 1491d16cbb6bSStefano Zampini 149206656605SStefano Zampini /* coarse basis functions */ 149306656605SStefano Zampini for (i=0;i<n_vertices;i++) { 149416f15bc4SStefano Zampini PetscScalar *y; 149516f15bc4SStefano Zampini 1496ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+lda_rhs*i);CHKERRQ(ierr); 149706656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 149806656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*i);CHKERRQ(ierr); 149906656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 150006656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 150106656605SStefano Zampini y[n_B*i+idx_V_B[i]] = 1.0; 150206656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 150306656605SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 150406656605SStefano Zampini 150506656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 15064f1b2e48SStefano Zampini PetscInt j; 15074f1b2e48SStefano Zampini 150806656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 150906656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*i);CHKERRQ(ierr); 151006656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 151106656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 151206656605SStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 15134f1b2e48SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) y[n_D*i+p0_lidx_I[j]] = 0.0; 151406656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 151506656605SStefano Zampini } 151606656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 151706656605SStefano Zampini } 151804708bb6SStefano Zampini /* if n_R == 0 the object is not destroyed */ 151904708bb6SStefano Zampini ierr = MatDestroy(&A_RV);CHKERRQ(ierr); 152006656605SStefano Zampini } 152106656605SStefano Zampini 152206656605SStefano Zampini if (n_constraints) { 152306656605SStefano Zampini Mat B; 152406656605SStefano Zampini 1525ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_constraints,work,&B);CHKERRQ(ierr); 152606656605SStefano Zampini ierr = MatScale(S_CC,m_one);CHKERRQ(ierr); 152780677318SStefano Zampini ierr = MatMatMult(local_auxmat2_R,S_CC,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B);CHKERRQ(ierr); 152806656605SStefano Zampini ierr = MatScale(S_CC,m_one);CHKERRQ(ierr); 152906656605SStefano Zampini if (n_vertices) { 153080677318SStefano Zampini if (isCHOL) { /* if we can solve the interior problem with cholesky, we should also be fine with transposing here */ 153180677318SStefano Zampini ierr = MatTranspose(S_CV,MAT_REUSE_MATRIX,&S_VC);CHKERRQ(ierr); 153280677318SStefano Zampini } else { 153380677318SStefano Zampini Mat S_VCt; 153480677318SStefano Zampini 1535ffd830a3SStefano Zampini if (lda_rhs != n_R) { 1536ffd830a3SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 1537ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_vertices,work,&B);CHKERRQ(ierr); 1538ffd830a3SStefano Zampini ierr = MatSeqDenseSetLDA(B,lda_rhs);CHKERRQ(ierr); 1539ffd830a3SStefano Zampini } 154080677318SStefano Zampini ierr = MatMatMult(A_VR,B,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&S_VCt);CHKERRQ(ierr); 154180677318SStefano Zampini ierr = MatCopy(S_VCt,S_VC,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 154280677318SStefano Zampini ierr = MatDestroy(&S_VCt);CHKERRQ(ierr); 154380677318SStefano Zampini } 154406656605SStefano Zampini } 154506656605SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 154606656605SStefano Zampini /* coarse basis functions */ 154706656605SStefano Zampini for (i=0;i<n_constraints;i++) { 154806656605SStefano Zampini PetscScalar *y; 154906656605SStefano Zampini 1550ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+lda_rhs*i);CHKERRQ(ierr); 155106656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 155206656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*(i+n_vertices));CHKERRQ(ierr); 155306656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 155406656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 155506656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 155606656605SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 155706656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 15584f1b2e48SStefano Zampini PetscInt j; 15594f1b2e48SStefano Zampini 156006656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 156106656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*(i+n_vertices));CHKERRQ(ierr); 156206656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 156306656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 156406656605SStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 15654f1b2e48SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) y[n_D*i+p0_lidx_I[j]] = 0.0; 156606656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 156706656605SStefano Zampini } 156806656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 156906656605SStefano Zampini } 157006656605SStefano Zampini } 157180677318SStefano Zampini if (n_constraints) { 157280677318SStefano Zampini ierr = MatDestroy(&local_auxmat2_R);CHKERRQ(ierr); 157380677318SStefano Zampini } 15744f1b2e48SStefano Zampini ierr = PetscFree(p0_lidx_I);CHKERRQ(ierr); 1575019a44ceSStefano Zampini ierr = MatDestroy(&B0_V);CHKERRQ(ierr); 1576019a44ceSStefano Zampini 157706656605SStefano Zampini /* compute other basis functions for non-symmetric problems */ 15783301b35fSStefano Zampini if (!pcbddc->symmetric_primal) { 1579ffd830a3SStefano Zampini Mat B_V=NULL,B_C=NULL; 1580ffd830a3SStefano Zampini PetscScalar *marray; 158106656605SStefano Zampini 158206656605SStefano Zampini if (n_constraints) { 1583ffd830a3SStefano Zampini Mat S_CCT,C_CRT; 158406656605SStefano Zampini 1585ffd830a3SStefano Zampini ierr = MatTranspose(C_CR,MAT_INITIAL_MATRIX,&C_CRT);CHKERRQ(ierr); 158606656605SStefano Zampini ierr = MatTranspose(S_CC,MAT_INITIAL_MATRIX,&S_CCT);CHKERRQ(ierr); 1587ffd830a3SStefano Zampini ierr = MatMatMult(C_CRT,S_CCT,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&B_C);CHKERRQ(ierr); 158816f15bc4SStefano Zampini ierr = MatDestroy(&S_CCT);CHKERRQ(ierr); 158906656605SStefano Zampini if (n_vertices) { 1590ffd830a3SStefano Zampini Mat S_VCT; 159106656605SStefano Zampini 159206656605SStefano Zampini ierr = MatTranspose(S_VC,MAT_INITIAL_MATRIX,&S_VCT);CHKERRQ(ierr); 1593ffd830a3SStefano Zampini ierr = MatMatMult(C_CRT,S_VCT,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&B_V);CHKERRQ(ierr); 159416f15bc4SStefano Zampini ierr = MatDestroy(&S_VCT);CHKERRQ(ierr); 159506656605SStefano Zampini } 1596ffd830a3SStefano Zampini ierr = MatDestroy(&C_CRT);CHKERRQ(ierr); 159706656605SStefano Zampini } 159816f15bc4SStefano Zampini if (n_vertices && n_R) { 1599ffd830a3SStefano Zampini PetscScalar *av,*marray; 1600ffd830a3SStefano Zampini const PetscInt *xadj,*adjncy; 1601ffd830a3SStefano Zampini PetscInt n; 1602ffd830a3SStefano Zampini PetscBool flg_row; 160306656605SStefano Zampini 1604ffd830a3SStefano Zampini /* B_V = B_V - A_VR^T */ 1605ffd830a3SStefano Zampini ierr = MatConvert(A_VR,MATSEQAIJ,MAT_REUSE_MATRIX,&A_VR);CHKERRQ(ierr); 1606ffd830a3SStefano Zampini ierr = MatGetRowIJ(A_VR,0,PETSC_FALSE,PETSC_FALSE,&n,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 1607ffd830a3SStefano Zampini ierr = MatSeqAIJGetArray(A_VR,&av);CHKERRQ(ierr); 1608ffd830a3SStefano Zampini ierr = MatDenseGetArray(B_V,&marray);CHKERRQ(ierr); 1609ffd830a3SStefano Zampini for (i=0;i<n;i++) { 1610ffd830a3SStefano Zampini PetscInt j; 1611ffd830a3SStefano Zampini for (j=xadj[i];j<xadj[i+1];j++) marray[i*n_R + adjncy[j]] -= av[j]; 1612ffd830a3SStefano Zampini } 1613ffd830a3SStefano Zampini ierr = MatDenseRestoreArray(B_V,&marray);CHKERRQ(ierr); 1614ffd830a3SStefano Zampini ierr = MatRestoreRowIJ(A_VR,0,PETSC_FALSE,PETSC_FALSE,&n,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 1615ffd830a3SStefano Zampini ierr = MatDestroy(&A_VR);CHKERRQ(ierr); 161606656605SStefano Zampini } 161706656605SStefano Zampini 1618ffd830a3SStefano Zampini /* currently there's no support for MatTransposeMatSolve(F,B,X) */ 1619ffd830a3SStefano Zampini ierr = MatDenseGetArray(B_V,&marray);CHKERRQ(ierr); 1620ffd830a3SStefano Zampini for (i=0;i<n_vertices;i++) { 1621ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,marray+i*n_R);CHKERRQ(ierr); 1622ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,work+i*n_R);CHKERRQ(ierr); 162306656605SStefano Zampini ierr = KSPSolveTranspose(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 162406656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 162506656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 162606656605SStefano Zampini } 1627ffd830a3SStefano Zampini ierr = MatDenseRestoreArray(B_V,&marray);CHKERRQ(ierr); 1628ffd830a3SStefano Zampini ierr = MatDenseGetArray(B_C,&marray);CHKERRQ(ierr); 1629ffd830a3SStefano Zampini for (i=n_vertices;i<n_constraints+n_vertices;i++) { 1630ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,marray+(i-n_vertices)*n_R);CHKERRQ(ierr); 1631ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,work+i*n_R);CHKERRQ(ierr); 1632ffd830a3SStefano Zampini ierr = KSPSolveTranspose(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 1633ffd830a3SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 1634ffd830a3SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 163506656605SStefano Zampini } 1636ffd830a3SStefano Zampini ierr = MatDenseRestoreArray(B_C,&marray);CHKERRQ(ierr); 163706656605SStefano Zampini /* coarse basis functions */ 163806656605SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 163906656605SStefano Zampini PetscScalar *y; 164006656605SStefano Zampini 1641ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*n_R);CHKERRQ(ierr); 164206656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_psi_B,&y);CHKERRQ(ierr); 164306656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*i);CHKERRQ(ierr); 164406656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 164506656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 164606656605SStefano Zampini if (i<n_vertices) { 164706656605SStefano Zampini y[n_B*i+idx_V_B[i]] = 1.0; 164806656605SStefano Zampini } 164906656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_psi_B,&y);CHKERRQ(ierr); 165006656605SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 165106656605SStefano Zampini 165206656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 165306656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_psi_D,&y);CHKERRQ(ierr); 165406656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*i);CHKERRQ(ierr); 165506656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 165606656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 165706656605SStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 165806656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_psi_D,&y);CHKERRQ(ierr); 165906656605SStefano Zampini } 166006656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 166106656605SStefano Zampini } 1662ffd830a3SStefano Zampini ierr = MatDestroy(&B_V);CHKERRQ(ierr); 1663ffd830a3SStefano Zampini ierr = MatDestroy(&B_C);CHKERRQ(ierr); 166406656605SStefano Zampini } 1665d62866d3SStefano Zampini /* free memory */ 166688ebb749SStefano Zampini ierr = PetscFree(idx_V_B);CHKERRQ(ierr); 166706656605SStefano Zampini ierr = MatDestroy(&S_VV);CHKERRQ(ierr); 166806656605SStefano Zampini ierr = MatDestroy(&S_CV);CHKERRQ(ierr); 166906656605SStefano Zampini ierr = MatDestroy(&S_VC);CHKERRQ(ierr); 167006656605SStefano Zampini ierr = MatDestroy(&S_CC);CHKERRQ(ierr); 1671d62866d3SStefano Zampini ierr = PetscFree(work);CHKERRQ(ierr); 1672d62866d3SStefano Zampini if (n_vertices) { 1673d62866d3SStefano Zampini ierr = MatDestroy(&A_VR);CHKERRQ(ierr); 1674d62866d3SStefano Zampini } 1675d62866d3SStefano Zampini if (n_constraints) { 1676d62866d3SStefano Zampini ierr = MatDestroy(&C_CR);CHKERRQ(ierr); 1677d62866d3SStefano Zampini } 167888ebb749SStefano Zampini /* Checking coarse_sub_mat and coarse basis functios */ 167988ebb749SStefano Zampini /* Symmetric case : It should be \Phi^{(j)^T} A^{(j)} \Phi^{(j)}=coarse_sub_mat */ 168088ebb749SStefano Zampini /* Non-symmetric case : It should be \Psi^{(j)^T} A^{(j)} \Phi^{(j)}=coarse_sub_mat */ 1681d12edf2fSStefano Zampini if (pcbddc->dbg_flag) { 168288ebb749SStefano Zampini Mat coarse_sub_mat; 168325084f0cSStefano Zampini Mat AUXMAT,TM1,TM2,TM3,TM4; 168488ebb749SStefano Zampini Mat coarse_phi_D,coarse_phi_B; 168588ebb749SStefano Zampini Mat coarse_psi_D,coarse_psi_B; 168688ebb749SStefano Zampini Mat A_II,A_BB,A_IB,A_BI; 16878bec7fa6SStefano Zampini Mat C_B,CPHI; 16888bec7fa6SStefano Zampini IS is_dummy; 16898bec7fa6SStefano Zampini Vec mones; 169088ebb749SStefano Zampini MatType checkmattype=MATSEQAIJ; 169188ebb749SStefano Zampini PetscReal real_value; 169288ebb749SStefano Zampini 169388ebb749SStefano Zampini ierr = MatConvert(pcis->A_II,checkmattype,MAT_INITIAL_MATRIX,&A_II);CHKERRQ(ierr); 169488ebb749SStefano Zampini ierr = MatConvert(pcis->A_IB,checkmattype,MAT_INITIAL_MATRIX,&A_IB);CHKERRQ(ierr); 169588ebb749SStefano Zampini ierr = MatConvert(pcis->A_BI,checkmattype,MAT_INITIAL_MATRIX,&A_BI);CHKERRQ(ierr); 169688ebb749SStefano Zampini ierr = MatConvert(pcis->A_BB,checkmattype,MAT_INITIAL_MATRIX,&A_BB);CHKERRQ(ierr); 169788ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_phi_D,checkmattype,MAT_INITIAL_MATRIX,&coarse_phi_D);CHKERRQ(ierr); 169888ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_phi_B,checkmattype,MAT_INITIAL_MATRIX,&coarse_phi_B);CHKERRQ(ierr); 1699ffd830a3SStefano Zampini if (!pcbddc->symmetric_primal) { 170088ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_psi_D,checkmattype,MAT_INITIAL_MATRIX,&coarse_psi_D);CHKERRQ(ierr); 170188ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_psi_B,checkmattype,MAT_INITIAL_MATRIX,&coarse_psi_B);CHKERRQ(ierr); 170288ebb749SStefano Zampini } 170388ebb749SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,pcbddc->local_primal_size,pcbddc->local_primal_size,coarse_submat_vals,&coarse_sub_mat);CHKERRQ(ierr); 170488ebb749SStefano Zampini 170525084f0cSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 17063301b35fSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Check coarse sub mat computation (symmetric %d)\n",pcbddc->symmetric_primal);CHKERRQ(ierr); 170725084f0cSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1708ffd830a3SStefano Zampini if (!pcbddc->symmetric_primal) { 170988ebb749SStefano Zampini ierr = MatMatMult(A_II,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 171088ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM1);CHKERRQ(ierr); 171188ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 171288ebb749SStefano Zampini ierr = MatMatMult(A_BB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 171388ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM2);CHKERRQ(ierr); 171488ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 171588ebb749SStefano Zampini ierr = MatMatMult(A_IB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 171688ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM3);CHKERRQ(ierr); 171788ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 171888ebb749SStefano Zampini ierr = MatMatMult(A_BI,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 171988ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM4);CHKERRQ(ierr); 172088ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 172188ebb749SStefano Zampini } else { 172288ebb749SStefano Zampini ierr = MatPtAP(A_II,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&TM1);CHKERRQ(ierr); 172388ebb749SStefano Zampini ierr = MatPtAP(A_BB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&TM2);CHKERRQ(ierr); 172488ebb749SStefano Zampini ierr = MatMatMult(A_IB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 172588ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_phi_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_phi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM4);CHKERRQ(ierr); 172988ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 173088ebb749SStefano Zampini } 173188ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM2,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 173288ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM3,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 173388ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM4,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 173488ebb749SStefano Zampini ierr = MatConvert(TM1,MATSEQDENSE,MAT_REUSE_MATRIX,&TM1);CHKERRQ(ierr); 17354f1b2e48SStefano Zampini if (pcbddc->benign_n) { 1736d12edf2fSStefano Zampini Mat B0_I,B0_B,B0_BPHI,B0_IPHI; 1737d12edf2fSStefano Zampini PetscScalar *data,*data2; 17384f1b2e48SStefano Zampini PetscInt j; 1739d12edf2fSStefano Zampini 17404f1b2e48SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcbddc->benign_n,0,1,&is_dummy);CHKERRQ(ierr); 17414f1b2e48SStefano Zampini ierr = MatGetSubMatrix(pcbddc->benign_B0,is_dummy,pcis->is_B_local,MAT_INITIAL_MATRIX,&B0_B); 17424f1b2e48SStefano Zampini ierr = MatGetSubMatrix(pcbddc->benign_B0,is_dummy,pcis->is_I_local,MAT_INITIAL_MATRIX,&B0_I); 1743d12edf2fSStefano Zampini ierr = MatMatMult(B0_B,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&B0_BPHI);CHKERRQ(ierr); 1744d12edf2fSStefano Zampini ierr = MatConvert(B0_BPHI,MATSEQDENSE,MAT_REUSE_MATRIX,&B0_BPHI);CHKERRQ(ierr); 1745d12edf2fSStefano Zampini ierr = MatDenseGetArray(TM1,&data);CHKERRQ(ierr); 1746d12edf2fSStefano Zampini ierr = MatDenseGetArray(B0_BPHI,&data2);CHKERRQ(ierr); 17474f1b2e48SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) { 17484f1b2e48SStefano Zampini PetscInt primal_idx = pcbddc->local_primal_size - pcbddc->benign_n + j; 1749d12edf2fSStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 17504f1b2e48SStefano Zampini data[primal_idx*pcbddc->local_primal_size+i] += data2[i*pcbddc->benign_n+j]; 17514f1b2e48SStefano Zampini data[i*pcbddc->local_primal_size+primal_idx] += data2[i*pcbddc->benign_n+j]; 17524f1b2e48SStefano Zampini } 1753d12edf2fSStefano Zampini } 1754d12edf2fSStefano Zampini ierr = MatDenseRestoreArray(TM1,&data);CHKERRQ(ierr); 1755d12edf2fSStefano Zampini ierr = MatDenseRestoreArray(B0_BPHI,&data2);CHKERRQ(ierr); 1756d12edf2fSStefano Zampini ierr = MatDestroy(&B0_B);CHKERRQ(ierr); 1757d12edf2fSStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 1758d12edf2fSStefano Zampini ierr = MatDestroy(&B0_BPHI);CHKERRQ(ierr); 1759d12edf2fSStefano Zampini ierr = MatMatMult(B0_I,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&B0_IPHI);CHKERRQ(ierr); 1760d12edf2fSStefano Zampini ierr = MatDestroy(&B0_I);CHKERRQ(ierr); 1761d12edf2fSStefano Zampini ierr = MatNorm(B0_IPHI,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 1762d12edf2fSStefano Zampini ierr = MatDestroy(&B0_IPHI);CHKERRQ(ierr); 1763d12edf2fSStefano Zampini } 1764d12edf2fSStefano Zampini #if 0 1765d12edf2fSStefano Zampini { 1766d12edf2fSStefano Zampini PetscViewer viewer; 1767d12edf2fSStefano Zampini char filename[256]; 1768ffd830a3SStefano Zampini sprintf(filename,"details_local_coarse_mat%d_level%d.m",PetscGlobalRank,pcbddc->current_level); 1769d12edf2fSStefano Zampini ierr = PetscViewerASCIIOpen(PETSC_COMM_SELF,filename,&viewer);CHKERRQ(ierr); 1770d12edf2fSStefano Zampini ierr = PetscViewerSetFormat(viewer,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 1771ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)coarse_sub_mat,"computed");CHKERRQ(ierr); 1772ffd830a3SStefano Zampini ierr = MatView(coarse_sub_mat,viewer);CHKERRQ(ierr); 1773ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)TM1,"projected");CHKERRQ(ierr); 1774d12edf2fSStefano Zampini ierr = MatView(TM1,viewer);CHKERRQ(ierr); 1775ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)pcbddc->coarse_phi_B,"phi_B");CHKERRQ(ierr); 1776ffd830a3SStefano Zampini ierr = MatView(pcbddc->coarse_phi_B,viewer);CHKERRQ(ierr); 1777ffd830a3SStefano Zampini if (pcbddc->coarse_phi_D) { 1778ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)pcbddc->coarse_phi_D,"phi_D");CHKERRQ(ierr); 1779ffd830a3SStefano Zampini ierr = MatView(pcbddc->coarse_phi_D,viewer);CHKERRQ(ierr); 1780ffd830a3SStefano Zampini } 1781ffd830a3SStefano Zampini if (pcbddc->coarse_psi_B) { 1782ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)pcbddc->coarse_psi_B,"psi_B");CHKERRQ(ierr); 1783ffd830a3SStefano Zampini ierr = MatView(pcbddc->coarse_psi_B,viewer);CHKERRQ(ierr); 1784ffd830a3SStefano Zampini } 1785ffd830a3SStefano Zampini if (pcbddc->coarse_phi_B) { 1786ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)pcbddc->coarse_psi_D,"psi_D");CHKERRQ(ierr); 1787ffd830a3SStefano Zampini ierr = MatView(pcbddc->coarse_psi_D,viewer);CHKERRQ(ierr); 1788ffd830a3SStefano Zampini } 1789d12edf2fSStefano Zampini ierr = PetscViewerDestroy(&viewer);CHKERRQ(ierr); 1790d12edf2fSStefano Zampini } 1791d12edf2fSStefano Zampini #endif 179281d9aea3SBarry Smith ierr = MatAXPY(TM1,m_one,coarse_sub_mat,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 17938bec7fa6SStefano Zampini ierr = MatNorm(TM1,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 17941575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 179506656605SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d matrix error % 1.14e\n",PetscGlobalRank,real_value);CHKERRQ(ierr); 17968bec7fa6SStefano Zampini 17978bec7fa6SStefano Zampini /* check constraints */ 17984f1b2e48SStefano Zampini if (!pcbddc->benign_n) { /* TODO: add benign case */ 17998bec7fa6SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcbddc->local_primal_size,0,1,&is_dummy);CHKERRQ(ierr); 18008bec7fa6SStefano Zampini ierr = MatGetSubMatrix(pcbddc->ConstraintMatrix,is_dummy,pcis->is_B_local,MAT_INITIAL_MATRIX,&C_B); 18018bec7fa6SStefano Zampini ierr = MatMatMult(C_B,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&CPHI);CHKERRQ(ierr); 18028bec7fa6SStefano Zampini ierr = MatCreateVecs(CPHI,&mones,NULL);CHKERRQ(ierr); 18038bec7fa6SStefano Zampini ierr = VecSet(mones,-1.0);CHKERRQ(ierr); 18048bec7fa6SStefano Zampini ierr = MatDiagonalSet(CPHI,mones,ADD_VALUES);CHKERRQ(ierr); 18058bec7fa6SStefano Zampini ierr = MatNorm(CPHI,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 1806bdae7319SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d phi constraints error % 1.14e\n",PetscGlobalRank,real_value);CHKERRQ(ierr); 1807ffd830a3SStefano Zampini if (!pcbddc->symmetric_primal) { 1808bdae7319SStefano Zampini ierr = MatMatMult(C_B,coarse_psi_B,MAT_REUSE_MATRIX,1.0,&CPHI);CHKERRQ(ierr); 1809bdae7319SStefano Zampini ierr = VecSet(mones,-1.0);CHKERRQ(ierr); 1810bdae7319SStefano Zampini ierr = MatDiagonalSet(CPHI,mones,ADD_VALUES);CHKERRQ(ierr); 1811bdae7319SStefano Zampini ierr = MatNorm(CPHI,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 1812bdae7319SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d psi constraints error % 1.14e\n",PetscGlobalRank,real_value);CHKERRQ(ierr); 181388ebb749SStefano Zampini } 18148bec7fa6SStefano Zampini ierr = MatDestroy(&C_B);CHKERRQ(ierr); 18158bec7fa6SStefano Zampini ierr = MatDestroy(&CPHI);CHKERRQ(ierr); 18168bec7fa6SStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 18178bec7fa6SStefano Zampini ierr = VecDestroy(&mones);CHKERRQ(ierr); 1818d12edf2fSStefano Zampini } 181925084f0cSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 182088ebb749SStefano Zampini ierr = MatDestroy(&A_II);CHKERRQ(ierr); 182188ebb749SStefano Zampini ierr = MatDestroy(&A_BB);CHKERRQ(ierr); 182288ebb749SStefano Zampini ierr = MatDestroy(&A_IB);CHKERRQ(ierr); 182388ebb749SStefano Zampini ierr = MatDestroy(&A_BI);CHKERRQ(ierr); 182488ebb749SStefano Zampini ierr = MatDestroy(&TM1);CHKERRQ(ierr); 182588ebb749SStefano Zampini ierr = MatDestroy(&TM2);CHKERRQ(ierr); 182688ebb749SStefano Zampini ierr = MatDestroy(&TM3);CHKERRQ(ierr); 182788ebb749SStefano Zampini ierr = MatDestroy(&TM4);CHKERRQ(ierr); 182888ebb749SStefano Zampini ierr = MatDestroy(&coarse_phi_D);CHKERRQ(ierr); 182988ebb749SStefano Zampini ierr = MatDestroy(&coarse_phi_B);CHKERRQ(ierr); 1830ffd830a3SStefano Zampini if (!pcbddc->symmetric_primal) { 183188ebb749SStefano Zampini ierr = MatDestroy(&coarse_psi_D);CHKERRQ(ierr); 183288ebb749SStefano Zampini ierr = MatDestroy(&coarse_psi_B);CHKERRQ(ierr); 183388ebb749SStefano Zampini } 183488ebb749SStefano Zampini ierr = MatDestroy(&coarse_sub_mat);CHKERRQ(ierr); 183588ebb749SStefano Zampini } 18368629588bSStefano Zampini /* get back data */ 18378629588bSStefano Zampini *coarse_submat_vals_n = coarse_submat_vals; 183888ebb749SStefano Zampini PetscFunctionReturn(0); 183988ebb749SStefano Zampini } 184088ebb749SStefano Zampini 184188ebb749SStefano Zampini #undef __FUNCT__ 1842d65f70fdSStefano Zampini #define __FUNCT__ "MatGetSubMatrixUnsorted" 1843d648f858SStefano Zampini PetscErrorCode MatGetSubMatrixUnsorted(Mat A, IS isrow, IS iscol, Mat* B) 1844aa0d41d4SStefano Zampini { 1845d65f70fdSStefano Zampini Mat *work_mat; 1846d65f70fdSStefano Zampini IS isrow_s,iscol_s; 1847d65f70fdSStefano Zampini PetscBool rsorted,csorted; 1848d65f70fdSStefano Zampini PetscInt rsize,*idxs_perm_r,csize,*idxs_perm_c; 1849aa0d41d4SStefano Zampini PetscErrorCode ierr; 1850aa0d41d4SStefano Zampini 1851aa0d41d4SStefano Zampini PetscFunctionBegin; 1852d65f70fdSStefano Zampini ierr = ISSorted(isrow,&rsorted);CHKERRQ(ierr); 1853d65f70fdSStefano Zampini ierr = ISSorted(iscol,&csorted);CHKERRQ(ierr); 1854d65f70fdSStefano Zampini ierr = ISGetLocalSize(isrow,&rsize);CHKERRQ(ierr); 1855d65f70fdSStefano Zampini ierr = ISGetLocalSize(iscol,&csize);CHKERRQ(ierr); 1856aa0d41d4SStefano Zampini 1857d65f70fdSStefano Zampini if (!rsorted) { 1858906d46d4SStefano Zampini const PetscInt *idxs; 1859906d46d4SStefano Zampini PetscInt *idxs_sorted,i; 1860aa0d41d4SStefano Zampini 1861d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_perm_r);CHKERRQ(ierr); 1862d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_sorted);CHKERRQ(ierr); 1863d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 1864d65f70fdSStefano Zampini idxs_perm_r[i] = i; 1865aa0d41d4SStefano Zampini } 1866d65f70fdSStefano Zampini ierr = ISGetIndices(isrow,&idxs);CHKERRQ(ierr); 1867d65f70fdSStefano Zampini ierr = PetscSortIntWithPermutation(rsize,idxs,idxs_perm_r);CHKERRQ(ierr); 1868d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 1869d65f70fdSStefano Zampini idxs_sorted[i] = idxs[idxs_perm_r[i]]; 1870aa0d41d4SStefano Zampini } 1871d65f70fdSStefano Zampini ierr = ISRestoreIndices(isrow,&idxs);CHKERRQ(ierr); 1872d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,rsize,idxs_sorted,PETSC_OWN_POINTER,&isrow_s);CHKERRQ(ierr); 1873d65f70fdSStefano Zampini } else { 1874d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)isrow);CHKERRQ(ierr); 1875d65f70fdSStefano Zampini isrow_s = isrow; 1876aa0d41d4SStefano Zampini } 1877906d46d4SStefano Zampini 1878d65f70fdSStefano Zampini if (!csorted) { 1879d65f70fdSStefano Zampini if (isrow == iscol) { 1880d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)isrow_s);CHKERRQ(ierr); 1881d65f70fdSStefano Zampini iscol_s = isrow_s; 1882d65f70fdSStefano Zampini } else { 1883d65f70fdSStefano Zampini const PetscInt *idxs; 1884d65f70fdSStefano Zampini PetscInt *idxs_sorted,i; 1885906d46d4SStefano Zampini 1886d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_perm_c);CHKERRQ(ierr); 1887d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_sorted);CHKERRQ(ierr); 1888d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 1889d65f70fdSStefano Zampini idxs_perm_c[i] = i; 1890d65f70fdSStefano Zampini } 1891d65f70fdSStefano Zampini ierr = ISGetIndices(iscol,&idxs);CHKERRQ(ierr); 1892d65f70fdSStefano Zampini ierr = PetscSortIntWithPermutation(csize,idxs,idxs_perm_c);CHKERRQ(ierr); 1893d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 1894d65f70fdSStefano Zampini idxs_sorted[i] = idxs[idxs_perm_c[i]]; 1895d65f70fdSStefano Zampini } 1896d65f70fdSStefano Zampini ierr = ISRestoreIndices(iscol,&idxs);CHKERRQ(ierr); 1897d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,csize,idxs_sorted,PETSC_OWN_POINTER,&iscol_s);CHKERRQ(ierr); 1898d65f70fdSStefano Zampini } 1899d65f70fdSStefano Zampini } else { 1900d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)iscol);CHKERRQ(ierr); 1901d65f70fdSStefano Zampini iscol_s = iscol; 1902d65f70fdSStefano Zampini } 1903d65f70fdSStefano Zampini 1904d648f858SStefano Zampini ierr = MatGetSubMatrices(A,1,&isrow_s,&iscol_s,MAT_INITIAL_MATRIX,&work_mat);CHKERRQ(ierr); 1905d65f70fdSStefano Zampini 1906d65f70fdSStefano Zampini if (!rsorted || !csorted) { 1907906d46d4SStefano Zampini Mat new_mat; 1908d65f70fdSStefano Zampini IS is_perm_r,is_perm_c; 1909906d46d4SStefano Zampini 1910d65f70fdSStefano Zampini if (!rsorted) { 1911d65f70fdSStefano Zampini PetscInt *idxs_r,i; 1912d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_r);CHKERRQ(ierr); 1913d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 1914d65f70fdSStefano Zampini idxs_r[idxs_perm_r[i]] = i; 1915906d46d4SStefano Zampini } 1916d65f70fdSStefano Zampini ierr = PetscFree(idxs_perm_r);CHKERRQ(ierr); 1917d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,rsize,idxs_r,PETSC_OWN_POINTER,&is_perm_r);CHKERRQ(ierr); 1918d65f70fdSStefano Zampini } else { 1919d65f70fdSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,rsize,0,1,&is_perm_r);CHKERRQ(ierr); 1920906d46d4SStefano Zampini } 1921d65f70fdSStefano Zampini ierr = ISSetPermutation(is_perm_r);CHKERRQ(ierr); 1922d65f70fdSStefano Zampini 1923d65f70fdSStefano Zampini if (!csorted) { 1924d65f70fdSStefano Zampini if (isrow_s == iscol_s) { 1925d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)is_perm_r);CHKERRQ(ierr); 1926d65f70fdSStefano Zampini is_perm_c = is_perm_r; 1927d65f70fdSStefano Zampini } else { 1928d65f70fdSStefano Zampini PetscInt *idxs_c,i; 1929d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_c);CHKERRQ(ierr); 1930d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 1931d65f70fdSStefano Zampini idxs_c[idxs_perm_c[i]] = i; 1932d65f70fdSStefano Zampini } 1933d65f70fdSStefano Zampini ierr = PetscFree(idxs_perm_c);CHKERRQ(ierr); 1934d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,csize,idxs_c,PETSC_OWN_POINTER,&is_perm_c);CHKERRQ(ierr); 1935d65f70fdSStefano Zampini } 1936d65f70fdSStefano Zampini } else { 1937d65f70fdSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,csize,0,1,&is_perm_c);CHKERRQ(ierr); 1938d65f70fdSStefano Zampini } 1939d65f70fdSStefano Zampini ierr = ISSetPermutation(is_perm_c);CHKERRQ(ierr); 1940d65f70fdSStefano Zampini 1941d65f70fdSStefano Zampini ierr = MatPermute(work_mat[0],is_perm_r,is_perm_c,&new_mat);CHKERRQ(ierr); 1942d65f70fdSStefano Zampini ierr = MatDestroy(&work_mat[0]);CHKERRQ(ierr); 1943d65f70fdSStefano Zampini work_mat[0] = new_mat; 1944d65f70fdSStefano Zampini ierr = ISDestroy(&is_perm_r);CHKERRQ(ierr); 1945d65f70fdSStefano Zampini ierr = ISDestroy(&is_perm_c);CHKERRQ(ierr); 1946d65f70fdSStefano Zampini } 1947d65f70fdSStefano Zampini 1948d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)work_mat[0]);CHKERRQ(ierr); 1949d65f70fdSStefano Zampini *B = work_mat[0]; 1950d65f70fdSStefano Zampini ierr = MatDestroyMatrices(1,&work_mat);CHKERRQ(ierr); 1951d65f70fdSStefano Zampini ierr = ISDestroy(&isrow_s);CHKERRQ(ierr); 1952d65f70fdSStefano Zampini ierr = ISDestroy(&iscol_s);CHKERRQ(ierr); 1953d65f70fdSStefano Zampini PetscFunctionReturn(0); 1954d65f70fdSStefano Zampini } 1955d65f70fdSStefano Zampini 1956d65f70fdSStefano Zampini #undef __FUNCT__ 19575e8657edSStefano Zampini #define __FUNCT__ "PCBDDCComputeLocalMatrix" 19585e8657edSStefano Zampini PetscErrorCode PCBDDCComputeLocalMatrix(PC pc, Mat ChangeOfBasisMatrix) 1959aa0d41d4SStefano Zampini { 1960aa0d41d4SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 19615e8657edSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 1962d65f70fdSStefano Zampini Mat new_mat; 19635e8657edSStefano Zampini IS is_local,is_global; 1964d65f70fdSStefano Zampini PetscInt local_size; 1965d65f70fdSStefano Zampini PetscBool isseqaij; 1966aa0d41d4SStefano Zampini PetscErrorCode ierr; 1967aa0d41d4SStefano Zampini 1968aa0d41d4SStefano Zampini PetscFunctionBegin; 1969aa0d41d4SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 19705e8657edSStefano Zampini ierr = MatGetSize(matis->A,&local_size,NULL);CHKERRQ(ierr); 19715e8657edSStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)matis->A),local_size,0,1,&is_local);CHKERRQ(ierr); 1972b087196eSStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(pc->pmat->rmap->mapping,is_local,&is_global);CHKERRQ(ierr); 1973aa0d41d4SStefano Zampini ierr = ISDestroy(&is_local);CHKERRQ(ierr); 1974d648f858SStefano Zampini ierr = MatGetSubMatrixUnsorted(ChangeOfBasisMatrix,is_global,is_global,&new_mat);CHKERRQ(ierr); 1975aa0d41d4SStefano Zampini ierr = ISDestroy(&is_global);CHKERRQ(ierr); 1976906d46d4SStefano Zampini 1977906d46d4SStefano Zampini /* check */ 1978906d46d4SStefano Zampini if (pcbddc->dbg_flag) { 1979906d46d4SStefano Zampini Vec x,x_change; 1980906d46d4SStefano Zampini PetscReal error; 1981906d46d4SStefano Zampini 19825e8657edSStefano Zampini ierr = MatCreateVecs(ChangeOfBasisMatrix,&x,&x_change);CHKERRQ(ierr); 1983906d46d4SStefano Zampini ierr = VecSetRandom(x,NULL);CHKERRQ(ierr); 19845e8657edSStefano Zampini ierr = MatMult(ChangeOfBasisMatrix,x,x_change);CHKERRQ(ierr); 1985e176bc59SStefano Zampini ierr = VecScatterBegin(matis->cctx,x,matis->x,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1986e176bc59SStefano Zampini ierr = VecScatterEnd(matis->cctx,x,matis->x,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1987d65f70fdSStefano Zampini ierr = MatMult(new_mat,matis->x,matis->y);CHKERRQ(ierr); 1988e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,matis->y,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1989e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,matis->y,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1990906d46d4SStefano Zampini ierr = VecAXPY(x,-1.0,x_change);CHKERRQ(ierr); 1991906d46d4SStefano Zampini ierr = VecNorm(x,NORM_INFINITY,&error);CHKERRQ(ierr); 1992906d46d4SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1993906d46d4SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Error global vs local change on N: %1.6e\n",error);CHKERRQ(ierr); 1994906d46d4SStefano Zampini ierr = VecDestroy(&x);CHKERRQ(ierr); 1995906d46d4SStefano Zampini ierr = VecDestroy(&x_change);CHKERRQ(ierr); 1996906d46d4SStefano Zampini } 1997906d46d4SStefano Zampini 199822d5777bSStefano Zampini /* TODO: HOW TO WORK WITH BAIJ and SBAIJ and SEQDENSE? */ 19999b28b3ffSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)matis->A,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 200022d5777bSStefano Zampini if (isseqaij) { 20011cf9b237SStefano Zampini Mat M; 20021cf9b237SStefano Zampini 20031cf9b237SStefano Zampini ierr = MatPtAP(matis->A,new_mat,MAT_INITIAL_MATRIX,2.0,&M);CHKERRQ(ierr); 20041cf9b237SStefano Zampini ierr = MatSeqAIJCompress(M,&pcbddc->local_mat);CHKERRQ(ierr); 20051cf9b237SStefano Zampini ierr = MatDestroy(&M);CHKERRQ(ierr); 2006aa0d41d4SStefano Zampini } else { 20071cf9b237SStefano Zampini Mat work_mat,M; 20081cf9b237SStefano Zampini 2009aa0d41d4SStefano Zampini ierr = MatConvert(matis->A,MATSEQAIJ,MAT_INITIAL_MATRIX,&work_mat);CHKERRQ(ierr); 20101cf9b237SStefano Zampini ierr = MatPtAP(work_mat,new_mat,MAT_INITIAL_MATRIX,2.0,&M);CHKERRQ(ierr); 20111cf9b237SStefano Zampini ierr = MatSeqAIJCompress(M,&pcbddc->local_mat);CHKERRQ(ierr); 20121cf9b237SStefano Zampini ierr = MatDestroy(&M);CHKERRQ(ierr); 2013aa0d41d4SStefano Zampini ierr = MatDestroy(&work_mat);CHKERRQ(ierr); 2014aa0d41d4SStefano Zampini } 20153301b35fSStefano Zampini if (matis->A->symmetric_set) { 20163301b35fSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_SYMMETRIC,matis->A->symmetric);CHKERRQ(ierr); 2017e496cd5dSStefano Zampini #if !defined(PETSC_USE_COMPLEX) 20183301b35fSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_HERMITIAN,matis->A->symmetric);CHKERRQ(ierr); 2019e496cd5dSStefano Zampini #endif 20203301b35fSStefano Zampini } 202145a1bb75SStefano Zampini /* 202245a1bb75SStefano Zampini ierr = PetscViewerSetFormat(PETSC_VIEWER_STDOUT_SELF,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 2023d65f70fdSStefano Zampini ierr = MatView(new_mat,(PetscViewer)0);CHKERRQ(ierr); 202445a1bb75SStefano Zampini */ 2025d65f70fdSStefano Zampini ierr = MatDestroy(&new_mat);CHKERRQ(ierr); 2026aa0d41d4SStefano Zampini PetscFunctionReturn(0); 2027aa0d41d4SStefano Zampini } 2028aa0d41d4SStefano Zampini 2029aa0d41d4SStefano Zampini #undef __FUNCT__ 2030a64d13efSStefano Zampini #define __FUNCT__ "PCBDDCSetUpLocalScatters" 20318ce42a96SStefano Zampini PetscErrorCode PCBDDCSetUpLocalScatters(PC pc) 2032a64d13efSStefano Zampini { 2033a64d13efSStefano Zampini PC_IS* pcis = (PC_IS*)(pc->data); 2034a64d13efSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 2035d62866d3SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 203653892102SStefano Zampini PetscInt *idx_R_local=NULL; 20373a50541eSStefano Zampini PetscInt n_vertices,i,j,n_R,n_D,n_B; 20383a50541eSStefano Zampini PetscInt vbs,bs; 20396816873aSStefano Zampini PetscBT bitmask=NULL; 2040a64d13efSStefano Zampini PetscErrorCode ierr; 2041a64d13efSStefano Zampini 2042a64d13efSStefano Zampini PetscFunctionBegin; 2043b23d619eSStefano Zampini /* 2044b23d619eSStefano Zampini No need to setup local scatters if 2045b23d619eSStefano Zampini - primal space is unchanged 2046b23d619eSStefano Zampini AND 2047b23d619eSStefano Zampini - we actually have locally some primal dofs (could not be true in multilevel or for isolated subdomains) 2048b23d619eSStefano Zampini AND 2049b23d619eSStefano Zampini - we are not in debugging mode (this is needed since there are Synchronized prints at the end of the subroutine 2050b23d619eSStefano Zampini */ 2051b23d619eSStefano Zampini if (!pcbddc->new_primal_space_local && pcbddc->local_primal_size && !pcbddc->dbg_flag) { 2052f4ddd8eeSStefano Zampini PetscFunctionReturn(0); 2053f4ddd8eeSStefano Zampini } 2054f4ddd8eeSStefano Zampini /* destroy old objects */ 2055f4ddd8eeSStefano Zampini ierr = ISDestroy(&pcbddc->is_R_local);CHKERRQ(ierr); 2056f4ddd8eeSStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_B);CHKERRQ(ierr); 2057f4ddd8eeSStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_D);CHKERRQ(ierr); 2058a64d13efSStefano Zampini /* Set Non-overlapping dimensions */ 2059b371cd4fSStefano Zampini n_B = pcis->n_B; 2060b371cd4fSStefano Zampini n_D = pcis->n - n_B; 2061b371cd4fSStefano Zampini n_vertices = pcbddc->n_vertices; 20623a50541eSStefano Zampini 2063a64d13efSStefano Zampini /* Dohrmann's notation: dofs splitted in R (Remaining: all dofs but the vertices) and V (Vertices) */ 20646816873aSStefano Zampini 206553892102SStefano Zampini /* create auxiliary bitmask and allocate workspace */ 20666816873aSStefano Zampini if (!sub_schurs->reuse_mumps) { 2067854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n-n_vertices,&idx_R_local);CHKERRQ(ierr); 2068a64d13efSStefano Zampini ierr = PetscBTCreate(pcis->n,&bitmask);CHKERRQ(ierr); 2069a64d13efSStefano Zampini for (i=0;i<n_vertices;i++) { 20700e6343abSStefano Zampini ierr = PetscBTSet(bitmask,pcbddc->local_primal_ref_node[i]);CHKERRQ(ierr); 2071a64d13efSStefano Zampini } 2072a64d13efSStefano Zampini 2073a64d13efSStefano Zampini for (i=0, n_R=0; i<pcis->n; i++) { 20744641a718SStefano Zampini if (!PetscBTLookup(bitmask,i)) { 20756816873aSStefano Zampini idx_R_local[n_R++] = i; 2076a64d13efSStefano Zampini } 2077a64d13efSStefano Zampini } 207853892102SStefano Zampini } else { /* A different ordering (already computed) is present if we are reusing MUMPS Schur solver */ 20796816873aSStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 20806816873aSStefano Zampini 208153892102SStefano Zampini ierr = ISGetIndices(reuse_mumps->is_R,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 208253892102SStefano Zampini ierr = ISGetLocalSize(reuse_mumps->is_R,&n_R);CHKERRQ(ierr); 20836816873aSStefano Zampini } 20843a50541eSStefano Zampini 20853a50541eSStefano Zampini /* Block code */ 20863a50541eSStefano Zampini vbs = 1; 20873a50541eSStefano Zampini ierr = MatGetBlockSize(pcbddc->local_mat,&bs);CHKERRQ(ierr); 20883a50541eSStefano Zampini if (bs>1 && !