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); 258339f8db1SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->local_mat);CHKERRQ(ierr); 259339f8db1SStefano Zampini pcbddc->benign_original_mat = pcbddc->local_mat; 2604f1b2e48SStefano Zampini /* if a local info on dofs is present, assumes the last field is represented by "pressures" 2614f1b2e48SStefano Zampini otherwise, it uses only zerodiagonal dofs (ok if the pressure block is all zero; it could fail if it is not) 2624f1b2e48SStefano Zampini Checks if all the pressure dofs in each subdomain have a zero diagonal 2634f1b2e48SStefano Zampini If not, a change of basis on pressures is not needed 2644f1b2e48SStefano Zampini since the local Schur complements are SPD 2654f1b2e48SStefano Zampini */ 2664f1b2e48SStefano Zampini has_null_pressures = PETSC_TRUE; 2674f1b2e48SStefano Zampini have_null = PETSC_TRUE; 26840fa8d13SStefano Zampini if (pcbddc->n_ISForDofsLocal) { 2694f1b2e48SStefano Zampini PetscInt npl,*idxs,p = pcbddc->n_ISForDofsLocal-1; 2704f1b2e48SStefano Zampini 2714f1b2e48SStefano Zampini /* Dofs splitting for BDDC cannot have PETSC_COMM_SELF, so create a sequential IS */ 2724f1b2e48SStefano Zampini ierr = ISGetLocalSize(pcbddc->ISForDofsLocal[p],&npl);CHKERRQ(ierr); 2734f1b2e48SStefano Zampini ierr = ISGetIndices(pcbddc->ISForDofsLocal[p],(const PetscInt**)&idxs);CHKERRQ(ierr); 2744f1b2e48SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,npl,idxs,PETSC_COPY_VALUES,&pressures);CHKERRQ(ierr); 275ba14f8e3SStefano Zampini ierr = ISRestoreIndices(pcbddc->ISForDofsLocal[p],(const PetscInt**)&idxs);CHKERRQ(ierr); 27640fa8d13SStefano Zampini ierr = ISSorted(pressures,&sorted);CHKERRQ(ierr); 27740fa8d13SStefano Zampini if (!sorted) { 27840fa8d13SStefano Zampini ierr = ISSort(pressures);CHKERRQ(ierr); 27940fa8d13SStefano Zampini } 28040fa8d13SStefano Zampini } else { 28140fa8d13SStefano Zampini pressures = NULL; 28240fa8d13SStefano Zampini } 2834f1b2e48SStefano Zampini ierr = MatGetLocalSize(pcbddc->benign_original_mat,&n,NULL);CHKERRQ(ierr); 2844f1b2e48SStefano Zampini /* TODO: add check for shared dofs and raise error */ 285339f8db1SStefano Zampini ierr = MatFindZeroDiagonals(pcbddc->benign_original_mat,&zerodiag);CHKERRQ(ierr); 286339f8db1SStefano Zampini ierr = ISSorted(zerodiag,&sorted);CHKERRQ(ierr); 287339f8db1SStefano Zampini if (!sorted) { 288339f8db1SStefano Zampini ierr = ISSort(zerodiag);CHKERRQ(ierr); 289339f8db1SStefano Zampini } 290339f8db1SStefano Zampini ierr = ISGetLocalSize(zerodiag,&nz);CHKERRQ(ierr); 2914f1b2e48SStefano Zampini if (!nz) { 2924f1b2e48SStefano Zampini if (n) have_null = PETSC_FALSE; 2934f1b2e48SStefano Zampini has_null_pressures = PETSC_FALSE; 29440fa8d13SStefano Zampini ierr = ISDestroy(&zerodiag);CHKERRQ(ierr); 29540fa8d13SStefano Zampini } 2964f1b2e48SStefano Zampini recompute_zerodiag = PETSC_FALSE; 2974f1b2e48SStefano Zampini /* in case disconnected subdomains info is present, split the pressures accordingly (otherwise the benign trick could fail) */ 2984f1b2e48SStefano Zampini zerodiag_subs = NULL; 2994f1b2e48SStefano Zampini pcbddc->benign_n = 0; 3004f1b2e48SStefano Zampini if (has_null_pressures) { 3014f1b2e48SStefano Zampini IS *subs; 3024f1b2e48SStefano Zampini PetscInt nsubs,i; 3034f1b2e48SStefano Zampini 3044f1b2e48SStefano Zampini subs = pcbddc->local_subs; 3054f1b2e48SStefano Zampini nsubs = pcbddc->n_local_subs; 3064f1b2e48SStefano Zampini if (nsubs > 1) { 3074f1b2e48SStefano Zampini ierr = PetscCalloc1(nsubs,&zerodiag_subs);CHKERRQ(ierr); 3084f1b2e48SStefano Zampini for (i=0;i<nsubs;i++) { 3094f1b2e48SStefano Zampini ISLocalToGlobalMapping l2g; 3104f1b2e48SStefano Zampini IS t_zerodiag_subs; 3114f1b2e48SStefano Zampini PetscInt nl; 3124f1b2e48SStefano Zampini 3134f1b2e48SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(subs[i],&l2g);CHKERRQ(ierr); 3144f1b2e48SStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(l2g,IS_GTOLM_DROP,zerodiag,&t_zerodiag_subs);CHKERRQ(ierr); 3154f1b2e48SStefano Zampini ierr = ISGetLocalSize(t_zerodiag_subs,&nl);CHKERRQ(ierr); 3164f1b2e48SStefano Zampini if (nl) { 3174f1b2e48SStefano Zampini PetscBool valid = PETSC_TRUE; 3184f1b2e48SStefano Zampini 3194f