(n_vertices%bs)) { 20893a50541eSStefano Zampini PetscBool is_blocked = PETSC_TRUE; 20903a50541eSStefano Zampini PetscInt *vary; 2091d3df7717SStefano Zampini if (!sub_schurs->reuse_mumps) { 2092785e854fSJed Brown ierr = PetscMalloc1(pcis->n/bs,&vary);CHKERRQ(ierr); 20933a50541eSStefano Zampini ierr = PetscMemzero(vary,pcis->n/bs*sizeof(PetscInt));CHKERRQ(ierr); 2094d3df7717SStefano Zampini /* Verify that the vertex indices correspond to each element in a block (code taken from sbaij2.c) */ 2095d3df7717SStefano 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 */ 20960e6343abSStefano Zampini for (i=0; i<n_vertices; i++) vary[pcbddc->local_primal_ref_node[i]/bs]++; 2097d3df7717SStefano Zampini for (i=0; i<pcis->n/bs; i++) { 20983a50541eSStefano Zampini if (vary[i]!=0 && vary[i]!=bs) { 20993a50541eSStefano Zampini is_blocked = PETSC_FALSE; 21003a50541eSStefano Zampini break; 21013a50541eSStefano Zampini } 21023a50541eSStefano Zampini } 2103d3df7717SStefano Zampini ierr = PetscFree(vary);CHKERRQ(ierr); 2104d3df7717SStefano Zampini } else { 2105d3df7717SStefano Zampini /* Verify directly the R set */ 2106d3df7717SStefano Zampini for (i=0; i<n_R/bs; i++) { 2107d3df7717SStefano Zampini PetscInt j,node=idx_R_local[bs*i]; 2108d3df7717SStefano Zampini for (j=1; j<bs; j++) { 2109d3df7717SStefano Zampini if (node != idx_R_local[bs*i+j]-j) { 2110d3df7717SStefano Zampini is_blocked = PETSC_FALSE; 2111d3df7717SStefano Zampini break; 2112d3df7717SStefano Zampini } 2113d3df7717SStefano Zampini } 2114d3df7717SStefano Zampini } 2115d3df7717SStefano Zampini } 21163a50541eSStefano Zampini if (is_blocked) { /* build compressed IS for R nodes (complement of vertices) */ 21173a50541eSStefano Zampini vbs = bs; 21183a50541eSStefano Zampini for (i=0;i<n_R/vbs;i++) { 21193a50541eSStefano Zampini idx_R_local[i] = idx_R_local[vbs*i]/vbs; 21203a50541eSStefano Zampini } 21213a50541eSStefano Zampini } 21223a50541eSStefano Zampini } 21233a50541eSStefano Zampini ierr = ISCreateBlock(PETSC_COMM_SELF,vbs,n_R/vbs,idx_R_local,PETSC_COPY_VALUES,&pcbddc->is_R_local);CHKERRQ(ierr); 212453892102SStefano Zampini if (sub_schurs->reuse_mumps) { 212553892102SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 212653892102SStefano Zampini 212753892102SStefano Zampini ierr = ISRestoreIndices(reuse_mumps->is_R,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 212853892102SStefano Zampini ierr = ISDestroy(&reuse_mumps->is_R);CHKERRQ(ierr); 212953892102SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->is_R_local);CHKERRQ(ierr); 213053892102SStefano Zampini reuse_mumps->is_R = pcbddc->is_R_local; 213153892102SStefano Zampini } else { 21323a50541eSStefano Zampini ierr = PetscFree(idx_R_local);CHKERRQ(ierr); 213353892102SStefano Zampini } 2134a64d13efSStefano Zampini 2135a64d13efSStefano Zampini /* print some info if requested */ 2136a64d13efSStefano Zampini if (pcbddc->dbg_flag) { 2137a64d13efSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 2138a64d13efSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 21391575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 2140a64d13efSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d local dimensions\n",PetscGlobalRank);CHKERRQ(ierr); 2141a64d13efSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"local_size = %d, dirichlet_size = %d, boundary_size = %d\n",pcis->n,n_D,n_B);CHKERRQ(ierr); 21424f1b2e48SStefano 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); 2143a64d13efSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 2144a64d13efSStefano Zampini } 2145a64d13efSStefano Zampini 2146a64d13efSStefano Zampini /* VecScatters pcbddc->R_to_B and (optionally) pcbddc->R_to_D */ 21476816873aSStefano Zampini if (!sub_schurs->reuse_mumps) { 21486816873aSStefano Zampini IS is_aux1,is_aux2; 21496816873aSStefano Zampini PetscInt *aux_array1,*aux_array2,*is_indices,*idx_R_local; 21506816873aSStefano Zampini 21513a50541eSStefano Zampini ierr = ISGetIndices(pcbddc->is_R_local,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 2152854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n_B-n_vertices,&aux_array1);CHKERRQ(ierr); 2153854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n_B-n_vertices,&aux_array2);CHKERRQ(ierr); 2154a64d13efSStefano Zampini ierr = ISGetIndices(pcis->is_I_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 21554641a718SStefano Zampini for (i=0; i<n_D; i++) { 21564641a718SStefano Zampini ierr = PetscBTSet(bitmask,is_indices[i]);CHKERRQ(ierr); 21574641a718SStefano Zampini } 2158a64d13efSStefano Zampini ierr = ISRestoreIndices(pcis->is_I_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 2159a64d13efSStefano Zampini for (i=0, j=0; i<n_R; i++) { 21604641a718SStefano Zampini if (!PetscBTLookup(bitmask,idx_R_local[i])) { 21614641a718SStefano Zampini aux_array1[j++] = i; 2162a64d13efSStefano Zampini } 2163a64d13efSStefano Zampini } 2164a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array1,PETSC_OWN_POINTER,&is_aux1);CHKERRQ(ierr); 2165a64d13efSStefano Zampini ierr = ISGetIndices(pcis->is_B_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 2166a64d13efSStefano Zampini for (i=0, j=0; i<n_B; i++) { 21674641a718SStefano Zampini if (!PetscBTLookup(bitmask,is_indices[i])) { 21684641a718SStefano Zampini aux_array2[j++] = i; 2169a64d13efSStefano Zampini } 2170a64d13efSStefano Zampini } 2171a64d13efSStefano Zampini ierr = ISRestoreIndices(pcis->is_B_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 2172a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array2,PETSC_OWN_POINTER,&is_aux2);CHKERRQ(ierr); 2173a64d13efSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,is_aux1,pcis->vec1_B,is_aux2,&pcbddc->R_to_B);CHKERRQ(ierr); 2174a64d13efSStefano Zampini ierr = ISDestroy(&is_aux1);CHKERRQ(ierr); 2175a64d13efSStefano Zampini ierr = ISDestroy(&is_aux2);CHKERRQ(ierr); 2176a64d13efSStefano Zampini 21778eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 2178785e854fSJed Brown ierr = PetscMalloc1(n_D,&aux_array1);CHKERRQ(ierr); 2179a64d13efSStefano Zampini for (i=0, j=0; i<n_R; i++) { 21804641a718SStefano Zampini if (PetscBTLookup(bitmask,idx_R_local[i])) { 21814641a718SStefano Zampini aux_array1[j++] = i; 2182a64d13efSStefano Zampini } 2183a64d13efSStefano Zampini } 2184a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array1,PETSC_OWN_POINTER,&is_aux1);CHKERRQ(ierr); 2185a64d13efSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,is_aux1,pcis->vec1_D,(IS)0,&pcbddc->R_to_D);CHKERRQ(ierr); 2186a64d13efSStefano Zampini ierr = ISDestroy(&is_aux1);CHKERRQ(ierr); 2187a64d13efSStefano Zampini } 21884641a718SStefano Zampini ierr = PetscBTDestroy(&bitmask);CHKERRQ(ierr); 21893a50541eSStefano Zampini ierr = ISRestoreIndices(pcbddc->is_R_local,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 2190d62866d3SStefano Zampini } else { 219153892102SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 21926816873aSStefano Zampini IS tis; 21936816873aSStefano Zampini PetscInt schur_size; 21946816873aSStefano Zampini 219553892102SStefano Zampini ierr = ISGetLocalSize(reuse_mumps->is_B,&schur_size);CHKERRQ(ierr); 21966816873aSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,schur_size,n_D,1,&tis);CHKERRQ(ierr); 219753892102SStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,tis,pcis->vec1_B,reuse_mumps->is_B,&pcbddc->R_to_B);CHKERRQ(ierr); 21986816873aSStefano Zampini ierr = ISDestroy(&tis);CHKERRQ(ierr); 21996816873aSStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 22006816873aSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,n_D,0,1,&tis);CHKERRQ(ierr); 22016816873aSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,tis,pcis->vec1_D,(IS)0,&pcbddc->R_to_D);CHKERRQ(ierr); 22026816873aSStefano Zampini ierr = ISDestroy(&tis);CHKERRQ(ierr); 2203d62866d3SStefano Zampini } 2204d62866d3SStefano Zampini } 2205a64d13efSStefano Zampini PetscFunctionReturn(0); 2206a64d13efSStefano Zampini } 2207a64d13efSStefano Zampini 2208304d26faSStefano Zampini 2209304d26faSStefano Zampini #undef __FUNCT__ 2210304d26faSStefano Zampini #define __FUNCT__ "PCBDDCSetUpLocalSolvers" 2211684f6988SStefano Zampini PetscErrorCode PCBDDCSetUpLocalSolvers(PC pc, PetscBool dirichlet, PetscBool neumann) 2212304d26faSStefano Zampini { 2213304d26faSStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 2214304d26faSStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 2215304d26faSStefano Zampini PC pc_temp; 2216304d26faSStefano Zampini Mat A_RR; 2217f4ddd8eeSStefano Zampini MatReuse reuse; 2218304d26faSStefano Zampini PetscScalar m_one = -1.0; 2219304d26faSStefano Zampini PetscReal value; 222004708bb6SStefano Zampini PetscInt n_D,n_R; 22219577ea80SStefano Zampini PetscBool use_exact,use_exact_reduced,issbaij; 2222304d26faSStefano Zampini PetscErrorCode ierr; 2223e604994aSStefano Zampini /* prefixes stuff */ 2224312be037SStefano Zampini char dir_prefix[256],neu_prefix[256],str_level[16]; 2225e604994aSStefano Zampini size_t len; 2226304d26faSStefano Zampini 2227304d26faSStefano Zampini PetscFunctionBegin; 2228304d26faSStefano Zampini 2229e604994aSStefano Zampini /* compute prefixes */ 2230e604994aSStefano Zampini ierr = PetscStrcpy(dir_prefix,"");CHKERRQ(ierr); 2231e604994aSStefano Zampini ierr = PetscStrcpy(neu_prefix,"");CHKERRQ(ierr); 2232e604994aSStefano Zampini if (!pcbddc->current_level) { 2233e604994aSStefano Zampini ierr = PetscStrcpy(dir_prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 2234e604994aSStefano Zampini ierr = PetscStrcpy(neu_prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 2235e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,"pc_bddc_dirichlet_");CHKERRQ(ierr); 2236e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,"pc_bddc_neumann_");CHKERRQ(ierr); 2237e604994aSStefano Zampini } else { 2238e604994aSStefano Zampini ierr = PetscStrcpy(str_level,"");CHKERRQ(ierr); 2239312be037SStefano Zampini sprintf(str_level,"l%d_",(int)(pcbddc->current_level)); 2240e604994aSStefano Zampini ierr = PetscStrlen(((PetscObject)pc)->prefix,&len);CHKERRQ(ierr); 2241e604994aSStefano Zampini len -= 15; /* remove "pc_bddc_coarse_" */ 2242312be037SStefano Zampini if (pcbddc->current_level>1) len -= 3; /* remove "lX_" with X level number */ 2243312be037SStefano Zampini if (pcbddc->current_level>10) len -= 1; /* remove another char from level number */ 224434d6797cSStefano Zampini ierr = PetscStrncpy(dir_prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 224534d6797cSStefano Zampini ierr = PetscStrncpy(neu_prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 2246e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,"pc_bddc_dirichlet_");CHKERRQ(ierr); 2247e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,"pc_bddc_neumann_");CHKERRQ(ierr); 2248e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,str_level);CHKERRQ(ierr); 2249e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,str_level);CHKERRQ(ierr); 2250e604994aSStefano Zampini } 2251e604994aSStefano Zampini 2252304d26faSStefano Zampini /* DIRICHLET PROBLEM */ 2253684f6988SStefano Zampini if (dirichlet) { 2254d5574798SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 22553301b35fSStefano Zampini if (pcbddc->local_mat->symmetric_set) { 22563301b35fSStefano Zampini ierr = MatSetOption(pcis->A_II,MAT_SYMMETRIC,pcbddc->local_mat->symmetric_set);CHKERRQ(ierr); 2257964fefecSStefano Zampini } 2258ac78edfcSStefano Zampini /* Matrix for Dirichlet problem is pcis->A_II */ 2259964fefecSStefano Zampini n_D = pcis->n - pcis->n_B; 2260304d26faSStefano Zampini if (!pcbddc->ksp_D) { /* create object if not yet build */ 2261304d26faSStefano Zampini ierr = KSPCreate(PETSC_COMM_SELF,&pcbddc->ksp_D);CHKERRQ(ierr); 2262304d26faSStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->ksp_D,(PetscObject)pc,1);CHKERRQ(ierr); 2263304d26faSStefano Zampini /* default */ 2264304d26faSStefano Zampini ierr = KSPSetType(pcbddc->ksp_D,KSPPREONLY);CHKERRQ(ierr); 2265e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->ksp_D,dir_prefix);CHKERRQ(ierr); 22669577ea80SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcis->A_II,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 2267304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_D,&pc_temp);CHKERRQ(ierr); 22689577ea80SStefano Zampini if (issbaij) { 22699577ea80SStefano Zampini ierr = PCSetType(pc_temp,PCCHOLESKY);CHKERRQ(ierr); 22709577ea80SStefano Zampini } else { 2271304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCLU);CHKERRQ(ierr); 22729577ea80SStefano Zampini } 2273304d26faSStefano Zampini /* Allow user's customization */ 2274304d26faSStefano Zampini ierr = KSPSetFromOptions(pcbddc->ksp_D);CHKERRQ(ierr); 2275304d26faSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 2276304d26faSStefano Zampini } 2277d1e9a80fSBarry Smith ierr = KSPSetOperators(pcbddc->ksp_D,pcis->A_II,pcis->A_II);CHKERRQ(ierr); 2278d62866d3SStefano Zampini if (sub_schurs->reuse_mumps) { 2279d62866d3SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 2280d62866d3SStefano Zampini 2281d62866d3SStefano Zampini ierr = KSPSetPC(pcbddc->ksp_D,reuse_mumps->interior_solver);CHKERRQ(ierr); 2282d5574798SStefano Zampini } 2283304d26faSStefano Zampini /* umfpack interface has a bug when matrix dimension is zero. TODO solve from umfpack interface */ 2284304d26faSStefano Zampini if (!n_D) { 2285304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_D,&pc_temp);CHKERRQ(ierr); 2286304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCNONE);CHKERRQ(ierr); 2287304d26faSStefano Zampini } 2288304d26faSStefano Zampini /* Set Up KSP for Dirichlet problem of BDDC */ 2289304d26faSStefano Zampini ierr = KSPSetUp(pcbddc->ksp_D);CHKERRQ(ierr); 2290304d26faSStefano Zampini /* set ksp_D into pcis data */ 2291304d26faSStefano Zampini ierr = KSPDestroy(&pcis->ksp_D);CHKERRQ(ierr); 2292304d26faSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->ksp_D);CHKERRQ(ierr); 2293304d26faSStefano Zampini pcis->ksp_D = pcbddc->ksp_D; 2294684f6988SStefano Zampini } 2295304d26faSStefano Zampini 2296304d26faSStefano Zampini /* NEUMANN PROBLEM */ 2297684f6988SStefano Zampini A_RR = 0; 2298684f6988SStefano Zampini if (neumann) { 2299d62866d3SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 230004708bb6SStefano Zampini PetscInt ibs,mbs; 230104708bb6SStefano Zampini PetscBool issbaij; 230204708bb6SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 2303f4ddd8eeSStefano Zampini /* Matrix for Neumann problem is A_RR -> we need to create/reuse it at this point */ 23048ce42a96SStefano Zampini ierr = ISGetSize(pcbddc->is_R_local,&n_R);CHKERRQ(ierr); 2305f4ddd8eeSStefano Zampini if (pcbddc->ksp_R) { /* already created ksp */ 2306f4ddd8eeSStefano Zampini PetscInt nn_R; 230781d9aea3SBarry Smith ierr = KSPGetOperators(pcbddc->ksp_R,NULL,&A_RR);CHKERRQ(ierr); 2308f4ddd8eeSStefano Zampini ierr = PetscObjectReference((PetscObject)A_RR);CHKERRQ(ierr); 2309f4ddd8eeSStefano Zampini ierr = MatGetSize(A_RR,&nn_R,NULL);CHKERRQ(ierr); 2310f4ddd8eeSStefano Zampini if (nn_R != n_R) { /* old ksp is not reusable, so reset it */ 2311f4ddd8eeSStefano Zampini ierr = KSPReset(pcbddc->ksp_R);CHKERRQ(ierr); 2312f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 2313f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 2314f4ddd8eeSStefano Zampini } else { /* same sizes, but nonzero pattern depend on primal vertices so it can be changed */ 2315727cdba6SStefano Zampini if (pcbddc->new_primal_space_local) { /* we are not sure the matrix will have the same nonzero pattern */ 2316f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 2317f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 2318f4ddd8eeSStefano Zampini } else { /* safe to reuse the matrix */ 2319f4ddd8eeSStefano Zampini reuse = MAT_REUSE_MATRIX; 2320f4ddd8eeSStefano Zampini } 2321f4ddd8eeSStefano Zampini } 2322f4ddd8eeSStefano Zampini /* last check */ 2323d1e9a80fSBarry Smith if (pc->flag == DIFFERENT_NONZERO_PATTERN) { 2324f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 2325f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 2326f4ddd8eeSStefano Zampini } 2327f4ddd8eeSStefano Zampini } else { /* first time, so we need to create the matrix */ 2328f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 2329f4ddd8eeSStefano Zampini } 2330f4ddd8eeSStefano Zampini /* extract A_RR */ 2331af732b37SStefano Zampini ierr = MatGetBlockSize(pcbddc->local_mat,&mbs);CHKERRQ(ierr); 2332af732b37SStefano Zampini ierr = ISGetBlockSize(pcbddc->is_R_local,&ibs);CHKERRQ(ierr); 233304708bb6SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcbddc->local_mat,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 233404708bb6SStefano Zampini if (ibs != mbs) { /* need to convert to SEQAIJ to extract any submatrix with is_R_local */ 233504708bb6SStefano Zampini if (matis->A == pcbddc->local_mat) { 233604708bb6SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 233704708bb6SStefano Zampini ierr = MatConvert(matis->A,MATSEQAIJ,MAT_INITIAL_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 2338af732b37SStefano Zampini } else { 233904708bb6SStefano Zampini ierr = MatConvert(pcbddc->local_mat,MATSEQAIJ,MAT_REUSE_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 23406816873aSStefano Zampini } 234104708bb6SStefano Zampini } else if (issbaij) { /* need to convert to BAIJ to get offdiagonal blocks */ 234204708bb6SStefano Zampini if (matis->A == pcbddc->local_mat) { 234304708bb6SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 234404708bb6SStefano Zampini ierr = MatConvert(matis->A,MATSEQBAIJ,MAT_INITIAL_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 234504708bb6SStefano Zampini } else { 234604708bb6SStefano Zampini ierr = MatConvert(pcbddc->local_mat,MATSEQBAIJ,MAT_REUSE_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 234704708bb6SStefano Zampini } 234804708bb6SStefano Zampini } 2349f4ddd8eeSStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,pcbddc->is_R_local,pcbddc->is_R_local,reuse,&A_RR);CHKERRQ(ierr); 23503301b35fSStefano Zampini if (pcbddc->local_mat->symmetric_set) { 23513301b35fSStefano Zampini ierr = MatSetOption(A_RR,MAT_SYMMETRIC,pcbddc->local_mat->symmetric_set);CHKERRQ(ierr); 23526816873aSStefano Zampini } 2353f4ddd8eeSStefano Zampini if (!pcbddc->ksp_R) { /* create object if not present */ 2354304d26faSStefano Zampini ierr = KSPCreate(PETSC_COMM_SELF,&pcbddc->ksp_R);CHKERRQ(ierr); 2355304d26faSStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->ksp_R,(PetscObject)pc,1);CHKERRQ(ierr); 2356304d26faSStefano Zampini /* default */ 2357304d26faSStefano Zampini ierr = KSPSetType(pcbddc->ksp_R,KSPPREONLY);CHKERRQ(ierr); 2358e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->ksp_R,neu_prefix);CHKERRQ(ierr); 2359304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_temp);CHKERRQ(ierr); 23609577ea80SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)A_RR,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 23619577ea80SStefano Zampini if (issbaij) { 23629577ea80SStefano Zampini ierr = PCSetType(pc_temp,PCCHOLESKY);CHKERRQ(ierr); 23639577ea80SStefano Zampini } else { 2364304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCLU);CHKERRQ(ierr); 23659577ea80SStefano Zampini } 2366304d26faSStefano Zampini /* Allow user's customization */ 2367304d26faSStefano Zampini ierr = KSPSetFromOptions(pcbddc->ksp_R);CHKERRQ(ierr); 2368304d26faSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 2369304d26faSStefano Zampini } 2370304d26faSStefano Zampini /* umfpack interface has a bug when matrix dimension is zero. TODO solve from umfpack interface */ 2371304d26faSStefano Zampini if (!n_R) { 2372304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_temp);CHKERRQ(ierr); 2373304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCNONE);CHKERRQ(ierr); 2374304d26faSStefano Zampini } 2375d62866d3SStefano Zampini /* Reuse MUMPS solver if it is present */ 2376d62866d3SStefano Zampini if (sub_schurs->reuse_mumps) { 2377d62866d3SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 2378d62866d3SStefano Zampini 2379d62866d3SStefano Zampini ierr = KSPSetPC(pcbddc->ksp_R,reuse_mumps->correction_solver);CHKERRQ(ierr); 2380d62866d3SStefano Zampini } 2381ffd830a3SStefano Zampini ierr = KSPSetOperators(pcbddc->ksp_R,A_RR,A_RR);CHKERRQ(ierr); 2382304d26faSStefano Zampini /* Set Up KSP for Neumann problem of BDDC */ 2383304d26faSStefano Zampini ierr = KSPSetUp(pcbddc->ksp_R);CHKERRQ(ierr); 2384684f6988SStefano Zampini } 23856816873aSStefano Zampini /* free Neumann problem's matrix */ 23866816873aSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 2387304d26faSStefano Zampini 2388304d26faSStefano Zampini /* check Dirichlet and Neumann solvers and adapt them if a nullspace correction is needed */ 23890fccc4e9SStefano Zampini if (pcbddc->NullSpace || pcbddc->dbg_flag) { 2390684f6988SStefano Zampini if (pcbddc->dbg_flag) { 2391684f6988SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 23921575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 2393684f6988SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 2394684f6988SStefano Zampini } 2395684f6988SStefano Zampini if (dirichlet) { /* Dirichlet */ 23960fccc4e9SStefano Zampini ierr = VecSetRandom(pcis->vec1_D,NULL);CHKERRQ(ierr); 23970fccc4e9SStefano Zampini ierr = MatMult(pcis->A_II,pcis->vec1_D,pcis->vec2_D);CHKERRQ(ierr); 23980fccc4e9SStefano Zampini ierr = KSPSolve(pcbddc->ksp_D,pcis->vec2_D,pcis->vec2_D);CHKERRQ(ierr); 23990fccc4e9SStefano Zampini ierr = VecAXPY(pcis->vec1_D,m_one,pcis->vec2_D);CHKERRQ(ierr); 24000fccc4e9SStefano Zampini ierr = VecNorm(pcis->vec1_D,NORM_INFINITY,&value);CHKERRQ(ierr); 2401304d26faSStefano Zampini /* need to be adapted? */ 2402b8ffe317SStefano Zampini use_exact = (PetscAbsReal(value) > 1.e-4 ? PETSC_FALSE : PETSC_TRUE); 2403b8ffe317SStefano Zampini ierr = MPI_Allreduce(&use_exact,&use_exact_reduced,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 2404b8ffe317SStefano Zampini ierr = PCBDDCSetUseExactDirichlet(pc,use_exact_reduced);CHKERRQ(ierr); 2405304d26faSStefano Zampini /* print info */ 2406304d26faSStefano Zampini if (pcbddc->dbg_flag) { 2407e604994aSStefano 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); 2408304d26faSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 2409304d26faSStefano Zampini } 2410b8ffe317SStefano Zampini if (pcbddc->NullSpace && !use_exact_reduced && !pcbddc->switch_static) { 2411298c0119SStefano Zampini ierr = PCBDDCNullSpaceAssembleCorrection(pc,PETSC_TRUE,pcis->is_I_local);CHKERRQ(ierr); 2412304d26faSStefano Zampini } 2413684f6988SStefano Zampini } 2414684f6988SStefano Zampini if (neumann) { /* Neumann */ 24156816873aSStefano Zampini ierr = KSPGetOperators(pcbddc->ksp_R,&A_RR,NULL);CHKERRQ(ierr); 24160fccc4e9SStefano Zampini ierr = VecSetRandom(pcbddc->vec1_R,NULL);CHKERRQ(ierr); 24170fccc4e9SStefano Zampini ierr = MatMult(A_RR,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 24180fccc4e9SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec2_R,pcbddc->vec2_R);CHKERRQ(ierr); 24190fccc4e9SStefano Zampini ierr = VecAXPY(pcbddc->vec1_R,m_one,pcbddc->vec2_R);CHKERRQ(ierr); 24200fccc4e9SStefano Zampini ierr = VecNorm(pcbddc->vec1_R,NORM_INFINITY,&value);CHKERRQ(ierr); 2421304d26faSStefano Zampini /* need to be adapted? */ 2422b8ffe317SStefano Zampini use_exact = (PetscAbsReal(value) > 1.e-4 ? PETSC_FALSE : PETSC_TRUE); 2423b8ffe317SStefano Zampini ierr = MPI_Allreduce(&use_exact,&use_exact_reduced,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 2424304d26faSStefano Zampini /* print info */ 2425304d26faSStefano Zampini if (pcbddc->dbg_flag) { 2426e604994aSStefano 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); 2427304d26faSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 2428304d26faSStefano Zampini } 2429b8ffe317SStefano Zampini if (pcbddc->NullSpace && !use_exact_reduced) { /* is it the right logic? */ 2430298c0119SStefano Zampini ierr = PCBDDCNullSpaceAssembleCorrection(pc,PETSC_FALSE,pcbddc->is_R_local);CHKERRQ(ierr); 2431304d26faSStefano Zampini } 24320fccc4e9SStefano Zampini } 2433684f6988SStefano Zampini } 2434304d26faSStefano Zampini PetscFunctionReturn(0); 2435304d26faSStefano Zampini } 2436304d26faSStefano Zampini 2437304d26faSStefano Zampini #undef __FUNCT__ 2438ba15a52eSStefano Zampini #define __FUNCT__ "PCBDDCSolveSubstructureCorrection" 243980677318SStefano Zampini static PetscErrorCode PCBDDCSolveSubstructureCorrection(PC pc, Vec inout_B, Vec inout_D, PetscBool applytranspose) 2440674ae819SStefano Zampini { 2441674ae819SStefano Zampini PetscErrorCode ierr; 2442674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 2443be83ff47SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 2444674ae819SStefano Zampini 2445674ae819SStefano Zampini PetscFunctionBegin; 2446be83ff47SStefano Zampini if (!sub_schurs->reuse_mumps) { 244780677318SStefano Zampini ierr = VecSet(pcbddc->vec1_R,0.);CHKERRQ(ierr); 244820c7b377SStefano Zampini } 244980677318SStefano Zampini if (!pcbddc->switch_static) { 245080677318SStefano Zampini if (applytranspose && pcbddc->local_auxmat1) { 245180677318SStefano Zampini ierr = MatMultTranspose(pcbddc->local_auxmat2,inout_B,pcbddc->vec1_C);CHKERRQ(ierr); 245280677318SStefano Zampini ierr = MatMultTransposeAdd(pcbddc->local_auxmat1,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr); 245320c7b377SStefano Zampini } 2454be83ff47SStefano Zampini if (!sub_schurs->reuse_mumps) { 245580677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 245680677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 245720c7b377SStefano Zampini } else { 2458be83ff47SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 2459be83ff47SStefano Zampini 246053892102SStefano Zampini ierr = VecScatterBegin(reuse_mumps->correction_scatter_B,inout_B,reuse_mumps->rhs_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 246153892102SStefano Zampini ierr = VecScatterEnd(reuse_mumps->correction_scatter_B,inout_B,reuse_mumps->rhs_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 246220c7b377SStefano Zampini } 2463be83ff47SStefano Zampini } else { 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); 246680677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,inout_D,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 246780677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,inout_D,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 246880677318SStefano Zampini if (applytranspose && pcbddc->local_auxmat1) { 246980677318SStefano Zampini ierr = MatMultTranspose(pcbddc->local_auxmat2,pcbddc->vec1_R,pcbddc->vec1_C);CHKERRQ(ierr); 247080677318SStefano Zampini ierr = MatMultTransposeAdd(pcbddc->local_auxmat1,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr); 247180677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 247280677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2473674ae819SStefano Zampini } 2474674ae819SStefano Zampini } 2475be83ff47SStefano Zampini if (!sub_schurs->reuse_mumps) { 247680677318SStefano Zampini if (applytranspose) { 247780677318SStefano Zampini ierr = KSPSolveTranspose(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 247880677318SStefano Zampini } else { 247980677318SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 248080677318SStefano Zampini } 2481be83ff47SStefano Zampini } else { 2482be83ff47SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 2483be83ff47SStefano Zampini 2484be83ff47SStefano Zampini if (applytranspose) { 24855a05ddb0SStefano Zampini ierr = MatFactorSolveSchurComplementTranspose(reuse_mumps->F,reuse_mumps->rhs_B,reuse_mumps->sol_B);CHKERRQ(ierr); 2486be83ff47SStefano Zampini } else { 24875a05ddb0SStefano Zampini ierr = MatFactorSolveSchurComplement(reuse_mumps->F,reuse_mumps->rhs_B,reuse_mumps->sol_B);CHKERRQ(ierr); 2488be83ff47SStefano Zampini } 2489be83ff47SStefano Zampini } 249080677318SStefano Zampini ierr = VecSet(inout_B,0.);CHKERRQ(ierr); 249180677318SStefano Zampini if (!pcbddc->switch_static) { 2492be83ff47SStefano Zampini if (!sub_schurs->reuse_mumps) { 249380677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 249480677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2495be83ff47SStefano Zampini } else { 2496be83ff47SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 2497be83ff47SStefano Zampini 249853892102SStefano Zampini ierr = VecScatterBegin(reuse_mumps->correction_scatter_B,reuse_mumps->sol_B,inout_B,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 249953892102SStefano Zampini ierr = VecScatterEnd(reuse_mumps->correction_scatter_B,reuse_mumps->sol_B,inout_B,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2500be83ff47SStefano Zampini } 250180677318SStefano Zampini if (!applytranspose && pcbddc->local_auxmat1) { 250280677318SStefano Zampini ierr = MatMult(pcbddc->local_auxmat1,inout_B,pcbddc->vec1_C);CHKERRQ(ierr); 250380677318SStefano Zampini ierr = MatMultAdd(pcbddc->local_auxmat2,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr); 250480677318SStefano Zampini } 250580677318SStefano Zampini } else { 250680677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 250780677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 250880677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 250980677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 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,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 251380677318SStefano Zampini } 251480677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 251580677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 251680677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 251780677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2518674ae819SStefano Zampini } 2519674ae819SStefano Zampini PetscFunctionReturn(0); 2520674ae819SStefano Zampini } 2521674ae819SStefano Zampini 2522dc359a40SStefano Zampini /* parameter apply transpose determines if the interface preconditioner should be applied transposed or not */ 2523674ae819SStefano Zampini #undef __FUNCT__ 2524674ae819SStefano Zampini #define __FUNCT__ "PCBDDCApplyInterfacePreconditioner" 2525dc359a40SStefano Zampini PetscErrorCode PCBDDCApplyInterfacePreconditioner(PC pc, PetscBool applytranspose) 2526674ae819SStefano Zampini { 2527674ae819SStefano Zampini PetscErrorCode ierr; 2528674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 2529674ae819SStefano Zampini PC_IS* pcis = (PC_IS*) (pc->data); 2530674ae819SStefano Zampini const PetscScalar zero = 0.0; 2531674ae819SStefano Zampini 2532674ae819SStefano Zampini PetscFunctionBegin; 2533dc359a40SStefano Zampini /* Application of PSI^T or PHI^T (depending on applytranspose, see comment above) */ 2534dc359a40SStefano Zampini if (applytranspose) { 2535674ae819SStefano Zampini ierr = MatMultTranspose(pcbddc->coarse_phi_B,pcis->vec1_B,pcbddc->vec1_P);CHKERRQ(ierr); 25368eeda7d8SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultTransposeAdd(pcbddc->coarse_phi_D,pcis->vec1_D,pcbddc->vec1_P,pcbddc->vec1_P);CHKERRQ(ierr); } 2537dc359a40SStefano Zampini } else { 2538674ae819SStefano Zampini ierr = MatMultTranspose(pcbddc->coarse_psi_B,pcis->vec1_B,pcbddc->vec1_P);CHKERRQ(ierr); 2539674ae819SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultTransposeAdd(pcbddc->coarse_psi_D,pcis->vec1_D,pcbddc->vec1_P,pcbddc->vec1_P);CHKERRQ(ierr); } 254015aaf578SStefano Zampini } 2541efc2fbd9SStefano Zampini 2542efc2fbd9SStefano Zampini /* add p0 to the last value of vec1_P holding the coarse dof relative to p0 */ 25434f1b2e48SStefano Zampini if (pcbddc->benign_n) { 2544efc2fbd9SStefano Zampini PetscScalar *array; 25454f1b2e48SStefano Zampini PetscInt j; 2546efc2fbd9SStefano Zampini 2547efc2fbd9SStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 25484f1b2e48SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) array[pcbddc->local_primal_size-pcbddc->benign_n+j] += pcbddc->benign_p0[j]; 2549efc2fbd9SStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 2550efc2fbd9SStefano Zampini } 2551efc2fbd9SStefano Zampini 255212edc857SStefano Zampini /* start communications from local primal nodes to rhs of coarse solver */ 255312edc857SStefano Zampini ierr = VecSet(pcbddc->coarse_vec,zero);CHKERRQ(ierr); 255412edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataBegin(pc,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 255512edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataEnd(pc,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 255612edc857SStefano Zampini 25579f00e9b4SStefano Zampini /* Coarse solution -> rhs and sol updated inside PCBDDCScattarCoarseDataBegin/End */ 255812edc857SStefano Zampini /* TODO remove null space when doing multilevel */ 255912edc857SStefano Zampini if (pcbddc->coarse_ksp) { 256051694757SStefano Zampini Mat coarse_mat; 2561964fefecSStefano Zampini Vec rhs,sol; 256251694757SStefano Zampini MatNullSpace nullsp; 2563964fefecSStefano Zampini 2564964fefecSStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&rhs);CHKERRQ(ierr); 2565964fefecSStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&sol);CHKERRQ(ierr); 256651694757SStefano Zampini ierr = KSPGetOperators(pcbddc->coarse_ksp,&coarse_mat,NULL);CHKERRQ(ierr); 256751694757SStefano Zampini ierr = MatGetNullSpace(coarse_mat,&nullsp);CHKERRQ(ierr); 256851694757SStefano Zampini if (nullsp) { 256951694757SStefano Zampini ierr = MatNullSpaceRemove(nullsp,rhs);CHKERRQ(ierr); 257051694757SStefano Zampini } 257112edc857SStefano Zampini if (applytranspose) { 2572964fefecSStefano Zampini ierr = KSPSolveTranspose(pcbddc->coarse_ksp,rhs,sol);CHKERRQ(ierr); 257312edc857SStefano Zampini } else { 2574964fefecSStefano Zampini ierr = KSPSolve(pcbddc->coarse_ksp,rhs,sol);CHKERRQ(ierr); 257512edc857SStefano Zampini } 257651694757SStefano Zampini if (nullsp) { 257751694757SStefano Zampini ierr = MatNullSpaceRemove(nullsp,sol);CHKERRQ(ierr); 257851694757SStefano Zampini } 257912edc857SStefano Zampini } 2580674ae819SStefano Zampini 2581674ae819SStefano Zampini /* Local solution on R nodes */ 258280677318SStefano Zampini if (pcis->n) { /* in/out pcbddc->vec1_B,pcbddc->vec1_D */ 258380677318SStefano Zampini ierr = PCBDDCSolveSubstructureCorrection(pc,pcis->vec1_B,pcis->vec1_D,applytranspose);CHKERRQ(ierr); 25849f00e9b4SStefano Zampini } 2585674ae819SStefano Zampini 25869f00e9b4SStefano Zampini /* communications from coarse sol to local primal nodes */ 25879f00e9b4SStefano Zampini ierr = PCBDDCScatterCoarseDataBegin(pc,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 258812edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataEnd(pc,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2589674ae819SStefano Zampini 2590674ae819SStefano Zampini /* Sum contributions from two levels */ 2591dc359a40SStefano Zampini if (applytranspose) { 2592dc359a40SStefano Zampini ierr = MatMultAdd(pcbddc->coarse_psi_B,pcbddc->vec1_P,pcis->vec1_B,pcis->vec1_B);CHKERRQ(ierr); 2593dc359a40SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultAdd(pcbddc->coarse_psi_D,pcbddc->vec1_P,pcis->vec1_D,pcis->vec1_D);CHKERRQ(ierr); } 2594dc359a40SStefano Zampini } else { 2595674ae819SStefano Zampini ierr = MatMultAdd(pcbddc->coarse_phi_B,pcbddc->vec1_P,pcis->vec1_B,pcis->vec1_B);CHKERRQ(ierr); 25968eeda7d8SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultAdd(pcbddc->coarse_phi_D,pcbddc->vec1_P,pcis->vec1_D,pcis->vec1_D);CHKERRQ(ierr); } 2597dc359a40SStefano Zampini } 2598efc2fbd9SStefano Zampini /* store p0 */ 25994f1b2e48SStefano Zampini if (pcbddc->benign_n) { 2600efc2fbd9SStefano Zampini PetscScalar *array; 26014f1b2e48SStefano Zampini PetscInt j; 2602efc2fbd9SStefano Zampini 2603efc2fbd9SStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 26044f1b2e48SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) pcbddc->benign_p0[j] = array[pcbddc->local_primal_size-pcbddc->benign_n+j]; 2605efc2fbd9SStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 2606efc2fbd9SStefano Zampini } 2607674ae819SStefano Zampini PetscFunctionReturn(0); 2608674ae819SStefano Zampini } 2609674ae819SStefano Zampini 2610674ae819SStefano Zampini #undef __FUNCT__ 2611674ae819SStefano Zampini #define __FUNCT__ "PCBDDCScatterCoarseDataBegin" 261212edc857SStefano Zampini PetscErrorCode PCBDDCScatterCoarseDataBegin(PC pc,InsertMode imode, ScatterMode smode) 2613674ae819SStefano Zampini { 2614674ae819SStefano Zampini PetscErrorCode ierr; 2615674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 261658da7f69SStefano Zampini PetscScalar *array; 261712edc857SStefano Zampini Vec from,to; 2618674ae819SStefano Zampini 2619674ae819SStefano Zampini PetscFunctionBegin; 262012edc857SStefano Zampini if (smode == SCATTER_REVERSE) { /* from global to local -> get data from coarse solution */ 262112edc857SStefano Zampini from = pcbddc->coarse_vec; 262212edc857SStefano Zampini to = pcbddc->vec1_P; 262312edc857SStefano Zampini if (pcbddc->coarse_ksp) { /* get array from coarse processes */ 262412edc857SStefano Zampini Vec tvec; 262558da7f69SStefano Zampini 262658da7f69SStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 262758da7f69SStefano Zampini ierr = VecResetArray(tvec);CHKERRQ(ierr); 262812edc857SStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 262958da7f69SStefano Zampini ierr = VecGetArray(tvec,&array);CHKERRQ(ierr); 263058da7f69SStefano Zampini ierr = VecPlaceArray(from,array);CHKERRQ(ierr); 263158da7f69SStefano Zampini ierr = VecRestoreArray(tvec,&array);CHKERRQ(ierr); 263212edc857SStefano Zampini } 263312edc857SStefano Zampini } else { /* from local to global -> put data in coarse right hand side */ 263412edc857SStefano Zampini from = pcbddc->vec1_P; 263512edc857SStefano Zampini to = pcbddc->coarse_vec; 263612edc857SStefano Zampini } 263712edc857SStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,from,to,imode,smode);CHKERRQ(ierr); 2638674ae819SStefano Zampini PetscFunctionReturn(0); 2639674ae819SStefano Zampini } 2640674ae819SStefano Zampini 2641674ae819SStefano Zampini #undef __FUNCT__ 2642674ae819SStefano Zampini #define __FUNCT__ "PCBDDCScatterCoarseDataEnd" 264312edc857SStefano Zampini PetscErrorCode PCBDDCScatterCoarseDataEnd(PC pc, InsertMode imode, ScatterMode smode) 2644674ae819SStefano Zampini { 2645674ae819SStefano Zampini PetscErrorCode ierr; 2646674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 264758da7f69SStefano Zampini PetscScalar *array; 264812edc857SStefano Zampini Vec from,to; 2649674ae819SStefano Zampini 2650674ae819SStefano Zampini PetscFunctionBegin; 265112edc857SStefano Zampini if (smode == SCATTER_REVERSE) { /* from global to local -> get data from coarse solution */ 265212edc857SStefano Zampini from = pcbddc->coarse_vec; 265312edc857SStefano Zampini to = pcbddc->vec1_P; 265412edc857SStefano Zampini } else { /* from local to global -> put data in coarse right hand side */ 265512edc857SStefano Zampini from = pcbddc->vec1_P; 265612edc857SStefano Zampini to = pcbddc->coarse_vec; 265712edc857SStefano Zampini } 265812edc857SStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,from,to,imode,smode);CHKERRQ(ierr); 265912edc857SStefano Zampini if (smode == SCATTER_FORWARD) { 266012edc857SStefano Zampini if (pcbddc->coarse_ksp) { /* get array from coarse processes */ 266112edc857SStefano Zampini Vec tvec; 266258da7f69SStefano Zampini 266312edc857SStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 266458da7f69SStefano Zampini ierr = VecGetArray(to,&array);CHKERRQ(ierr); 266558da7f69SStefano Zampini ierr = VecPlaceArray(tvec,array);CHKERRQ(ierr); 266658da7f69SStefano Zampini ierr = VecRestoreArray(to,&array);CHKERRQ(ierr); 266758da7f69SStefano Zampini } 266858da7f69SStefano Zampini } else { 266958da7f69SStefano Zampini if (pcbddc->coarse_ksp) { /* restore array of pcbddc->coarse_vec */ 267058da7f69SStefano Zampini ierr = VecResetArray(from);CHKERRQ(ierr); 267112edc857SStefano Zampini } 267212edc857SStefano Zampini } 2673674ae819SStefano Zampini PetscFunctionReturn(0); 2674674ae819SStefano Zampini } 2675674ae819SStefano Zampini 2676984c4197SStefano Zampini /* uncomment for testing purposes */ 2677984c4197SStefano Zampini /* #define PETSC_MISSING_LAPACK_GESVD 1 */ 2678674ae819SStefano Zampini #undef __FUNCT__ 2679674ae819SStefano Zampini #define __FUNCT__ "PCBDDCConstraintsSetUp" 2680674ae819SStefano Zampini PetscErrorCode PCBDDCConstraintsSetUp(PC pc) 2681674ae819SStefano Zampini { 2682674ae819SStefano Zampini PetscErrorCode ierr; 2683674ae819SStefano Zampini PC_IS* pcis = (PC_IS*)(pc->data); 2684674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 2685674ae819SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 2686984c4197SStefano Zampini /* one and zero */ 2687984c4197SStefano Zampini PetscScalar one=1.0,zero=0.0; 2688984c4197SStefano Zampini /* space to store constraints and their local indices */ 26899162d606SStefano Zampini PetscScalar *constraints_data; 26909162d606SStefano Zampini PetscInt *constraints_idxs,*constraints_idxs_B; 26919162d606SStefano Zampini PetscInt *constraints_idxs_ptr,*constraints_data_ptr; 26929162d606SStefano Zampini PetscInt *constraints_n; 2693984c4197SStefano Zampini /* iterators */ 2694b3d85658SStefano Zampini PetscInt i,j,k,total_counts,total_counts_cc,cum; 2695984c4197SStefano Zampini /* BLAS integers */ 2696e310c8b4SStefano Zampini PetscBLASInt lwork,lierr; 2697e310c8b4SStefano Zampini PetscBLASInt Blas_N,Blas_M,Blas_K,Blas_one=1; 2698c4303822SStefano Zampini PetscBLASInt Blas_LDA,Blas_LDB,Blas_LDC; 2699727cdba6SStefano Zampini /* reuse */ 27000e6343abSStefano Zampini PetscInt olocal_primal_size,olocal_primal_size_cc; 27010e6343abSStefano Zampini PetscInt *olocal_primal_ref_node,*olocal_primal_ref_mult; 2702984c4197SStefano Zampini /* change of basis */ 2703b3d85658SStefano Zampini PetscBool qr_needed; 27049162d606SStefano Zampini PetscBT change_basis,qr_needed_idx; 2705984c4197SStefano Zampini /* auxiliary stuff */ 270664efe560SStefano Zampini PetscInt *nnz,*is_indices; 27078a0068c3SStefano Zampini PetscInt ncc; 2708984c4197SStefano Zampini /* some quantities */ 270945a1bb75SStefano Zampini PetscInt n_vertices,total_primal_vertices,valid_constraints; 2710a58a30b4SStefano Zampini PetscInt size_of_constraint,max_size_of_constraint=0,max_constraints,temp_constraints; 2711984c4197SStefano Zampini 2712674ae819SStefano Zampini PetscFunctionBegin; 27138e61c736SStefano Zampini /* Destroy Mat objects computed previously */ 27148e61c736SStefano Zampini ierr = MatDestroy(&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 27158e61c736SStefano Zampini ierr = MatDestroy(&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 2716088faed8SStefano Zampini /* save info on constraints from previous setup (if any) */ 2717088faed8SStefano Zampini olocal_primal_size = pcbddc->local_primal_size; 27180e6343abSStefano Zampini olocal_primal_size_cc = pcbddc->local_primal_size_cc; 27190e6343abSStefano Zampini ierr = PetscMalloc2(olocal_primal_size_cc,&olocal_primal_ref_node,olocal_primal_size_cc,&olocal_primal_ref_mult);CHKERRQ(ierr); 27200e6343abSStefano Zampini ierr = PetscMemcpy(olocal_primal_ref_node,pcbddc->local_primal_ref_node,olocal_primal_size_cc*sizeof(PetscInt));CHKERRQ(ierr); 27210e6343abSStefano Zampini ierr = PetscMemcpy(olocal_primal_ref_mult,pcbddc->local_primal_ref_mult,olocal_primal_size_cc*sizeof(PetscInt));CHKERRQ(ierr); 27220e6343abSStefano Zampini ierr = PetscFree2(pcbddc->local_primal_ref_node,pcbddc->local_primal_ref_mult);CHKERRQ(ierr); 2723088faed8SStefano Zampini ierr = PetscFree(pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 2724cf5a6209SStefano Zampini 2725cf5a6209SStefano Zampini /* print some info */ 2726cf5a6209SStefano Zampini if (pcbddc->dbg_flag) { 2727cf5a6209SStefano Zampini IS vertices; 2728cf5a6209SStefano Zampini PetscInt nv,nedges,nfaces; 2729cf5a6209SStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,&nfaces,NULL,&nedges,NULL,&vertices);CHKERRQ(ierr); 2730cf5a6209SStefano Zampini ierr = ISGetSize(vertices,&nv);CHKERRQ(ierr); 2731cf5a6209SStefano Zampini ierr = ISDestroy(&vertices);CHKERRQ(ierr); 27321575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 2733cf5a6209SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"--------------------------------------------------------------\n");CHKERRQ(ierr); 2734cf5a6209SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate vertices (%d)\n",PetscGlobalRank,nv,pcbddc->use_vertices);CHKERRQ(ierr); 2735fd14bc51SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate edges (%d)\n",PetscGlobalRank,nedges,pcbddc->use_edges);CHKERRQ(ierr); 2736fd14bc51SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate faces (%d)\n",PetscGlobalRank,nfaces,pcbddc->use_faces);CHKERRQ(ierr); 2737cf5a6209SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 27381575c14dSBarry Smith ierr = PetscViewerASCIIPopSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 2739cf5a6209SStefano Zampini } 2740cf5a6209SStefano Zampini 2741cf5a6209SStefano Zampini if (!pcbddc->adaptive_selection) { 27429162d606SStefano Zampini IS ISForVertices,*ISForFaces,*ISForEdges; 2743cf5a6209SStefano Zampini MatNullSpace nearnullsp; 2744cf5a6209SStefano Zampini const Vec *nearnullvecs; 2745cf5a6209SStefano Zampini Vec *localnearnullsp; 2746cf5a6209SStefano Zampini PetscScalar *array; 2747cf5a6209SStefano Zampini PetscInt n_ISForFaces,n_ISForEdges,nnsp_size; 2748cf5a6209SStefano Zampini PetscBool nnsp_has_cnst; 2749674ae819SStefano Zampini /* LAPACK working arrays for SVD or POD */ 2750b3d85658SStefano Zampini PetscBool skip_lapack,boolforchange; 2751674ae819SStefano Zampini PetscScalar *work; 2752674ae819SStefano Zampini PetscReal *singular_vals; 2753674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX) 2754674ae819SStefano Zampini PetscReal *rwork; 2755674ae819SStefano Zampini #endif 2756674ae819SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 2757674ae819SStefano Zampini PetscScalar *temp_basis,*correlation_mat; 2758674ae819SStefano Zampini #else 2759964fefecSStefano Zampini PetscBLASInt dummy_int=1; 2760964fefecSStefano Zampini PetscScalar dummy_scalar=1.; 2761674ae819SStefano Zampini #endif 2762674ae819SStefano Zampini 2763674ae819SStefano Zampini /* Get index sets for faces, edges and vertices from graph */ 2764d06fc5fdSStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,&n_ISForFaces,&ISForFaces,&n_ISForEdges,&ISForEdges,&ISForVertices);CHKERRQ(ierr); 2765d06fc5fdSStefano Zampini /* free unneeded index sets */ 2766d06fc5fdSStefano Zampini if (!pcbddc->use_vertices) { 2767d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForVertices);CHKERRQ(ierr); 2768674ae819SStefano Zampini } 2769d06fc5fdSStefano Zampini if (!pcbddc->use_edges) { 2770d06fc5fdSStefano Zampini for (i=0;i<n_ISForEdges;i++) { 2771d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForEdges[i]);CHKERRQ(ierr); 2772d06fc5fdSStefano Zampini } 2773d06fc5fdSStefano Zampini ierr = PetscFree(ISForEdges);CHKERRQ(ierr); 2774d06fc5fdSStefano Zampini n_ISForEdges = 0; 2775d06fc5fdSStefano Zampini } 2776d06fc5fdSStefano Zampini if (!pcbddc->use_faces) { 2777d06fc5fdSStefano Zampini for (i=0;i<n_ISForFaces;i++) { 2778d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForFaces[i]);CHKERRQ(ierr); 2779d06fc5fdSStefano Zampini } 2780d06fc5fdSStefano Zampini ierr = PetscFree(ISForFaces);CHKERRQ(ierr); 2781d06fc5fdSStefano Zampini n_ISForFaces = 0; 2782d06fc5fdSStefano Zampini } 278370022509SStefano Zampini 278470022509SStefano Zampini #if defined(PETSC_USE_DEBUG) 278570022509SStefano Zampini /* HACK: when solving singular problems not using vertices, a change of basis is mandatory. 278670022509SStefano Zampini Also use_change_of_basis should be consistent among processors */ 278770022509SStefano Zampini if (pcbddc->NullSpace) { 278870022509SStefano Zampini PetscBool tbool[2],gbool[2]; 278970022509SStefano Zampini 279070022509SStefano Zampini if (!ISForVertices && !pcbddc->user_ChangeOfBasisMatrix) { 2791b8ffe317SStefano Zampini pcbddc->use_change_of_basis = PETSC_TRUE; 2792d06fc5fdSStefano Zampini if (!ISForEdges) { 2793d06fc5fdSStefano Zampini pcbddc->use_change_on_faces = PETSC_TRUE; 2794d06fc5fdSStefano Zampini } 2795b8ffe317SStefano Zampini } 2796d06fc5fdSStefano Zampini tbool[0] = pcbddc->use_change_of_basis; 2797d06fc5fdSStefano Zampini tbool[1] = pcbddc->use_change_on_faces; 2798d06fc5fdSStefano Zampini ierr = MPI_Allreduce(tbool,gbool,2,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 2799d06fc5fdSStefano Zampini pcbddc->use_change_of_basis = gbool[0]; 2800d06fc5fdSStefano Zampini pcbddc->use_change_on_faces = gbool[1]; 280198a51de6SStefano Zampini } 280270022509SStefano Zampini #endif 280308122e43SStefano Zampini 2804674ae819SStefano Zampini /* check if near null space is attached to global mat */ 2805674ae819SStefano Zampini ierr = MatGetNearNullSpace(pc->pmat,&nearnullsp);CHKERRQ(ierr); 2806674ae819SStefano Zampini if (nearnullsp) { 2807674ae819SStefano Zampini ierr = MatNullSpaceGetVecs(nearnullsp,&nnsp_has_cnst,&nnsp_size,&nearnullvecs);CHKERRQ(ierr); 2808f4ddd8eeSStefano Zampini /* remove any stored info */ 2809f4ddd8eeSStefano Zampini ierr = MatNullSpaceDestroy(&pcbddc->onearnullspace);CHKERRQ(ierr); 2810f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->onearnullvecs_state);CHKERRQ(ierr); 2811f4ddd8eeSStefano Zampini /* store information for BDDC solver reuse */ 2812f4ddd8eeSStefano Zampini ierr = PetscObjectReference((PetscObject)nearnullsp);CHKERRQ(ierr); 2813f4ddd8eeSStefano Zampini pcbddc->onearnullspace = nearnullsp; 2814473ba861SJed Brown ierr = PetscMalloc1(nnsp_size,&pcbddc->onearnullvecs_state);CHKERRQ(ierr); 2815f4ddd8eeSStefano Zampini for (i=0;i<nnsp_size;i++) { 2816f4ddd8eeSStefano Zampini ierr = PetscObjectStateGet((PetscObject)nearnullvecs[i],&pcbddc->onearnullvecs_state[i]);CHKERRQ(ierr); 2817f4ddd8eeSStefano Zampini } 2818984c4197SStefano Zampini } else { /* if near null space is not provided BDDC uses constants by default */ 2819984c4197SStefano Zampini nnsp_size = 0; 2820674ae819SStefano Zampini nnsp_has_cnst = PETSC_TRUE; 2821674ae819SStefano Zampini } 2822984c4197SStefano Zampini /* get max number of constraints on a single cc */ 2823984c4197SStefano Zampini max_constraints = nnsp_size; 2824984c4197SStefano Zampini if (nnsp_has_cnst) max_constraints++; 2825984c4197SStefano Zampini 2826674ae819SStefano Zampini /* 2827674ae819SStefano Zampini Evaluate maximum storage size needed by the procedure 28289162d606SStefano Zampini - Indices for connected component i stored at "constraints_idxs + constraints_idxs_ptr[i]" 28299162d606SStefano Zampini - Values for constraints on connected component i stored at "constraints_data + constraints_data_ptr[i]" 28309162d606SStefano Zampini There can be multiple constraints per connected component 2831674ae819SStefano Zampini */ 2832674ae819SStefano Zampini n_vertices = 0; 2833674ae819SStefano Zampini if (ISForVertices) { 2834674ae819SStefano Zampini ierr = ISGetSize(ISForVertices,&n_vertices);CHKERRQ(ierr); 2835674ae819SStefano Zampini } 28369162d606SStefano Zampini ncc = n_vertices+n_ISForFaces+n_ISForEdges; 28379162d606SStefano Zampini ierr = PetscMalloc3(ncc+1,&constraints_idxs_ptr,ncc+1,&constraints_data_ptr,ncc,&constraints_n);CHKERRQ(ierr); 28389162d606SStefano Zampini 28399162d606SStefano Zampini total_counts = n_ISForFaces+n_ISForEdges; 28409162d606SStefano Zampini total_counts *= max_constraints; 2841674ae819SStefano Zampini total_counts += n_vertices; 28424641a718SStefano Zampini ierr = PetscBTCreate(total_counts,&change_basis);CHKERRQ(ierr); 28439162d606SStefano Zampini 2844674ae819SStefano Zampini total_counts = 0; 2845674ae819SStefano Zampini max_size_of_constraint = 0; 2846674ae819SStefano Zampini for (i=0;i<n_ISForEdges+n_ISForFaces;i++) { 28479162d606SStefano Zampini IS used_is; 2848674ae819SStefano Zampini if (i<n_ISForEdges) { 28499162d606SStefano Zampini used_is = ISForEdges[i]; 2850674ae819SStefano Zampini } else { 28519162d606SStefano Zampini used_is = ISForFaces[i-n_ISForEdges]; 2852674ae819SStefano Zampini } 28539162d606SStefano Zampini ierr = ISGetSize(used_is,&j);CHKERRQ(ierr); 2854674ae819SStefano Zampini total_counts += j; 2855674ae819SStefano Zampini max_size_of_constraint = PetscMax(j,max_size_of_constraint); 2856674ae819SStefano Zampini } 28579162d606SStefano 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); 28589162d606SStefano Zampini 2859984c4197SStefano Zampini /* get local part of global near null space vectors */ 2860785e854fSJed Brown ierr = PetscMalloc1(nnsp_size,&localnearnullsp);CHKERRQ(ierr); 2861984c4197SStefano Zampini for (k=0;k<nnsp_size;k++) { 2862984c4197SStefano Zampini ierr = VecDuplicate(pcis->vec1_N,&localnearnullsp[k]);CHKERRQ(ierr); 2863e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,nearnullvecs[k],localnearnullsp[k],INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2864e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,nearnullvecs[k],localnearnullsp[k],INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2865984c4197SStefano Zampini } 2866674ae819SStefano Zampini 2867242a89d7SStefano Zampini /* whether or not to skip lapack calls */ 2868242a89d7SStefano Zampini skip_lapack = PETSC_TRUE; 2869a773dcb8SStefano Zampini if (n_ISForFaces+n_ISForEdges && max_constraints > 1 && !pcbddc->use_nnsp_true) skip_lapack = PETSC_FALSE; 2870242a89d7SStefano Zampini 2871984c4197SStefano Zampini /* First we issue queries to allocate optimal workspace for LAPACKgesvd (or LAPACKsyev if SVD is missing) */ 2872a773dcb8SStefano Zampini if (!skip_lapack) { 2873674ae819SStefano Zampini PetscScalar temp_work; 2874911cabfeSStefano Zampini 2875674ae819SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 2876984c4197SStefano Zampini /* Proper Orthogonal Decomposition (POD) using the snapshot method */ 2877785e854fSJed Brown ierr = PetscMalloc1(max_constraints*max_constraints,&correlation_mat);CHKERRQ(ierr); 2878785e854fSJed Brown ierr = PetscMalloc1(max_constraints,&singular_vals);CHKERRQ(ierr); 2879785e854fSJed Brown ierr = PetscMalloc1(max_size_of_constraint*max_constraints,&temp_basis);CHKERRQ(ierr); 2880674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX) 2881785e854fSJed Brown ierr = PetscMalloc1(3*max_constraints,&rwork);CHKERRQ(ierr); 2882674ae819SStefano Zampini #endif 2883674ae819SStefano Zampini /* now we evaluate the optimal workspace using query with lwork=-1 */ 2884c8244a33SStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_N);CHKERRQ(ierr); 2885c8244a33SStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_LDA);CHKERRQ(ierr); 2886674ae819SStefano Zampini lwork = -1; 2887674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 2888674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 2889c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,&temp_work,&lwork,&lierr)); 2890674ae819SStefano Zampini #else 2891c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,&temp_work,&lwork,rwork,&lierr)); 2892674ae819SStefano Zampini #endif 2893674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2894984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to SYEV Lapack routine %d",(int)lierr); 2895674ae819SStefano Zampini #else /* on missing GESVD */ 2896674ae819SStefano Zampini /* SVD */ 2897674ae819SStefano Zampini PetscInt max_n,min_n; 2898674ae819SStefano Zampini max_n = max_size_of_constraint; 2899984c4197SStefano Zampini min_n = max_constraints; 2900984c4197SStefano Zampini if (max_size_of_constraint < max_constraints) { 2901674ae819SStefano Zampini min_n = max_size_of_constraint; 2902984c4197SStefano Zampini max_n = max_constraints; 2903674ae819SStefano Zampini } 2904785e854fSJed Brown ierr = PetscMalloc1(min_n,&singular_vals);CHKERRQ(ierr); 2905674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX) 2906785e854fSJed Brown ierr = PetscMalloc1(5*min_n,&rwork);CHKERRQ(ierr); 2907674ae819SStefano Zampini #endif 2908674ae819SStefano Zampini /* now we evaluate the optimal workspace using query with lwork=-1 */ 2909674ae819SStefano Zampini lwork = -1; 2910e310c8b4SStefano Zampini ierr = PetscBLASIntCast(max_n,&Blas_M);CHKERRQ(ierr); 2911e310c8b4SStefano Zampini ierr = PetscBLASIntCast(min_n,&Blas_N);CHKERRQ(ierr); 2912b7d8b9f8SStefano Zampini ierr = PetscBLASIntCast(max_n,&Blas_LDA);CHKERRQ(ierr); 2913674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 2914674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 29159162d606SStefano 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)); 2916674ae819SStefano Zampini #else 29179162d606SStefano 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)); 2918674ae819SStefano Zampini #endif 2919674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2920984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to GESVD Lapack routine %d",(int)lierr); 2921984c4197SStefano Zampini #endif /* on missing GESVD */ 2922674ae819SStefano Zampini /* Allocate optimal workspace */ 2923674ae819SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(temp_work),&lwork);CHKERRQ(ierr); 2924854ce69bSBarry Smith ierr = PetscMalloc1(lwork,&work);CHKERRQ(ierr); 2925674ae819SStefano Zampini } 2926674ae819SStefano Zampini /* Now we can loop on constraining sets */ 2927674ae819SStefano Zampini total_counts = 0; 29289162d606SStefano Zampini constraints_idxs_ptr[0] = 0; 29299162d606SStefano Zampini constraints_data_ptr[0] = 0; 2930674ae819SStefano Zampini /* vertices */ 29319162d606SStefano Zampini if (n_vertices) { 2932674ae819SStefano Zampini ierr = ISGetIndices(ISForVertices,(const PetscInt**)&is_indices);CHKERRQ(ierr); 29339162d606SStefano Zampini ierr = PetscMemcpy(constraints_idxs,is_indices,n_vertices*sizeof(PetscInt));CHKERRQ(ierr); 2934674ae819SStefano Zampini for (i=0;i<n_vertices;i++) { 29359162d606SStefano Zampini constraints_n[total_counts] = 1; 29369162d606SStefano Zampini constraints_data[total_counts] = 1.0; 29379162d606SStefano Zampini constraints_idxs_ptr[total_counts+1] = constraints_idxs_ptr[total_counts]+1; 29389162d606SStefano Zampini constraints_data_ptr[total_counts+1] = constraints_data_ptr[total_counts]+1; 2939674ae819SStefano Zampini total_counts++; 2940674ae819SStefano Zampini } 2941674ae819SStefano Zampini ierr = ISRestoreIndices(ISForVertices,(const PetscInt**)&is_indices);CHKERRQ(ierr); 2942674ae819SStefano Zampini n_vertices = total_counts; 2943674ae819SStefano Zampini } 2944984c4197SStefano Zampini 2945674ae819SStefano Zampini /* edges and faces */ 29469162d606SStefano Zampini total_counts_cc = total_counts; 2947911cabfeSStefano Zampini for (ncc=0;ncc<n_ISForEdges+n_ISForFaces;ncc++) { 29489162d606SStefano Zampini IS used_is; 29499162d606SStefano Zampini PetscBool idxs_copied = PETSC_FALSE; 29509162d606SStefano Zampini 2951911cabfeSStefano Zampini if (ncc<n_ISForEdges) { 29529162d606SStefano Zampini used_is = ISForEdges[ncc]; 2953984c4197SStefano Zampini boolforchange = pcbddc->use_change_of_basis; /* change or not the basis on the edge */ 2954674ae819SStefano Zampini } else { 29559162d606SStefano Zampini used_is = ISForFaces[ncc-n_ISForEdges]; 2956984c4197SStefano Zampini boolforchange = (PetscBool)(pcbddc->use_change_of_basis && pcbddc->use_change_on_faces); /* change or not the basis on the face */ 2957674ae819SStefano Zampini } 2958674ae819SStefano Zampini temp_constraints = 0; /* zero the number of constraints I have on this conn comp */ 29599162d606SStefano Zampini 29609162d606SStefano Zampini ierr = ISGetSize(used_is,&size_of_constraint);CHKERRQ(ierr); 29619162d606SStefano Zampini ierr = ISGetIndices(used_is,(const PetscInt**)&is_indices);CHKERRQ(ierr); 2962984c4197SStefano Zampini /* change of basis should not be performed on local periodic nodes */ 2963984c4197SStefano Zampini if (pcbddc->mat_graph->mirrors && pcbddc->mat_graph->mirrors[is_indices[0]]) boolforchange = PETSC_FALSE; 2964674ae819SStefano Zampini if (nnsp_has_cnst) { 29655b08dc53SStefano Zampini PetscScalar quad_value; 29669162d606SStefano Zampini 29679162d606SStefano Zampini ierr = PetscMemcpy(constraints_idxs + constraints_idxs_ptr[total_counts_cc],is_indices,size_of_constraint*sizeof(PetscInt));CHKERRQ(ierr); 29689162d606SStefano Zampini idxs_copied = PETSC_TRUE; 29699162d606SStefano Zampini 2970a773dcb8SStefano Zampini if (!pcbddc->use_nnsp_true) { 2971674ae819SStefano Zampini quad_value = (PetscScalar)(1.0/PetscSqrtReal((PetscReal)size_of_constraint)); 2972a773dcb8SStefano Zampini } else { 2973a773dcb8SStefano Zampini quad_value = 1.0; 2974a773dcb8SStefano Zampini } 2975674ae819SStefano Zampini for (j=0;j<size_of_constraint;j++) { 29769162d606SStefano Zampini constraints_data[constraints_data_ptr[total_counts_cc]+j] = quad_value; 2977674ae819SStefano Zampini } 29789162d606SStefano Zampini temp_constraints++; 2979674ae819SStefano Zampini total_counts++; 2980674ae819SStefano Zampini } 2981674ae819SStefano Zampini for (k=0;k<nnsp_size;k++) { 2982984c4197SStefano Zampini PetscReal real_value; 29839162d606SStefano Zampini PetscScalar *ptr_to_data; 29849162d606SStefano Zampini 2985984c4197SStefano Zampini ierr = VecGetArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 29869162d606SStefano Zampini ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]+temp_constraints*size_of_constraint]; 2987674ae819SStefano Zampini for (j=0;j<size_of_constraint;j++) { 29889162d606SStefano Zampini ptr_to_data[j] = array[is_indices[j]]; 2989674ae819SStefano Zampini } 2990984c4197SStefano Zampini ierr = VecRestoreArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 2991984c4197SStefano Zampini /* check if array is null on the connected component */ 2992e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 29939162d606SStefano Zampini PetscStackCallBLAS("BLASasum",real_value = BLASasum_(&Blas_N,ptr_to_data,&Blas_one)); 29945b08dc53SStefano Zampini if (real_value > 0.0) { /* keep indices and values */ 2995674ae819SStefano Zampini temp_constraints++; 2996674ae819SStefano Zampini total_counts++; 29979162d606SStefano Zampini if (!idxs_copied) { 29989162d606SStefano Zampini ierr = PetscMemcpy(constraints_idxs + constraints_idxs_ptr[total_counts_cc],is_indices,size_of_constraint*sizeof(PetscInt));CHKERRQ(ierr); 29999162d606SStefano Zampini idxs_copied = PETSC_TRUE; 3000674ae819SStefano Zampini } 3001674ae819SStefano Zampini } 30029162d606SStefano Zampini } 30039162d606SStefano Zampini ierr = ISRestoreIndices(used_is,(const PetscInt**)&is_indices);CHKERRQ(ierr); 300445a1bb75SStefano Zampini valid_constraints = temp_constraints; 3005eb97c9d2SStefano Zampini if (!pcbddc->use_nnsp_true && temp_constraints) { 3006a773dcb8SStefano Zampini if (temp_constraints == 1) { /* just normalize the constraint */ 30079162d606SStefano Zampini PetscScalar norm,*ptr_to_data; 30089162d606SStefano Zampini 30099162d606SStefano Zampini ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]]; 3010a773dcb8SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 30119162d606SStefano Zampini PetscStackCallBLAS("BLASdot",norm = BLASdot_(&Blas_N,ptr_to_data,&Blas_one,ptr_to_data,&Blas_one)); 3012a773dcb8SStefano Zampini norm = 1.0/PetscSqrtReal(PetscRealPart(norm)); 30139162d606SStefano Zampini PetscStackCallBLAS("BLASscal",BLASscal_(&Blas_N,&norm,ptr_to_data,&Blas_one)); 3014a773dcb8SStefano Zampini } else { /* perform SVD */ 3015984c4197SStefano Zampini PetscReal tol = 1.0e-8; /* tolerance for retaining eigenmodes */ 30169162d606SStefano Zampini PetscScalar *ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]]; 3017674ae819SStefano Zampini 3018674ae819SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 3019984c4197SStefano Zampini /* SVD: Y = U*S*V^H -> U (eigenvectors of Y*Y^H) = Y*V*(S)^\dag 3020984c4197SStefano Zampini POD: Y^H*Y = V*D*V^H, D = S^H*S -> U = Y*V*D^(-1/2) 3021984c4197SStefano Zampini -> When PETSC_USE_COMPLEX and PETSC_MISSING_LAPACK_GESVD are defined 3022984c4197SStefano Zampini the constraints basis will differ (by a complex factor with absolute value equal to 1) 3023984c4197SStefano Zampini from that computed using LAPACKgesvd 3024984c4197SStefano Zampini -> This is due to a different computation of eigenvectors in LAPACKheev 3025984c4197SStefano Zampini -> The quality of the POD-computed basis will be the same */ 3026984c4197SStefano Zampini ierr = PetscMemzero(correlation_mat,temp_constraints*temp_constraints*sizeof(PetscScalar));CHKERRQ(ierr); 3027674ae819SStefano Zampini /* Store upper triangular part of correlation matrix */ 3028e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 3029984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 3030674ae819SStefano Zampini for (j=0;j<temp_constraints;j++) { 3031674ae819SStefano Zampini for (k=0;k<j+1;k++) { 30329162d606SStefano 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)); 3033674ae819SStefano Zampini } 3034674ae819SStefano Zampini } 3035e310c8b4SStefano Zampini /* compute eigenvalues and eigenvectors of correlation matrix */ 3036e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 3037e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_LDA);CHKERRQ(ierr); 3038674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 3039c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,work,&lwork,&lierr)); 3040674ae819SStefano Zampini #else 3041c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,work,&lwork,rwork,&lierr)); 3042674ae819SStefano Zampini #endif 3043674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 3044984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYEV Lapack routine %d",(int)lierr); 3045984c4197SStefano Zampini /* retain eigenvalues greater than tol: note that LAPACKsyev gives eigs in ascending order */ 3046674ae819SStefano Zampini j = 0; 3047984c4197SStefano Zampini while (j < temp_constraints && singular_vals[j] < tol) j++; 3048674ae819SStefano Zampini total_counts = total_counts-j; 304945a1bb75SStefano Zampini valid_constraints = temp_constraints-j; 3050e310c8b4SStefano Zampini /* scale and copy POD basis into used quadrature memory */ 3051c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 3052c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 3053c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_K);CHKERRQ(ierr); 3054c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 3055c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_LDB);CHKERRQ(ierr); 3056c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDC);CHKERRQ(ierr); 3057674ae819SStefano Zampini if (j<temp_constraints) { 3058984c4197SStefano Zampini PetscInt ii; 3059984c4197SStefano Zampini for (k=j;k<temp_constraints;k++) singular_vals[k] = 1.0/PetscSqrtReal(singular_vals[k]); 3060674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 30619162d606SStefano 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)); 3062674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 3063984c4197SStefano Zampini for (k=0;k<temp_constraints-j;k++) { 3064674ae819SStefano Zampini for (ii=0;ii<size_of_constraint;ii++) { 30659162d606SStefano 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]; 3066674ae819SStefano Zampini } 3067674ae819SStefano Zampini } 3068674ae819SStefano Zampini } 3069674ae819SStefano Zampini #else /* on missing GESVD */ 3070e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 3071e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 3072b7d8b9f8SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 3073674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 3074674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 30759162d606SStefano 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)); 3076674ae819SStefano Zampini #else 30779162d606SStefano 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)); 3078674ae819SStefano Zampini #endif 3079984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GESVD Lapack routine %d",(int)lierr); 3080674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 3081984c4197SStefano Zampini /* retain eigenvalues greater than tol: note that LAPACKgesvd gives eigs in descending order */ 3082e310c8b4SStefano Zampini k = temp_constraints; 3083e310c8b4SStefano Zampini if (k > size_of_constraint) k = size_of_constraint; 3084674ae819SStefano Zampini j = 0; 3085e310c8b4SStefano Zampini while (j < k && singular_vals[k-j-1] < tol) j++; 308645a1bb75SStefano Zampini valid_constraints = k-j; 3087911cabfeSStefano Zampini total_counts = total_counts-temp_constraints+valid_constraints; 3088984c4197SStefano Zampini #endif /* on missing GESVD */ 3089674ae819SStefano Zampini } 3090a773dcb8SStefano Zampini } 30919162d606SStefano Zampini /* update pointers information */ 30929162d606SStefano Zampini if (valid_constraints) { 30939162d606SStefano Zampini constraints_n[total_counts_cc] = valid_constraints; 30949162d606SStefano Zampini constraints_idxs_ptr[total_counts_cc+1] = constraints_idxs_ptr[total_counts_cc]+size_of_constraint; 30959162d606SStefano Zampini constraints_data_ptr[total_counts_cc+1] = constraints_data_ptr[total_counts_cc]+size_of_constraint*valid_constraints; 30969162d606SStefano Zampini /* set change_of_basis flag */ 309745a1bb75SStefano Zampini if (boolforchange) { 3098b3d85658SStefano Zampini PetscBTSet(change_basis,total_counts_cc); 30999162d606SStefano Zampini } 3100b3d85658SStefano Zampini total_counts_cc++; 310145a1bb75SStefano Zampini } 310245a1bb75SStefano Zampini } 3103984c4197SStefano Zampini /* free workspace */ 31048f1c130eSStefano Zampini if (!skip_lapack) { 3105984c4197SStefano Zampini ierr = PetscFree(work);CHKERRQ(ierr); 3106984c4197SStefano Zampini #if defined(PETSC_USE_COMPLEX) 3107984c4197SStefano Zampini ierr = PetscFree(rwork);CHKERRQ(ierr); 3108984c4197SStefano Zampini #endif 3109984c4197SStefano Zampini ierr = PetscFree(singular_vals);CHKERRQ(ierr); 3110984c4197SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 3111984c4197SStefano Zampini ierr = PetscFree(correlation_mat);CHKERRQ(ierr); 3112984c4197SStefano Zampini ierr = PetscFree(temp_basis);CHKERRQ(ierr); 3113984c4197SStefano Zampini #endif 3114984c4197SStefano Zampini } 3115984c4197SStefano Zampini for (k=0;k<nnsp_size;k++) { 3116984c4197SStefano Zampini ierr = VecDestroy(&localnearnullsp[k]);CHKERRQ(ierr); 3117984c4197SStefano Zampini } 3118984c4197SStefano Zampini ierr = PetscFree(localnearnullsp);CHKERRQ(ierr); 3119cf5a6209SStefano Zampini /* free index sets of faces, edges and vertices */ 3120cf5a6209SStefano Zampini for (i=0;i<n_ISForFaces;i++) { 3121cf5a6209SStefano Zampini ierr = ISDestroy(&ISForFaces[i]);CHKERRQ(ierr); 3122cf5a6209SStefano Zampini } 3123cf5a6209SStefano Zampini if (n_ISForFaces) { 3124cf5a6209SStefano Zampini ierr = PetscFree(ISForFaces);CHKERRQ(ierr); 3125cf5a6209SStefano Zampini } 3126cf5a6209SStefano Zampini for (i=0;i<n_ISForEdges;i++) { 3127cf5a6209SStefano Zampini ierr = ISDestroy(&ISForEdges[i]);CHKERRQ(ierr); 3128cf5a6209SStefano Zampini } 3129cf5a6209SStefano Zampini if (n_ISForEdges) { 3130cf5a6209SStefano Zampini ierr = PetscFree(ISForEdges);CHKERRQ(ierr); 3131cf5a6209SStefano Zampini } 3132cf5a6209SStefano Zampini ierr = ISDestroy(&ISForVertices);CHKERRQ(ierr); 313308122e43SStefano Zampini } else { 313408122e43SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 3135984c4197SStefano Zampini 313608122e43SStefano Zampini total_counts = 0; 313708122e43SStefano Zampini n_vertices = 0; 3138d62866d3SStefano Zampini if (sub_schurs->is_vertices && pcbddc->use_vertices) { 3139d62866d3SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_vertices,&n_vertices);CHKERRQ(ierr); 314008122e43SStefano Zampini } 314108122e43SStefano Zampini max_constraints = 0; 31429162d606SStefano Zampini total_counts_cc = 0; 314308122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs+n_vertices;i++) { 314408122e43SStefano Zampini total_counts += pcbddc->adaptive_constraints_n[i]; 31459162d606SStefano Zampini if (pcbddc->adaptive_constraints_n[i]) total_counts_cc++; 314608122e43SStefano Zampini max_constraints = PetscMax(max_constraints,pcbddc->adaptive_constraints_n[i]); 314708122e43SStefano Zampini } 31489162d606SStefano Zampini constraints_idxs_ptr = pcbddc->adaptive_constraints_idxs_ptr; 31499162d606SStefano Zampini constraints_data_ptr = pcbddc->adaptive_constraints_data_ptr; 31509162d606SStefano Zampini constraints_idxs = pcbddc->adaptive_constraints_idxs; 31519162d606SStefano Zampini constraints_data = pcbddc->adaptive_constraints_data; 315274d5cdf7SStefano Zampini /* constraints_n differs from pcbddc->adaptive_constraints_n */ 31539162d606SStefano Zampini ierr = PetscMalloc1(total_counts_cc,&constraints_n);CHKERRQ(ierr); 31549162d606SStefano Zampini total_counts_cc = 0; 31559162d606SStefano Zampini for (i=0;i<sub_schurs->n_subs+n_vertices;i++) { 31569162d606SStefano Zampini if (pcbddc->adaptive_constraints_n[i]) { 31579162d606SStefano Zampini constraints_n[total_counts_cc++] = pcbddc->adaptive_constraints_n[i]; 315808122e43SStefano Zampini } 315908122e43SStefano Zampini } 31609162d606SStefano Zampini #if 0 31619162d606SStefano Zampini printf("Found %d totals (%d)\n",total_counts_cc,total_counts); 31629162d606SStefano Zampini for (i=0;i<total_counts_cc;i++) { 31639162d606SStefano Zampini printf("const %d, start %d",i,constraints_idxs_ptr[i]); 31649162d606SStefano Zampini printf(" end %d:\n",constraints_idxs_ptr[i+1]); 31659162d606SStefano Zampini for (j=constraints_idxs_ptr[i];j<constraints_idxs_ptr[i+1];j++) { 31669162d606SStefano Zampini printf(" %d",constraints_idxs[j]); 31679162d606SStefano Zampini } 31689162d606SStefano Zampini printf("\n"); 31699162d606SStefano Zampini printf("number of cc: %d\n",constraints_n[i]); 31709162d606SStefano Zampini } 31711b968477SStefano Zampini for (i=0;i<n_vertices;i++) { 31728bec7fa6SStefano Zampini PetscPrintf(PETSC_COMM_SELF,"[%d] vertex %d, n %d\n",PetscGlobalRank,i,pcbddc->adaptive_constraints_n[i]); 31731b968477SStefano Zampini } 31741b968477SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 31758bec7fa6SStefano 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]); 31761b968477SStefano Zampini } 317708122e43SStefano Zampini #endif 317808122e43SStefano Zampini 31798bec7fa6SStefano Zampini max_size_of_constraint = 0; 31809162d606SStefano 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]); 31819162d606SStefano Zampini ierr = PetscMalloc1(constraints_idxs_ptr[total_counts_cc],&constraints_idxs_B);CHKERRQ(ierr); 318208122e43SStefano Zampini /* Change of basis */ 3183b3d85658SStefano Zampini ierr = PetscBTCreate(total_counts_cc,&change_basis);CHKERRQ(ierr); 318408122e43SStefano Zampini if (pcbddc->use_change_of_basis) { 318508122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 318608122e43SStefano Zampini if (PetscBTLookup(sub_schurs->is_edge,i) || pcbddc->use_change_on_faces) { 3187b3d85658SStefano Zampini ierr = PetscBTSet(change_basis,i+n_vertices);CHKERRQ(ierr); 318808122e43SStefano Zampini } 318908122e43SStefano Zampini } 319008122e43SStefano Zampini } 319108122e43SStefano Zampini } 3192984c4197SStefano Zampini pcbddc->local_primal_size = total_counts; 31934f1b2e48SStefano Zampini ierr = PetscMalloc1(pcbddc->local_primal_size+pcbddc->benign_n,&pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 319408122e43SStefano Zampini 31959162d606SStefano Zampini /* map constraints_idxs in boundary numbering */ 31969162d606SStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,constraints_idxs_ptr[total_counts_cc],constraints_idxs,&i,constraints_idxs_B);CHKERRQ(ierr); 31979162d606SStefano Zampini if (i != constraints_idxs_ptr[total_counts_cc]) { 31989162d606SStefano 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); 319908122e43SStefano Zampini } 3200674ae819SStefano Zampini 3201674ae819SStefano Zampini /* Create constraint matrix */ 3202674ae819SStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 320316f15bc4SStefano Zampini ierr = MatSetType(pcbddc->ConstraintMatrix,MATAIJ);CHKERRQ(ierr); 3204984c4197SStefano Zampini ierr = MatSetSizes(pcbddc->ConstraintMatrix,pcbddc->local_primal_size,pcis->n,pcbddc->local_primal_size,pcis->n);CHKERRQ(ierr); 3205984c4197SStefano Zampini 3206984c4197SStefano Zampini /* find primal_dofs: subdomain corners plus dofs selected as primal after change of basis */ 3207a717540cSStefano Zampini /* determine if a QR strategy is needed for change of basis */ 3208a717540cSStefano Zampini qr_needed = PETSC_FALSE; 320974d5cdf7SStefano Zampini ierr = PetscBTCreate(total_counts_cc,&qr_needed_idx);CHKERRQ(ierr); 3210984c4197SStefano Zampini total_primal_vertices=0; 3211b3d85658SStefano Zampini pcbddc->local_primal_size_cc = 0; 32129162d606SStefano Zampini for (i=0;i<total_counts_cc;i++) { 32139162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 3214984c4197SStefano Zampini if (size_of_constraint == 1) { 32159162d606SStefano Zampini pcbddc->primal_indices_local_idxs[total_primal_vertices++] = constraints_idxs[constraints_idxs_ptr[i]]; 3216b3d85658SStefano Zampini pcbddc->local_primal_size_cc += 1; 321764efe560SStefano Zampini } else if (PetscBTLookup(change_basis,i)) { 32189162d606SStefano Zampini for (k=0;k<constraints_n[i];k++) { 32199162d606SStefano Zampini pcbddc->primal_indices_local_idxs[total_primal_vertices++] = constraints_idxs[constraints_idxs_ptr[i]+k]; 3220a717540cSStefano Zampini } 3221b3d85658SStefano Zampini pcbddc->local_primal_size_cc += constraints_n[i]; 322274d5cdf7SStefano Zampini if (constraints_n[i] > 1 || pcbddc->use_qr_single || pcbddc->faster_deluxe) { 3223a717540cSStefano Zampini PetscBTSet(qr_needed_idx,i); 3224a717540cSStefano Zampini qr_needed = PETSC_TRUE; 3225a717540cSStefano Zampini } 3226fa434743SStefano Zampini } else { 3227b3d85658SStefano Zampini pcbddc->local_primal_size_cc += 1; 3228fa434743SStefano Zampini } 3229a717540cSStefano Zampini } 3230b371cd4fSStefano Zampini /* note that the local variable n_vertices used below stores the number of pointwise constraints */ 3231b371cd4fSStefano Zampini pcbddc->n_vertices = total_primal_vertices; 3232674ae819SStefano Zampini /* permute indices in order to have a sorted set of vertices */ 323370022509SStefano Zampini ierr = PetscSortInt(total_primal_vertices,pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 3234b3d85658SStefano Zampini 32354f1b2e48SStefano 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); 32360e6343abSStefano Zampini ierr = PetscMemcpy(pcbddc->local_primal_ref_node,pcbddc->primal_indices_local_idxs,total_primal_vertices*sizeof(PetscInt));CHKERRQ(ierr); 32370e6343abSStefano Zampini for (i=0;i<total_primal_vertices;i++) pcbddc->local_primal_ref_mult[i] = 1; 3238984c4197SStefano Zampini 3239984c4197SStefano Zampini /* nonzero structure of constraint matrix */ 324074d5cdf7SStefano Zampini /* and get reference dof for local constraints */ 3241785e854fSJed Brown ierr = PetscMalloc1(pcbddc->local_primal_size,&nnz);CHKERRQ(ierr); 3242984c4197SStefano Zampini for (i=0;i<total_primal_vertices;i++) nnz[i] = 1; 324374d5cdf7SStefano Zampini 3244984c4197SStefano Zampini j = total_primal_vertices; 324574d5cdf7SStefano Zampini total_counts = total_primal_vertices; 3246b3d85658SStefano Zampini cum = total_primal_vertices; 32479162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 32484641a718SStefano Zampini if (!PetscBTLookup(change_basis,i)) { 3249b3d85658SStefano Zampini pcbddc->local_primal_ref_node[cum] = constraints_idxs[constraints_idxs_ptr[i]]; 3250b3d85658SStefano Zampini pcbddc->local_primal_ref_mult[cum] = constraints_n[i]; 3251b3d85658SStefano Zampini cum++; 32529162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 325374d5cdf7SStefano Zampini for (k=0;k<constraints_n[i];k++) { 325474d5cdf7SStefano Zampini pcbddc->primal_indices_local_idxs[total_counts++] = constraints_idxs[constraints_idxs_ptr[i]+k]; 325574d5cdf7SStefano Zampini nnz[j+k] = size_of_constraint; 325674d5cdf7SStefano Zampini } 32579162d606SStefano Zampini j += constraints_n[i]; 3258674ae819SStefano Zampini } 3259674ae819SStefano Zampini } 3260674ae819SStefano Zampini ierr = MatSeqAIJSetPreallocation(pcbddc->ConstraintMatrix,0,nnz);CHKERRQ(ierr); 3261674ae819SStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 3262088faed8SStefano Zampini 3263674ae819SStefano Zampini /* set values in constraint matrix */ 3264984c4197SStefano Zampini for (i=0;i<total_primal_vertices;i++) { 32650e6343abSStefano Zampini ierr = MatSetValue(pcbddc->ConstraintMatrix,i,pcbddc->local_primal_ref_node[i],1.0,INSERT_VALUES);CHKERRQ(ierr); 3266674ae819SStefano Zampini } 3267984c4197SStefano Zampini total_counts = total_primal_vertices; 32689162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 32694641a718SStefano Zampini if (!PetscBTLookup(change_basis,i)) { 32709162d606SStefano Zampini PetscInt *cols; 32719162d606SStefano Zampini 32729162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 32739162d606SStefano Zampini cols = constraints_idxs+constraints_idxs_ptr[i]; 32749162d606SStefano Zampini for (k=0;k<constraints_n[i];k++) { 32759162d606SStefano Zampini PetscInt row = total_counts+k; 32769162d606SStefano Zampini PetscScalar *vals; 32779162d606SStefano Zampini 32789162d606SStefano Zampini vals = constraints_data+constraints_data_ptr[i]+k*size_of_constraint; 32799162d606SStefano Zampini ierr = MatSetValues(pcbddc->ConstraintMatrix,1,&row,size_of_constraint,cols,vals,INSERT_VALUES);CHKERRQ(ierr); 32809162d606SStefano Zampini } 32819162d606SStefano Zampini total_counts += constraints_n[i]; 3282674ae819SStefano Zampini } 3283674ae819SStefano Zampini } 3284674ae819SStefano Zampini /* assembling */ 3285674ae819SStefano Zampini ierr = MatAssemblyBegin(pcbddc->ConstraintMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3286674ae819SStefano Zampini ierr = MatAssemblyEnd(pcbddc->ConstraintMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3287088faed8SStefano Zampini 3288984c4197SStefano Zampini /* 328945a1bb75SStefano Zampini ierr = PetscViewerSetFormat(PETSC_VIEWER_STDOUT_SELF,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 3290984c4197SStefano Zampini ierr = MatView(pcbddc->ConstraintMatrix,(PetscViewer)0);CHKERRQ(ierr); 3291984c4197SStefano Zampini */ 3292674ae819SStefano Zampini /* Create matrix for change of basis. We don't need it in case pcbddc->use_change_of_basis is FALSE */ 3293674ae819SStefano Zampini if (pcbddc->use_change_of_basis) { 3294026de310SStefano Zampini /* dual and primal dofs on a single cc */ 3295984c4197SStefano Zampini PetscInt dual_dofs,primal_dofs; 3296984c4197SStefano Zampini /* working stuff for GEQRF */ 329781d9aea3SBarry Smith PetscScalar *qr_basis,*qr_tau = NULL,*qr_work,lqr_work_t; 3298984c4197SStefano Zampini PetscBLASInt lqr_work; 3299984c4197SStefano Zampini /* working stuff for UNGQR */ 3300984c4197SStefano Zampini PetscScalar *gqr_work,lgqr_work_t; 3301984c4197SStefano Zampini PetscBLASInt lgqr_work; 3302984c4197SStefano Zampini /* working stuff for TRTRS */ 3303984c4197SStefano Zampini PetscScalar *trs_rhs; 33043f08241aSStefano Zampini PetscBLASInt Blas_NRHS; 3305984c4197SStefano Zampini /* pointers for values insertion into change of basis matrix */ 3306984c4197SStefano Zampini PetscInt *start_rows,*start_cols; 3307984c4197SStefano Zampini PetscScalar *start_vals; 3308984c4197SStefano Zampini /* working stuff for values insertion */ 33094641a718SStefano Zampini PetscBT is_primal; 331064efe560SStefano Zampini PetscInt *aux_primal_numbering_B; 3311906d46d4SStefano Zampini /* matrix sizes */ 3312906d46d4SStefano Zampini PetscInt global_size,local_size; 3313906d46d4SStefano Zampini /* temporary change of basis */ 3314906d46d4SStefano Zampini Mat localChangeOfBasisMatrix; 3315cf5a6209SStefano Zampini /* extra space for debugging */ 3316cf5a6209SStefano Zampini PetscScalar *dbg_work; 3317984c4197SStefano Zampini 3318906d46d4SStefano Zampini /* local temporary change of basis acts on local interfaces -> dimension is n_B x n_B */ 3319906d46d4SStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&localChangeOfBasisMatrix);CHKERRQ(ierr); 332016f15bc4SStefano Zampini ierr = MatSetType(localChangeOfBasisMatrix,MATAIJ);CHKERRQ(ierr); 3321bbb9e6c6SStefano Zampini ierr = MatSetSizes(localChangeOfBasisMatrix,pcis->n,pcis->n,pcis->n,pcis->n);CHKERRQ(ierr); 3322906d46d4SStefano Zampini /* nonzeros for local mat */ 3323bbb9e6c6SStefano Zampini ierr = PetscMalloc1(pcis->n,&nnz);CHKERRQ(ierr); 3324bbb9e6c6SStefano Zampini for (i=0;i<pcis->n;i++) nnz[i]=1; 33259162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 3326a717540cSStefano Zampini if (PetscBTLookup(change_basis,i)) { 33279162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 3328a717540cSStefano Zampini if (PetscBTLookup(qr_needed_idx,i)) { 33299162d606SStefano Zampini for (j=0;j<size_of_constraint;j++) nnz[constraints_idxs[constraints_idxs_ptr[i]+j]] = size_of_constraint; 3330a717540cSStefano Zampini } else { 33319162d606SStefano Zampini nnz[constraints_idxs[constraints_idxs_ptr[i]]] = size_of_constraint; 33329162d606SStefano Zampini for (j=1;j<size_of_constraint;j++) nnz[constraints_idxs[constraints_idxs_ptr[i]+j]] = 2; 3333a717540cSStefano Zampini } 3334a717540cSStefano Zampini } 3335a717540cSStefano Zampini } 3336906d46d4SStefano Zampini ierr = MatSeqAIJSetPreallocation(localChangeOfBasisMatrix,0,nnz);CHKERRQ(ierr); 3337bbb9e6c6SStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 3338a717540cSStefano Zampini /* Set initial identity in the matrix */ 3339bbb9e6c6SStefano Zampini for (i=0;i<pcis->n;i++) { 3340906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,i,i,1.0,INSERT_VALUES);CHKERRQ(ierr); 3341a717540cSStefano Zampini } 3342a717540cSStefano Zampini 3343a717540cSStefano Zampini if (pcbddc->dbg_flag) { 3344a717540cSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"--------------------------------------------------------------\n");CHKERRQ(ierr); 3345a717540cSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Checking change of basis computation for subdomain %04d\n",PetscGlobalRank);CHKERRQ(ierr); 3346a717540cSStefano Zampini } 3347a717540cSStefano Zampini 3348a717540cSStefano Zampini 3349a717540cSStefano Zampini /* Now we loop on the constraints which need a change of basis */ 3350a717540cSStefano Zampini /* 3351a717540cSStefano Zampini Change of basis matrix is evaluated similarly to the FIRST APPROACH in 3352a717540cSStefano Zampini Klawonn and Widlund, Dual-primal FETI-DP methods for linear elasticity, (see Sect 6.2.1) 3353a717540cSStefano Zampini 3354a6b551f4SStefano Zampini Basic blocks of change of basis matrix T computed by 3355a717540cSStefano Zampini 3356a6b551f4SStefano 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) 3357a6b551f4SStefano Zampini 3358a6b551f4SStefano Zampini | 1 0 ... 0 s_1/S | 3359a6b551f4SStefano Zampini | 0 1 ... 0 s_2/S | 3360a717540cSStefano Zampini | ... | 3361a6b551f4SStefano Zampini | 0 ... 1 s_{n-1}/S | 3362a6b551f4SStefano Zampini | -s_1/s_n ... -s_{n-1}/s_n s_n/S | 3363a717540cSStefano Zampini 3364a6b551f4SStefano Zampini with S = \sum_{i=1}^n s_i^2 3365a6b551f4SStefano Zampini NOTE: in the above example, the primal dof is the last one of the edge in LOCAL ordering 3366a6b551f4SStefano Zampini in the current implementation, the primal dof is the first one of the edge in GLOBAL ordering 3367a6b551f4SStefano Zampini 3368a6b551f4SStefano Zampini - QR decomposition of constraints otherwise 3369a717540cSStefano Zampini */ 3370a717540cSStefano Zampini if (qr_needed) { 3371984c4197SStefano Zampini /* space to store Q */ 3372854ce69bSBarry Smith ierr = PetscMalloc1(max_size_of_constraint*max_size_of_constraint,&qr_basis);CHKERRQ(ierr); 3373984c4197SStefano Zampini /* first we issue queries for optimal work */ 33743f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_M);CHKERRQ(ierr); 33753f08241aSStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_N);CHKERRQ(ierr); 33763f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 3377984c4197SStefano Zampini lqr_work = -1; 33783f08241aSStefano Zampini PetscStackCallBLAS("LAPACKgeqrf",LAPACKgeqrf_(&Blas_M,&Blas_N,qr_basis,&Blas_LDA,qr_tau,&lqr_work_t,&lqr_work,&lierr)); 3379984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to GEQRF Lapack routine %d",(int)lierr); 3380984c4197SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lqr_work_t),&lqr_work);CHKERRQ(ierr); 3381785e854fSJed Brown ierr = PetscMalloc1((PetscInt)PetscRealPart(lqr_work_t),&qr_work);CHKERRQ(ierr); 3382984c4197SStefano Zampini lgqr_work = -1; 33833f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_M);CHKERRQ(ierr); 33843f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_N);CHKERRQ(ierr); 33853f08241aSStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_K);CHKERRQ(ierr); 33863f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 33873f08241aSStefano Zampini if (Blas_K>Blas_M) Blas_K=Blas_M; /* adjust just for computing optimal work */ 33883f08241aSStefano Zampini PetscStackCallBLAS("LAPACKungqr",LAPACKungqr_(&Blas_M,&Blas_N,&Blas_K,qr_basis,&Blas_LDA,qr_tau,&lgqr_work_t,&lgqr_work,&lierr)); 3389984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to UNGQR Lapack routine %d",(int)lierr); 3390984c4197SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lgqr_work_t),&lgqr_work);CHKERRQ(ierr); 3391785e854fSJed Brown ierr = PetscMalloc1((PetscInt)PetscRealPart(lgqr_work_t),&gqr_work);CHKERRQ(ierr); 3392984c4197SStefano Zampini /* array to store scaling factors for reflectors */ 3393785e854fSJed Brown ierr = PetscMalloc1(max_constraints,&qr_tau);CHKERRQ(ierr); 3394984c4197SStefano Zampini /* array to store rhs and solution of triangular solver */ 3395785e854fSJed Brown ierr = PetscMalloc1(max_constraints*max_constraints,&trs_rhs);CHKERRQ(ierr); 3396a717540cSStefano Zampini /* allocating workspace for check */ 3397a717540cSStefano Zampini if (pcbddc->dbg_flag) { 3398cf5a6209SStefano Zampini ierr = PetscMalloc1(max_size_of_constraint*(max_constraints+max_size_of_constraint),&dbg_work);CHKERRQ(ierr); 3399a717540cSStefano Zampini } 3400a717540cSStefano Zampini } 3401984c4197SStefano Zampini /* array to store whether a node is primal or not */ 34024641a718SStefano Zampini ierr = PetscBTCreate(pcis->n_B,&is_primal);CHKERRQ(ierr); 3403473ba861SJed Brown ierr = PetscMalloc1(total_primal_vertices,&aux_primal_numbering_B);CHKERRQ(ierr); 34040e6343abSStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,total_primal_vertices,pcbddc->local_primal_ref_node,&i,aux_primal_numbering_B);CHKERRQ(ierr); 340539e2fb2aSStefano Zampini if (i != total_primal_vertices) { 340639e2fb2aSStefano Zampini SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_SUP,"Error in boundary numbering for BDDC vertices! %d != %d\n",total_primal_vertices,i); 34074641a718SStefano Zampini } 340839e2fb2aSStefano Zampini for (i=0;i<total_primal_vertices;i++) { 340939e2fb2aSStefano Zampini ierr = PetscBTSet(is_primal,aux_primal_numbering_B[i]);CHKERRQ(ierr); 341039e2fb2aSStefano Zampini } 341139e2fb2aSStefano Zampini ierr = PetscFree(aux_primal_numbering_B);CHKERRQ(ierr); 3412984c4197SStefano Zampini 3413a717540cSStefano Zampini /* loop on constraints and see whether or not they need a change of basis and compute it */ 34149162d606SStefano Zampini for (total_counts=n_vertices;total_counts<total_counts_cc;total_counts++) { 34159162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[total_counts+1]-constraints_idxs_ptr[total_counts]; 34164641a718SStefano Zampini if (PetscBTLookup(change_basis,total_counts)) { 3417984c4197SStefano Zampini /* get constraint info */ 34189162d606SStefano Zampini primal_dofs = constraints_n[total_counts]; 3419984c4197SStefano Zampini dual_dofs = size_of_constraint-primal_dofs; 3420984c4197SStefano Zampini 3421984c4197SStefano Zampini if (pcbddc->dbg_flag) { 34229162d606SStefano 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); 3423674ae819SStefano Zampini } 3424984c4197SStefano Zampini 3425fa434743SStefano Zampini if (PetscBTLookup(qr_needed_idx,total_counts)) { /* QR */ 3426a717540cSStefano Zampini 3427a717540cSStefano Zampini /* copy quadrature constraints for change of basis check */ 3428a717540cSStefano Zampini if (pcbddc->dbg_flag) { 34299162d606SStefano Zampini ierr = PetscMemcpy(dbg_work,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 3430a717540cSStefano Zampini } 3431984c4197SStefano Zampini /* copy temporary constraints into larger work vector (in order to store all columns of Q) */ 34329162d606SStefano Zampini ierr = PetscMemcpy(qr_basis,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 3433984c4197SStefano Zampini 3434984c4197SStefano Zampini /* compute QR decomposition of constraints */ 34353f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 34363f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 34373f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 3438674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 34393f08241aSStefano Zampini PetscStackCallBLAS("LAPACKgeqrf",LAPACKgeqrf_(&Blas_M,&Blas_N,qr_basis,&Blas_LDA,qr_tau,qr_work,&lqr_work,&lierr)); 3440984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GEQRF Lapack routine %d",(int)lierr); 3441674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 3442984c4197SStefano Zampini 3443984c4197SStefano Zampini /* explictly compute R^-T */ 3444984c4197SStefano Zampini ierr = PetscMemzero(trs_rhs,primal_dofs*primal_dofs*sizeof(*trs_rhs));CHKERRQ(ierr); 3445984c4197SStefano Zampini for (j=0;j<primal_dofs;j++) trs_rhs[j*(primal_dofs+1)] = 1.0; 34463f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 34473f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_NRHS);CHKERRQ(ierr); 34483f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 34493f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDB);CHKERRQ(ierr); 3450984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 34513f08241aSStefano Zampini PetscStackCallBLAS("LAPACKtrtrs",LAPACKtrtrs_("U","T","N",&Blas_N,&Blas_NRHS,qr_basis,&Blas_LDA,trs_rhs,&Blas_LDB,&lierr)); 3452984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in TRTRS Lapack routine %d",(int)lierr); 3453984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 3454984c4197SStefano Zampini 3455a717540cSStefano Zampini /* explicitly compute all columns of Q (Q = [Q1 | Q2] ) overwriting QR factorization in qr_basis */ 34563f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 34573f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 34583f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_K);CHKERRQ(ierr); 34593f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 3460984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 34613f08241aSStefano Zampini PetscStackCallBLAS("LAPACKungqr",LAPACKungqr_(&Blas_M,&Blas_N,&Blas_K,qr_basis,&Blas_LDA,qr_tau,gqr_work,&lgqr_work,&lierr)); 3462984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in UNGQR Lapack routine %d",(int)lierr); 3463984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 3464984c4197SStefano Zampini 3465984c4197SStefano Zampini /* first primal_dofs columns of Q need to be re-scaled in order to be unitary w.r.t constraints 3466984c4197SStefano Zampini i.e. C_{pxn}*Q_{nxn} should be equal to [I_pxp | 0_pxd] (see check below) 3467984c4197SStefano Zampini where n=size_of_constraint, p=primal_dofs, d=dual_dofs (n=p+d), I and 0 identity and null matrix resp. */ 34683f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 34693f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 34703f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_K);CHKERRQ(ierr); 34713f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 34723f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDB);CHKERRQ(ierr); 34733f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDC);CHKERRQ(ierr); 3474984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 34759162d606SStefano 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)); 3476984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 34779162d606SStefano Zampini ierr = PetscMemcpy(qr_basis,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 3478984c4197SStefano Zampini 3479984c4197SStefano Zampini /* insert values in change of basis matrix respecting global ordering of new primal dofs */ 34809162d606SStefano Zampini start_rows = &constraints_idxs[constraints_idxs_ptr[total_counts]]; 3481984c4197SStefano Zampini /* insert cols for primal dofs */ 3482984c4197SStefano Zampini for (j=0;j<primal_dofs;j++) { 3483984c4197SStefano Zampini start_vals = &qr_basis[j*size_of_constraint]; 34849162d606SStefano Zampini start_cols = &constraints_idxs[constraints_idxs_ptr[total_counts]+j]; 3485906d46d4SStefano Zampini ierr = MatSetValues(localChangeOfBasisMatrix,size_of_constraint,start_rows,1,start_cols,start_vals,INSERT_VALUES);CHKERRQ(ierr); 3486984c4197SStefano Zampini } 3487984c4197SStefano Zampini /* insert cols for dual dofs */ 3488984c4197SStefano Zampini for (j=0,k=0;j<dual_dofs;k++) { 34899162d606SStefano Zampini if (!PetscBTLookup(is_primal,constraints_idxs_B[constraints_idxs_ptr[total_counts]+k])) { 3490984c4197SStefano Zampini start_vals = &qr_basis[(primal_dofs+j)*size_of_constraint]; 34919162d606SStefano Zampini start_cols = &constraints_idxs[constraints_idxs_ptr[total_counts]+k]; 3492906d46d4SStefano Zampini ierr = MatSetValues(localChangeOfBasisMatrix,size_of_constraint,start_rows,1,start_cols,start_vals,INSERT_VALUES);CHKERRQ(ierr); 3493984c4197SStefano Zampini j++; 3494674ae819SStefano Zampini } 3495674ae819SStefano Zampini } 3496984c4197SStefano Zampini 3497984c4197SStefano Zampini /* check change of basis */ 3498984c4197SStefano Zampini if (pcbddc->dbg_flag) { 3499984c4197SStefano Zampini PetscInt ii,jj; 3500984c4197SStefano Zampini PetscBool valid_qr=PETSC_TRUE; 3501c4303822SStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_M);CHKERRQ(ierr); 3502c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 3503c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_K);CHKERRQ(ierr); 3504c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 3505c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDB);CHKERRQ(ierr); 3506c4303822SStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDC);CHKERRQ(ierr); 3507984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 3508cf5a6209SStefano 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)); 3509984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 3510984c4197SStefano Zampini for (jj=0;jj<size_of_constraint;jj++) { 3511984c4197SStefano Zampini for (ii=0;ii<primal_dofs;ii++) { 3512cf5a6209SStefano Zampini if (ii != jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]) > 1.e-12) valid_qr = PETSC_FALSE; 3513cf5a6209SStefano 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; 3514674ae819SStefano Zampini } 3515674ae819SStefano Zampini } 3516984c4197SStefano Zampini if (!valid_qr) { 351722d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> wrong change of basis!\n");CHKERRQ(ierr); 3518984c4197SStefano Zampini for (jj=0;jj<size_of_constraint;jj++) { 3519984c4197SStefano Zampini for (ii=0;ii<primal_dofs;ii++) { 3520cf5a6209SStefano Zampini if (ii != jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]) > 1.e-12) { 3521cf5a6209SStefano 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])); 3522674ae819SStefano Zampini } 3523cf5a6209SStefano Zampini if (ii == jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]-1.0) > 1.e-12) { 3524cf5a6209SStefano 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])); 3525984c4197SStefano Zampini } 3526984c4197SStefano Zampini } 3527984c4197SStefano Zampini } 3528674ae819SStefano Zampini } else { 352922d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> right change of basis!\n");CHKERRQ(ierr); 3530674ae819SStefano Zampini } 3531674ae819SStefano Zampini } 3532a717540cSStefano Zampini } else { /* simple transformation block */ 3533a717540cSStefano Zampini PetscInt row,col; 3534a6b551f4SStefano Zampini PetscScalar val,norm; 3535a6b551f4SStefano Zampini 3536a6b551f4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 35379162d606SStefano 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)); 3538a717540cSStefano Zampini for (j=0;j<size_of_constraint;j++) { 35399162d606SStefano Zampini PetscInt row_B = constraints_idxs_B[constraints_idxs_ptr[total_counts]+j]; 35409162d606SStefano Zampini row = constraints_idxs[constraints_idxs_ptr[total_counts]+j]; 3541bbb9e6c6SStefano Zampini if (!PetscBTLookup(is_primal,row_B)) { 35429162d606SStefano Zampini col = constraints_idxs[constraints_idxs_ptr[total_counts]]; 3543906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,row,1.0,INSERT_VALUES);CHKERRQ(ierr); 35449162d606SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,col,constraints_data[constraints_data_ptr[total_counts]+j]/norm,INSERT_VALUES);CHKERRQ(ierr); 3545a717540cSStefano Zampini } else { 3546a717540cSStefano Zampini for (k=0;k<size_of_constraint;k++) { 35479162d606SStefano Zampini col = constraints_idxs[constraints_idxs_ptr[total_counts]+k]; 3548a717540cSStefano Zampini if (row != col) { 35499162d606SStefano Zampini val = -constraints_data[constraints_data_ptr[total_counts]+k]/constraints_data[constraints_data_ptr[total_counts]]; 3550a717540cSStefano Zampini } else { 35519162d606SStefano Zampini val = constraints_data[constraints_data_ptr[total_counts]]/norm; 3552a717540cSStefano Zampini } 3553906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,col,val,INSERT_VALUES);CHKERRQ(ierr); 3554a717540cSStefano Zampini } 3555a717540cSStefano Zampini } 3556a717540cSStefano Zampini } 355798a51de6SStefano Zampini if (pcbddc->dbg_flag) { 355822d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> using standard change of basis\n");CHKERRQ(ierr); 3559a717540cSStefano Zampini } 3560674ae819SStefano Zampini } 3561984c4197SStefano Zampini } else { 3562984c4197SStefano Zampini if (pcbddc->dbg_flag) { 35639162d606SStefano 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); 3564674ae819SStefano Zampini } 3565674ae819SStefano Zampini } 3566674ae819SStefano Zampini } 3567a717540cSStefano Zampini 3568a717540cSStefano Zampini /* free workspace */ 3569a717540cSStefano Zampini if (qr_needed) { 3570984c4197SStefano Zampini if (pcbddc->dbg_flag) { 3571cf5a6209SStefano Zampini ierr = PetscFree(dbg_work);CHKERRQ(ierr); 3572984c4197SStefano Zampini } 3573984c4197SStefano Zampini ierr = PetscFree(trs_rhs);CHKERRQ(ierr); 3574984c4197SStefano Zampini ierr = PetscFree(qr_tau);CHKERRQ(ierr); 3575984c4197SStefano Zampini ierr = PetscFree(qr_work);CHKERRQ(ierr); 3576984c4197SStefano Zampini ierr = PetscFree(gqr_work);CHKERRQ(ierr); 3577984c4197SStefano Zampini ierr = PetscFree(qr_basis);CHKERRQ(ierr); 3578674ae819SStefano Zampini } 3579a717540cSStefano Zampini ierr = PetscBTDestroy(&is_primal);CHKERRQ(ierr); 3580906d46d4SStefano Zampini ierr = MatAssemblyBegin(localChangeOfBasisMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3581906d46d4SStefano Zampini ierr = MatAssemblyEnd(localChangeOfBasisMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3582906d46d4SStefano Zampini 3583906d46d4SStefano Zampini /* assembling of global change of variable */ 3584bbb9e6c6SStefano Zampini { 3585bbb9e6c6SStefano Zampini Mat tmat; 358616f15bc4SStefano Zampini PetscInt bs; 358716f15bc4SStefano Zampini 3588906d46d4SStefano Zampini ierr = VecGetSize(pcis->vec1_global,&global_size);CHKERRQ(ierr); 3589906d46d4SStefano Zampini ierr = VecGetLocalSize(pcis->vec1_global,&local_size);CHKERRQ(ierr); 3590bbb9e6c6SStefano Zampini ierr = MatDuplicate(pc->pmat,MAT_DO_NOT_COPY_VALUES,&tmat);CHKERRQ(ierr); 3591bbb9e6c6SStefano Zampini ierr = MatISSetLocalMat(tmat,localChangeOfBasisMatrix);CHKERRQ(ierr); 3592bbb9e6c6SStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)pc),&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 3593bbb9e6c6SStefano Zampini ierr = MatSetType(pcbddc->ChangeOfBasisMatrix,MATAIJ);CHKERRQ(ierr); 359416f15bc4SStefano Zampini ierr = MatGetBlockSize(pc->pmat,&bs);CHKERRQ(ierr); 359516f15bc4SStefano Zampini ierr = MatSetBlockSize(pcbddc->ChangeOfBasisMatrix,bs);CHKERRQ(ierr); 3596906d46d4SStefano Zampini ierr = MatSetSizes(pcbddc->ChangeOfBasisMatrix,local_size,local_size,global_size,global_size);CHKERRQ(ierr); 3597bbb9e6c6SStefano Zampini ierr = MatISSetMPIXAIJPreallocation_Private(tmat,pcbddc->ChangeOfBasisMatrix,PETSC_TRUE);CHKERRQ(ierr); 3598bbb9e6c6SStefano Zampini ierr = MatISGetMPIXAIJ(tmat,MAT_REUSE_MATRIX,&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 3599bbb9e6c6SStefano Zampini ierr = MatDestroy(&tmat);CHKERRQ(ierr); 3600bbb9e6c6SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 3601bbb9e6c6SStefano Zampini ierr = VecSet(pcis->vec1_N,1.0);CHKERRQ(ierr); 3602e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 3603e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 3604bbb9e6c6SStefano Zampini ierr = VecReciprocal(pcis->vec1_global);CHKERRQ(ierr); 3605bbb9e6c6SStefano Zampini ierr = MatDiagonalScale(pcbddc->ChangeOfBasisMatrix,pcis->vec1_global,NULL);CHKERRQ(ierr); 3606906d46d4SStefano Zampini } 3607906d46d4SStefano Zampini /* check */ 3608906d46d4SStefano Zampini if (pcbddc->dbg_flag) { 3609906d46d4SStefano Zampini PetscReal error; 3610906d46d4SStefano Zampini Vec x,x_change; 3611906d46d4SStefano Zampini 3612906d46d4SStefano Zampini ierr = VecDuplicate(pcis->vec1_global,&x);CHKERRQ(ierr); 3613906d46d4SStefano Zampini ierr = VecDuplicate(pcis->vec1_global,&x_change);CHKERRQ(ierr); 3614906d46d4SStefano Zampini ierr = VecSetRandom(x,NULL);CHKERRQ(ierr); 3615906d46d4SStefano Zampini ierr = VecCopy(x,pcis->vec1_global);CHKERRQ(ierr); 3616e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,x,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 3617e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,x,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 3618bbb9e6c6SStefano Zampini ierr = MatMult(localChangeOfBasisMatrix,pcis->vec1_N,pcis->vec2_N);CHKERRQ(ierr); 3619e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec2_N,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 3620e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec2_N,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 3621906d46d4SStefano Zampini ierr = MatMult(pcbddc->ChangeOfBasisMatrix,pcis->vec1_global,x_change);CHKERRQ(ierr); 3622906d46d4SStefano Zampini ierr = VecAXPY(x,-1.0,x_change);CHKERRQ(ierr); 3623906d46d4SStefano Zampini ierr = VecNorm(x,NORM_INFINITY,&error);CHKERRQ(ierr); 3624906d46d4SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 3625bbb9e6c6SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Error global vs local change: %1.6e\n",error);CHKERRQ(ierr); 3626906d46d4SStefano Zampini ierr = VecDestroy(&x);CHKERRQ(ierr); 3627906d46d4SStefano Zampini ierr = VecDestroy(&x_change);CHKERRQ(ierr); 3628906d46d4SStefano Zampini } 3629b96c3477SStefano Zampini 3630b96c3477SStefano Zampini /* adapt sub_schurs computed (if any) */ 3631b96c3477SStefano Zampini if (pcbddc->use_deluxe_scaling) { 3632b96c3477SStefano Zampini PCBDDCSubSchurs sub_schurs=pcbddc->sub_schurs; 3633b96c3477SStefano Zampini if (sub_schurs->S_Ej_all) { 3634ac632422SStefano Zampini Mat S_new,tmat; 3635b087196eSStefano Zampini ISLocalToGlobalMapping NtoSall; 3636b087196eSStefano Zampini IS is_all_N,is_V,is_V_Sall; 3637b087196eSStefano Zampini const PetscScalar *array; 3638b087196eSStefano Zampini const PetscInt *idxs_V,*idxs_all; 3639b087196eSStefano Zampini PetscInt i,n_V; 3640bbb9e6c6SStefano Zampini 3641bbb9e6c6SStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(pcis->BtoNmap,sub_schurs->is_Ej_all,&is_all_N);CHKERRQ(ierr); 36426816873aSStefano Zampini ierr = MatGetSubMatrix(localChangeOfBasisMatrix,is_all_N,is_all_N,MAT_INITIAL_MATRIX,&tmat);CHKERRQ(ierr); 3643b087196eSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,pcbddc->n_vertices,pcbddc->local_primal_ref_node,PETSC_COPY_VALUES,&is_V);CHKERRQ(ierr); 3644b087196eSStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(is_all_N,&NtoSall);CHKERRQ(ierr); 3645b087196eSStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(NtoSall,IS_GTOLM_DROP,is_V,&is_V_Sall);CHKERRQ(ierr); 3646b087196eSStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&NtoSall);CHKERRQ(ierr); 3647bbb9e6c6SStefano Zampini ierr = ISDestroy(&is_all_N);CHKERRQ(ierr); 3648b087196eSStefano Zampini ierr = ISDestroy(&is_V);CHKERRQ(ierr); 3649ac632422SStefano Zampini ierr = MatPtAP(sub_schurs->S_Ej_all,tmat,MAT_INITIAL_MATRIX,1.0,&S_new);CHKERRQ(ierr); 3650b96c3477SStefano Zampini ierr = MatDestroy(&sub_schurs->S_Ej_all);CHKERRQ(ierr); 3651ac632422SStefano Zampini ierr = PetscObjectReference((PetscObject)S_new);CHKERRQ(ierr); 3652b087196eSStefano Zampini ierr = MatZeroRowsColumnsIS(S_new,is_V_Sall,1.,NULL,NULL);CHKERRQ(ierr); 3653b087196eSStefano Zampini ierr = ISGetLocalSize(is_V_Sall,&n_V);CHKERRQ(ierr); 3654b087196eSStefano Zampini ierr = ISGetIndices(is_V_Sall,&idxs_V);CHKERRQ(ierr); 3655b087196eSStefano Zampini ierr = ISGetIndices(sub_schurs->is_Ej_all,&idxs_all);CHKERRQ(ierr); 3656b087196eSStefano Zampini ierr = VecGetArrayRead(pcis->D,&array);CHKERRQ(ierr); 3657b087196eSStefano Zampini for (i=0;i<n_V;i++) { 3658b087196eSStefano Zampini PetscScalar val; 3659b087196eSStefano Zampini PetscInt idx; 3660b087196eSStefano Zampini 3661b087196eSStefano Zampini idx = idxs_V[i]; 3662b087196eSStefano Zampini val = array[idxs_all[idxs_V[i]]]; 3663b087196eSStefano Zampini ierr = MatSetValue(S_new,idx,idx,val,INSERT_VALUES);CHKERRQ(ierr); 3664b087196eSStefano Zampini } 3665b087196eSStefano Zampini ierr = MatAssemblyBegin(S_new,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3666b087196eSStefano Zampini ierr = MatAssemblyEnd(S_new,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3667ac632422SStefano Zampini sub_schurs->S_Ej_all = S_new; 3668ac632422SStefano Zampini ierr = MatDestroy(&S_new);CHKERRQ(ierr); 3669ac632422SStefano Zampini if (sub_schurs->sum_S_Ej_all) { 3670ac632422SStefano Zampini ierr = MatPtAP(sub_schurs->sum_S_Ej_all,tmat,MAT_INITIAL_MATRIX,1.0,&S_new);CHKERRQ(ierr); 3671b96c3477SStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_all);CHKERRQ(ierr); 3672ac632422SStefano Zampini ierr = PetscObjectReference((PetscObject)S_new);CHKERRQ(ierr); 3673b087196eSStefano Zampini ierr = MatZeroRowsColumnsIS(S_new,is_V_Sall,1.,NULL,NULL);CHKERRQ(ierr); 3674ac632422SStefano Zampini sub_schurs->sum_S_Ej_all = S_new; 3675ac632422SStefano Zampini ierr = MatDestroy(&S_new);CHKERRQ(ierr); 3676ac632422SStefano Zampini } 3677b087196eSStefano Zampini ierr = VecRestoreArrayRead(pcis->D,&array);CHKERRQ(ierr); 3678b087196eSStefano Zampini ierr = ISRestoreIndices(sub_schurs->is_Ej_all,&idxs_all);CHKERRQ(ierr); 3679b087196eSStefano Zampini ierr = ISRestoreIndices(is_V_Sall,&idxs_V);CHKERRQ(ierr); 3680b96c3477SStefano Zampini ierr = MatDestroy(&tmat);CHKERRQ(ierr); 3681b087196eSStefano Zampini ierr = ISDestroy(&is_V_Sall);CHKERRQ(ierr); 3682b96c3477SStefano Zampini } 3683b96c3477SStefano Zampini } 3684906d46d4SStefano Zampini ierr = MatDestroy(&localChangeOfBasisMatrix);CHKERRQ(ierr); 3685906d46d4SStefano Zampini } else if (pcbddc->user_ChangeOfBasisMatrix) { 3686b9b85e73SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->user_ChangeOfBasisMatrix);CHKERRQ(ierr); 3687b9b85e73SStefano Zampini pcbddc->ChangeOfBasisMatrix = pcbddc->user_ChangeOfBasisMatrix; 3688b9b85e73SStefano Zampini } 3689906d46d4SStefano Zampini 3690906d46d4SStefano Zampini /* set up change of basis context */ 3691906d46d4SStefano Zampini if (pcbddc->ChangeOfBasisMatrix) { 3692906d46d4SStefano Zampini PCBDDCChange_ctx change_ctx; 3693906d46d4SStefano Zampini 3694906d46d4SStefano Zampini if (!pcbddc->new_global_mat) { 3695906d46d4SStefano Zampini PetscInt global_size,local_size; 3696906d46d4SStefano Zampini 3697906d46d4SStefano Zampini ierr = VecGetSize(pcis->vec1_global,&global_size);CHKERRQ(ierr); 3698906d46d4SStefano Zampini ierr = VecGetLocalSize(pcis->vec1_global,&local_size);CHKERRQ(ierr); 3699906d46d4SStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)pc),&pcbddc->new_global_mat);CHKERRQ(ierr); 3700906d46d4SStefano Zampini ierr = MatSetSizes(pcbddc->new_global_mat,local_size,local_size,global_size,global_size);CHKERRQ(ierr); 3701906d46d4SStefano Zampini ierr = MatSetType(pcbddc->new_global_mat,MATSHELL);CHKERRQ(ierr); 3702906d46d4SStefano Zampini ierr = MatShellSetOperation(pcbddc->new_global_mat,MATOP_MULT,(void (*)(void))PCBDDCMatMult_Private);CHKERRQ(ierr); 3703906d46d4SStefano Zampini ierr = MatShellSetOperation(pcbddc->new_global_mat,MATOP_MULT_TRANSPOSE,(void (*)(void))PCBDDCMatMultTranspose_Private);CHKERRQ(ierr); 3704906d46d4SStefano Zampini ierr = PetscNew(&change_ctx);CHKERRQ(ierr); 3705906d46d4SStefano Zampini ierr = MatShellSetContext(pcbddc->new_global_mat,change_ctx);CHKERRQ(ierr); 3706906d46d4SStefano Zampini } else { 3707906d46d4SStefano Zampini ierr = MatShellGetContext(pcbddc->new_global_mat,&change_ctx);CHKERRQ(ierr); 3708906d46d4SStefano Zampini ierr = MatDestroy(&change_ctx->global_change);CHKERRQ(ierr); 3709906d46d4SStefano Zampini ierr = VecDestroyVecs(2,&change_ctx->work);CHKERRQ(ierr); 3710906d46d4SStefano Zampini } 3711906d46d4SStefano Zampini if (!pcbddc->user_ChangeOfBasisMatrix) { 3712906d46d4SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 3713906d46d4SStefano Zampini change_ctx->global_change = pcbddc->ChangeOfBasisMatrix; 3714906d46d4SStefano Zampini } else { 3715906d46d4SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->user_ChangeOfBasisMatrix);CHKERRQ(ierr); 3716906d46d4SStefano Zampini change_ctx->global_change = pcbddc->user_ChangeOfBasisMatrix; 3717906d46d4SStefano Zampini } 3718906d46d4SStefano Zampini ierr = VecDuplicateVecs(pcis->vec1_global,2,&change_ctx->work);CHKERRQ(ierr); 3719906d46d4SStefano Zampini ierr = MatSetUp(pcbddc->new_global_mat);CHKERRQ(ierr); 3720906d46d4SStefano Zampini ierr = MatAssemblyBegin(pcbddc->new_global_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3721906d46d4SStefano Zampini ierr = MatAssemblyEnd(pcbddc->new_global_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3722*a8661815SStefano Zampini } else { 3723*a8661815SStefano Zampini ierr = MatDestroy(&pcbddc->new_global_mat);CHKERRQ(ierr); 3724b9b85e73SStefano Zampini } 3725a717540cSStefano Zampini 37264f1b2e48SStefano Zampini /* add pressure dofs to set of primal nodes for numbering purposes */ 37274f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 37284f1b2e48SStefano Zampini pcbddc->local_primal_ref_node[pcbddc->local_primal_size_cc] = pcbddc->benign_p0_lidx[i]; 37294f1b2e48SStefano Zampini pcbddc->primal_indices_local_idxs[pcbddc->local_primal_size] = pcbddc->benign_p0_lidx[i]; 3730019a44ceSStefano Zampini pcbddc->local_primal_ref_mult[pcbddc->local_primal_size_cc] = 1; 3731019a44ceSStefano Zampini pcbddc->local_primal_size_cc++; 3732019a44ceSStefano Zampini pcbddc->local_primal_size++; 3733019a44ceSStefano Zampini } 3734019a44ceSStefano Zampini 3735019a44ceSStefano Zampini /* check if a new primal space has been introduced (also take into account benign trick) */ 3736727cdba6SStefano Zampini pcbddc->new_primal_space_local = PETSC_TRUE; 3737727cdba6SStefano Zampini if (olocal_primal_size == pcbddc->local_primal_size) { 37389f47a83aSStefano 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); 3739c1c8e736SStefano Zampini pcbddc->new_primal_space_local = (PetscBool)(!pcbddc->new_primal_space_local); 37400e6343abSStefano Zampini if (!pcbddc->new_primal_space_local) { 37419f47a83aSStefano 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); 3742727cdba6SStefano Zampini pcbddc->new_primal_space_local = (PetscBool)(!pcbddc->new_primal_space_local); 3743727cdba6SStefano Zampini } 37440e6343abSStefano Zampini } 37450e6343abSStefano Zampini ierr = PetscFree2(olocal_primal_ref_node,olocal_primal_ref_mult);CHKERRQ(ierr); 3746727cdba6SStefano Zampini /* new_primal_space will be used for numbering of coarse dofs, so it should be the same across all subdomains */ 3747727cdba6SStefano Zampini ierr = MPI_Allreduce(&pcbddc->new_primal_space_local,&pcbddc->new_primal_space,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 3748727cdba6SStefano Zampini 3749a717540cSStefano Zampini /* flush dbg viewer */ 3750b8ffe317SStefano Zampini if (pcbddc->dbg_flag) { 3751b8ffe317SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 3752b8ffe317SStefano Zampini } 3753a717540cSStefano Zampini 3754e310c8b4SStefano Zampini /* free workspace */ 3755a717540cSStefano Zampini ierr = PetscBTDestroy(&qr_needed_idx);CHKERRQ(ierr); 37564641a718SStefano Zampini ierr = PetscBTDestroy(&change_basis);CHKERRQ(ierr); 375708122e43SStefano Zampini if (!pcbddc->adaptive_selection) { 37589162d606SStefano Zampini ierr = PetscFree3(constraints_idxs_ptr,constraints_data_ptr,constraints_n);CHKERRQ(ierr); 37599162d606SStefano Zampini ierr = PetscFree3(constraints_data,constraints_idxs,constraints_idxs_B);CHKERRQ(ierr); 376008122e43SStefano Zampini } else { 37619162d606SStefano Zampini ierr = PetscFree5(pcbddc->adaptive_constraints_n, 37629162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr, 37639162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr, 376408122e43SStefano Zampini pcbddc->adaptive_constraints_idxs, 376508122e43SStefano Zampini pcbddc->adaptive_constraints_data);CHKERRQ(ierr); 37669162d606SStefano Zampini ierr = PetscFree(constraints_n);CHKERRQ(ierr); 37679162d606SStefano Zampini ierr = PetscFree(constraints_idxs_B);CHKERRQ(ierr); 376808122e43SStefano Zampini } 3769674ae819SStefano Zampini PetscFunctionReturn(0); 3770674ae819SStefano Zampini } 3771674ae819SStefano Zampini 3772674ae819SStefano Zampini #undef __FUNCT__ 3773674ae819SStefano Zampini #define __FUNCT__ "PCBDDCAnalyzeInterface" 3774674ae819SStefano Zampini PetscErrorCode PCBDDCAnalyzeInterface(PC pc) 3775674ae819SStefano Zampini { 3776674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 3777674ae819SStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 3778674ae819SStefano Zampini Mat_IS *matis = (Mat_IS*)pc->pmat->data; 37797fb0e2dbSStefano Zampini PetscInt ierr,i,vertex_size,N; 3780674ae819SStefano Zampini PetscViewer viewer=pcbddc->dbg_viewer; 3781674ae819SStefano Zampini 3782674ae819SStefano Zampini PetscFunctionBegin; 37838e61c736SStefano Zampini /* Reset previously computed graph */ 37848e61c736SStefano Zampini ierr = PCBDDCGraphReset(pcbddc->mat_graph);CHKERRQ(ierr); 3785674ae819SStefano Zampini /* Init local Graph struct */ 37867fb0e2dbSStefano Zampini ierr = MatGetSize(pc->pmat,&N,NULL);CHKERRQ(ierr); 37873bbff08aSStefano Zampini ierr = PCBDDCGraphInit(pcbddc->mat_graph,pcis->mapping,N);CHKERRQ(ierr); 3788674ae819SStefano Zampini 3789575ad6abSStefano Zampini /* Check validity of the csr graph passed in by the user */ 37905099eff2SStefano Zampini if (pcbddc->mat_graph->nvtxs_csr && pcbddc->mat_graph->nvtxs_csr != pcbddc->mat_graph->nvtxs) { 37915099eff2SStefano 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); 3792575ad6abSStefano Zampini } 37939577ea80SStefano Zampini 3794674ae819SStefano Zampini /* Set default CSR adjacency of local dofs if not provided by the user with PCBDDCSetLocalAdjacencyGraph */ 37954d379d7bSStefano Zampini if (!pcbddc->mat_graph->xadj || !pcbddc->mat_graph->adjncy) { 37964d379d7bSStefano Zampini PetscInt *xadj,*adjncy; 37974d379d7bSStefano Zampini PetscInt nvtxs; 3798e496cd5dSStefano Zampini PetscBool flg_row=PETSC_FALSE; 3799674ae819SStefano Zampini 38004d379d7bSStefano Zampini if (pcbddc->use_local_adj) { 38012fffb893SStefano Zampini 38022fffb893SStefano Zampini ierr = MatGetRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 38032fffb893SStefano Zampini if (flg_row) { 38044d379d7bSStefano Zampini ierr = PCBDDCSetLocalAdjacencyGraph(pc,nvtxs,xadj,adjncy,PETSC_COPY_VALUES);CHKERRQ(ierr); 3805b96c3477SStefano Zampini pcbddc->computed_rowadj = PETSC_TRUE; 38062fffb893SStefano Zampini } 38072fffb893SStefano Zampini ierr = MatRestoreRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 38089b28b941SStefano 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 */ 38094d379d7bSStefano Zampini IS is_dummy; 38104d379d7bSStefano Zampini ISLocalToGlobalMapping l2gmap_dummy; 38114d379d7bSStefano Zampini PetscInt j,sum; 38124d379d7bSStefano Zampini PetscInt *cxadj,*cadjncy; 38134d379d7bSStefano Zampini const PetscInt *idxs; 38144d379d7bSStefano Zampini PCBDDCGraph graph; 38154d379d7bSStefano Zampini PetscBT is_on_boundary; 38164d379d7bSStefano Zampini 38174d379d7bSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcis->n,0,1,&is_dummy);CHKERRQ(ierr); 38184d379d7bSStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(is_dummy,&l2gmap_dummy);CHKERRQ(ierr); 38194d379d7bSStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 38204d379d7bSStefano Zampini ierr = PCBDDCGraphCreate(&graph);CHKERRQ(ierr); 38217fb0e2dbSStefano Zampini ierr = PCBDDCGraphInit(graph,l2gmap_dummy,pcis->n);CHKERRQ(ierr); 38224d379d7bSStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&l2gmap_dummy);CHKERRQ(ierr); 3823e496cd5dSStefano Zampini ierr = MatGetRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 3824e496cd5dSStefano Zampini if (flg_row) { 38254d379d7bSStefano Zampini graph->xadj = xadj; 38264d379d7bSStefano Zampini graph->adjncy = adjncy; 3827e496cd5dSStefano Zampini } 38284d379d7bSStefano Zampini ierr = PCBDDCGraphSetUp(graph,1,NULL,NULL,0,NULL,NULL);CHKERRQ(ierr); 38294d379d7bSStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(graph);CHKERRQ(ierr); 3830e496cd5dSStefano Zampini ierr = MatRestoreRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 38314d379d7bSStefano Zampini 38324d379d7bSStefano Zampini if (pcbddc->dbg_flag) { 38339b28b941SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"[%d] Found %d subdomains (local size %d)\n",PetscGlobalRank,graph->ncc,pcis->n);CHKERRQ(ierr); 38344d379d7bSStefano Zampini for (i=0;i<graph->ncc;i++) { 38354d379d7bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"[%d] %d cc size %d\n",PetscGlobalRank,i,graph->cptr[i+1]-graph->cptr[i]);CHKERRQ(ierr); 38364d379d7bSStefano Zampini } 38374d379d7bSStefano Zampini } 38384d379d7bSStefano Zampini 3839e496cd5dSStefano Zampini ierr = PetscBTCreate(pcis->n,&is_on_boundary);CHKERRQ(ierr); 38404d379d7bSStefano Zampini ierr = ISGetIndices(pcis->is_B_local,&idxs);CHKERRQ(ierr); 38414d379d7bSStefano Zampini for (i=0;i<pcis->n_B;i++) { 38424d379d7bSStefano Zampini ierr = PetscBTSet(is_on_boundary,idxs[i]);CHKERRQ(ierr); 38434d379d7bSStefano Zampini } 38444d379d7bSStefano Zampini ierr = ISRestoreIndices(pcis->is_B_local,&idxs);CHKERRQ(ierr); 38454d379d7bSStefano Zampini 3846e496cd5dSStefano Zampini ierr = PetscCalloc1(pcis->n+1,&cxadj);CHKERRQ(ierr); 38474d379d7bSStefano Zampini sum = 0; 38484d379d7bSStefano Zampini for (i=0;i<graph->ncc;i++) { 38494d379d7bSStefano Zampini PetscInt sizecc = 0; 38504d379d7bSStefano Zampini for (j=graph->cptr[i];j<graph->cptr[i+1];j++) { 38514d379d7bSStefano Zampini if (PetscBTLookup(is_on_boundary,graph->queue[j])) { 38524d379d7bSStefano Zampini sizecc++; 38534d379d7bSStefano Zampini } 38544d379d7bSStefano Zampini } 38554d379d7bSStefano Zampini for (j=graph->cptr[i];j<graph->cptr[i+1];j++) { 38564d379d7bSStefano Zampini if (PetscBTLookup(is_on_boundary,graph->queue[j])) { 38574d379d7bSStefano Zampini cxadj[graph->queue[j]] = sizecc; 38584d379d7bSStefano Zampini } 38594d379d7bSStefano Zampini } 38604d379d7bSStefano Zampini sum += sizecc*sizecc; 38614d379d7bSStefano Zampini } 38624d379d7bSStefano Zampini ierr = PetscMalloc1(sum,&cadjncy);CHKERRQ(ierr); 38634d379d7bSStefano Zampini sum = 0; 3864e496cd5dSStefano Zampini for (i=0;i<pcis->n;i++) { 38654d379d7bSStefano Zampini PetscInt temp = cxadj[i]; 38664d379d7bSStefano Zampini cxadj[i] = sum; 38674d379d7bSStefano Zampini sum += temp; 38684d379d7bSStefano Zampini } 3869e496cd5dSStefano Zampini cxadj[pcis->n] = sum; 38704d379d7bSStefano Zampini for (i=0;i<graph->ncc;i++) { 38714d379d7bSStefano Zampini for (j=graph->cptr[i];j<graph->cptr[i+1];j++) { 38724d379d7bSStefano Zampini if (PetscBTLookup(is_on_boundary,graph->queue[j])) { 38734d379d7bSStefano Zampini PetscInt k,sizecc = 0; 38744d379d7bSStefano Zampini for (k=graph->cptr[i];k<graph->cptr[i+1];k++) { 38754d379d7bSStefano Zampini if (PetscBTLookup(is_on_boundary,graph->queue[k])) { 38764d379d7bSStefano Zampini cadjncy[cxadj[graph->queue[j]]+sizecc]=graph->queue[k]; 38774d379d7bSStefano Zampini sizecc++; 38784d379d7bSStefano Zampini } 38794d379d7bSStefano Zampini } 38804d379d7bSStefano Zampini } 38814d379d7bSStefano Zampini } 38824d379d7bSStefano Zampini } 38839b28b941SStefano Zampini if (sum) { 3884e496cd5dSStefano Zampini ierr = PCBDDCSetLocalAdjacencyGraph(pc,pcis->n,cxadj,cadjncy,PETSC_OWN_POINTER);CHKERRQ(ierr); 38854d379d7bSStefano Zampini } else { 38864d379d7bSStefano Zampini ierr = PetscFree(cxadj);CHKERRQ(ierr); 38874d379d7bSStefano Zampini ierr = PetscFree(cadjncy);CHKERRQ(ierr); 38884d379d7bSStefano Zampini } 38894d379d7bSStefano Zampini graph->xadj = 0; 38904d379d7bSStefano Zampini graph->adjncy = 0; 38914d379d7bSStefano Zampini ierr = PCBDDCGraphDestroy(&graph);CHKERRQ(ierr); 38924d379d7bSStefano Zampini ierr = PetscBTDestroy(&is_on_boundary);CHKERRQ(ierr); 38934d379d7bSStefano Zampini } 3894674ae819SStefano Zampini } 38959b28b941SStefano Zampini if (pcbddc->dbg_flag) { 38969b28b941SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 3897674ae819SStefano Zampini } 3898674ae819SStefano Zampini 389963602bcaSStefano Zampini /* Set default dofs' splitting if no information has been provided by the user with PCBDDCSetDofsSplitting or PCBDDCSetDofsSplittingLocal */ 3900674ae819SStefano Zampini vertex_size = 1; 390163602bcaSStefano Zampini if (pcbddc->user_provided_isfordofs) { 390263602bcaSStefano Zampini if (pcbddc->n_ISForDofs) { /* need to convert from global to local and remove references to global dofs splitting */ 390395ecbf38SStefano Zampini ierr = PetscMalloc1(pcbddc->n_ISForDofs,&pcbddc->ISForDofsLocal);CHKERRQ(ierr); 390463602bcaSStefano Zampini for (i=0;i<pcbddc->n_ISForDofs;i++) { 3905e176bc59SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->rctx,pcis->vec1_global,pcis->vec1_N,pcbddc->ISForDofs[i],&pcbddc->ISForDofsLocal[i]);CHKERRQ(ierr); 390663602bcaSStefano Zampini ierr = ISDestroy(&pcbddc->ISForDofs[i]);CHKERRQ(ierr); 3907674ae819SStefano Zampini } 390863602bcaSStefano Zampini pcbddc->n_ISForDofsLocal = pcbddc->n_ISForDofs; 390963602bcaSStefano Zampini pcbddc->n_ISForDofs = 0; 391063602bcaSStefano Zampini ierr = PetscFree(pcbddc->ISForDofs);CHKERRQ(ierr); 3911674ae819SStefano Zampini } 391263602bcaSStefano Zampini /* mat block size as vertex size (used for elasticity with rigid body modes as nearnullspace) */ 3913674ae819SStefano Zampini ierr = MatGetBlockSize(matis->A,&vertex_size);CHKERRQ(ierr); 391463602bcaSStefano Zampini } else { 391563602bcaSStefano Zampini if (!pcbddc->n_ISForDofsLocal) { /* field split not present, create it in local ordering */ 391663602bcaSStefano Zampini ierr = MatGetBlockSize(pc->pmat,&pcbddc->n_ISForDofsLocal);CHKERRQ(ierr); 3917854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->n_ISForDofsLocal,&pcbddc->ISForDofsLocal);CHKERRQ(ierr); 391863602bcaSStefano Zampini for (i=0;i<pcbddc->n_ISForDofsLocal;i++) { 391963602bcaSStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)pc),pcis->n/pcbddc->n_ISForDofsLocal,i,pcbddc->n_ISForDofsLocal,&pcbddc->ISForDofsLocal[i]);CHKERRQ(ierr); 392063602bcaSStefano Zampini } 392163602bcaSStefano Zampini } 3922674ae819SStefano Zampini } 3923674ae819SStefano Zampini 3924674ae819SStefano Zampini /* Setup of Graph */ 3925785d1243SStefano Zampini if (!pcbddc->DirichletBoundariesLocal && pcbddc->DirichletBoundaries) { /* need to convert from global to local */ 3926e176bc59SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->rctx,pcis->vec1_global,pcis->vec1_N,pcbddc->DirichletBoundaries,&pcbddc->DirichletBoundariesLocal);CHKERRQ(ierr); 3927785d1243SStefano Zampini } 3928785d1243SStefano Zampini if (!pcbddc->NeumannBoundariesLocal && pcbddc->NeumannBoundaries) { /* need to convert from global to local */ 3929e176bc59SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->rctx,pcis->vec1_global,pcis->vec1_N,pcbddc->NeumannBoundaries,&pcbddc->NeumannBoundariesLocal);CHKERRQ(ierr); 3930785d1243SStefano Zampini } 393130368db7SStefano Zampini if (!pcbddc->user_primal_vertices_local && pcbddc->user_primal_vertices) { /* need to convert from global to local */ 393230368db7SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->rctx,pcis->vec1_global,pcis->vec1_N,pcbddc->user_primal_vertices,&pcbddc->user_primal_vertices_local);CHKERRQ(ierr); 393330368db7SStefano Zampini } 393430368db7SStefano Zampini ierr = PCBDDCGraphSetUp(pcbddc->mat_graph,vertex_size,pcbddc->NeumannBoundariesLocal,pcbddc->DirichletBoundariesLocal,pcbddc->n_ISForDofsLocal,pcbddc->ISForDofsLocal,pcbddc->user_primal_vertices_local);CHKERRQ(ierr); 3935674ae819SStefano Zampini 39364f1b2e48SStefano Zampini /* attach info on disconnected subdomains if present */ 39374f1b2e48SStefano Zampini if (pcbddc->n_local_subs) { 39384f1b2e48SStefano Zampini PetscInt *local_subs; 39394f1b2e48SStefano Zampini 39404f1b2e48SStefano Zampini ierr = PetscMalloc1(N,&local_subs);CHKERRQ(ierr); 39414f1b2e48SStefano Zampini for (i=0;i<pcbddc->n_local_subs;i++) { 39424f1b2e48SStefano Zampini const PetscInt *idxs; 39434f1b2e48SStefano Zampini PetscInt nl,j; 39444f1b2e48SStefano Zampini 39454f1b2e48SStefano Zampini ierr = ISGetLocalSize(pcbddc->local_subs[i],&nl);CHKERRQ(ierr); 39464f1b2e48SStefano Zampini ierr = ISGetIndices(pcbddc->local_subs[i],&idxs);CHKERRQ(ierr); 39474f1b2e48SStefano Zampini for (j=0;j<nl;j++) { 39484f1b2e48SStefano Zampini local_subs[idxs[j]] = i; 39494f1b2e48SStefano Zampini } 39504f1b2e48SStefano Zampini ierr = ISRestoreIndices(pcbddc->local_subs[i],&idxs);CHKERRQ(ierr); 39514f1b2e48SStefano Zampini } 39524f1b2e48SStefano Zampini pcbddc->mat_graph->n_local_subs = pcbddc->n_local_subs; 39534f1b2e48SStefano Zampini pcbddc->mat_graph->local_subs = local_subs; 39544f1b2e48SStefano Zampini } 39554f1b2e48SStefano Zampini 3956674ae819SStefano Zampini /* Graph's connected components analysis */ 3957674ae819SStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(pcbddc->mat_graph);CHKERRQ(ierr); 3958674ae819SStefano Zampini 3959674ae819SStefano Zampini /* print some info to stdout */ 3960674ae819SStefano Zampini if (pcbddc->dbg_flag) { 3961302440fdSBarry Smith ierr = PCBDDCGraphASCIIView(pcbddc->mat_graph,pcbddc->dbg_flag,viewer);CHKERRQ(ierr); 3962674ae819SStefano Zampini } 3963fb180af4SStefano Zampini 3964fb180af4SStefano Zampini /* mark topography has done */ 3965fb180af4SStefano Zampini pcbddc->recompute_topography = PETSC_FALSE; 3966674ae819SStefano Zampini PetscFunctionReturn(0); 3967674ae819SStefano Zampini } 3968674ae819SStefano Zampini 3969dc456d91SStefano Zampini /* given an index sets possibly with holes, renumbers the indexes removing the holes */ 3970674ae819SStefano Zampini #undef __FUNCT__ 3971674ae819SStefano Zampini #define __FUNCT__ "PCBDDCSubsetNumbering" 3972dc456d91SStefano Zampini PetscErrorCode PCBDDCSubsetNumbering(IS subset, IS subset_mult, PetscInt *N_n, IS *subset_n) 3973674ae819SStefano Zampini { 3974dc456d91SStefano Zampini PetscSF sf; 3975dc456d91SStefano Zampini PetscLayout map; 3976dc456d91SStefano Zampini const PetscInt *idxs; 3977dc456d91SStefano Zampini PetscInt *leaf_data,*root_data,*gidxs; 3978dc456d91SStefano Zampini PetscInt N,n,i,lbounds[2],gbounds[2],Nl; 3979dc456d91SStefano Zampini PetscInt n_n,nlocals,start,first_index; 3980dc456d91SStefano Zampini PetscMPIInt commsize; 3981674ae819SStefano Zampini PetscBool first_found; 3982674ae819SStefano Zampini PetscErrorCode ierr; 3983674ae819SStefano Zampini 3984674ae819SStefano Zampini PetscFunctionBegin; 3985dc456d91SStefano Zampini ierr = ISGetLocalSize(subset,&n);CHKERRQ(ierr); 3986dc456d91SStefano Zampini if (subset_mult) { 3987dc456d91SStefano Zampini PetscCheckSameComm(subset,1,subset_mult,2); 3988dc456d91SStefano Zampini ierr = ISGetLocalSize(subset,&i);CHKERRQ(ierr); 3989dc456d91SStefano Zampini if (i != n) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Local subset and multiplicity sizes don't match! %d != %d",n,i); 3990674ae819SStefano Zampini } 3991dc456d91SStefano Zampini /* create workspace layout for computing global indices of subset */ 3992dc456d91SStefano Zampini ierr = ISGetIndices(subset,&idxs);CHKERRQ(ierr); 3993dc456d91SStefano Zampini lbounds[0] = lbounds[1] = 0; 3994dc456d91SStefano Zampini for (i=0;i<n;i++) { 3995dc456d91SStefano Zampini if (idxs[i] < lbounds[0]) lbounds[0] = idxs[i]; 3996dc456d91SStefano Zampini else if (idxs[i] > lbounds[1]) lbounds[1] = idxs[i]; 3997674ae819SStefano Zampini } 3998dc456d91SStefano Zampini lbounds[0] = -lbounds[0]; 3999dc456d91SStefano Zampini ierr = MPI_Allreduce(lbounds,gbounds,2,MPIU_INT,MPI_MAX,PetscObjectComm((PetscObject)subset));CHKERRQ(ierr); 4000dc456d91SStefano Zampini gbounds[0] = -gbounds[0]; 4001dc456d91SStefano Zampini N = gbounds[1] - gbounds[0] + 1; 4002dc456d91SStefano Zampini ierr = PetscLayoutCreate(PetscObjectComm((PetscObject)subset),&map);CHKERRQ(ierr); 4003dc456d91SStefano Zampini ierr = PetscLayoutSetBlockSize(map,1);CHKERRQ(ierr); 4004dc456d91SStefano Zampini ierr = PetscLayoutSetSize(map,N);CHKERRQ(ierr); 4005dc456d91SStefano Zampini ierr = PetscLayoutSetUp(map);CHKERRQ(ierr); 4006dc456d91SStefano Zampini ierr = PetscLayoutGetLocalSize(map,&Nl);CHKERRQ(ierr); 4007dc456d91SStefano Zampini 4008dc456d91SStefano Zampini /* create sf : leaf_data == multiplicity of indexes, root data == global index in layout */ 4009dc456d91SStefano Zampini ierr = PetscMalloc2(n,&leaf_data,Nl,&root_data);CHKERRQ(ierr); 4010dc456d91SStefano Zampini if (subset_mult) { 4011dc456d91SStefano Zampini const PetscInt* idxs_mult; 4012dc456d91SStefano Zampini 4013dc456d91SStefano Zampini ierr = ISGetIndices(subset_mult,&idxs_mult);CHKERRQ(ierr); 4014dc456d91SStefano Zampini ierr = PetscMemcpy(leaf_data,idxs_mult,n*sizeof(PetscInt));CHKERRQ(ierr); 4015dc456d91SStefano Zampini ierr = ISRestoreIndices(subset_mult,&idxs_mult);CHKERRQ(ierr); 4016674ae819SStefano Zampini } else { 4017dc456d91SStefano Zampini for (i=0;i<n;i++) leaf_data[i] = 1; 4018674ae819SStefano Zampini } 4019dc456d91SStefano Zampini /* local size of new subset */ 4020dc456d91SStefano Zampini n_n = 0; 4021dc456d91SStefano Zampini for (i=0;i<n;i++) n_n += leaf_data[i]; 4022dc456d91SStefano Zampini 4023dc456d91SStefano Zampini /* global indexes in layout */ 4024dc456d91SStefano Zampini ierr = PetscMalloc1(n_n,&gidxs);CHKERRQ(ierr); /* allocating possibly extra space in gidxs which will be used later */ 4025dc456d91SStefano Zampini for (i=0;i<n;i++) gidxs[i] = idxs[i] - gbounds[0]; 4026dc456d91SStefano Zampini ierr = ISRestoreIndices(subset,&idxs);CHKERRQ(ierr); 4027dc456d91SStefano Zampini ierr = PetscSFCreate(PetscObjectComm((PetscObject)subset),&sf);CHKERRQ(ierr); 4028dc456d91SStefano Zampini ierr = PetscSFSetGraphLayout(sf,map,n,NULL,PETSC_COPY_VALUES,gidxs);CHKERRQ(ierr); 4029dc456d91SStefano Zampini ierr = PetscLayoutDestroy(&map);CHKERRQ(ierr); 4030dc456d91SStefano Zampini 4031dc456d91SStefano Zampini /* reduce from leaves to roots */ 4032dc456d91SStefano Zampini ierr = PetscMemzero(root_data,Nl*sizeof(PetscInt));CHKERRQ(ierr); 403364a8e5bcSStefano Zampini ierr = PetscSFReduceBegin(sf,MPIU_INT,leaf_data,root_data,MPI_MAX);CHKERRQ(ierr); 403464a8e5bcSStefano Zampini ierr = PetscSFReduceEnd(sf,MPIU_INT,leaf_data,root_data,MPI_MAX);CHKERRQ(ierr); 4035dc456d91SStefano Zampini 4036dc456d91SStefano Zampini /* count indexes in local part of layout */ 4037674ae819SStefano Zampini nlocals = 0; 4038674ae819SStefano Zampini first_index = -1; 4039674ae819SStefano Zampini first_found = PETSC_FALSE; 4040dc456d91SStefano Zampini for (i=0;i<Nl;i++) { 4041dc456d91SStefano Zampini if (!first_found && root_data[i]) { 4042674ae819SStefano Zampini first_found = PETSC_TRUE; 4043674ae819SStefano Zampini first_index = i; 4044674ae819SStefano Zampini } 4045dc456d91SStefano Zampini nlocals += root_data[i]; 4046674ae819SStefano Zampini } 4047dc456d91SStefano Zampini 4048dc456d91SStefano Zampini /* cumulative of number of indexes and size of subset without holes */ 40495fa240b1SStefano Zampini #if defined(PETSC_HAVE_MPI_EXSCAN) 4050dc456d91SStefano Zampini start = 0; 405164a8e5bcSStefano Zampini ierr = MPI_Exscan(&nlocals,&start,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)subset));CHKERRQ(ierr); 40525fa240b1SStefano Zampini #else 405364a8e5bcSStefano Zampini ierr = MPI_Scan(&nlocals,&start,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)subset));CHKERRQ(ierr); 40545fa240b1SStefano Zampini start = start-nlocals; 40555fa240b1SStefano Zampini #endif 40565fa240b1SStefano Zampini 4057dc456d91SStefano Zampini if (N_n) { /* compute total size of new subset if requested */ 4058dc456d91SStefano Zampini *N_n = start + nlocals; 4059dc456d91SStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)subset),&commsize);CHKERRQ(ierr); 4060dc456d91SStefano Zampini ierr = MPI_Bcast(N_n,1,MPIU_INT,commsize-1,PetscObjectComm((PetscObject)subset));CHKERRQ(ierr); 4061674ae819SStefano Zampini } 40625fa240b1SStefano Zampini 40635fa240b1SStefano Zampini /* adapt root data with cumulative */ 4064674ae819SStefano Zampini if (first_found) { 4065dc456d91SStefano Zampini PetscInt old_index; 4066dc456d91SStefano Zampini 4067dc456d91SStefano Zampini root_data[first_index] += start; 4068674ae819SStefano Zampini old_index = first_index; 4069dc456d91SStefano Zampini for (i=first_index+1;i<Nl;i++) { 4070dc456d91SStefano Zampini if (root_data[i]) { 4071dc456d91SStefano Zampini root_data[i] += root_data[old_index]; 4072674ae819SStefano Zampini old_index = i; 4073674ae819SStefano Zampini } 4074674ae819SStefano Zampini } 4075674ae819SStefano Zampini } 4076dc456d91SStefano Zampini 4077dc456d91SStefano Zampini /* from roots to leaves */ 4078dc456d91SStefano Zampini ierr = PetscSFBcastBegin(sf,MPIU_INT,root_data,leaf_data);CHKERRQ(ierr); 4079dc456d91SStefano Zampini ierr = PetscSFBcastEnd(sf,MPIU_INT,root_data,leaf_data);CHKERRQ(ierr); 4080dc456d91SStefano Zampini ierr = PetscSFDestroy(&sf);CHKERRQ(ierr); 4081dc456d91SStefano Zampini 4082dc456d91SStefano Zampini /* create new IS with global indexes without holes */ 4083dc456d91SStefano Zampini if (subset_mult) { 4084dc456d91SStefano Zampini const PetscInt* idxs_mult; 4085dc456d91SStefano Zampini PetscInt cum; 4086dc456d91SStefano Zampini 4087dc456d91SStefano Zampini cum = 0; 4088dc456d91SStefano Zampini ierr = ISGetIndices(subset_mult,&idxs_mult);CHKERRQ(ierr); 4089dc456d91SStefano Zampini for (i=0;i<n;i++) { 4090dc456d91SStefano Zampini PetscInt j; 4091dc456d91SStefano Zampini for (j=0;j<idxs_mult[i];j++) gidxs[cum++] = leaf_data[i] - idxs_mult[i] + j; 4092674ae819SStefano Zampini } 4093dc456d91SStefano Zampini ierr = ISRestoreIndices(subset_mult,&idxs_mult);CHKERRQ(ierr); 4094674ae819SStefano Zampini } else { 4095dc456d91SStefano Zampini for (i=0;i<n;i++) { 4096dc456d91SStefano Zampini gidxs[i] = leaf_data[i]-1; 4097674ae819SStefano Zampini } 4098674ae819SStefano Zampini } 4099dc456d91SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)subset),n_n,gidxs,PETSC_OWN_POINTER,subset_n);CHKERRQ(ierr); 4100dc456d91SStefano Zampini ierr = PetscFree2(leaf_data,root_data);CHKERRQ(ierr); 4101674ae819SStefano Zampini PetscFunctionReturn(0); 4102674ae819SStefano Zampini } 41039a7d3425SStefano Zampini 41049a7d3425SStefano Zampini #undef __FUNCT__ 41059a7d3425SStefano Zampini #define __FUNCT__ "PCBDDCOrthonormalizeVecs" 41069a7d3425SStefano Zampini PetscErrorCode PCBDDCOrthonormalizeVecs(PetscInt n, Vec vecs[]) 41079a7d3425SStefano Zampini { 41089a7d3425SStefano Zampini PetscInt i,j; 41099a7d3425SStefano Zampini PetscScalar *alphas; 41109a7d3425SStefano Zampini PetscErrorCode ierr; 41119a7d3425SStefano Zampini 41129a7d3425SStefano Zampini PetscFunctionBegin; 41139a7d3425SStefano Zampini /* this implements stabilized Gram-Schmidt */ 4114785e854fSJed Brown ierr = PetscMalloc1(n,&alphas);CHKERRQ(ierr); 41159a7d3425SStefano Zampini for (i=0;i<n;i++) { 41169a7d3425SStefano Zampini ierr = VecNormalize(vecs[i],NULL);CHKERRQ(ierr); 41179a7d3425SStefano Zampini if (i<n) { ierr = VecMDot(vecs[i],n-i-1,&vecs[i+1],&alphas[i+1]);CHKERRQ(ierr); } 41189a7d3425SStefano Zampini for (j=i+1;j<n;j++) { ierr = VecAXPY(vecs[j],PetscConj(-alphas[j]),vecs[i]);CHKERRQ(ierr); } 41199a7d3425SStefano Zampini } 41209a7d3425SStefano Zampini ierr = PetscFree(alphas);CHKERRQ(ierr); 41219a7d3425SStefano Zampini PetscFunctionReturn(0); 41229a7d3425SStefano Zampini } 41239a7d3425SStefano Zampini 4124e7931f94SStefano Zampini #undef __FUNCT__ 412570cf5478SStefano Zampini #define __FUNCT__ "MatISGetSubassemblingPattern" 4126b0c7d250SStefano Zampini PetscErrorCode MatISGetSubassemblingPattern(Mat mat, PetscInt n_subdomains, PetscInt redprocs, IS* is_sends) 4127e7931f94SStefano Zampini { 412852e5ac9dSStefano Zampini IS ranks_send_to; 4129e7931f94SStefano Zampini PetscInt n_neighs,*neighs,*n_shared,**shared; 4130e7931f94SStefano Zampini PetscMPIInt size,rank,color; 413152e5ac9dSStefano Zampini PetscInt *xadj,*adjncy; 413252e5ac9dSStefano Zampini PetscInt *adjncy_wgt,*v_wgt,*ranks_send_to_idx; 41333837a79fSStefano Zampini PetscInt i,local_size,threshold=0; 41342b510759SStefano Zampini PetscBool use_vwgt=PETSC_FALSE,use_square=PETSC_FALSE; 4135e7931f94SStefano Zampini PetscSubcomm subcomm; 413652e5ac9dSStefano Zampini PetscErrorCode ierr; 4137a57a6d2fSStefano Zampini 4138e7931f94SStefano Zampini PetscFunctionBegin; 41392b510759SStefano Zampini ierr = PetscOptionsGetBool(NULL,"-matis_partitioning_use_square",&use_square,NULL);CHKERRQ(ierr); 41402b510759SStefano Zampini ierr = PetscOptionsGetBool(NULL,"-matis_partitioning_use_vwgt",&use_vwgt,NULL);CHKERRQ(ierr); 41412b510759SStefano Zampini ierr = PetscOptionsGetInt(NULL,"-matis_partitioning_threshold",&threshold,NULL);CHKERRQ(ierr); 4142e7931f94SStefano Zampini 4143e7931f94SStefano Zampini /* Get info on mapping */ 41443bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetSize(mat->rmap->mapping,&local_size);CHKERRQ(ierr); 41453bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetInfo(mat->rmap->mapping,&n_neighs,&neighs,&n_shared,&shared);CHKERRQ(ierr); 4146e7931f94SStefano Zampini 4147e7931f94SStefano Zampini /* build local CSR graph of subdomains' connectivity */ 4148785e854fSJed Brown ierr = PetscMalloc1(2,&xadj);CHKERRQ(ierr); 4149e7931f94SStefano Zampini xadj[0] = 0; 4150e7931f94SStefano Zampini xadj[1] = PetscMax(n_neighs-1,0); 4151785e854fSJed Brown ierr = PetscMalloc1(xadj[1],&adjncy);CHKERRQ(ierr); 4152785e854fSJed Brown ierr = PetscMalloc1(xadj[1],&adjncy_wgt);CHKERRQ(ierr); 4153e7931f94SStefano Zampini 41542b510759SStefano Zampini if (threshold) { 4155d023bfaeSStefano Zampini PetscInt xadj_count = 0; 41562b510759SStefano Zampini for (i=1;i<n_neighs;i++) { 4157d023bfaeSStefano Zampini if (n_shared[i] > threshold) { 4158d023bfaeSStefano Zampini adjncy[xadj_count] = neighs[i]; 4159d023bfaeSStefano Zampini adjncy_wgt[xadj_count] = n_shared[i]; 4160d023bfaeSStefano Zampini xadj_count++; 4161e7931f94SStefano Zampini } 4162e7931f94SStefano Zampini } 4163d023bfaeSStefano Zampini xadj[1] = xadj_count; 4164c8587f34SStefano Zampini } else { 4165e7931f94SStefano Zampini if (xadj[1]) { 4166e7931f94SStefano Zampini ierr = PetscMemcpy(adjncy,&neighs[1],xadj[1]*sizeof(*adjncy));CHKERRQ(ierr); 4167e7931f94SStefano Zampini ierr = PetscMemcpy(adjncy_wgt,&n_shared[1],xadj[1]*sizeof(*adjncy_wgt));CHKERRQ(ierr); 4168c8587f34SStefano Zampini } 4169e7931f94SStefano Zampini } 41703bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingRestoreInfo(mat->rmap->mapping,&n_neighs,&neighs,&n_shared,&shared);CHKERRQ(ierr); 4171e7931f94SStefano Zampini if (use_square) { 4172e7931f94SStefano Zampini for (i=0;i<xadj[1];i++) { 4173e7931f94SStefano Zampini adjncy_wgt[i] = adjncy_wgt[i]*adjncy_wgt[i]; 4174e7931f94SStefano Zampini } 4175e7931f94SStefano Zampini } 4176e7931f94SStefano Zampini ierr = PetscSortIntWithArray(xadj[1],adjncy,adjncy_wgt);CHKERRQ(ierr); 4177e7931f94SStefano Zampini 41783837a79fSStefano Zampini ierr = PetscMalloc1(1,&ranks_send_to_idx);CHKERRQ(ierr); 4179e7931f94SStefano Zampini 4180e7931f94SStefano Zampini /* 4181e7931f94SStefano Zampini Restrict work on active processes only. 4182e7931f94SStefano Zampini */ 4183e7931f94SStefano Zampini ierr = PetscSubcommCreate(PetscObjectComm((PetscObject)mat),&subcomm);CHKERRQ(ierr); 4184e7931f94SStefano Zampini ierr = PetscSubcommSetNumber(subcomm,2);CHKERRQ(ierr); /* 2 groups, active process and not active processes */ 4185e7931f94SStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)mat),&rank);CHKERRQ(ierr); 41862b510759SStefano Zampini ierr = PetscMPIIntCast(!local_size,&color);CHKERRQ(ierr); 4187d3531aaaSJed Brown ierr = PetscSubcommSetTypeGeneral(subcomm,color,rank);CHKERRQ(ierr); 4188e7931f94SStefano Zampini if (color) { 4189e7931f94SStefano Zampini ierr = PetscFree(xadj);CHKERRQ(ierr); 4190e7931f94SStefano Zampini ierr = PetscFree(adjncy);CHKERRQ(ierr); 4191e7931f94SStefano Zampini ierr = PetscFree(adjncy_wgt);CHKERRQ(ierr); 4192c8587f34SStefano Zampini } else { 419352e5ac9dSStefano Zampini Mat subdomain_adj; 419452e5ac9dSStefano Zampini IS new_ranks,new_ranks_contig; 419552e5ac9dSStefano Zampini MatPartitioning partitioner; 419652e5ac9dSStefano Zampini PetscInt prank,rstart=0,rend=0; 419752e5ac9dSStefano Zampini PetscInt *is_indices,*oldranks; 4198b0c7d250SStefano Zampini PetscBool aggregate; 4199b0c7d250SStefano Zampini 4200306c2d5bSBarry Smith ierr = MPI_Comm_size(PetscSubcommChild(subcomm),&size);CHKERRQ(ierr); 4201785e854fSJed Brown ierr = PetscMalloc1(size,&oldranks);CHKERRQ(ierr); 4202e7931f94SStefano Zampini prank = rank; 4203306c2d5bSBarry Smith ierr = MPI_Allgather(&prank,1,MPIU_INT,oldranks,1,MPIU_INT,PetscSubcommChild(subcomm));CHKERRQ(ierr); 42048002ef2cSStefano Zampini /* 4205e7931f94SStefano Zampini for (i=0;i<size;i++) { 4206e7931f94SStefano Zampini PetscPrintf(subcomm->comm,"oldranks[%d] = %d\n",i,oldranks[i]); 4207c8587f34SStefano Zampini } 42088002ef2cSStefano Zampini */ 4209e7931f94SStefano Zampini for (i=0;i<xadj[1];i++) { 4210e7931f94SStefano Zampini ierr = PetscFindInt(adjncy[i],size,oldranks,&adjncy[i]);CHKERRQ(ierr); 4211c8587f34SStefano Zampini } 4212e7931f94SStefano Zampini ierr = PetscSortIntWithArray(xadj[1],adjncy,adjncy_wgt);CHKERRQ(ierr); 4213b0c7d250SStefano Zampini aggregate = ((redprocs > 0 && redprocs < size) ? PETSC_TRUE : PETSC_FALSE); 4214b0c7d250SStefano Zampini if (aggregate) { 4215b0c7d250SStefano Zampini PetscInt lrows,row,ncols,*cols; 4216b0c7d250SStefano Zampini PetscMPIInt nrank; 4217b0c7d250SStefano Zampini PetscScalar *vals; 4218b0c7d250SStefano Zampini 4219b0c7d250SStefano Zampini ierr = MPI_Comm_rank(PetscSubcommChild(subcomm),&nrank);CHKERRQ(ierr); 4220b0c7d250SStefano Zampini lrows = 0; 4221b0c7d250SStefano Zampini if (nrank<redprocs) { 4222b0c7d250SStefano Zampini lrows = size/redprocs; 4223b0c7d250SStefano Zampini if (nrank<size%redprocs) lrows++; 4224b0c7d250SStefano Zampini } 42255fa240b1SStefano Zampini ierr = MatCreateAIJ(PetscSubcommChild(subcomm),lrows,lrows,size,size,50,NULL,50,NULL,&subdomain_adj);CHKERRQ(ierr); 4226b0c7d250SStefano Zampini ierr = MatGetOwnershipRange(subdomain_adj,&rstart,&rend);CHKERRQ(ierr); 4227b0c7d250SStefano Zampini ierr = MatSetOption(subdomain_adj,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_FALSE);CHKERRQ(ierr); 4228b0c7d250SStefano Zampini ierr = MatSetOption(subdomain_adj,MAT_NEW_NONZERO_ALLOCATION_ERR,PETSC_FALSE);CHKERRQ(ierr); 4229b0c7d250SStefano Zampini row = nrank; 4230b0c7d250SStefano Zampini ncols = xadj[1]-xadj[0]; 4231b0c7d250SStefano Zampini cols = adjncy; 4232b0c7d250SStefano Zampini ierr = PetscMalloc1(ncols,&vals);CHKERRQ(ierr); 4233b0c7d250SStefano Zampini for (i=0;i<ncols;i++) vals[i] = adjncy_wgt[i]; 4234b0c7d250SStefano Zampini ierr = MatSetValues(subdomain_adj,1,&row,ncols,cols,vals,INSERT_VALUES);CHKERRQ(ierr); 4235b0c7d250SStefano Zampini ierr = MatAssemblyBegin(subdomain_adj,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4236b0c7d250SStefano Zampini ierr = MatAssemblyEnd(subdomain_adj,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 423752e5ac9dSStefano Zampini ierr = PetscFree(xadj);CHKERRQ(ierr); 423852e5ac9dSStefano Zampini ierr = PetscFree(adjncy);CHKERRQ(ierr); 423952e5ac9dSStefano Zampini ierr = PetscFree(adjncy_wgt);CHKERRQ(ierr); 4240b0c7d250SStefano Zampini ierr = PetscFree(vals);CHKERRQ(ierr); 4241b0c7d250SStefano Zampini } else { 4242306c2d5bSBarry Smith ierr = MatCreateMPIAdj(PetscSubcommChild(subcomm),1,(PetscInt)size,xadj,adjncy,adjncy_wgt,&subdomain_adj);CHKERRQ(ierr); 4243b0c7d250SStefano Zampini } 424422b6e8a2SStefano Zampini /* ierr = MatView(subdomain_adj,0);CHKERRQ(ierr); */ 4245e7931f94SStefano Zampini 4246e7931f94SStefano Zampini /* Partition */ 4247306c2d5bSBarry Smith ierr = MatPartitioningCreate(PetscSubcommChild(subcomm),&partitioner);CHKERRQ(ierr); 4248e7931f94SStefano Zampini ierr = MatPartitioningSetAdjacency(partitioner,subdomain_adj);CHKERRQ(ierr); 4249e7931f94SStefano Zampini if (use_vwgt) { 42503837a79fSStefano Zampini ierr = PetscMalloc1(1,&v_wgt);CHKERRQ(ierr); 4251e7931f94SStefano Zampini v_wgt[0] = local_size; 4252e7931f94SStefano Zampini ierr = MatPartitioningSetVertexWeights(partitioner,v_wgt);CHKERRQ(ierr); 4253c8587f34SStefano Zampini } 425428143c3dSStefano Zampini n_subdomains = PetscMin((PetscInt)size,n_subdomains); 425528143c3dSStefano Zampini ierr = MatPartitioningSetNParts(partitioner,n_subdomains);CHKERRQ(ierr); 4256e7931f94SStefano Zampini ierr = MatPartitioningSetFromOptions(partitioner);CHKERRQ(ierr); 4257e7931f94SStefano Zampini ierr = MatPartitioningApply(partitioner,&new_ranks);CHKERRQ(ierr); 425822b6e8a2SStefano Zampini /* ierr = MatPartitioningView(partitioner,0);CHKERRQ(ierr); */ 4259e7931f94SStefano Zampini 426052e5ac9dSStefano Zampini /* renumber new_ranks to avoid "holes" in new set of processors */ 426152e5ac9dSStefano Zampini ierr = PCBDDCSubsetNumbering(new_ranks,NULL,NULL,&new_ranks_contig);CHKERRQ(ierr); 426252e5ac9dSStefano Zampini ierr = ISDestroy(&new_ranks);CHKERRQ(ierr); 426352e5ac9dSStefano Zampini ierr = ISGetIndices(new_ranks_contig,(const PetscInt**)&is_indices);CHKERRQ(ierr); 4264b0c7d250SStefano Zampini if (!redprocs) { 4265b0c7d250SStefano Zampini ranks_send_to_idx[0] = oldranks[is_indices[0]]; 426628143c3dSStefano Zampini } else { 4267b0c7d250SStefano Zampini PetscInt idxs[1]; 4268b0c7d250SStefano Zampini PetscMPIInt tag; 4269b0c7d250SStefano Zampini MPI_Request *reqs; 4270b0c7d250SStefano Zampini 4271b0c7d250SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)subdomain_adj,&tag);CHKERRQ(ierr); 4272b0c7d250SStefano Zampini ierr = PetscMalloc1(rend-rstart,&reqs);CHKERRQ(ierr); 4273b0c7d250SStefano Zampini for (i=rstart;i<rend;i++) { 4274b0c7d250SStefano Zampini ierr = MPI_Isend(is_indices+i-rstart,1,MPIU_INT,i,tag,PetscSubcommChild(subcomm),&reqs[i-rstart]);CHKERRQ(ierr); 427528143c3dSStefano Zampini } 4276b0c7d250SStefano Zampini ierr = MPI_Recv(idxs,1,MPIU_INT,MPI_ANY_SOURCE,tag,PetscSubcommChild(subcomm),MPI_STATUS_IGNORE);CHKERRQ(ierr); 4277b0c7d250SStefano Zampini ierr = MPI_Waitall(rend-rstart,reqs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 4278b0c7d250SStefano Zampini ierr = PetscFree(reqs);CHKERRQ(ierr); 4279b0c7d250SStefano Zampini ranks_send_to_idx[0] = oldranks[idxs[0]]; 4280e7931f94SStefano Zampini } 428152e5ac9dSStefano Zampini ierr = ISRestoreIndices(new_ranks_contig,(const PetscInt**)&is_indices);CHKERRQ(ierr); 4282e7931f94SStefano Zampini /* clean up */ 4283e7931f94SStefano Zampini ierr = PetscFree(oldranks);CHKERRQ(ierr); 428452e5ac9dSStefano Zampini ierr = ISDestroy(&new_ranks_contig);CHKERRQ(ierr); 4285e7931f94SStefano Zampini ierr = MatDestroy(&subdomain_adj);CHKERRQ(ierr); 4286e7931f94SStefano Zampini ierr = MatPartitioningDestroy(&partitioner);CHKERRQ(ierr); 4287e7931f94SStefano Zampini } 4288e7931f94SStefano Zampini ierr = PetscSubcommDestroy(&subcomm);CHKERRQ(ierr); 4289e7931f94SStefano Zampini 4290e7931f94SStefano Zampini /* assemble parallel IS for sends */ 4291e7931f94SStefano Zampini i = 1; 4292e7931f94SStefano Zampini if (color) i=0; 4293e7931f94SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)mat),i,ranks_send_to_idx,PETSC_OWN_POINTER,&ranks_send_to);CHKERRQ(ierr); 4294e7931f94SStefano Zampini /* get back IS */ 4295e7931f94SStefano Zampini *is_sends = ranks_send_to; 4296e7931f94SStefano Zampini PetscFunctionReturn(0); 4297e7931f94SStefano Zampini } 4298e7931f94SStefano Zampini 4299e7931f94SStefano Zampini typedef enum {MATDENSE_PRIVATE=0,MATAIJ_PRIVATE,MATBAIJ_PRIVATE,MATSBAIJ_PRIVATE}MatTypePrivate; 4300e7931f94SStefano Zampini 4301e7931f94SStefano Zampini #undef __FUNCT__ 4302e7931f94SStefano Zampini #define __FUNCT__ "MatISSubassemble" 430353a05cb3SStefano 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[]) 4304e7931f94SStefano Zampini { 430570cf5478SStefano Zampini Mat local_mat; 4306e7931f94SStefano Zampini IS is_sends_internal; 43079d30be91SStefano Zampini PetscInt rows,cols,new_local_rows; 430828143c3dSStefano Zampini PetscInt i,bs,buf_size_idxs,buf_size_idxs_is,buf_size_vals; 43099d30be91SStefano Zampini PetscBool ismatis,isdense,newisdense,destroy_mat; 4310e7931f94SStefano Zampini ISLocalToGlobalMapping l2gmap; 4311e7931f94SStefano Zampini PetscInt* l2gmap_indices; 4312e7931f94SStefano Zampini const PetscInt* is_indices; 4313e7931f94SStefano Zampini MatType new_local_type; 4314e7931f94SStefano Zampini /* buffers */ 4315e7931f94SStefano Zampini PetscInt *ptr_idxs,*send_buffer_idxs,*recv_buffer_idxs; 431628143c3dSStefano Zampini PetscInt *ptr_idxs_is,*send_buffer_idxs_is,*recv_buffer_idxs_is; 43179d30be91SStefano Zampini PetscInt *recv_buffer_idxs_local; 4318e7931f94SStefano Zampini PetscScalar *ptr_vals,*send_buffer_vals,*recv_buffer_vals; 4319e7931f94SStefano Zampini /* MPI */ 432028143c3dSStefano Zampini MPI_Comm comm,comm_n; 432128143c3dSStefano Zampini PetscSubcomm subcomm; 4322e7931f94SStefano Zampini PetscMPIInt n_sends,n_recvs,commsize; 432328143c3dSStefano Zampini PetscMPIInt *iflags,*ilengths_idxs,*ilengths_vals,*ilengths_idxs_is; 432428143c3dSStefano Zampini PetscMPIInt *onodes,*onodes_is,*olengths_idxs,*olengths_idxs_is,*olengths_vals; 432528143c3dSStefano Zampini PetscMPIInt len,tag_idxs,tag_idxs_is,tag_vals,source_dest; 432628143c3dSStefano Zampini MPI_Request *send_req_idxs,*send_req_idxs_is,*send_req_vals; 432728143c3dSStefano Zampini MPI_Request *recv_req_idxs,*recv_req_idxs_is,*recv_req_vals; 4328e7931f94SStefano Zampini PetscErrorCode ierr; 4329e7931f94SStefano Zampini 4330e7931f94SStefano Zampini PetscFunctionBegin; 433128143c3dSStefano Zampini /* TODO: add missing checks */ 433228143c3dSStefano Zampini PetscValidLogicalCollectiveInt(mat,n_subdomains,3); 433328143c3dSStefano Zampini PetscValidLogicalCollectiveBool(mat,restrict_comm,4); 433428143c3dSStefano Zampini PetscValidLogicalCollectiveEnum(mat,reuse,5); 433528143c3dSStefano Zampini PetscValidLogicalCollectiveInt(mat,nis,7); 4336e7931f94SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)mat,MATIS,&ismatis);CHKERRQ(ierr); 433728143c3dSStefano Zampini if (!ismatis) SETERRQ1(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot use %s on a matrix object which is not of type MATIS",__FUNCT__); 4338e7931f94SStefano Zampini ierr = MatISGetLocalMat(mat,&local_mat);CHKERRQ(ierr); 4339e7931f94SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)local_mat,MATSEQDENSE,&isdense);CHKERRQ(ierr); 4340e7931f94SStefano Zampini if (!isdense) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Currently cannot subassemble MATIS when local matrix type is not of type SEQDENSE"); 4341e7931f94SStefano Zampini ierr = MatGetSize(local_mat,&rows,&cols);CHKERRQ(ierr); 4342e7931f94SStefano Zampini if (rows != cols) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Local MATIS matrices should be square"); 434328143c3dSStefano Zampini if (reuse == MAT_REUSE_MATRIX && *mat_n) { 434470cf5478SStefano Zampini PetscInt mrows,mcols,mnrows,mncols; 434570cf5478SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)*mat_n,MATIS,&ismatis);CHKERRQ(ierr); 434628143c3dSStefano Zampini if (!ismatis) SETERRQ(PetscObjectComm((PetscObject)*mat_n),PETSC_ERR_SUP,"Cannot reuse a matrix which is not of type MATIS"); 434770cf5478SStefano Zampini ierr = MatGetSize(mat,&mrows,&mcols);CHKERRQ(ierr); 434870cf5478SStefano Zampini ierr = MatGetSize(*mat_n,&mnrows,&mncols);CHKERRQ(ierr); 434970cf5478SStefano Zampini if (mrows != mnrows) SETERRQ2(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot reuse matrix! Wrong number of rows %D != %D",mrows,mnrows); 435070cf5478SStefano Zampini if (mcols != mncols) SETERRQ2(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot reuse matrix! Wrong number of cols %D != %D",mcols,mncols); 435170cf5478SStefano Zampini } 4352e7931f94SStefano Zampini ierr = MatGetBlockSize(local_mat,&bs);CHKERRQ(ierr); 4353e7931f94SStefano Zampini PetscValidLogicalCollectiveInt(mat,bs,0); 4354e7931f94SStefano Zampini /* prepare IS for sending if not provided */ 4355e7931f94SStefano Zampini if (!is_sends) { 435628143c3dSStefano Zampini if (!n_subdomains) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"You should specify either an IS or a target number of subdomains"); 4357b0c7d250SStefano Zampini ierr = MatISGetSubassemblingPattern(mat,n_subdomains,0,&is_sends_internal);CHKERRQ(ierr); 4358c8587f34SStefano Zampini } else { 4359e7931f94SStefano Zampini ierr = PetscObjectReference((PetscObject)is_sends);CHKERRQ(ierr); 4360e7931f94SStefano Zampini is_sends_internal = is_sends; 4361c8587f34SStefano Zampini } 4362e7931f94SStefano Zampini 4363e7931f94SStefano Zampini /* get comm */ 4364a316fed8SStefano Zampini ierr = PetscObjectGetComm((PetscObject)mat,&comm);CHKERRQ(ierr); 4365e7931f94SStefano Zampini 4366e7931f94SStefano Zampini /* compute number of sends */ 4367e7931f94SStefano Zampini ierr = ISGetLocalSize(is_sends_internal,&i);CHKERRQ(ierr); 4368e7931f94SStefano Zampini ierr = PetscMPIIntCast(i,&n_sends);CHKERRQ(ierr); 4369e7931f94SStefano Zampini 4370e7931f94SStefano Zampini /* compute number of receives */ 4371e7931f94SStefano Zampini ierr = MPI_Comm_size(comm,&commsize);CHKERRQ(ierr); 4372785e854fSJed Brown ierr = PetscMalloc1(commsize,&iflags);CHKERRQ(ierr); 4373e7931f94SStefano Zampini ierr = PetscMemzero(iflags,commsize*sizeof(*iflags));CHKERRQ(ierr); 4374e7931f94SStefano Zampini ierr = ISGetIndices(is_sends_internal,&is_indices);CHKERRQ(ierr); 4375e7931f94SStefano Zampini for (i=0;i<n_sends;i++) iflags[is_indices[i]] = 1; 4376e7931f94SStefano Zampini ierr = PetscGatherNumberOfMessages(comm,iflags,NULL,&n_recvs);CHKERRQ(ierr); 4377e7931f94SStefano Zampini ierr = PetscFree(iflags);CHKERRQ(ierr); 4378e7931f94SStefano Zampini 437928143c3dSStefano Zampini /* restrict comm if requested */ 438028143c3dSStefano Zampini subcomm = 0; 438128143c3dSStefano Zampini destroy_mat = PETSC_FALSE; 438228143c3dSStefano Zampini if (restrict_comm) { 4383779c1cceSStefano Zampini PetscMPIInt color,subcommsize; 4384779c1cceSStefano Zampini 438528143c3dSStefano Zampini color = 0; 438653a05cb3SStefano Zampini if (restrict_full) { 438753a05cb3SStefano Zampini if (!n_recvs) color = 1; /* processes not receiving anything will not partecipate in new comm (full restriction) */ 438853a05cb3SStefano Zampini } else { 438953a05cb3SStefano Zampini if (!n_recvs && n_sends) color = 1; /* just those processes that are sending but not receiving anything will not partecipate in new comm */ 439053a05cb3SStefano Zampini } 439128143c3dSStefano Zampini ierr = MPI_Allreduce(&color,&subcommsize,1,MPI_INT,MPI_SUM,comm);CHKERRQ(ierr); 439228143c3dSStefano Zampini subcommsize = commsize - subcommsize; 439328143c3dSStefano Zampini /* check if reuse has been requested */ 439428143c3dSStefano Zampini if (reuse == MAT_REUSE_MATRIX) { 439528143c3dSStefano Zampini if (*mat_n) { 439628143c3dSStefano Zampini PetscMPIInt subcommsize2; 439728143c3dSStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)*mat_n),&subcommsize2);CHKERRQ(ierr); 439828143c3dSStefano Zampini if (subcommsize != subcommsize2) SETERRQ2(PetscObjectComm((PetscObject)*mat_n),PETSC_ERR_PLIB,"Cannot reuse matrix! wrong subcomm size %d != %d",subcommsize,subcommsize2); 439928143c3dSStefano Zampini comm_n = PetscObjectComm((PetscObject)*mat_n); 440028143c3dSStefano Zampini } else { 440128143c3dSStefano Zampini comm_n = PETSC_COMM_SELF; 440228143c3dSStefano Zampini } 440328143c3dSStefano Zampini } else { /* MAT_INITIAL_MATRIX */ 4404779c1cceSStefano Zampini PetscMPIInt rank; 4405779c1cceSStefano Zampini 4406779c1cceSStefano Zampini ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 440728143c3dSStefano Zampini ierr = PetscSubcommCreate(comm,&subcomm);CHKERRQ(ierr); 440828143c3dSStefano Zampini ierr = PetscSubcommSetNumber(subcomm,2);CHKERRQ(ierr); 440928143c3dSStefano Zampini ierr = PetscSubcommSetTypeGeneral(subcomm,color,rank);CHKERRQ(ierr); 4410306c2d5bSBarry Smith comm_n = PetscSubcommChild(subcomm); 441128143c3dSStefano Zampini } 441228143c3dSStefano Zampini /* flag to destroy *mat_n if not significative */ 441328143c3dSStefano Zampini if (color) destroy_mat = PETSC_TRUE; 441428143c3dSStefano Zampini } else { 441528143c3dSStefano Zampini comm_n = comm; 441628143c3dSStefano Zampini } 441728143c3dSStefano Zampini 4418e7931f94SStefano Zampini /* prepare send/receive buffers */ 4419785e854fSJed Brown ierr = PetscMalloc1(commsize,&ilengths_idxs);CHKERRQ(ierr); 4420e7931f94SStefano Zampini ierr = PetscMemzero(ilengths_idxs,commsize*sizeof(*ilengths_idxs));CHKERRQ(ierr); 4421785e854fSJed Brown ierr = PetscMalloc1(commsize,&ilengths_vals);CHKERRQ(ierr); 4422e7931f94SStefano Zampini ierr = PetscMemzero(ilengths_vals,commsize*sizeof(*ilengths_vals));CHKERRQ(ierr); 442328143c3dSStefano Zampini if (nis) { 4424854ce69bSBarry Smith ierr = PetscCalloc1(commsize,&ilengths_idxs_is);CHKERRQ(ierr); 442528143c3dSStefano Zampini } 4426e7931f94SStefano Zampini 442728143c3dSStefano Zampini /* Get data from local matrices */ 4428e7931f94SStefano Zampini if (!isdense) { 4429a26c9d0eSStefano Zampini SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Subassembling of AIJ local matrices not yet implemented"); 4430e7931f94SStefano Zampini /* TODO: See below some guidelines on how to prepare the local buffers */ 4431e7931f94SStefano Zampini /* 4432e7931f94SStefano Zampini send_buffer_vals should contain the raw values of the local matrix 4433e7931f94SStefano Zampini send_buffer_idxs should contain: 4434e7931f94SStefano Zampini - MatType_PRIVATE type 4435e7931f94SStefano Zampini - PetscInt size_of_l2gmap 4436e7931f94SStefano Zampini - PetscInt global_row_indices[size_of_l2gmap] 4437e7931f94SStefano Zampini - PetscInt all_other_info_which_is_needed_to_compute_preallocation_and_set_values 4438e7931f94SStefano Zampini */ 4439e7931f94SStefano Zampini } else { 4440e7931f94SStefano Zampini ierr = MatDenseGetArray(local_mat,&send_buffer_vals);CHKERRQ(ierr); 44413bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetSize(mat->rmap->mapping,&i);CHKERRQ(ierr); 4442854ce69bSBarry Smith ierr = PetscMalloc1(i+2,&send_buffer_idxs);CHKERRQ(ierr); 4443e7931f94SStefano Zampini send_buffer_idxs[0] = (PetscInt)MATDENSE_PRIVATE; 4444e7931f94SStefano Zampini send_buffer_idxs[1] = i; 44453bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetIndices(mat->rmap->mapping,(const PetscInt**)&ptr_idxs);CHKERRQ(ierr); 4446e7931f94SStefano Zampini ierr = PetscMemcpy(&send_buffer_idxs[2],ptr_idxs,i*sizeof(PetscInt));CHKERRQ(ierr); 44473bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingRestoreIndices(mat->rmap->mapping,(const PetscInt**)&ptr_idxs);CHKERRQ(ierr); 4448e7931f94SStefano Zampini ierr = PetscMPIIntCast(i,&len);CHKERRQ(ierr); 4449e7931f94SStefano Zampini for (i=0;i<n_sends;i++) { 4450e7931f94SStefano Zampini ilengths_vals[is_indices[i]] = len*len; 4451e7931f94SStefano Zampini ilengths_idxs[is_indices[i]] = len+2; 4452c8587f34SStefano Zampini } 4453c8587f34SStefano Zampini } 4454e7931f94SStefano Zampini ierr = PetscGatherMessageLengths2(comm,n_sends,n_recvs,ilengths_idxs,ilengths_vals,&onodes,&olengths_idxs,&olengths_vals);CHKERRQ(ierr); 445528143c3dSStefano Zampini /* additional is (if any) */ 445628143c3dSStefano Zampini if (nis) { 445728143c3dSStefano Zampini PetscMPIInt psum; 445828143c3dSStefano Zampini PetscInt j; 445928143c3dSStefano Zampini for (j=0,psum=0;j<nis;j++) { 446028143c3dSStefano Zampini PetscInt plen; 446128143c3dSStefano Zampini ierr = ISGetLocalSize(isarray[j],&plen);CHKERRQ(ierr); 446228143c3dSStefano Zampini ierr = PetscMPIIntCast(plen,&len);CHKERRQ(ierr); 446328143c3dSStefano Zampini psum += len+1; /* indices + lenght */ 446428143c3dSStefano Zampini } 4465854ce69bSBarry Smith ierr = PetscMalloc1(psum,&send_buffer_idxs_is);CHKERRQ(ierr); 446628143c3dSStefano Zampini for (j=0,psum=0;j<nis;j++) { 446728143c3dSStefano Zampini PetscInt plen; 446828143c3dSStefano Zampini const PetscInt *is_array_idxs; 446928143c3dSStefano Zampini ierr = ISGetLocalSize(isarray[j],&plen);CHKERRQ(ierr); 447028143c3dSStefano Zampini send_buffer_idxs_is[psum] = plen; 447128143c3dSStefano Zampini ierr = ISGetIndices(isarray[j],&is_array_idxs);CHKERRQ(ierr); 447228143c3dSStefano Zampini ierr = PetscMemcpy(&send_buffer_idxs_is[psum+1],is_array_idxs,plen*sizeof(PetscInt));CHKERRQ(ierr); 447328143c3dSStefano Zampini ierr = ISRestoreIndices(isarray[j],&is_array_idxs);CHKERRQ(ierr); 447428143c3dSStefano Zampini psum += plen+1; /* indices + lenght */ 447528143c3dSStefano Zampini } 447628143c3dSStefano Zampini for (i=0;i<n_sends;i++) { 447728143c3dSStefano Zampini ilengths_idxs_is[is_indices[i]] = psum; 447828143c3dSStefano Zampini } 447928143c3dSStefano Zampini ierr = PetscGatherMessageLengths(comm,n_sends,n_recvs,ilengths_idxs_is,&onodes_is,&olengths_idxs_is);CHKERRQ(ierr); 448028143c3dSStefano Zampini } 448128143c3dSStefano Zampini 4482e7931f94SStefano Zampini buf_size_idxs = 0; 4483e7931f94SStefano Zampini buf_size_vals = 0; 448428143c3dSStefano Zampini buf_size_idxs_is = 0; 4485e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 4486e7931f94SStefano Zampini buf_size_idxs += (PetscInt)olengths_idxs[i]; 4487e7931f94SStefano Zampini buf_size_vals += (PetscInt)olengths_vals[i]; 448828143c3dSStefano Zampini if (nis) buf_size_idxs_is += (PetscInt)olengths_idxs_is[i]; 4489e7931f94SStefano Zampini } 4490785e854fSJed Brown ierr = PetscMalloc1(buf_size_idxs,&recv_buffer_idxs);CHKERRQ(ierr); 4491785e854fSJed Brown ierr = PetscMalloc1(buf_size_vals,&recv_buffer_vals);CHKERRQ(ierr); 449295ecbf38SStefano Zampini ierr = PetscMalloc1(buf_size_idxs_is,&recv_buffer_idxs_is);CHKERRQ(ierr); 4493e7931f94SStefano Zampini 4494e7931f94SStefano Zampini /* get new tags for clean communications */ 4495e7931f94SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_idxs);CHKERRQ(ierr); 4496e7931f94SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_vals);CHKERRQ(ierr); 449728143c3dSStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_idxs_is);CHKERRQ(ierr); 4498e7931f94SStefano Zampini 4499e7931f94SStefano Zampini /* allocate for requests */ 4500785e854fSJed Brown ierr = PetscMalloc1(n_sends,&send_req_idxs);CHKERRQ(ierr); 4501785e854fSJed Brown ierr = PetscMalloc1(n_sends,&send_req_vals);CHKERRQ(ierr); 450295ecbf38SStefano Zampini ierr = PetscMalloc1(n_sends,&send_req_idxs_is);CHKERRQ(ierr); 4503785e854fSJed Brown ierr = PetscMalloc1(n_recvs,&recv_req_idxs);CHKERRQ(ierr); 4504785e854fSJed Brown ierr = PetscMalloc1(n_recvs,&recv_req_vals);CHKERRQ(ierr); 450595ecbf38SStefano Zampini ierr = PetscMalloc1(n_recvs,&recv_req_idxs_is);CHKERRQ(ierr); 4506e7931f94SStefano Zampini 4507e7931f94SStefano Zampini /* communications */ 4508e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 4509e7931f94SStefano Zampini ptr_vals = recv_buffer_vals; 451028143c3dSStefano Zampini ptr_idxs_is = recv_buffer_idxs_is; 4511e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 4512e7931f94SStefano Zampini source_dest = onodes[i]; 4513e7931f94SStefano Zampini ierr = MPI_Irecv(ptr_idxs,olengths_idxs[i],MPIU_INT,source_dest,tag_idxs,comm,&recv_req_idxs[i]);CHKERRQ(ierr); 4514e7931f94SStefano Zampini ierr = MPI_Irecv(ptr_vals,olengths_vals[i],MPIU_SCALAR,source_dest,tag_vals,comm,&recv_req_vals[i]);CHKERRQ(ierr); 4515e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 4516e7931f94SStefano Zampini ptr_vals += olengths_vals[i]; 451728143c3dSStefano Zampini if (nis) { 451828143c3dSStefano 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); 451928143c3dSStefano Zampini ptr_idxs_is += olengths_idxs_is[i]; 452028143c3dSStefano Zampini } 4521e7931f94SStefano Zampini } 4522e7931f94SStefano Zampini for (i=0;i<n_sends;i++) { 4523e7931f94SStefano Zampini ierr = PetscMPIIntCast(is_indices[i],&source_dest);CHKERRQ(ierr); 4524e7931f94SStefano Zampini ierr = MPI_Isend(send_buffer_idxs,ilengths_idxs[source_dest],MPIU_INT,source_dest,tag_idxs,comm,&send_req_idxs[i]);CHKERRQ(ierr); 4525e7931f94SStefano Zampini ierr = MPI_Isend(send_buffer_vals,ilengths_vals[source_dest],MPIU_SCALAR,source_dest,tag_vals,comm,&send_req_vals[i]);CHKERRQ(ierr); 452628143c3dSStefano Zampini if (nis) { 452728143c3dSStefano 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); 452828143c3dSStefano Zampini } 4529e7931f94SStefano Zampini } 4530e7931f94SStefano Zampini ierr = ISRestoreIndices(is_sends_internal,&is_indices);CHKERRQ(ierr); 4531e7931f94SStefano Zampini ierr = ISDestroy(&is_sends_internal);CHKERRQ(ierr); 4532e7931f94SStefano Zampini 4533e7931f94SStefano Zampini /* assemble new l2g map */ 4534e7931f94SStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_idxs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 4535e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 45369d30be91SStefano Zampini new_local_rows = 0; 4537e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 45389d30be91SStefano Zampini new_local_rows += *(ptr_idxs+1); /* second element is the local size of the l2gmap */ 4539e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 4540e7931f94SStefano Zampini } 45419d30be91SStefano Zampini ierr = PetscMalloc1(new_local_rows,&l2gmap_indices);CHKERRQ(ierr); 4542e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 45439d30be91SStefano Zampini new_local_rows = 0; 4544e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 45459d30be91SStefano Zampini ierr = PetscMemcpy(&l2gmap_indices[new_local_rows],ptr_idxs+2,(*(ptr_idxs+1))*sizeof(PetscInt));CHKERRQ(ierr); 45469d30be91SStefano Zampini new_local_rows += *(ptr_idxs+1); /* second element is the local size of the l2gmap */ 4547e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 4548e7931f94SStefano Zampini } 45499d30be91SStefano Zampini ierr = PetscSortRemoveDupsInt(&new_local_rows,l2gmap_indices);CHKERRQ(ierr); 45509d30be91SStefano Zampini ierr = ISLocalToGlobalMappingCreate(comm_n,1,new_local_rows,l2gmap_indices,PETSC_COPY_VALUES,&l2gmap);CHKERRQ(ierr); 4551e7931f94SStefano Zampini ierr = PetscFree(l2gmap_indices);CHKERRQ(ierr); 4552e7931f94SStefano Zampini 4553e7931f94SStefano Zampini /* infer new local matrix type from received local matrices type */ 4554e7931f94SStefano Zampini /* currently if all local matrices are of type X, then the resulting matrix will be of type X, except for the dense case */ 4555e7931f94SStefano 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) */ 4556e7931f94SStefano Zampini if (n_recvs) { 455728143c3dSStefano Zampini MatTypePrivate new_local_type_private = (MatTypePrivate)send_buffer_idxs[0]; 4558e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 4559e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 4560e7931f94SStefano Zampini if ((PetscInt)new_local_type_private != *ptr_idxs) { 4561e7931f94SStefano Zampini new_local_type_private = MATAIJ_PRIVATE; 4562e7931f94SStefano Zampini break; 4563e7931f94SStefano Zampini } 4564e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 4565e7931f94SStefano Zampini } 4566e7931f94SStefano Zampini switch (new_local_type_private) { 456728143c3dSStefano Zampini case MATDENSE_PRIVATE: 456828143c3dSStefano Zampini if (n_recvs>1) { /* subassembling of dense matrices does not give a dense matrix! */ 4569e7931f94SStefano Zampini new_local_type = MATSEQAIJ; 4570e7931f94SStefano Zampini bs = 1; 457128143c3dSStefano Zampini } else { /* if I receive only 1 dense matrix */ 457228143c3dSStefano Zampini new_local_type = MATSEQDENSE; 457328143c3dSStefano Zampini bs = 1; 457428143c3dSStefano Zampini } 4575e7931f94SStefano Zampini break; 4576e7931f94SStefano Zampini case MATAIJ_PRIVATE: 4577e7931f94SStefano Zampini new_local_type = MATSEQAIJ; 4578e7931f94SStefano Zampini bs = 1; 4579e7931f94SStefano Zampini break; 4580e7931f94SStefano Zampini case MATBAIJ_PRIVATE: 4581e7931f94SStefano Zampini new_local_type = MATSEQBAIJ; 4582e7931f94SStefano Zampini break; 4583e7931f94SStefano Zampini case MATSBAIJ_PRIVATE: 4584e7931f94SStefano Zampini new_local_type = MATSEQSBAIJ; 4585e7931f94SStefano Zampini break; 4586e7931f94SStefano Zampini default: 45879d30be91SStefano Zampini SETERRQ2(comm,PETSC_ERR_SUP,"Unsupported private type %d in %s",new_local_type_private,__FUNCT__); 4588e7931f94SStefano Zampini break; 4589e7931f94SStefano Zampini } 459028143c3dSStefano Zampini } else { /* by default, new_local_type is seqdense */ 459128143c3dSStefano Zampini new_local_type = MATSEQDENSE; 459228143c3dSStefano Zampini bs = 1; 4593e7931f94SStefano Zampini } 4594e7931f94SStefano Zampini 459570cf5478SStefano Zampini /* create MATIS object if needed */ 459670cf5478SStefano Zampini if (reuse == MAT_INITIAL_MATRIX) { 4597e7931f94SStefano Zampini ierr = MatGetSize(mat,&rows,&cols);CHKERRQ(ierr); 4598e176bc59SStefano Zampini ierr = MatCreateIS(comm_n,bs,PETSC_DECIDE,PETSC_DECIDE,rows,cols,l2gmap,NULL,mat_n);CHKERRQ(ierr); 459970cf5478SStefano Zampini } else { 460070cf5478SStefano Zampini /* it also destroys the local matrices */ 460170cf5478SStefano Zampini ierr = MatSetLocalToGlobalMapping(*mat_n,l2gmap,l2gmap);CHKERRQ(ierr); 460270cf5478SStefano Zampini } 460370cf5478SStefano Zampini ierr = MatISGetLocalMat(*mat_n,&local_mat);CHKERRQ(ierr); 4604e7931f94SStefano Zampini ierr = MatSetType(local_mat,new_local_type);CHKERRQ(ierr); 46059d30be91SStefano Zampini 46069d30be91SStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_vals,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 46079d30be91SStefano Zampini 46089d30be91SStefano Zampini /* Global to local map of received indices */ 46099d30be91SStefano Zampini ierr = PetscMalloc1(buf_size_idxs,&recv_buffer_idxs_local);CHKERRQ(ierr); /* needed for values insertion */ 46109d30be91SStefano Zampini ierr = ISGlobalToLocalMappingApply(l2gmap,IS_GTOLM_MASK,buf_size_idxs,recv_buffer_idxs,&i,recv_buffer_idxs_local);CHKERRQ(ierr); 46119d30be91SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&l2gmap);CHKERRQ(ierr); 46129d30be91SStefano Zampini 46139d30be91SStefano Zampini /* restore attributes -> type of incoming data and its size */ 46149d30be91SStefano Zampini buf_size_idxs = 0; 46159d30be91SStefano Zampini for (i=0;i<n_recvs;i++) { 46169d30be91SStefano Zampini recv_buffer_idxs_local[buf_size_idxs] = recv_buffer_idxs[buf_size_idxs]; 46179d30be91SStefano Zampini recv_buffer_idxs_local[buf_size_idxs+1] = recv_buffer_idxs[buf_size_idxs+1]; 46189d30be91SStefano Zampini buf_size_idxs += (PetscInt)olengths_idxs[i]; 46199d30be91SStefano Zampini } 46209d30be91SStefano Zampini ierr = PetscFree(recv_buffer_idxs);CHKERRQ(ierr); 46219d30be91SStefano Zampini 46229d30be91SStefano Zampini /* set preallocation */ 46239d30be91SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)local_mat,MATSEQDENSE,&newisdense);CHKERRQ(ierr); 46249d30be91SStefano Zampini if (!newisdense) { 46259d30be91SStefano Zampini PetscInt *new_local_nnz=0; 46269d30be91SStefano Zampini 46279d30be91SStefano Zampini ptr_vals = recv_buffer_vals; 46289d30be91SStefano Zampini ptr_idxs = recv_buffer_idxs_local; 46299d30be91SStefano Zampini if (n_recvs) { 46309d30be91SStefano Zampini ierr = PetscCalloc1(new_local_rows,&new_local_nnz);CHKERRQ(ierr); 46319d30be91SStefano Zampini } 46329d30be91SStefano Zampini for (i=0;i<n_recvs;i++) { 46339d30be91SStefano Zampini PetscInt j; 46349d30be91SStefano Zampini if (*ptr_idxs == (PetscInt)MATDENSE_PRIVATE) { /* preallocation provided for dense case only */ 46359d30be91SStefano Zampini for (j=0;j<*(ptr_idxs+1);j++) { 46369d30be91SStefano Zampini new_local_nnz[*(ptr_idxs+2+j)] += *(ptr_idxs+1); 46379d30be91SStefano Zampini } 46389d30be91SStefano Zampini } else { 46399d30be91SStefano Zampini /* TODO */ 46409d30be91SStefano Zampini } 46419d30be91SStefano Zampini ptr_idxs += olengths_idxs[i]; 46429d30be91SStefano Zampini } 46439d30be91SStefano Zampini if (new_local_nnz) { 46449d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] = PetscMin(new_local_nnz[i],new_local_rows); 46459d30be91SStefano Zampini ierr = MatSeqAIJSetPreallocation(local_mat,0,new_local_nnz);CHKERRQ(ierr); 46469d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] /= bs; 46479d30be91SStefano Zampini ierr = MatSeqBAIJSetPreallocation(local_mat,bs,0,new_local_nnz);CHKERRQ(ierr); 46489d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] = PetscMax(new_local_nnz[i]-i,0); 46499d30be91SStefano Zampini ierr = MatSeqSBAIJSetPreallocation(local_mat,bs,0,new_local_nnz);CHKERRQ(ierr); 46509d30be91SStefano Zampini } else { 46519d30be91SStefano Zampini ierr = MatSetUp(local_mat);CHKERRQ(ierr); 46529d30be91SStefano Zampini } 46539d30be91SStefano Zampini ierr = PetscFree(new_local_nnz);CHKERRQ(ierr); 46549d30be91SStefano Zampini } else { 46559d30be91SStefano Zampini ierr = MatSetUp(local_mat);CHKERRQ(ierr); 46569d30be91SStefano Zampini } 4657e7931f94SStefano Zampini 4658e7931f94SStefano Zampini /* set values */ 4659e7931f94SStefano Zampini ptr_vals = recv_buffer_vals; 46609d30be91SStefano Zampini ptr_idxs = recv_buffer_idxs_local; 4661e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 4662e7931f94SStefano Zampini if (*ptr_idxs == (PetscInt)MATDENSE_PRIVATE) { /* values insertion provided for dense case only */ 4663e7931f94SStefano Zampini ierr = MatSetOption(local_mat,MAT_ROW_ORIENTED,PETSC_FALSE);CHKERRQ(ierr); 46649d30be91SStefano Zampini ierr = MatSetValues(local_mat,*(ptr_idxs+1),ptr_idxs+2,*(ptr_idxs+1),ptr_idxs+2,ptr_vals,ADD_VALUES);CHKERRQ(ierr); 4665e7931f94SStefano Zampini ierr = MatAssemblyBegin(local_mat,MAT_FLUSH_ASSEMBLY);CHKERRQ(ierr); 4666e7931f94SStefano Zampini ierr = MatAssemblyEnd(local_mat,MAT_FLUSH_ASSEMBLY);CHKERRQ(ierr); 4667e7931f94SStefano Zampini ierr = MatSetOption(local_mat,MAT_ROW_ORIENTED,PETSC_TRUE);CHKERRQ(ierr); 466828143c3dSStefano Zampini } else { 466928143c3dSStefano Zampini /* TODO */ 4670e7931f94SStefano Zampini } 4671e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 4672e7931f94SStefano Zampini ptr_vals += olengths_vals[i]; 4673e7931f94SStefano Zampini } 4674e7931f94SStefano Zampini ierr = MatAssemblyBegin(local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4675e7931f94SStefano Zampini ierr = MatAssemblyEnd(local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 467670cf5478SStefano Zampini ierr = MatAssemblyBegin(*mat_n,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 467770cf5478SStefano Zampini ierr = MatAssemblyEnd(*mat_n,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 46789d30be91SStefano Zampini ierr = PetscFree(recv_buffer_vals);CHKERRQ(ierr); 46799d30be91SStefano Zampini ierr = PetscFree(recv_buffer_idxs_local);CHKERRQ(ierr); 4680e7931f94SStefano Zampini 4681dfd14d43SStefano Zampini #if 0 468228143c3dSStefano Zampini if (!restrict_comm) { /* check */ 4683e7931f94SStefano Zampini Vec lvec,rvec; 4684e7931f94SStefano Zampini PetscReal infty_error; 4685e7931f94SStefano Zampini 46862a7a6963SBarry Smith ierr = MatCreateVecs(mat,&rvec,&lvec);CHKERRQ(ierr); 4687e7931f94SStefano Zampini ierr = VecSetRandom(rvec,NULL);CHKERRQ(ierr); 4688e7931f94SStefano Zampini ierr = MatMult(mat,rvec,lvec);CHKERRQ(ierr); 4689e7931f94SStefano Zampini ierr = VecScale(lvec,-1.0);CHKERRQ(ierr); 469070cf5478SStefano Zampini ierr = MatMultAdd(*mat_n,rvec,lvec,lvec);CHKERRQ(ierr); 4691e7931f94SStefano Zampini ierr = VecNorm(lvec,NORM_INFINITY,&infty_error);CHKERRQ(ierr); 4692e7931f94SStefano Zampini ierr = PetscPrintf(PetscObjectComm((PetscObject)mat),"Infinity error subassembling %1.6e\n",infty_error); 4693e7931f94SStefano Zampini ierr = VecDestroy(&rvec);CHKERRQ(ierr); 4694e7931f94SStefano Zampini ierr = VecDestroy(&lvec);CHKERRQ(ierr); 4695e7931f94SStefano Zampini } 469628143c3dSStefano Zampini #endif 4697e7931f94SStefano Zampini 469828143c3dSStefano Zampini /* assemble new additional is (if any) */ 469928143c3dSStefano Zampini if (nis) { 470028143c3dSStefano Zampini PetscInt **temp_idxs,*count_is,j,psum; 470128143c3dSStefano Zampini 470228143c3dSStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_idxs_is,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 4703854ce69bSBarry Smith ierr = PetscCalloc1(nis,&count_is);CHKERRQ(ierr); 470428143c3dSStefano Zampini ptr_idxs = recv_buffer_idxs_is; 470528143c3dSStefano Zampini psum = 0; 470628143c3dSStefano Zampini for (i=0;i<n_recvs;i++) { 470728143c3dSStefano Zampini for (j=0;j<nis;j++) { 470828143c3dSStefano Zampini PetscInt plen = *(ptr_idxs); /* first element is the local size of IS's indices */ 470928143c3dSStefano Zampini count_is[j] += plen; /* increment counting of buffer for j-th IS */ 471028143c3dSStefano Zampini psum += plen; 471128143c3dSStefano Zampini ptr_idxs += plen+1; /* shift pointer to received data */ 471228143c3dSStefano Zampini } 471328143c3dSStefano Zampini } 4714854ce69bSBarry Smith ierr = PetscMalloc1(nis,&temp_idxs);CHKERRQ(ierr); 4715854ce69bSBarry Smith ierr = PetscMalloc1(psum,&temp_idxs[0]);CHKERRQ(ierr); 471628143c3dSStefano Zampini for (i=1;i<nis;i++) { 471728143c3dSStefano Zampini temp_idxs[i] = temp_idxs[i-1]+count_is[i-1]; 471828143c3dSStefano Zampini } 471928143c3dSStefano Zampini ierr = PetscMemzero(count_is,nis*sizeof(PetscInt));CHKERRQ(ierr); 472028143c3dSStefano Zampini ptr_idxs = recv_buffer_idxs_is; 472128143c3dSStefano Zampini for (i=0;i<n_recvs;i++) { 472228143c3dSStefano Zampini for (j=0;j<nis;j++) { 472328143c3dSStefano Zampini PetscInt plen = *(ptr_idxs); /* first element is the local size of IS's indices */ 472428143c3dSStefano Zampini ierr = PetscMemcpy(&temp_idxs[j][count_is[j]],ptr_idxs+1,plen*sizeof(PetscInt));CHKERRQ(ierr); 472528143c3dSStefano Zampini count_is[j] += plen; /* increment starting point of buffer for j-th IS */ 472628143c3dSStefano Zampini ptr_idxs += plen+1; /* shift pointer to received data */ 472728143c3dSStefano Zampini } 472828143c3dSStefano Zampini } 472928143c3dSStefano Zampini for (i=0;i<nis;i++) { 473028143c3dSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 473128143c3dSStefano Zampini ierr = PetscSortRemoveDupsInt(&count_is[i],temp_idxs[i]);CHKERRQ(ierr);CHKERRQ(ierr); 473228143c3dSStefano Zampini ierr = ISCreateGeneral(comm_n,count_is[i],temp_idxs[i],PETSC_COPY_VALUES,&isarray[i]);CHKERRQ(ierr); 473328143c3dSStefano Zampini } 473428143c3dSStefano Zampini ierr = PetscFree(count_is);CHKERRQ(ierr); 473528143c3dSStefano Zampini ierr = PetscFree(temp_idxs[0]);CHKERRQ(ierr); 473628143c3dSStefano Zampini ierr = PetscFree(temp_idxs);CHKERRQ(ierr); 473728143c3dSStefano Zampini } 4738e7931f94SStefano Zampini /* free workspace */ 473928143c3dSStefano Zampini ierr = PetscFree(recv_buffer_idxs_is);CHKERRQ(ierr); 4740e7931f94SStefano Zampini ierr = MPI_Waitall(n_sends,send_req_idxs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 4741e7931f94SStefano Zampini ierr = PetscFree(send_buffer_idxs);CHKERRQ(ierr); 4742e7931f94SStefano Zampini ierr = MPI_Waitall(n_sends,send_req_vals,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 4743e7931f94SStefano Zampini if (isdense) { 4744e7931f94SStefano Zampini ierr = MatISGetLocalMat(mat,&local_mat);CHKERRQ(ierr); 4745e7931f94SStefano Zampini ierr = MatDenseRestoreArray(local_mat,&send_buffer_vals);CHKERRQ(ierr); 4746e7931f94SStefano Zampini } else { 4747e7931f94SStefano Zampini /* ierr = PetscFree(send_buffer_vals);CHKERRQ(ierr); */ 4748e7931f94SStefano Zampini } 474928143c3dSStefano Zampini if (nis) { 475028143c3dSStefano Zampini ierr = MPI_Waitall(n_sends,send_req_idxs_is,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 475128143c3dSStefano Zampini ierr = PetscFree(send_buffer_idxs_is);CHKERRQ(ierr); 475228143c3dSStefano Zampini } 4753e7931f94SStefano Zampini ierr = PetscFree(recv_req_idxs);CHKERRQ(ierr); 4754e7931f94SStefano Zampini ierr = PetscFree(recv_req_vals);CHKERRQ(ierr); 475528143c3dSStefano Zampini ierr = PetscFree(recv_req_idxs_is);CHKERRQ(ierr); 4756e7931f94SStefano Zampini ierr = PetscFree(send_req_idxs);CHKERRQ(ierr); 4757e7931f94SStefano Zampini ierr = PetscFree(send_req_vals);CHKERRQ(ierr); 475828143c3dSStefano Zampini ierr = PetscFree(send_req_idxs_is);CHKERRQ(ierr); 4759e7931f94SStefano Zampini ierr = PetscFree(ilengths_vals);CHKERRQ(ierr); 4760e7931f94SStefano Zampini ierr = PetscFree(ilengths_idxs);CHKERRQ(ierr); 4761e7931f94SStefano Zampini ierr = PetscFree(olengths_vals);CHKERRQ(ierr); 4762e7931f94SStefano Zampini ierr = PetscFree(olengths_idxs);CHKERRQ(ierr); 4763e7931f94SStefano Zampini ierr = PetscFree(onodes);CHKERRQ(ierr); 476428143c3dSStefano Zampini if (nis) { 476528143c3dSStefano Zampini ierr = PetscFree(ilengths_idxs_is);CHKERRQ(ierr); 476628143c3dSStefano Zampini ierr = PetscFree(olengths_idxs_is);CHKERRQ(ierr); 476728143c3dSStefano Zampini ierr = PetscFree(onodes_is);CHKERRQ(ierr); 476828143c3dSStefano Zampini } 476928143c3dSStefano Zampini ierr = PetscSubcommDestroy(&subcomm);CHKERRQ(ierr); 477028143c3dSStefano Zampini if (destroy_mat) { /* destroy mat is true only if restrict comm is true and process will not partecipate */ 477128143c3dSStefano Zampini ierr = MatDestroy(mat_n);CHKERRQ(ierr); 477228143c3dSStefano Zampini for (i=0;i<nis;i++) { 477328143c3dSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 477428143c3dSStefano Zampini } 477553a05cb3SStefano Zampini *mat_n = NULL; 477628143c3dSStefano Zampini } 4777e7931f94SStefano Zampini PetscFunctionReturn(0); 4778e7931f94SStefano Zampini } 4779a57a6d2fSStefano Zampini 478012edc857SStefano Zampini /* temporary hack into ksp private data structure */ 4781af0996ceSBarry Smith #include <petsc/private/kspimpl.h> 478212edc857SStefano Zampini 4783c8587f34SStefano Zampini #undef __FUNCT__ 4784c8587f34SStefano Zampini #define __FUNCT__ "PCBDDCSetUpCoarseSolver" 4785c8587f34SStefano Zampini PetscErrorCode PCBDDCSetUpCoarseSolver(PC pc,PetscScalar* coarse_submat_vals) 4786c8587f34SStefano Zampini { 4787c8587f34SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 4788c8587f34SStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 478920a2ab83SStefano Zampini Mat coarse_mat,coarse_mat_is,coarse_submat_dense; 47909881197aSStefano Zampini MatNullSpace CoarseNullSpace=NULL; 479120a2ab83SStefano Zampini ISLocalToGlobalMapping coarse_islg; 47926e683305SStefano Zampini IS coarse_is,*isarray; 47936e683305SStefano Zampini PetscInt i,im_active=-1,active_procs=-1; 479430368db7SStefano Zampini PetscInt nis,nisdofs,nisneu,nisvert; 4795f9eb5b7dSStefano Zampini PC pc_temp; 4796c8587f34SStefano Zampini PCType coarse_pc_type; 4797c8587f34SStefano Zampini KSPType coarse_ksp_type; 4798f9eb5b7dSStefano Zampini PetscBool multilevel_requested,multilevel_allowed; 47994f3a063dSStefano Zampini PetscBool isredundant,isbddc,isnn,coarse_reuse; 48006e683305SStefano Zampini Mat t_coarse_mat_is; 48016e683305SStefano Zampini PetscInt void_procs,ncoarse_ml,ncoarse_ds,ncoarse; 48026e683305SStefano Zampini PetscMPIInt all_procs; 480374e2c79eSStefano Zampini PetscBool csin_ml,csin_ds,csin,csin_type_simple,redist; 480468457ee5SStefano Zampini PetscBool compute_vecs = PETSC_FALSE; 480522bc73bbSStefano Zampini PetscScalar *array; 48069881197aSStefano Zampini PetscErrorCode ierr; 4807fdc09c96SStefano Zampini 4808c8587f34SStefano Zampini PetscFunctionBegin; 4809c8587f34SStefano Zampini /* Assign global numbering to coarse dofs */ 481068457ee5SStefano 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 */ 4811fa7f1dd8SStefano Zampini PetscInt ocoarse_size; 48125a75c04eSSatish Balay compute_vecs = PETSC_TRUE; 4813fa7f1dd8SStefano Zampini ocoarse_size = pcbddc->coarse_size; 4814f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->global_primal_indices);CHKERRQ(ierr); 4815f4ddd8eeSStefano Zampini ierr = PCBDDCComputePrimalNumbering(pc,&pcbddc->coarse_size,&pcbddc->global_primal_indices);CHKERRQ(ierr); 4816f4ddd8eeSStefano Zampini /* see if we can avoid some work */ 4817fa7f1dd8SStefano Zampini if (pcbddc->coarse_ksp) { /* coarse ksp has already been created */ 481851bea450SStefano Zampini /* if the coarse size is different or we are using adaptive selection, better to not reuse the coarse matrix */ 481951bea450SStefano Zampini if (ocoarse_size != pcbddc->coarse_size || pcbddc->adaptive_selection) { 4820dc4bcba2SStefano Zampini PC pc; 4821dc4bcba2SStefano Zampini PetscBool isbddc; 4822dc4bcba2SStefano Zampini 4823dc4bcba2SStefano Zampini /* temporary workaround since PCBDDC does not have a reset method so far */ 4824dc4bcba2SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&pc);CHKERRQ(ierr); 4825dc4bcba2SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc,PCBDDC,&isbddc);CHKERRQ(ierr); 4826dc4bcba2SStefano Zampini if (isbddc) { 4827dc4bcba2SStefano Zampini ierr = PCDestroy(&pc);CHKERRQ(ierr); 4828dc4bcba2SStefano Zampini } 4829727cdba6SStefano Zampini ierr = KSPReset(pcbddc->coarse_ksp);CHKERRQ(ierr); 4830fa7f1dd8SStefano Zampini coarse_reuse = PETSC_FALSE; 4831fa7f1dd8SStefano Zampini } else { /* we can safely reuse already computed coarse matrix */ 4832fa7f1dd8SStefano Zampini coarse_reuse = PETSC_TRUE; 4833f4ddd8eeSStefano Zampini } 4834fa7f1dd8SStefano Zampini } else { /* there's no coarse ksp, so we need to create the coarse matrix too */ 4835fa7f1dd8SStefano Zampini coarse_reuse = PETSC_FALSE; 4836f4ddd8eeSStefano Zampini } 483770cf5478SStefano Zampini /* reset any subassembling information */ 483870cf5478SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling);CHKERRQ(ierr); 48396e683305SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 48406e683305SStefano Zampini } else { /* primal space is unchanged, so we can reuse coarse matrix */ 4841fa7f1dd8SStefano Zampini coarse_reuse = PETSC_TRUE; 4842f4ddd8eeSStefano Zampini } 4843c8587f34SStefano Zampini 48446e683305SStefano Zampini /* count "active" (i.e. with positive local size) and "void" processes */ 48452b510759SStefano Zampini im_active = !!(pcis->n); 48462b510759SStefano Zampini ierr = MPI_Allreduce(&im_active,&active_procs,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 48476e683305SStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)pc),&all_procs);CHKERRQ(ierr); 48486e683305SStefano Zampini void_procs = all_procs-active_procs; 48496e683305SStefano Zampini csin_type_simple = PETSC_TRUE; 485074e2c79eSStefano Zampini redist = PETSC_FALSE; 485122bc73bbSStefano Zampini if (pcbddc->current_level && void_procs) { 48526e683305SStefano Zampini csin_ml = PETSC_TRUE; 48536e683305SStefano Zampini ncoarse_ml = void_procs; 4854779c1cceSStefano Zampini /* it has no sense to redistribute on a set of processors larger than the number of active processes */ 4855779c1cceSStefano Zampini if (pcbddc->redistribute_coarse > 0 && pcbddc->redistribute_coarse < active_procs) { 48566e683305SStefano Zampini csin_ds = PETSC_TRUE; 485718a45a71SStefano Zampini ncoarse_ds = pcbddc->redistribute_coarse; 485818a45a71SStefano Zampini redist = PETSC_TRUE; 485918a45a71SStefano Zampini } else { 48606e683305SStefano Zampini csin_ds = PETSC_TRUE; 4861779c1cceSStefano Zampini ncoarse_ds = active_procs; 4862779c1cceSStefano Zampini redist = PETSC_TRUE; 486318a45a71SStefano Zampini } 48646e683305SStefano Zampini } else { 48656e683305SStefano Zampini csin_ml = PETSC_FALSE; 48666e683305SStefano Zampini ncoarse_ml = all_procs; 48676e683305SStefano Zampini if (void_procs) { 48686e683305SStefano Zampini csin_ds = PETSC_TRUE; 48696e683305SStefano Zampini ncoarse_ds = void_procs; 48706e683305SStefano Zampini csin_type_simple = PETSC_FALSE; 48716e683305SStefano Zampini } else { 4872779c1cceSStefano Zampini if (pcbddc->redistribute_coarse > 0 && pcbddc->redistribute_coarse < all_procs) { 487374e2c79eSStefano Zampini csin_ds = PETSC_TRUE; 487474e2c79eSStefano Zampini ncoarse_ds = pcbddc->redistribute_coarse; 487574e2c79eSStefano Zampini redist = PETSC_TRUE; 487674e2c79eSStefano Zampini } else { 48776e683305SStefano Zampini csin_ds = PETSC_FALSE; 48786e683305SStefano Zampini ncoarse_ds = all_procs; 48796e683305SStefano Zampini } 48806e683305SStefano Zampini } 488174e2c79eSStefano Zampini } 48826e683305SStefano Zampini 48836e683305SStefano Zampini /* 48846e683305SStefano Zampini test if we can go multilevel: three conditions must be satisfied: 48856e683305SStefano Zampini - we have not exceeded the number of levels requested 48866e683305SStefano Zampini - we can actually subassemble the active processes 48876e683305SStefano Zampini - we can find a suitable number of MPI processes where we can place the subassembled problem 48886e683305SStefano Zampini */ 48896e683305SStefano Zampini multilevel_allowed = PETSC_FALSE; 48906e683305SStefano Zampini multilevel_requested = PETSC_FALSE; 48916e683305SStefano Zampini if (pcbddc->current_level < pcbddc->max_levels) { 48926e683305SStefano Zampini multilevel_requested = PETSC_TRUE; 48936e683305SStefano Zampini if (active_procs/pcbddc->coarsening_ratio < 2 || ncoarse_ml/pcbddc->coarsening_ratio < 2) { 4894f9eb5b7dSStefano Zampini multilevel_allowed = PETSC_FALSE; 48952b510759SStefano Zampini } else { 4896f9eb5b7dSStefano Zampini multilevel_allowed = PETSC_TRUE; 4897c8587f34SStefano Zampini } 4898c8587f34SStefano Zampini } 48996e683305SStefano Zampini /* determine number of process partecipating to coarse solver */ 49006e683305SStefano Zampini if (multilevel_allowed) { 49016e683305SStefano Zampini ncoarse = ncoarse_ml; 49026e683305SStefano Zampini csin = csin_ml; 490358da7f69SStefano Zampini redist = PETSC_FALSE; 49046e683305SStefano Zampini } else { 49056e683305SStefano Zampini ncoarse = ncoarse_ds; 49066e683305SStefano Zampini csin = csin_ds; 49076e683305SStefano Zampini } 4908e7931f94SStefano Zampini 4909abbbba34SStefano Zampini /* creates temporary l2gmap and IS for coarse indexes */ 4910abbbba34SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),pcbddc->local_primal_size,pcbddc->global_primal_indices,PETSC_COPY_VALUES,&coarse_is);CHKERRQ(ierr); 4911abbbba34SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(coarse_is,&coarse_islg);CHKERRQ(ierr); 4912abbbba34SStefano Zampini 4913abbbba34SStefano Zampini /* creates temporary MATIS object for coarse matrix */ 491422bc73bbSStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,pcbddc->local_primal_size,pcbddc->local_primal_size,NULL,&coarse_submat_dense);CHKERRQ(ierr); 491522bc73bbSStefano Zampini ierr = MatDenseGetArray(coarse_submat_dense,&array);CHKERRQ(ierr); 491622bc73bbSStefano Zampini ierr = PetscMemcpy(array,coarse_submat_vals,sizeof(*coarse_submat_vals)*pcbddc->local_primal_size*pcbddc->local_primal_size);CHKERRQ(ierr); 491722bc73bbSStefano Zampini ierr = MatDenseRestoreArray(coarse_submat_dense,&array);CHKERRQ(ierr); 4918e176bc59SStefano 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); 49196e683305SStefano Zampini ierr = MatISSetLocalMat(t_coarse_mat_is,coarse_submat_dense);CHKERRQ(ierr); 49206e683305SStefano Zampini ierr = MatAssemblyBegin(t_coarse_mat_is,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 49216e683305SStefano Zampini ierr = MatAssemblyEnd(t_coarse_mat_is,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4922abbbba34SStefano Zampini ierr = MatDestroy(&coarse_submat_dense);CHKERRQ(ierr); 4923abbbba34SStefano Zampini 49246e683305SStefano Zampini /* compute dofs splitting and neumann boundaries for coarse dofs */ 492530368db7SStefano Zampini if (multilevel_allowed && (pcbddc->n_ISForDofsLocal || pcbddc->NeumannBoundariesLocal || (pcbddc->benign_saddle_point && pcbddc->user_primal_vertices_local))) { /* protects from unneded computations */ 49266e683305SStefano Zampini PetscInt *tidxs,*tidxs2,nout,tsize,i; 49276e683305SStefano Zampini const PetscInt *idxs; 49286e683305SStefano Zampini ISLocalToGlobalMapping tmap; 49296e683305SStefano Zampini 49306e683305SStefano Zampini /* create map between primal indices (in local representative ordering) and local primal numbering */ 49310be93d54SStefano Zampini ierr = ISLocalToGlobalMappingCreate(PETSC_COMM_SELF,1,pcbddc->local_primal_size,pcbddc->primal_indices_local_idxs,PETSC_COPY_VALUES,&tmap);CHKERRQ(ierr); 49326e683305SStefano Zampini /* allocate space for temporary storage */ 4933854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->local_primal_size,&tidxs);CHKERRQ(ierr); 4934854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->local_primal_size,&tidxs2);CHKERRQ(ierr); 49356e683305SStefano Zampini /* allocate for IS array */ 49366e683305SStefano Zampini nisdofs = pcbddc->n_ISForDofsLocal; 49376e683305SStefano Zampini nisneu = !!pcbddc->NeumannBoundariesLocal; 493830368db7SStefano Zampini nisvert = 0; 493930368db7SStefano Zampini if (pcbddc->benign_saddle_point && pcbddc->user_primal_vertices_local) { 494030368db7SStefano Zampini nisvert = 1; 494130368db7SStefano Zampini } 494230368db7SStefano Zampini nis = nisdofs + nisneu + nisvert; 4943854ce69bSBarry Smith ierr = PetscMalloc1(nis,&isarray);CHKERRQ(ierr); 49446e683305SStefano Zampini /* dofs splitting */ 49456e683305SStefano Zampini for (i=0;i<nisdofs;i++) { 49466e683305SStefano Zampini /* ierr = ISView(pcbddc->ISForDofsLocal[i],0);CHKERRQ(ierr); */ 49476e683305SStefano Zampini ierr = ISGetLocalSize(pcbddc->ISForDofsLocal[i],&tsize);CHKERRQ(ierr); 49486e683305SStefano Zampini ierr = ISGetIndices(pcbddc->ISForDofsLocal[i],&idxs);CHKERRQ(ierr); 49496e683305SStefano Zampini ierr = ISGlobalToLocalMappingApply(tmap,IS_GTOLM_DROP,tsize,idxs,&nout,tidxs);CHKERRQ(ierr); 49506e683305SStefano Zampini ierr = ISRestoreIndices(pcbddc->ISForDofsLocal[i],&idxs);CHKERRQ(ierr); 49516e683305SStefano Zampini ierr = ISLocalToGlobalMappingApply(coarse_islg,nout,tidxs,tidxs2);CHKERRQ(ierr); 495230368db7SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),nout,tidxs2,PETSC_COPY_VALUES,&isarray[i]);CHKERRQ(ierr); 49536e683305SStefano Zampini /* ierr = ISView(isarray[i],0);CHKERRQ(ierr); */ 49546e683305SStefano Zampini } 49556e683305SStefano Zampini /* neumann boundaries */ 49566e683305SStefano Zampini if (pcbddc->NeumannBoundariesLocal) { 49576e683305SStefano Zampini /* ierr = ISView(pcbddc->NeumannBoundariesLocal,0);CHKERRQ(ierr); */ 49586e683305SStefano Zampini ierr = ISGetLocalSize(pcbddc->NeumannBoundariesLocal,&tsize);CHKERRQ(ierr); 49596e683305SStefano Zampini ierr = ISGetIndices(pcbddc->NeumannBoundariesLocal,&idxs);CHKERRQ(ierr); 49606e683305SStefano Zampini ierr = ISGlobalToLocalMappingApply(tmap,IS_GTOLM_DROP,tsize,idxs,&nout,tidxs);CHKERRQ(ierr); 49616e683305SStefano Zampini ierr = ISRestoreIndices(pcbddc->NeumannBoundariesLocal,&idxs);CHKERRQ(ierr); 49626e683305SStefano Zampini ierr = ISLocalToGlobalMappingApply(coarse_islg,nout,tidxs,tidxs2);CHKERRQ(ierr); 496330368db7SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),nout,tidxs2,PETSC_COPY_VALUES,&isarray[nisdofs]);CHKERRQ(ierr); 49646e683305SStefano Zampini /* ierr = ISView(isarray[nisdofs],0);CHKERRQ(ierr); */ 49656e683305SStefano Zampini } 496630368db7SStefano Zampini /* primal vertices (benign) */ 496730368db7SStefano Zampini if (pcbddc->benign_saddle_point && pcbddc->user_primal_vertices_local) { 496830368db7SStefano Zampini ierr = ISGetLocalSize(pcbddc->user_primal_vertices_local,&tsize);CHKERRQ(ierr); 496930368db7SStefano Zampini ierr = ISGetIndices(pcbddc->user_primal_vertices_local,&idxs);CHKERRQ(ierr); 497030368db7SStefano Zampini ierr = ISGlobalToLocalMappingApply(tmap,IS_GTOLM_DROP,tsize,idxs,&nout,tidxs);CHKERRQ(ierr); 497130368db7SStefano Zampini ierr = ISRestoreIndices(pcbddc->user_primal_vertices_local,&idxs);CHKERRQ(ierr); 497230368db7SStefano Zampini ierr = ISLocalToGlobalMappingApply(coarse_islg,nout,tidxs,tidxs2);CHKERRQ(ierr); 497330368db7SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),nout,tidxs2,PETSC_COPY_VALUES,&isarray[nis-1]);CHKERRQ(ierr); 497430368db7SStefano Zampini } 49756e683305SStefano Zampini /* free memory */ 49766e683305SStefano Zampini ierr = PetscFree(tidxs);CHKERRQ(ierr); 49776e683305SStefano Zampini ierr = PetscFree(tidxs2);CHKERRQ(ierr); 49786e683305SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&tmap);CHKERRQ(ierr); 49796e683305SStefano Zampini } else { 49806e683305SStefano Zampini nis = 0; 49816e683305SStefano Zampini nisdofs = 0; 49826e683305SStefano Zampini nisneu = 0; 498330368db7SStefano Zampini nisvert = 0; 49846e683305SStefano Zampini isarray = NULL; 49856e683305SStefano Zampini } 49866e683305SStefano Zampini /* destroy no longer needed map */ 49876e683305SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&coarse_islg);CHKERRQ(ierr); 49886e683305SStefano Zampini 49896e683305SStefano Zampini /* restrict on coarse candidates (if needed) */ 49906e683305SStefano Zampini coarse_mat_is = NULL; 49916e683305SStefano Zampini if (csin) { 49926e683305SStefano Zampini if (!pcbddc->coarse_subassembling_init ) { /* creates subassembling init pattern if not present */ 499374e2c79eSStefano Zampini if (redist) { 499474e2c79eSStefano Zampini PetscMPIInt rank; 4995779c1cceSStefano Zampini PetscInt spc,n_spc_p1,dest[1],destsize; 499674e2c79eSStefano Zampini 499774e2c79eSStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)pc),&rank);CHKERRQ(ierr); 499858da7f69SStefano Zampini spc = active_procs/ncoarse; 499958da7f69SStefano Zampini n_spc_p1 = active_procs%ncoarse; 5000779c1cceSStefano Zampini if (im_active) { 5001779c1cceSStefano Zampini destsize = 1; 500274e2c79eSStefano Zampini if (rank > n_spc_p1*(spc+1)-1) { 500374e2c79eSStefano Zampini dest[0] = n_spc_p1+(rank-(n_spc_p1*(spc+1)))/spc; 500474e2c79eSStefano Zampini } else { 500574e2c79eSStefano Zampini dest[0] = rank/(spc+1); 500674e2c79eSStefano Zampini } 500774e2c79eSStefano Zampini } else { 5008779c1cceSStefano Zampini destsize = 0; 50096e683305SStefano Zampini } 5010779c1cceSStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),destsize,dest,PETSC_COPY_VALUES,&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 5011779c1cceSStefano Zampini } else if (csin_type_simple) { 50126e683305SStefano Zampini PetscMPIInt rank; 50136e683305SStefano Zampini PetscInt issize,isidx; 5014779c1cceSStefano Zampini 50156e683305SStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)pc),&rank);CHKERRQ(ierr); 50166e683305SStefano Zampini if (im_active) { 50176e683305SStefano Zampini issize = 1; 50186e683305SStefano Zampini isidx = (PetscInt)rank; 50196e683305SStefano Zampini } else { 50206e683305SStefano Zampini issize = 0; 50216e683305SStefano Zampini isidx = -1; 50226e683305SStefano Zampini } 50236e683305SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),issize,&isidx,PETSC_COPY_VALUES,&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 5024779c1cceSStefano Zampini } else { /* get a suitable subassembling pattern from MATIS code */ 5025b0c7d250SStefano Zampini ierr = MatISGetSubassemblingPattern(t_coarse_mat_is,ncoarse,pcbddc->coarse_adj_red,&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 50266e683305SStefano Zampini } 5027779c1cceSStefano Zampini 5028779c1cceSStefano Zampini /* we need to shift on coarse candidates either if we are not redistributing or we are redistributing and we have enough void processes */ 5029779c1cceSStefano Zampini if (!redist || ncoarse <= void_procs) { 5030779c1cceSStefano Zampini PetscInt ncoarse_cand,tissize,*nisindices; 5031779c1cceSStefano Zampini PetscInt *coarse_candidates; 5032779c1cceSStefano Zampini const PetscInt* tisindices; 5033779c1cceSStefano Zampini 5034779c1cceSStefano Zampini /* get coarse candidates' ranks in pc communicator */ 5035779c1cceSStefano Zampini ierr = PetscMalloc1(all_procs,&coarse_candidates);CHKERRQ(ierr); 5036779c1cceSStefano Zampini ierr = MPI_Allgather(&im_active,1,MPIU_INT,coarse_candidates,1,MPIU_INT,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 5037779c1cceSStefano Zampini for (i=0,ncoarse_cand=0;i<all_procs;i++) { 5038779c1cceSStefano Zampini if (!coarse_candidates[i]) { 5039779c1cceSStefano Zampini coarse_candidates[ncoarse_cand++]=i; 5040779c1cceSStefano Zampini } 5041779c1cceSStefano Zampini } 5042779c1cceSStefano Zampini if (ncoarse_cand < ncoarse) SETERRQ2(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"This should not happen! %d < %d",ncoarse_cand,ncoarse); 5043779c1cceSStefano Zampini 5044779c1cceSStefano Zampini 50456e683305SStefano Zampini if (pcbddc->dbg_flag) { 50466e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 50476e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Subassembling pattern init (before shift)\n");CHKERRQ(ierr); 50486e683305SStefano Zampini ierr = ISView(pcbddc->coarse_subassembling_init,pcbddc->dbg_viewer);CHKERRQ(ierr); 50496e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Coarse candidates\n");CHKERRQ(ierr); 5050779c1cceSStefano Zampini for (i=0;i<ncoarse_cand;i++) { 50516e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"%d ",coarse_candidates[i]);CHKERRQ(ierr); 50526e683305SStefano Zampini } 50536e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"\n");CHKERRQ(ierr); 50546e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 50556e683305SStefano Zampini } 50566e683305SStefano Zampini /* shift the pattern on coarse candidates */ 50576e683305SStefano Zampini ierr = ISGetLocalSize(pcbddc->coarse_subassembling_init,&tissize);CHKERRQ(ierr); 50586e683305SStefano Zampini ierr = ISGetIndices(pcbddc->coarse_subassembling_init,&tisindices);CHKERRQ(ierr); 5059854ce69bSBarry Smith ierr = PetscMalloc1(tissize,&nisindices);CHKERRQ(ierr); 50606e683305SStefano Zampini for (i=0;i<tissize;i++) nisindices[i] = coarse_candidates[tisindices[i]]; 50616e683305SStefano Zampini ierr = ISRestoreIndices(pcbddc->coarse_subassembling_init,&tisindices);CHKERRQ(ierr); 50626e683305SStefano Zampini ierr = ISGeneralSetIndices(pcbddc->coarse_subassembling_init,tissize,nisindices,PETSC_OWN_POINTER);CHKERRQ(ierr); 50636e683305SStefano Zampini ierr = PetscFree(coarse_candidates);CHKERRQ(ierr); 50646e683305SStefano Zampini } 50656e683305SStefano Zampini if (pcbddc->dbg_flag) { 50666e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 50676e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Subassembling pattern init\n");CHKERRQ(ierr); 50686e683305SStefano Zampini ierr = ISView(pcbddc->coarse_subassembling_init,pcbddc->dbg_viewer);CHKERRQ(ierr); 50696e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 50706e683305SStefano Zampini } 5071779c1cceSStefano Zampini } 50726e683305SStefano Zampini /* get temporary coarse mat in IS format restricted on coarse procs (plus additional index sets of isarray) */ 507353a05cb3SStefano Zampini if (multilevel_allowed) { /* we need to keep tracking of void processes for future placements */ 507453a05cb3SStefano 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); 507553a05cb3SStefano Zampini } else { /* this is the last level, so use just receiving processes in subcomm */ 507653a05cb3SStefano 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); 507753a05cb3SStefano Zampini } 50786e683305SStefano Zampini } else { 50796e683305SStefano Zampini if (pcbddc->dbg_flag) { 50806e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 50816e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Subassembling pattern init not needed\n");CHKERRQ(ierr); 50826e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 50836e683305SStefano Zampini } 50846e683305SStefano Zampini ierr = PetscObjectReference((PetscObject)t_coarse_mat_is);CHKERRQ(ierr); 50856e683305SStefano Zampini coarse_mat_is = t_coarse_mat_is; 50866e683305SStefano Zampini } 50876e683305SStefano Zampini 50886e683305SStefano Zampini /* create local to global scatters for coarse problem */ 508968457ee5SStefano Zampini if (compute_vecs) { 50906e683305SStefano Zampini PetscInt lrows; 50916e683305SStefano Zampini ierr = VecDestroy(&pcbddc->coarse_vec);CHKERRQ(ierr); 50926e683305SStefano Zampini if (coarse_mat_is) { 50936e683305SStefano Zampini ierr = MatGetLocalSize(coarse_mat_is,&lrows,NULL);CHKERRQ(ierr); 50946e683305SStefano Zampini } else { 50956e683305SStefano Zampini lrows = 0; 50966e683305SStefano Zampini } 50976e683305SStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pc),&pcbddc->coarse_vec);CHKERRQ(ierr); 50986e683305SStefano Zampini ierr = VecSetSizes(pcbddc->coarse_vec,lrows,PETSC_DECIDE);CHKERRQ(ierr); 50996e683305SStefano Zampini ierr = VecSetType(pcbddc->coarse_vec,VECSTANDARD);CHKERRQ(ierr); 51006e683305SStefano Zampini ierr = VecScatterDestroy(&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 51016e683305SStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_P,NULL,pcbddc->coarse_vec,coarse_is,&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 51026e683305SStefano Zampini } 51036e683305SStefano Zampini ierr = ISDestroy(&coarse_is);CHKERRQ(ierr); 51046e683305SStefano Zampini ierr = MatDestroy(&t_coarse_mat_is);CHKERRQ(ierr); 5105c8587f34SStefano Zampini 5106f9eb5b7dSStefano Zampini /* set defaults for coarse KSP and PC */ 5107f9eb5b7dSStefano Zampini if (multilevel_allowed) { 5108f9eb5b7dSStefano Zampini coarse_ksp_type = KSPRICHARDSON; 5109f9eb5b7dSStefano Zampini coarse_pc_type = PCBDDC; 5110f9eb5b7dSStefano Zampini } else { 5111f9eb5b7dSStefano Zampini coarse_ksp_type = KSPPREONLY; 5112f9eb5b7dSStefano Zampini coarse_pc_type = PCREDUNDANT; 5113c8587f34SStefano Zampini } 5114c8587f34SStefano Zampini 51156e683305SStefano Zampini /* print some info if requested */ 51166e683305SStefano Zampini if (pcbddc->dbg_flag) { 51176e683305SStefano Zampini if (!multilevel_allowed) { 51186e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 51196e683305SStefano Zampini if (multilevel_requested) { 51206e683305SStefano 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); 51216e683305SStefano Zampini } else if (pcbddc->max_levels) { 51226e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Maximum number of requested levels reached (%d)\n",pcbddc->max_levels);CHKERRQ(ierr); 51236e683305SStefano Zampini } 51246e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 51256e683305SStefano Zampini } 51266e683305SStefano Zampini } 51276e683305SStefano Zampini 5128f9eb5b7dSStefano Zampini /* create the coarse KSP object only once with defaults */ 51296e683305SStefano Zampini if (coarse_mat_is) { 51306e683305SStefano Zampini MatReuse coarse_mat_reuse; 51316a1308c2SStefano Zampini PetscViewer dbg_viewer = NULL; 51326e683305SStefano Zampini if (pcbddc->dbg_flag) { 51336e683305SStefano Zampini dbg_viewer = PETSC_VIEWER_STDOUT_(PetscObjectComm((PetscObject)coarse_mat_is)); 51346e683305SStefano Zampini ierr = PetscViewerASCIIAddTab(dbg_viewer,2*pcbddc->current_level);CHKERRQ(ierr); 51356e683305SStefano Zampini } 5136f9eb5b7dSStefano Zampini if (!pcbddc->coarse_ksp) { 5137312be037SStefano Zampini char prefix[256],str_level[16]; 5138e604994aSStefano Zampini size_t len; 51396e683305SStefano Zampini ierr = KSPCreate(PetscObjectComm((PetscObject)coarse_mat_is),&pcbddc->coarse_ksp);CHKERRQ(ierr); 5140422a814eSBarry Smith ierr = KSPSetErrorIfNotConverged(pcbddc->coarse_ksp,pc->erroriffailure);CHKERRQ(ierr); 5141c8587f34SStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->coarse_ksp,(PetscObject)pc,1);CHKERRQ(ierr); 5142f9eb5b7dSStefano Zampini ierr = KSPSetTolerances(pcbddc->coarse_ksp,PETSC_DEFAULT,PETSC_DEFAULT,PETSC_DEFAULT,1);CHKERRQ(ierr); 51435f76c7aeSStefano Zampini ierr = KSPSetOperators(pcbddc->coarse_ksp,coarse_mat_is,coarse_mat_is);CHKERRQ(ierr); 5144c8587f34SStefano Zampini ierr = KSPSetType(pcbddc->coarse_ksp,coarse_ksp_type);CHKERRQ(ierr); 51456e683305SStefano Zampini ierr = KSPSetNormType(pcbddc->coarse_ksp,KSP_NORM_NONE);CHKERRQ(ierr); 5146c8587f34SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&pc_temp);CHKERRQ(ierr); 5147c8587f34SStefano Zampini ierr = PCSetType(pc_temp,coarse_pc_type);CHKERRQ(ierr); 5148e604994aSStefano Zampini /* prefix */ 5149e604994aSStefano Zampini ierr = PetscStrcpy(prefix,"");CHKERRQ(ierr); 5150e604994aSStefano Zampini ierr = PetscStrcpy(str_level,"");CHKERRQ(ierr); 5151e604994aSStefano Zampini if (!pcbddc->current_level) { 5152e604994aSStefano Zampini ierr = PetscStrcpy(prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 5153e604994aSStefano Zampini ierr = PetscStrcat(prefix,"pc_bddc_coarse_");CHKERRQ(ierr); 5154c8587f34SStefano Zampini } else { 5155e604994aSStefano Zampini ierr = PetscStrlen(((PetscObject)pc)->prefix,&len);CHKERRQ(ierr); 5156312be037SStefano Zampini if (pcbddc->current_level>1) len -= 3; /* remove "lX_" with X level number */ 5157312be037SStefano Zampini if (pcbddc->current_level>10) len -= 1; /* remove another char from level number */ 515834d6797cSStefano Zampini ierr = PetscStrncpy(prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 5159312be037SStefano Zampini sprintf(str_level,"l%d_",(int)(pcbddc->current_level)); 5160e604994aSStefano Zampini ierr = PetscStrcat(prefix,str_level);CHKERRQ(ierr); 5161e604994aSStefano Zampini } 5162e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->coarse_ksp,prefix);CHKERRQ(ierr); 51633e3c6dadSStefano Zampini /* propagate BDDC info to the next level (these are dummy calls if pc_temp is not of type PCBDDC) */ 51643e3c6dadSStefano Zampini ierr = PCBDDCSetLevel(pc_temp,pcbddc->current_level+1);CHKERRQ(ierr); 51653e3c6dadSStefano Zampini ierr = PCBDDCSetCoarseningRatio(pc_temp,pcbddc->coarsening_ratio);CHKERRQ(ierr); 51663e3c6dadSStefano Zampini ierr = PCBDDCSetLevels(pc_temp,pcbddc->max_levels);CHKERRQ(ierr); 5167f9eb5b7dSStefano Zampini /* allow user customization */ 5168f9eb5b7dSStefano Zampini ierr = KSPSetFromOptions(pcbddc->coarse_ksp);CHKERRQ(ierr); 51693e3c6dadSStefano Zampini } 51703e3c6dadSStefano Zampini /* propagate BDDC info to the next level (these are dummy calls if pc_temp is not of type PCBDDC) */ 517151bea450SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&pc_temp);CHKERRQ(ierr); 51723e3c6dadSStefano Zampini if (nisdofs) { 51733e3c6dadSStefano Zampini ierr = PCBDDCSetDofsSplitting(pc_temp,nisdofs,isarray);CHKERRQ(ierr); 51743e3c6dadSStefano Zampini for (i=0;i<nisdofs;i++) { 51753e3c6dadSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 51763e3c6dadSStefano Zampini } 51773e3c6dadSStefano Zampini } 51783e3c6dadSStefano Zampini if (nisneu) { 51793e3c6dadSStefano Zampini ierr = PCBDDCSetNeumannBoundaries(pc_temp,isarray[nisdofs]);CHKERRQ(ierr); 51803e3c6dadSStefano Zampini ierr = ISDestroy(&isarray[nisdofs]);CHKERRQ(ierr); 5181312be037SStefano Zampini } 518230368db7SStefano Zampini if (nisvert) { 518330368db7SStefano Zampini ierr = PCBDDCSetPrimalVerticesIS(pc_temp,isarray[nis-1]);CHKERRQ(ierr); 518430368db7SStefano Zampini ierr = ISDestroy(&isarray[nis-1]);CHKERRQ(ierr); 518530368db7SStefano Zampini } 5186f9eb5b7dSStefano Zampini 5187f9eb5b7dSStefano Zampini /* get some info after set from options */ 5188f9eb5b7dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCNN,&isnn);CHKERRQ(ierr); 5189f9eb5b7dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCBDDC,&isbddc);CHKERRQ(ierr); 51904f3a063dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCREDUNDANT,&isredundant);CHKERRQ(ierr); 51916e683305SStefano Zampini if (isbddc && !multilevel_allowed) { /* multilevel can only be requested via pc_bddc_set_levels */ 5192f9eb5b7dSStefano Zampini ierr = PCSetType(pc_temp,coarse_pc_type);CHKERRQ(ierr); 5193f9eb5b7dSStefano Zampini isbddc = PETSC_FALSE; 5194f9eb5b7dSStefano Zampini } 519539f0f02cSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 51964f3a063dSStefano Zampini if (isredundant) { 51974f3a063dSStefano Zampini KSP inner_ksp; 51984f3a063dSStefano Zampini PC inner_pc; 51994f3a063dSStefano Zampini ierr = PCRedundantGetKSP(pc_temp,&inner_ksp);CHKERRQ(ierr); 52004f3a063dSStefano Zampini ierr = KSPGetPC(inner_ksp,&inner_pc);CHKERRQ(ierr); 52014f3a063dSStefano Zampini ierr = PCFactorSetReuseFill(inner_pc,PETSC_TRUE);CHKERRQ(ierr); 52024f3a063dSStefano Zampini } 5203f9eb5b7dSStefano Zampini 5204f9eb5b7dSStefano Zampini /* assemble coarse matrix */ 5205fa7f1dd8SStefano Zampini if (coarse_reuse) { 520681d9aea3SBarry Smith ierr = KSPGetOperators(pcbddc->coarse_ksp,&coarse_mat,NULL);CHKERRQ(ierr); 5207fa7f1dd8SStefano Zampini ierr = PetscObjectReference((PetscObject)coarse_mat);CHKERRQ(ierr); 52086e683305SStefano Zampini coarse_mat_reuse = MAT_REUSE_MATRIX; 5209fa7f1dd8SStefano Zampini } else { 52106e683305SStefano Zampini coarse_mat_reuse = MAT_INITIAL_MATRIX; 5211fa7f1dd8SStefano Zampini } 5212c8587f34SStefano Zampini if (isbddc || isnn) { 521322bc73bbSStefano Zampini if (pcbddc->coarsening_ratio > 1) { 521470cf5478SStefano Zampini if (!pcbddc->coarse_subassembling) { /* subassembling info is not present */ 5215b0c7d250SStefano Zampini ierr = MatISGetSubassemblingPattern(coarse_mat_is,active_procs/pcbddc->coarsening_ratio,pcbddc->coarse_adj_red,&pcbddc->coarse_subassembling);CHKERRQ(ierr); 521622b6e8a2SStefano Zampini if (pcbddc->dbg_flag) { 52176e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 52186e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Subassembling pattern\n");CHKERRQ(ierr); 52196e683305SStefano Zampini ierr = ISView(pcbddc->coarse_subassembling,dbg_viewer);CHKERRQ(ierr); 52206e683305SStefano Zampini ierr = PetscViewerFlush(dbg_viewer);CHKERRQ(ierr); 522122b6e8a2SStefano Zampini } 522270cf5478SStefano Zampini } 522353a05cb3SStefano Zampini ierr = MatISSubassemble(coarse_mat_is,pcbddc->coarse_subassembling,0,PETSC_FALSE,PETSC_FALSE,coarse_mat_reuse,&coarse_mat,0,NULL);CHKERRQ(ierr); 522470cf5478SStefano Zampini } else { 522522bc73bbSStefano Zampini ierr = PetscObjectReference((PetscObject)coarse_mat_is);CHKERRQ(ierr); 522622bc73bbSStefano Zampini coarse_mat = coarse_mat_is; 522722bc73bbSStefano Zampini } 522822bc73bbSStefano Zampini } else { 52292e1e5fa4SStefano Zampini ierr = MatISGetMPIXAIJ(coarse_mat_is,coarse_mat_reuse,&coarse_mat);CHKERRQ(ierr); 5230c8587f34SStefano Zampini } 5231c8587f34SStefano Zampini ierr = MatDestroy(&coarse_mat_is);CHKERRQ(ierr); 5232c8587f34SStefano Zampini 52333301b35fSStefano Zampini /* propagate symmetry info of coarse matrix */ 52345a16e3a0SStefano Zampini ierr = MatSetOption(coarse_mat,MAT_STRUCTURALLY_SYMMETRIC,PETSC_TRUE);CHKERRQ(ierr); 52353301b35fSStefano Zampini if (pc->pmat->symmetric_set) { 52363301b35fSStefano Zampini ierr = MatSetOption(coarse_mat,MAT_SYMMETRIC,pc->pmat->symmetric);CHKERRQ(ierr); 52373301b35fSStefano Zampini } 52383301b35fSStefano Zampini if (pc->pmat->hermitian_set) { 52393301b35fSStefano Zampini ierr = MatSetOption(coarse_mat,MAT_HERMITIAN,pc->pmat->hermitian);CHKERRQ(ierr); 52403301b35fSStefano Zampini } 52413301b35fSStefano Zampini if (pc->pmat->spd_set) { 52423301b35fSStefano Zampini ierr = MatSetOption(coarse_mat,MAT_SPD,pc->pmat->spd);CHKERRQ(ierr); 52433301b35fSStefano Zampini } 52446e683305SStefano Zampini /* set operators */ 52455f76c7aeSStefano Zampini ierr = KSPSetOperators(pcbddc->coarse_ksp,coarse_mat,coarse_mat);CHKERRQ(ierr); 52466e683305SStefano Zampini if (pcbddc->dbg_flag) { 52476e683305SStefano Zampini ierr = PetscViewerASCIISubtractTab(dbg_viewer,2*pcbddc->current_level);CHKERRQ(ierr); 52486e683305SStefano Zampini } 52496e683305SStefano Zampini } else { /* processes non partecipating to coarse solver (if any) */ 52506e683305SStefano Zampini coarse_mat = 0; 52516e683305SStefano Zampini } 52526e683305SStefano Zampini ierr = PetscFree(isarray);CHKERRQ(ierr); 5253b1ecc7b1SStefano Zampini #if 0 5254b9b85e73SStefano Zampini { 5255b9b85e73SStefano Zampini PetscViewer viewer; 5256b9b85e73SStefano Zampini char filename[256]; 5257b1ecc7b1SStefano Zampini sprintf(filename,"coarse_mat_level%d.m",pcbddc->current_level); 5258b1ecc7b1SStefano Zampini ierr = PetscViewerASCIIOpen(PetscObjectComm((PetscObject)coarse_mat),filename,&viewer);CHKERRQ(ierr); 5259b9b85e73SStefano Zampini ierr = PetscViewerSetFormat(viewer,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 5260b9b85e73SStefano Zampini ierr = MatView(coarse_mat,viewer);CHKERRQ(ierr); 5261b9b85e73SStefano Zampini ierr = PetscViewerDestroy(&viewer);CHKERRQ(ierr); 5262b9b85e73SStefano Zampini } 5263b9b85e73SStefano Zampini #endif 5264c8587f34SStefano Zampini 5265c8587f34SStefano Zampini /* Compute coarse null space (special handling by BDDC only) */ 5266298c0119SStefano Zampini #if 0 5267c8587f34SStefano Zampini if (pcbddc->NullSpace) { 5268c8587f34SStefano Zampini ierr = PCBDDCNullSpaceAssembleCoarse(pc,coarse_mat,&CoarseNullSpace);CHKERRQ(ierr); 526998a51de6SStefano Zampini } 5270298c0119SStefano Zampini #endif 5271b0f5fe93SStefano Zampini /* hack */ 527298a51de6SStefano Zampini if (pcbddc->coarse_ksp) { 527398a51de6SStefano Zampini Vec crhs,csol; 527404708bb6SStefano Zampini 5275f347579bSStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&csol);CHKERRQ(ierr); 5276f347579bSStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&crhs);CHKERRQ(ierr); 5277f347579bSStefano Zampini if (!csol) { 52782a7a6963SBarry Smith ierr = MatCreateVecs(coarse_mat,&((pcbddc->coarse_ksp)->vec_sol),NULL);CHKERRQ(ierr); 5279f9eb5b7dSStefano Zampini } 5280f347579bSStefano Zampini if (!crhs) { 52812a7a6963SBarry Smith ierr = MatCreateVecs(coarse_mat,NULL,&((pcbddc->coarse_ksp)->vec_rhs));CHKERRQ(ierr); 5282f347579bSStefano Zampini } 5283b0f5fe93SStefano Zampini } 5284b0f5fe93SStefano Zampini 5285b0f5fe93SStefano Zampini /* compute null space for coarse solver if the benign trick has been requested */ 5286b0f5fe93SStefano Zampini if (pcbddc->benign_null) { 5287b0f5fe93SStefano Zampini 5288b0f5fe93SStefano Zampini ierr = VecSet(pcbddc->vec1_P,0.);CHKERRQ(ierr); 52894f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 52904f1b2e48SStefano Zampini ierr = VecSetValue(pcbddc->vec1_P,pcbddc->local_primal_size-pcbddc->benign_n+i,1.0,INSERT_VALUES);CHKERRQ(ierr); 52914f1b2e48SStefano Zampini } 5292b0f5fe93SStefano Zampini ierr = VecAssemblyBegin(pcbddc->vec1_P);CHKERRQ(ierr); 5293b0f5fe93SStefano Zampini ierr = VecAssemblyEnd(pcbddc->vec1_P);CHKERRQ(ierr); 5294b0f5fe93SStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 5295b0f5fe93SStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 5296b0f5fe93SStefano Zampini if (coarse_mat) { 5297b0f5fe93SStefano Zampini Vec nullv; 5298b0f5fe93SStefano Zampini PetscScalar *array,*array2; 5299b0f5fe93SStefano Zampini PetscInt nl; 5300b0f5fe93SStefano Zampini 5301b0f5fe93SStefano Zampini ierr = MatCreateVecs(coarse_mat,&nullv,NULL);CHKERRQ(ierr); 5302b0f5fe93SStefano Zampini ierr = VecGetLocalSize(nullv,&nl);CHKERRQ(ierr); 5303b0f5fe93SStefano Zampini ierr = VecGetArrayRead(pcbddc->coarse_vec,(const PetscScalar**)&array);CHKERRQ(ierr); 5304b0f5fe93SStefano Zampini ierr = VecGetArray(nullv,&array2);CHKERRQ(ierr); 5305b0f5fe93SStefano Zampini ierr = PetscMemcpy(array2,array,nl*sizeof(*array));CHKERRQ(ierr); 5306b0f5fe93SStefano Zampini ierr = VecRestoreArray(nullv,&array2);CHKERRQ(ierr); 5307b0f5fe93SStefano Zampini ierr = VecRestoreArrayRead(pcbddc->coarse_vec,(const PetscScalar**)&array);CHKERRQ(ierr); 5308b0f5fe93SStefano Zampini ierr = VecNormalize(nullv,NULL);CHKERRQ(ierr); 5309b0f5fe93SStefano Zampini ierr = MatNullSpaceCreate(PetscObjectComm((PetscObject)coarse_mat),PETSC_FALSE,1,&nullv,&CoarseNullSpace);CHKERRQ(ierr); 5310b0f5fe93SStefano Zampini ierr = VecDestroy(&nullv);CHKERRQ(ierr); 5311b0f5fe93SStefano Zampini } 5312b0f5fe93SStefano Zampini } 5313b0f5fe93SStefano Zampini 5314b0f5fe93SStefano Zampini if (pcbddc->coarse_ksp) { 5315b0f5fe93SStefano Zampini PetscBool ispreonly; 5316b0f5fe93SStefano Zampini 5317b0f5fe93SStefano Zampini if (CoarseNullSpace) { 5318b0f5fe93SStefano Zampini PetscBool isnull; 5319b0f5fe93SStefano Zampini ierr = MatNullSpaceTest(CoarseNullSpace,coarse_mat,&isnull);CHKERRQ(ierr); 5320b0f5fe93SStefano Zampini if (isnull) { 532130368db7SStefano Zampini if (isbddc && !pcbddc->benign_saddle_point) { 5322b0f5fe93SStefano Zampini ierr = PCBDDCSetNullSpace(pc_temp,CoarseNullSpace);CHKERRQ(ierr); 5323b0f5fe93SStefano Zampini } else { 5324b0f5fe93SStefano Zampini ierr = MatSetNullSpace(coarse_mat,CoarseNullSpace);CHKERRQ(ierr); 5325b0f5fe93SStefano Zampini } 5326b0f5fe93SStefano Zampini } else { 5327b0f5fe93SStefano Zampini ierr = MatNullSpaceDestroy(&CoarseNullSpace);CHKERRQ(ierr); 5328b0f5fe93SStefano Zampini } 5329b0f5fe93SStefano Zampini } 5330b0f5fe93SStefano Zampini /* setup coarse ksp */ 5331b0f5fe93SStefano Zampini ierr = KSPSetUp(pcbddc->coarse_ksp);CHKERRQ(ierr); 5332cbcc2c2aSStefano Zampini /* Check coarse problem if in debug mode or if solving with an iterative method */ 5333cbcc2c2aSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcbddc->coarse_ksp,KSPPREONLY,&ispreonly);CHKERRQ(ierr); 53346e683305SStefano Zampini if (pcbddc->dbg_flag || (!ispreonly && pcbddc->use_coarse_estimates) ) { 5335c8587f34SStefano Zampini KSP check_ksp; 53362b510759SStefano Zampini KSPType check_ksp_type; 5337c8587f34SStefano Zampini PC check_pc; 53386e683305SStefano Zampini Vec check_vec,coarse_vec; 53396a1308c2SStefano Zampini PetscReal abs_infty_error,infty_error,lambda_min=1.0,lambda_max=1.0; 53402b510759SStefano Zampini PetscInt its; 53416e683305SStefano Zampini PetscBool compute_eigs; 53426e683305SStefano Zampini PetscReal *eigs_r,*eigs_c; 53436e683305SStefano Zampini PetscInt neigs; 53448e185a42SStefano Zampini const char *prefix; 5345c8587f34SStefano Zampini 53462b510759SStefano Zampini /* Create ksp object suitable for estimation of extreme eigenvalues */ 53476e683305SStefano Zampini ierr = KSPCreate(PetscObjectComm((PetscObject)pcbddc->coarse_ksp),&check_ksp);CHKERRQ(ierr); 5348422a814eSBarry Smith ierr = KSPSetErrorIfNotConverged(pcbddc->coarse_ksp,pc->erroriffailure);CHKERRQ(ierr); 534923ee1639SBarry Smith ierr = KSPSetOperators(check_ksp,coarse_mat,coarse_mat);CHKERRQ(ierr); 5350f4ddd8eeSStefano Zampini ierr = KSPSetTolerances(check_ksp,1.e-12,1.e-12,PETSC_DEFAULT,pcbddc->coarse_size);CHKERRQ(ierr); 53512b510759SStefano Zampini if (ispreonly) { 53522b510759SStefano Zampini check_ksp_type = KSPPREONLY; 53536e683305SStefano Zampini compute_eigs = PETSC_FALSE; 53542b510759SStefano Zampini } else { 5355cbcc2c2aSStefano Zampini check_ksp_type = KSPGMRES; 53566e683305SStefano Zampini compute_eigs = PETSC_TRUE; 5357c8587f34SStefano Zampini } 5358c8587f34SStefano Zampini ierr = KSPSetType(check_ksp,check_ksp_type);CHKERRQ(ierr); 53596e683305SStefano Zampini ierr = KSPSetComputeSingularValues(check_ksp,compute_eigs);CHKERRQ(ierr); 53606e683305SStefano Zampini ierr = KSPSetComputeEigenvalues(check_ksp,compute_eigs);CHKERRQ(ierr); 53616e683305SStefano Zampini ierr = KSPGMRESSetRestart(check_ksp,pcbddc->coarse_size+1);CHKERRQ(ierr); 5362a7dc3881SStefano Zampini ierr = KSPGetOptionsPrefix(pcbddc->coarse_ksp,&prefix);CHKERRQ(ierr); 5363a7dc3881SStefano Zampini ierr = KSPSetOptionsPrefix(check_ksp,prefix);CHKERRQ(ierr); 5364a7dc3881SStefano Zampini ierr = KSPAppendOptionsPrefix(check_ksp,"check_");CHKERRQ(ierr); 5365a7dc3881SStefano Zampini ierr = KSPSetFromOptions(check_ksp);CHKERRQ(ierr); 5366c8587f34SStefano Zampini ierr = KSPSetUp(check_ksp);CHKERRQ(ierr); 5367c8587f34SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&check_pc);CHKERRQ(ierr); 5368c8587f34SStefano Zampini ierr = KSPSetPC(check_ksp,check_pc);CHKERRQ(ierr); 5369c8587f34SStefano Zampini /* create random vec */ 53706e683305SStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&coarse_vec);CHKERRQ(ierr); 53716e683305SStefano Zampini ierr = VecDuplicate(coarse_vec,&check_vec);CHKERRQ(ierr); 5372c8587f34SStefano Zampini ierr = VecSetRandom(check_vec,NULL);CHKERRQ(ierr); 5373c8587f34SStefano Zampini if (CoarseNullSpace) { 5374c8587f34SStefano Zampini ierr = MatNullSpaceRemove(CoarseNullSpace,check_vec);CHKERRQ(ierr); 5375c8587f34SStefano Zampini } 53766e683305SStefano Zampini ierr = MatMult(coarse_mat,check_vec,coarse_vec);CHKERRQ(ierr); 5377c8587f34SStefano Zampini /* solve coarse problem */ 53786e683305SStefano Zampini ierr = KSPSolve(check_ksp,coarse_vec,coarse_vec);CHKERRQ(ierr); 5379c8587f34SStefano Zampini if (CoarseNullSpace) { 53806e683305SStefano Zampini ierr = MatNullSpaceRemove(CoarseNullSpace,coarse_vec);CHKERRQ(ierr); 5381c8587f34SStefano Zampini } 5382cbcc2c2aSStefano Zampini /* set eigenvalue estimation if preonly has not been requested */ 53836e683305SStefano Zampini if (compute_eigs) { 5384854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->coarse_size+1,&eigs_r);CHKERRQ(ierr); 5385854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->coarse_size+1,&eigs_c);CHKERRQ(ierr); 53866e683305SStefano Zampini ierr = KSPComputeEigenvalues(check_ksp,pcbddc->coarse_size+1,eigs_r,eigs_c,&neigs);CHKERRQ(ierr); 53876e683305SStefano Zampini lambda_max = eigs_r[neigs-1]; 53886e683305SStefano Zampini lambda_min = eigs_r[0]; 53896e683305SStefano Zampini if (pcbddc->use_coarse_estimates) { 53906e683305SStefano Zampini if (lambda_max>lambda_min) { 5391cbcc2c2aSStefano Zampini ierr = KSPChebyshevSetEigenvalues(pcbddc->coarse_ksp,lambda_max,lambda_min);CHKERRQ(ierr); 5392cbcc2c2aSStefano Zampini ierr = KSPRichardsonSetScale(pcbddc->coarse_ksp,2.0/(lambda_max+lambda_min));CHKERRQ(ierr); 5393cbcc2c2aSStefano Zampini } 5394c8587f34SStefano Zampini } 5395c8587f34SStefano Zampini } 5396cbcc2c2aSStefano Zampini 5397c8587f34SStefano Zampini /* check coarse problem residual error */ 53986e683305SStefano Zampini if (pcbddc->dbg_flag) { 53996e683305SStefano Zampini PetscViewer dbg_viewer = PETSC_VIEWER_STDOUT_(PetscObjectComm((PetscObject)pcbddc->coarse_ksp)); 54006e683305SStefano Zampini ierr = PetscViewerASCIIAddTab(dbg_viewer,2*(pcbddc->current_level+1));CHKERRQ(ierr); 54016e683305SStefano Zampini ierr = VecAXPY(check_vec,-1.0,coarse_vec);CHKERRQ(ierr); 5402c8587f34SStefano Zampini ierr = VecNorm(check_vec,NORM_INFINITY,&infty_error);CHKERRQ(ierr); 54036e683305SStefano Zampini ierr = MatMult(coarse_mat,check_vec,coarse_vec);CHKERRQ(ierr); 54046e683305SStefano Zampini ierr = VecNorm(coarse_vec,NORM_INFINITY,&abs_infty_error);CHKERRQ(ierr); 5405c8587f34SStefano Zampini ierr = VecDestroy(&check_vec);CHKERRQ(ierr); 5406779c1cceSStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem details (use estimates %d)\n",pcbddc->use_coarse_estimates);CHKERRQ(ierr); 54076e683305SStefano Zampini ierr = PetscObjectPrintClassNamePrefixType((PetscObject)(pcbddc->coarse_ksp),dbg_viewer);CHKERRQ(ierr); 54086e683305SStefano Zampini ierr = PetscObjectPrintClassNamePrefixType((PetscObject)(check_pc),dbg_viewer);CHKERRQ(ierr); 54096e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem exact infty_error : %1.6e\n",infty_error);CHKERRQ(ierr); 54106e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem residual infty_error: %1.6e\n",abs_infty_error);CHKERRQ(ierr); 5411b0f5fe93SStefano Zampini if (CoarseNullSpace) { 5412b0f5fe93SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem is singular\n");CHKERRQ(ierr); 5413b0f5fe93SStefano Zampini } 54146e683305SStefano Zampini if (compute_eigs) { 54156e683305SStefano Zampini PetscReal lambda_max_s,lambda_min_s; 5416deec49d1SStefano Zampini ierr = KSPGetType(check_ksp,&check_ksp_type);CHKERRQ(ierr); 5417c8587f34SStefano Zampini ierr = KSPGetIterationNumber(check_ksp,&its);CHKERRQ(ierr); 54186e683305SStefano Zampini ierr = KSPComputeExtremeSingularValues(check_ksp,&lambda_max_s,&lambda_min_s);CHKERRQ(ierr); 54196e683305SStefano 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); 54206e683305SStefano Zampini for (i=0;i<neigs;i++) { 54216e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"%1.6e %1.6ei\n",eigs_r[i],eigs_c[i]);CHKERRQ(ierr); 5422c8587f34SStefano Zampini } 54236e683305SStefano Zampini } 54246e683305SStefano Zampini ierr = PetscViewerFlush(dbg_viewer);CHKERRQ(ierr); 54256e683305SStefano Zampini ierr = PetscViewerASCIISubtractTab(dbg_viewer,2*(pcbddc->current_level+1));CHKERRQ(ierr); 54266e683305SStefano Zampini } 5427c8587f34SStefano Zampini ierr = KSPDestroy(&check_ksp);CHKERRQ(ierr); 54286e683305SStefano Zampini if (compute_eigs) { 54296e683305SStefano Zampini ierr = PetscFree(eigs_r);CHKERRQ(ierr); 54306e683305SStefano Zampini ierr = PetscFree(eigs_c);CHKERRQ(ierr); 5431c8587f34SStefano Zampini } 54326e683305SStefano Zampini } 54336e683305SStefano Zampini } 5434cbcc2c2aSStefano Zampini /* print additional info */ 5435cbcc2c2aSStefano Zampini if (pcbddc->dbg_flag) { 54366e683305SStefano Zampini /* waits until all processes reaches this point */ 54376e683305SStefano Zampini ierr = PetscBarrier((PetscObject)pc);CHKERRQ(ierr); 5438cbcc2c2aSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Coarse solver setup completed at level %d\n",pcbddc->current_level);CHKERRQ(ierr); 5439cbcc2c2aSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 5440cbcc2c2aSStefano Zampini } 5441cbcc2c2aSStefano Zampini 54422b510759SStefano Zampini /* free memory */ 5443c8587f34SStefano Zampini ierr = MatNullSpaceDestroy(&CoarseNullSpace);CHKERRQ(ierr); 5444fdc635d7SStefano Zampini ierr = MatDestroy(&coarse_mat);CHKERRQ(ierr); 5445c8587f34SStefano Zampini PetscFunctionReturn(0); 5446c8587f34SStefano Zampini } 5447674ae819SStefano Zampini 5448f34684f1SStefano Zampini #undef __FUNCT__ 5449f34684f1SStefano Zampini #define __FUNCT__ "PCBDDCComputePrimalNumbering" 5450f34684f1SStefano Zampini PetscErrorCode PCBDDCComputePrimalNumbering(PC pc,PetscInt* coarse_size_n,PetscInt** local_primal_indices_n) 5451f34684f1SStefano Zampini { 5452f34684f1SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 5453f34684f1SStefano Zampini PC_IS* pcis = (PC_IS*)pc->data; 5454f34684f1SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 5455dc456d91SStefano Zampini IS subset,subset_mult,subset_n; 5456dc456d91SStefano Zampini PetscInt local_size,coarse_size=0; 545773be2a3aSStefano Zampini PetscInt *local_primal_indices=NULL; 5458dc456d91SStefano Zampini const PetscInt *t_local_primal_indices; 5459f34684f1SStefano Zampini PetscErrorCode ierr; 5460f34684f1SStefano Zampini 5461f34684f1SStefano Zampini PetscFunctionBegin; 5462f34684f1SStefano Zampini /* Compute global number of coarse dofs */ 54630e6343abSStefano Zampini if (pcbddc->local_primal_size && !pcbddc->local_primal_ref_node) { 54640e6343abSStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"BDDC ConstraintsSetUp should be called first"); 5465727cdba6SStefano Zampini } 5466dc456d91SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)(pc->pmat)),pcbddc->local_primal_size_cc,pcbddc->local_primal_ref_node,PETSC_COPY_VALUES,&subset_n);CHKERRQ(ierr); 54673bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(pcis->mapping,subset_n,&subset);CHKERRQ(ierr); 5468dc456d91SStefano Zampini ierr = ISDestroy(&subset_n);CHKERRQ(ierr); 5469dc456d91SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)(pc->pmat)),pcbddc->local_primal_size_cc,pcbddc->local_primal_ref_mult,PETSC_COPY_VALUES,&subset_mult);CHKERRQ(ierr); 5470dc456d91SStefano Zampini ierr = PCBDDCSubsetNumbering(subset,subset_mult,&coarse_size,&subset_n);CHKERRQ(ierr); 5471dc456d91SStefano Zampini ierr = ISDestroy(&subset);CHKERRQ(ierr); 5472dc456d91SStefano Zampini ierr = ISDestroy(&subset_mult);CHKERRQ(ierr); 5473dc456d91SStefano Zampini ierr = ISGetLocalSize(subset_n,&local_size);CHKERRQ(ierr); 54740e6343abSStefano Zampini if (local_size != pcbddc->local_primal_size) { 54750e6343abSStefano Zampini SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Invalid number of local primal indices computed %d != %d",local_size,pcbddc->local_primal_size); 54760e6343abSStefano Zampini } 5477dc456d91SStefano Zampini ierr = PetscMalloc1(local_size,&local_primal_indices);CHKERRQ(ierr); 5478dc456d91SStefano Zampini ierr = ISGetIndices(subset_n,&t_local_primal_indices);CHKERRQ(ierr); 5479dc456d91SStefano Zampini ierr = PetscMemcpy(local_primal_indices,t_local_primal_indices,local_size*sizeof(PetscInt));CHKERRQ(ierr); 5480dc456d91SStefano Zampini ierr = ISRestoreIndices(subset_n,&t_local_primal_indices);CHKERRQ(ierr); 5481dc456d91SStefano Zampini ierr = ISDestroy(&subset_n);CHKERRQ(ierr); 5482f34684f1SStefano Zampini 5483f34684f1SStefano Zampini /* check numbering */ 5484f34684f1SStefano Zampini if (pcbddc->dbg_flag) { 5485019a44ceSStefano Zampini PetscScalar coarsesum,*array,*array2; 5486dc456d91SStefano Zampini PetscInt i; 5487b9b85e73SStefano Zampini PetscBool set_error = PETSC_FALSE,set_error_reduced = PETSC_FALSE; 5488f34684f1SStefano Zampini 5489f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 5490f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 5491f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Check coarse indices\n");CHKERRQ(ierr); 54921575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 5493019a44ceSStefano Zampini /* counter */ 5494019a44ceSStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 5495019a44ceSStefano Zampini ierr = VecSet(pcis->vec1_N,1.0);CHKERRQ(ierr); 5496019a44ceSStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 5497019a44ceSStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 5498019a44ceSStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_global,pcis->vec2_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 5499019a44ceSStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_global,pcis->vec2_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 5500f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_N,0.0);CHKERRQ(ierr); 5501f34684f1SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 5502727cdba6SStefano Zampini ierr = VecSetValue(pcis->vec1_N,pcbddc->primal_indices_local_idxs[i],1.0,INSERT_VALUES);CHKERRQ(ierr); 5503f34684f1SStefano Zampini } 5504f34684f1SStefano Zampini ierr = VecAssemblyBegin(pcis->vec1_N);CHKERRQ(ierr); 5505f34684f1SStefano Zampini ierr = VecAssemblyEnd(pcis->vec1_N);CHKERRQ(ierr); 5506f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 5507e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 5508e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 5509e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_global,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 5510e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_global,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 5511f34684f1SStefano Zampini ierr = VecGetArray(pcis->vec1_N,&array);CHKERRQ(ierr); 5512019a44ceSStefano Zampini ierr = VecGetArray(pcis->vec2_N,&array2);CHKERRQ(ierr); 5513f34684f1SStefano Zampini for (i=0;i<pcis->n;i++) { 5514019a44ceSStefano Zampini if (array[i] != 0.0 && array[i] != array2[i]) { 551575c01103SStefano Zampini PetscInt owned = (PetscInt)PetscRealPart(array[i]); 551675c01103SStefano Zampini PetscInt neigh = (PetscInt)PetscRealPart(array2[i]); 5517b9b85e73SStefano Zampini set_error = PETSC_TRUE; 5518019a44ceSStefano 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); 5519f34684f1SStefano Zampini } 5520f34684f1SStefano Zampini } 5521019a44ceSStefano Zampini ierr = VecRestoreArray(pcis->vec2_N,&array2);CHKERRQ(ierr); 5522b9b85e73SStefano Zampini ierr = MPI_Allreduce(&set_error,&set_error_reduced,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 5523f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 5524f34684f1SStefano Zampini for (i=0;i<pcis->n;i++) { 5525f34684f1SStefano Zampini if (PetscRealPart(array[i]) > 0.0) array[i] = 1.0/PetscRealPart(array[i]); 5526f34684f1SStefano Zampini } 5527f34684f1SStefano Zampini ierr = VecRestoreArray(pcis->vec1_N,&array);CHKERRQ(ierr); 5528f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 5529e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 5530e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 5531f34684f1SStefano Zampini ierr = VecSum(pcis->vec1_global,&coarsesum);CHKERRQ(ierr); 5532f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Size of coarse problem is %d (%lf)\n",coarse_size,PetscRealPart(coarsesum));CHKERRQ(ierr); 5533b9b85e73SStefano Zampini if (pcbddc->dbg_flag > 1 || set_error_reduced) { 5534ca8b9ea9SStefano Zampini PetscInt *gidxs; 5535ca8b9ea9SStefano Zampini 5536ca8b9ea9SStefano Zampini ierr = PetscMalloc1(pcbddc->local_primal_size,&gidxs);CHKERRQ(ierr); 55373bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingApply(pcis->mapping,pcbddc->local_primal_size,pcbddc->primal_indices_local_idxs,gidxs);CHKERRQ(ierr); 5538f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Distribution of local primal indices\n");CHKERRQ(ierr); 5539f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 5540f34684f1SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d\n",PetscGlobalRank);CHKERRQ(ierr); 5541f34684f1SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 55424bc2dc4bSStefano 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); 5543f34684f1SStefano Zampini } 5544f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 5545ca8b9ea9SStefano Zampini ierr = PetscFree(gidxs);CHKERRQ(ierr); 5546f34684f1SStefano Zampini } 5547f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 55481575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 5549302440fdSBarry Smith if (set_error_reduced) SETERRQ(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"BDDC Numbering of coarse dofs failed"); 5550f34684f1SStefano Zampini } 55518bec7fa6SStefano Zampini /* ierr = PetscPrintf(PetscObjectComm((PetscObject)pc),"Size of coarse problem is %d\n",coarse_size);CHKERRQ(ierr); */ 5552f34684f1SStefano Zampini /* get back data */ 5553f34684f1SStefano Zampini *coarse_size_n = coarse_size; 5554f34684f1SStefano Zampini *local_primal_indices_n = local_primal_indices; 5555674ae819SStefano Zampini PetscFunctionReturn(0); 5556674ae819SStefano Zampini } 5557674ae819SStefano Zampini 5558e456f2a8SStefano Zampini #undef __FUNCT__ 5559e456f2a8SStefano Zampini #define __FUNCT__ "PCBDDCGlobalToLocal" 5560a7dc3881SStefano Zampini PetscErrorCode PCBDDCGlobalToLocal(VecScatter g2l_ctx,Vec gwork, Vec lwork, IS globalis, IS* localis) 5561e456f2a8SStefano Zampini { 5562e456f2a8SStefano Zampini IS localis_t; 5563a7dc3881SStefano Zampini PetscInt i,lsize,*idxs,n; 5564e456f2a8SStefano Zampini PetscScalar *vals; 5565e456f2a8SStefano Zampini PetscErrorCode ierr; 5566e456f2a8SStefano Zampini 5567e456f2a8SStefano Zampini PetscFunctionBegin; 5568a7dc3881SStefano Zampini /* get indices in local ordering exploiting local to global map */ 5569e456f2a8SStefano Zampini ierr = ISGetLocalSize(globalis,&lsize);CHKERRQ(ierr); 5570854ce69bSBarry Smith ierr = PetscMalloc1(lsize,&vals);CHKERRQ(ierr); 5571e456f2a8SStefano Zampini for (i=0;i<lsize;i++) vals[i] = 1.0; 5572e456f2a8SStefano Zampini ierr = ISGetIndices(globalis,(const PetscInt**)&idxs);CHKERRQ(ierr); 5573a7dc3881SStefano Zampini ierr = VecSet(gwork,0.0);CHKERRQ(ierr); 5574a7dc3881SStefano Zampini ierr = VecSet(lwork,0.0);CHKERRQ(ierr); 55751035eff8SStefano Zampini if (idxs) { /* multilevel guard */ 5576a7dc3881SStefano Zampini ierr = VecSetValues(gwork,lsize,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 55771035eff8SStefano Zampini } 5578a7dc3881SStefano Zampini ierr = VecAssemblyBegin(gwork);CHKERRQ(ierr); 5579e456f2a8SStefano Zampini ierr = ISRestoreIndices(globalis,(const PetscInt**)&idxs);CHKERRQ(ierr); 5580e456f2a8SStefano Zampini ierr = PetscFree(vals);CHKERRQ(ierr); 5581a7dc3881SStefano Zampini ierr = VecAssemblyEnd(gwork);CHKERRQ(ierr); 5582a7dc3881SStefano Zampini /* now compute set in local ordering */ 5583a7dc3881SStefano Zampini ierr = VecScatterBegin(g2l_ctx,gwork,lwork,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 5584a7dc3881SStefano Zampini ierr = VecScatterEnd(g2l_ctx,gwork,lwork,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 5585a7dc3881SStefano Zampini ierr = VecGetArrayRead(lwork,(const PetscScalar**)&vals);CHKERRQ(ierr); 5586a7dc3881SStefano Zampini ierr = VecGetSize(lwork,&n);CHKERRQ(ierr); 5587a7dc3881SStefano Zampini for (i=0,lsize=0;i<n;i++) { 5588ff92baa0SMatthew G. Knepley if (PetscRealPart(vals[i]) > 0.5) { 5589e456f2a8SStefano Zampini lsize++; 5590e456f2a8SStefano Zampini } 5591e456f2a8SStefano Zampini } 5592854ce69bSBarry Smith ierr = PetscMalloc1(lsize,&idxs);CHKERRQ(ierr); 5593a7dc3881SStefano Zampini for (i=0,lsize=0;i<n;i++) { 5594ff92baa0SMatthew G. Knepley if (PetscRealPart(vals[i]) > 0.5) { 5595e456f2a8SStefano Zampini idxs[lsize++] = i; 5596e456f2a8SStefano Zampini } 5597e456f2a8SStefano Zampini } 5598a7dc3881SStefano Zampini ierr = VecRestoreArrayRead(lwork,(const PetscScalar**)&vals);CHKERRQ(ierr); 5599a7dc3881SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)gwork),lsize,idxs,PETSC_OWN_POINTER,&localis_t);CHKERRQ(ierr); 5600e456f2a8SStefano Zampini *localis = localis_t; 5601e456f2a8SStefano Zampini PetscFunctionReturn(0); 5602e456f2a8SStefano Zampini } 5603906d46d4SStefano Zampini 5604906d46d4SStefano Zampini /* the next two functions will be called in KSPMatMult if a change of basis has been requested */ 5605906d46d4SStefano Zampini #undef __FUNCT__ 5606906d46d4SStefano Zampini #define __FUNCT__ "PCBDDCMatMult_Private" 5607906d46d4SStefano Zampini static PetscErrorCode PCBDDCMatMult_Private(Mat A, Vec x, Vec y) 5608906d46d4SStefano Zampini { 5609906d46d4SStefano Zampini PCBDDCChange_ctx change_ctx; 5610906d46d4SStefano Zampini PetscErrorCode ierr; 5611906d46d4SStefano Zampini 5612906d46d4SStefano Zampini PetscFunctionBegin; 5613906d46d4SStefano Zampini ierr = MatShellGetContext(A,&change_ctx);CHKERRQ(ierr); 5614906d46d4SStefano Zampini ierr = MatMult(change_ctx->global_change,x,change_ctx->work[0]);CHKERRQ(ierr); 5615906d46d4SStefano Zampini ierr = MatMult(change_ctx->original_mat,change_ctx->work[0],change_ctx->work[1]);CHKERRQ(ierr); 5616906d46d4SStefano Zampini ierr = MatMultTranspose(change_ctx->global_change,change_ctx->work[1],y);CHKERRQ(ierr); 5617906d46d4SStefano Zampini PetscFunctionReturn(0); 5618906d46d4SStefano Zampini } 5619906d46d4SStefano Zampini 5620906d46d4SStefano Zampini #undef __FUNCT__ 5621906d46d4SStefano Zampini #define __FUNCT__ "PCBDDCMatMultTranspose_Private" 5622906d46d4SStefano Zampini static PetscErrorCode PCBDDCMatMultTranspose_Private(Mat A, Vec x, Vec y) 5623906d46d4SStefano Zampini { 5624906d46d4SStefano Zampini PCBDDCChange_ctx change_ctx; 5625906d46d4SStefano Zampini PetscErrorCode ierr; 5626906d46d4SStefano Zampini 5627906d46d4SStefano Zampini PetscFunctionBegin; 5628906d46d4SStefano Zampini ierr = MatShellGetContext(A,&change_ctx);CHKERRQ(ierr); 5629906d46d4SStefano Zampini ierr = MatMult(change_ctx->global_change,x,change_ctx->work[0]);CHKERRQ(ierr); 5630906d46d4SStefano Zampini ierr = MatMultTranspose(change_ctx->original_mat,change_ctx->work[0],change_ctx->work[1]);CHKERRQ(ierr); 5631906d46d4SStefano Zampini ierr = MatMultTranspose(change_ctx->global_change,change_ctx->work[1],y);CHKERRQ(ierr); 5632906d46d4SStefano Zampini PetscFunctionReturn(0); 5633906d46d4SStefano Zampini } 5634b96c3477SStefano Zampini 5635b96c3477SStefano Zampini #undef __FUNCT__ 5636b96c3477SStefano Zampini #define __FUNCT__ "PCBDDCSetUpSubSchurs" 563708122e43SStefano Zampini PetscErrorCode PCBDDCSetUpSubSchurs(PC pc) 5638b96c3477SStefano Zampini { 5639a64f4aa4SStefano Zampini PC_IS *pcis=(PC_IS*)pc->data; 5640b96c3477SStefano Zampini PC_BDDC *pcbddc=(PC_BDDC*)pc->data; 5641b96c3477SStefano Zampini PCBDDCSubSchurs sub_schurs=pcbddc->sub_schurs; 5642a64f4aa4SStefano Zampini Mat S_j; 5643b96c3477SStefano Zampini PetscInt *used_xadj,*used_adjncy; 5644b96c3477SStefano Zampini PetscBool free_used_adj; 5645b96c3477SStefano Zampini PetscErrorCode ierr; 5646b96c3477SStefano Zampini 5647b96c3477SStefano Zampini PetscFunctionBegin; 5648b96c3477SStefano Zampini /* decide the adjacency to be used for determining internal problems for local schur on subsets */ 5649b96c3477SStefano Zampini free_used_adj = PETSC_FALSE; 565008122e43SStefano Zampini if (pcbddc->sub_schurs_layers == -1) { 5651b96c3477SStefano Zampini used_xadj = NULL; 5652b96c3477SStefano Zampini used_adjncy = NULL; 5653b96c3477SStefano Zampini } else { 565408122e43SStefano Zampini if (pcbddc->sub_schurs_use_useradj && pcbddc->mat_graph->xadj) { 565508122e43SStefano Zampini used_xadj = pcbddc->mat_graph->xadj; 565608122e43SStefano Zampini used_adjncy = pcbddc->mat_graph->adjncy; 565708122e43SStefano Zampini } else if (pcbddc->computed_rowadj) { 5658b96c3477SStefano Zampini used_xadj = pcbddc->mat_graph->xadj; 5659b96c3477SStefano Zampini used_adjncy = pcbddc->mat_graph->adjncy; 5660b96c3477SStefano Zampini } else { 56612fffb893SStefano Zampini PetscBool flg_row=PETSC_FALSE; 5662b96c3477SStefano Zampini const PetscInt *xadj,*adjncy; 5663b96c3477SStefano Zampini PetscInt nvtxs; 5664b96c3477SStefano Zampini 56652fffb893SStefano Zampini ierr = MatGetRowIJ(pcbddc->local_mat,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 56662fffb893SStefano Zampini if (flg_row) { 5667b96c3477SStefano Zampini ierr = PetscMalloc2(nvtxs+1,&used_xadj,xadj[nvtxs],&used_adjncy);CHKERRQ(ierr); 5668b96c3477SStefano Zampini ierr = PetscMemcpy(used_xadj,xadj,(nvtxs+1)*sizeof(*xadj));CHKERRQ(ierr); 5669b96c3477SStefano Zampini ierr = PetscMemcpy(used_adjncy,adjncy,(xadj[nvtxs])*sizeof(*adjncy));CHKERRQ(ierr); 5670b96c3477SStefano Zampini free_used_adj = PETSC_TRUE; 56712fffb893SStefano Zampini } else { 56722fffb893SStefano Zampini pcbddc->sub_schurs_layers = -1; 56732fffb893SStefano Zampini used_xadj = NULL; 56742fffb893SStefano Zampini used_adjncy = NULL; 56752fffb893SStefano Zampini } 56762fffb893SStefano Zampini ierr = MatRestoreRowIJ(pcbddc->local_mat,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 5677b96c3477SStefano Zampini } 5678b96c3477SStefano Zampini } 5679d5574798SStefano Zampini 5680d5574798SStefano Zampini /* setup sub_schurs data */ 5681a64f4aa4SStefano Zampini ierr = MatCreateSchurComplement(pcis->A_II,pcis->A_II,pcis->A_IB,pcis->A_BI,pcis->A_BB,&S_j);CHKERRQ(ierr); 5682a64f4aa4SStefano Zampini if (!sub_schurs->use_mumps) { 5683a64f4aa4SStefano Zampini /* pcbddc->ksp_D up to date only if not using MUMPS */ 5684a64f4aa4SStefano Zampini ierr = MatSchurComplementSetKSP(S_j,pcbddc->ksp_D);CHKERRQ(ierr); 5685683d3df6SStefano 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);CHKERRQ(ierr); 5686a64f4aa4SStefano Zampini } else { 56876816873aSStefano Zampini PetscBool reuse_solvers = (PetscBool)!pcbddc->use_change_of_basis; 568804708bb6SStefano Zampini PetscBool isseqaij; 56895feab87aSStefano Zampini if (!pcbddc->use_vertices && reuse_solvers) { 56905feab87aSStefano Zampini PetscInt n_vertices; 56915feab87aSStefano Zampini 56925feab87aSStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_vertices,&n_vertices);CHKERRQ(ierr); 56932034aafcSStefano Zampini reuse_solvers = (PetscBool)!n_vertices; 56945feab87aSStefano Zampini } 569504708bb6SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcbddc->local_mat,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 569604708bb6SStefano Zampini if (!isseqaij) { 569704708bb6SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 569804708bb6SStefano Zampini if (matis->A == pcbddc->local_mat) { 569904708bb6SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 570004708bb6SStefano Zampini ierr = MatConvert(matis->A,MATSEQAIJ,MAT_INITIAL_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 570104708bb6SStefano Zampini } else { 570204708bb6SStefano Zampini ierr = MatConvert(pcbddc->local_mat,MATSEQAIJ,MAT_REUSE_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 570304708bb6SStefano Zampini } 570404708bb6SStefano Zampini } 5705683d3df6SStefano 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);CHKERRQ(ierr); 5706a64f4aa4SStefano Zampini } 5707a64f4aa4SStefano Zampini ierr = MatDestroy(&S_j);CHKERRQ(ierr); 5708b96c3477SStefano Zampini 5709b96c3477SStefano Zampini /* free adjacency */ 5710b96c3477SStefano Zampini if (free_used_adj) { 5711b96c3477SStefano Zampini ierr = PetscFree2(used_xadj,used_adjncy);CHKERRQ(ierr); 5712b96c3477SStefano Zampini } 5713b96c3477SStefano Zampini PetscFunctionReturn(0); 5714b96c3477SStefano Zampini } 5715b96c3477SStefano Zampini 5716b96c3477SStefano Zampini #undef __FUNCT__ 5717b96c3477SStefano Zampini #define __FUNCT__ "PCBDDCInitSubSchurs" 571808122e43SStefano Zampini PetscErrorCode PCBDDCInitSubSchurs(PC pc) 5719b96c3477SStefano Zampini { 5720b96c3477SStefano Zampini PC_IS *pcis=(PC_IS*)pc->data; 5721b96c3477SStefano Zampini PC_BDDC *pcbddc=(PC_BDDC*)pc->data; 5722b96c3477SStefano Zampini PCBDDCSubSchurs sub_schurs=pcbddc->sub_schurs; 5723b96c3477SStefano Zampini PCBDDCGraph graph; 5724b96c3477SStefano Zampini PetscErrorCode ierr; 5725b96c3477SStefano Zampini 5726b96c3477SStefano Zampini PetscFunctionBegin; 5727b96c3477SStefano Zampini /* attach interface graph for determining subsets */ 572808122e43SStefano Zampini if (pcbddc->sub_schurs_rebuild) { /* in case rebuild has been requested, it uses a graph generated only by the neighbouring information */ 57293301b35fSStefano Zampini IS verticesIS,verticescomm; 57303301b35fSStefano Zampini PetscInt vsize,*idxs; 5731b96c3477SStefano Zampini 5732b96c3477SStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,NULL,NULL,NULL,NULL,&verticesIS);CHKERRQ(ierr); 57333301b35fSStefano Zampini ierr = ISGetSize(verticesIS,&vsize);CHKERRQ(ierr); 57343301b35fSStefano Zampini ierr = ISGetIndices(verticesIS,(const PetscInt**)&idxs);CHKERRQ(ierr); 57353301b35fSStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),vsize,idxs,PETSC_COPY_VALUES,&verticescomm);CHKERRQ(ierr); 57363301b35fSStefano Zampini ierr = ISRestoreIndices(verticesIS,(const PetscInt**)&idxs);CHKERRQ(ierr); 57373301b35fSStefano Zampini ierr = ISDestroy(&verticesIS);CHKERRQ(ierr); 5738b96c3477SStefano Zampini ierr = PCBDDCGraphCreate(&graph);CHKERRQ(ierr); 57397fb0e2dbSStefano Zampini ierr = PCBDDCGraphInit(graph,pcbddc->mat_graph->l2gmap,pcbddc->mat_graph->nvtxs_global);CHKERRQ(ierr); 57403301b35fSStefano Zampini ierr = PCBDDCGraphSetUp(graph,0,NULL,pcbddc->DirichletBoundariesLocal,0,NULL,verticescomm);CHKERRQ(ierr); 57413301b35fSStefano Zampini ierr = ISDestroy(&verticescomm);CHKERRQ(ierr); 5742b96c3477SStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(graph);CHKERRQ(ierr); 5743b96c3477SStefano Zampini /* 5744b96c3477SStefano Zampini if (pcbddc->dbg_flag) { 5745b96c3477SStefano Zampini ierr = PCBDDCGraphASCIIView(graph,pcbddc->dbg_flag,pcbddc->dbg_viewer);CHKERRQ(ierr); 5746b96c3477SStefano Zampini } 5747b96c3477SStefano Zampini */ 5748b96c3477SStefano Zampini } else { 5749b96c3477SStefano Zampini graph = pcbddc->mat_graph; 5750b96c3477SStefano Zampini } 5751b96c3477SStefano Zampini 5752b96c3477SStefano Zampini /* sub_schurs init */ 5753a64f4aa4SStefano Zampini ierr = PCBDDCSubSchursInit(sub_schurs,pcis->is_I_local,pcis->is_B_local,graph,pcis->BtoNmap);CHKERRQ(ierr); 5754a64f4aa4SStefano Zampini 5755b96c3477SStefano Zampini /* free graph struct */ 575608122e43SStefano Zampini if (pcbddc->sub_schurs_rebuild) { 5757b96c3477SStefano Zampini ierr = PCBDDCGraphDestroy(&graph);CHKERRQ(ierr); 5758b96c3477SStefano Zampini } 5759b96c3477SStefano Zampini PetscFunctionReturn(0); 5760b96c3477SStefano Zampini } 5761fa34dd3eSStefano Zampini 5762fa34dd3eSStefano Zampini #undef __FUNCT__ 5763fa34dd3eSStefano Zampini #define __FUNCT__ "PCBDDCCheckOperator" 5764fa34dd3eSStefano Zampini PetscErrorCode PCBDDCCheckOperator(PC pc) 5765fa34dd3eSStefano Zampini { 5766fa34dd3eSStefano Zampini PC_IS *pcis=(PC_IS*)pc->data; 5767fa34dd3eSStefano Zampini PC_BDDC *pcbddc=(PC_BDDC*)pc->data; 5768fa34dd3eSStefano Zampini PetscErrorCode ierr; 5769fa34dd3eSStefano Zampini 5770fa34dd3eSStefano Zampini PetscFunctionBegin; 5771fa34dd3eSStefano Zampini if (pcbddc->n_vertices == pcbddc->local_primal_size) { 5772fa34dd3eSStefano Zampini IS zerodiag = NULL; 57734f1b2e48SStefano Zampini Mat S_j,B0_B=NULL; 5774fa34dd3eSStefano Zampini Vec dummy_vec=NULL,vec_check_B,vec_scale_P; 57754f1b2e48SStefano Zampini PetscScalar *p0_check,*array,*array2; 577675c01103SStefano Zampini PetscReal norm; 5777fa34dd3eSStefano Zampini PetscInt i; 5778fa34dd3eSStefano Zampini 5779fa34dd3eSStefano Zampini /* B0 and B0_B */ 5780fa34dd3eSStefano Zampini if (zerodiag) { 5781fa34dd3eSStefano Zampini IS dummy; 5782fa34dd3eSStefano Zampini 57834f1b2e48SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcbddc->benign_n,0,1,&dummy);CHKERRQ(ierr); 57844f1b2e48SStefano Zampini ierr = MatGetSubMatrix(pcbddc->benign_B0,dummy,pcis->is_B_local,MAT_INITIAL_MATRIX,&B0_B);CHKERRQ(ierr); 5785fa34dd3eSStefano Zampini ierr = MatCreateVecs(B0_B,NULL,&dummy_vec);CHKERRQ(ierr); 5786fa34dd3eSStefano Zampini ierr = ISDestroy(&dummy);CHKERRQ(ierr); 5787fa34dd3eSStefano Zampini } 5788fa34dd3eSStefano Zampini /* I need a primal vector to scale primal nodes since BDDC sums contibutions */ 5789fa34dd3eSStefano Zampini ierr = VecDuplicate(pcbddc->vec1_P,&vec_scale_P);CHKERRQ(ierr); 5790fa34dd3eSStefano Zampini ierr = VecSet(pcbddc->vec1_P,1.0);CHKERRQ(ierr); 5791fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 5792fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 5793fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,vec_scale_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 5794fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,vec_scale_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 5795fa34dd3eSStefano Zampini ierr = VecReciprocal(vec_scale_P);CHKERRQ(ierr); 5796fa34dd3eSStefano Zampini /* S_j */ 5797fa34dd3eSStefano Zampini ierr = MatCreateSchurComplement(pcis->A_II,pcis->A_II,pcis->A_IB,pcis->A_BI,pcis->A_BB,&S_j);CHKERRQ(ierr); 5798fa34dd3eSStefano Zampini ierr = MatSchurComplementSetKSP(S_j,pcbddc->ksp_D);CHKERRQ(ierr); 5799fa34dd3eSStefano Zampini 5800fa34dd3eSStefano Zampini /* mimic vector in \widetilde{W}_\Gamma */ 5801fa34dd3eSStefano Zampini ierr = VecSetRandom(pcis->vec1_N,NULL);CHKERRQ(ierr); 5802fa34dd3eSStefano Zampini /* continuous in primal space */ 5803fa34dd3eSStefano Zampini ierr = VecSetRandom(pcbddc->coarse_vec,NULL);CHKERRQ(ierr); 5804fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,pcbddc->vec1_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 5805fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,pcbddc->vec1_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 5806fa34dd3eSStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 58074f1b2e48SStefano Zampini ierr = PetscCalloc1(pcbddc->benign_n,&p0_check);CHKERRQ(ierr); 58084f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) p0_check[i] = array[pcbddc->local_primal_size-pcbddc->benign_n+i]; 5809fa34dd3eSStefano Zampini ierr = VecSetValues(pcis->vec1_N,pcbddc->local_primal_size,pcbddc->local_primal_ref_node,array,INSERT_VALUES);CHKERRQ(ierr); 5810fa34dd3eSStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 5811fa34dd3eSStefano Zampini ierr = VecAssemblyBegin(pcis->vec1_N);CHKERRQ(ierr); 5812fa34dd3eSStefano Zampini ierr = VecAssemblyEnd(pcis->vec1_N);CHKERRQ(ierr); 5813fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcis->N_to_B,pcis->vec1_N,pcis->vec2_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 5814fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcis->N_to_B,pcis->vec1_N,pcis->vec2_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 5815fa34dd3eSStefano Zampini ierr = VecDuplicate(pcis->vec2_B,&vec_check_B);CHKERRQ(ierr); 5816fa34dd3eSStefano Zampini ierr = VecCopy(pcis->vec2_B,vec_check_B);CHKERRQ(ierr); 5817fa34dd3eSStefano Zampini 5818fa34dd3eSStefano Zampini /* assemble rhs for coarse problem */ 5819fa34dd3eSStefano Zampini /* widetilde{S}_\Gamma w_\Gamma + \widetilde{B0}^T_B p0 */ 5820fa34dd3eSStefano Zampini /* local with Schur */ 5821fa34dd3eSStefano Zampini ierr = MatMult(S_j,pcis->vec2_B,pcis->vec1_B);CHKERRQ(ierr); 5822fa34dd3eSStefano Zampini if (zerodiag) { 5823fa34dd3eSStefano Zampini ierr = VecGetArray(dummy_vec,&array);CHKERRQ(ierr); 58244f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) array[i] = p0_check[i]; 5825fa34dd3eSStefano Zampini ierr = VecRestoreArray(dummy_vec,&array);CHKERRQ(ierr); 5826fa34dd3eSStefano Zampini ierr = MatMultTransposeAdd(B0_B,dummy_vec,pcis->vec1_B,pcis->vec1_B);CHKERRQ(ierr); 5827fa34dd3eSStefano Zampini } 5828fa34dd3eSStefano Zampini /* sum on primal nodes the local contributions */ 5829fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcis->N_to_B,pcis->vec1_B,pcis->vec1_N,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 5830fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcis->N_to_B,pcis->vec1_B,pcis->vec1_N,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 5831fa34dd3eSStefano Zampini ierr = VecGetArray(pcis->vec1_N,&array);CHKERRQ(ierr); 5832fa34dd3eSStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array2);CHKERRQ(ierr); 5833fa34dd3eSStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) array2[i] = array[pcbddc->local_primal_ref_node[i]]; 5834fa34dd3eSStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array2);CHKERRQ(ierr); 5835fa34dd3eSStefano Zampini ierr = VecRestoreArray(pcis->vec1_N,&array);CHKERRQ(ierr); 5836fa34dd3eSStefano Zampini ierr = VecSet(pcbddc->coarse_vec,0.);CHKERRQ(ierr); 5837fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 5838fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 5839fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,pcbddc->vec1_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 5840fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,pcbddc->vec1_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 5841fa34dd3eSStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 5842fa34dd3eSStefano Zampini /* scale primal nodes (BDDC sums contibutions) */ 5843fa34dd3eSStefano Zampini ierr = VecPointwiseMult(pcbddc->vec1_P,vec_scale_P,pcbddc->vec1_P);CHKERRQ(ierr); 5844fa34dd3eSStefano Zampini ierr = VecSetValues(pcis->vec1_N,pcbddc->local_primal_size,pcbddc->local_primal_ref_node,array,INSERT_VALUES);CHKERRQ(ierr); 5845fa34dd3eSStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 5846fa34dd3eSStefano Zampini ierr = VecAssemblyBegin(pcis->vec1_N);CHKERRQ(ierr); 5847fa34dd3eSStefano Zampini ierr = VecAssemblyEnd(pcis->vec1_N);CHKERRQ(ierr); 5848fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcis->N_to_B,pcis->vec1_N,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 5849fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcis->N_to_B,pcis->vec1_N,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 5850fa34dd3eSStefano Zampini /* global: \widetilde{B0}_B w_\Gamma */ 5851fa34dd3eSStefano Zampini if (zerodiag) { 5852fa34dd3eSStefano Zampini ierr = MatMult(B0_B,pcis->vec2_B,dummy_vec);CHKERRQ(ierr); 5853fa34dd3eSStefano Zampini ierr = VecGetArray(dummy_vec,&array);CHKERRQ(ierr); 58544f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) pcbddc->benign_p0[i] = array[i]; 5855fa34dd3eSStefano Zampini ierr = VecRestoreArray(dummy_vec,&array);CHKERRQ(ierr); 5856fa34dd3eSStefano Zampini } 5857fa34dd3eSStefano Zampini /* BDDC */ 5858fa34dd3eSStefano Zampini ierr = VecSet(pcis->vec1_D,0.);CHKERRQ(ierr); 5859fa34dd3eSStefano Zampini ierr = PCBDDCApplyInterfacePreconditioner(pc,PETSC_FALSE);CHKERRQ(ierr); 5860fa34dd3eSStefano Zampini 5861fa34dd3eSStefano Zampini ierr = VecCopy(pcis->vec1_B,pcis->vec2_B);CHKERRQ(ierr); 5862fa34dd3eSStefano Zampini ierr = VecAXPY(pcis->vec1_B,-1.0,vec_check_B);CHKERRQ(ierr); 5863fa34dd3eSStefano Zampini ierr = VecNorm(pcis->vec1_B,NORM_INFINITY,&norm);CHKERRQ(ierr); 5864fa34dd3eSStefano Zampini PetscPrintf(PETSC_COMM_SELF,"[%d] BDDC local error is %1.4e\n",PetscGlobalRank,norm); 58654f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 58664f1b2e48SStefano Zampini PetscPrintf(PETSC_COMM_SELF,"[%d] BDDC p0[%d] error is %1.4e\n",PetscGlobalRank,i,PetscAbsScalar(pcbddc->benign_p0[i]-p0_check[i])); 5867fa34dd3eSStefano Zampini } 58684f1b2e48SStefano Zampini ierr = PetscFree(p0_check);CHKERRQ(ierr); 5869fa34dd3eSStefano Zampini ierr = VecDestroy(&vec_scale_P);CHKERRQ(ierr); 5870fa34dd3eSStefano Zampini ierr = VecDestroy(&vec_check_B);CHKERRQ(ierr); 5871fa34dd3eSStefano Zampini ierr = VecDestroy(&dummy_vec);CHKERRQ(ierr); 5872fa34dd3eSStefano Zampini ierr = MatDestroy(&S_j);CHKERRQ(ierr); 5873fa34dd3eSStefano Zampini ierr = MatDestroy(&B0_B);CHKERRQ(ierr); 5874fa34dd3eSStefano Zampini } 5875fa34dd3eSStefano Zampini PetscFunctionReturn(0); 5876fa34dd3eSStefano Zampini } 5877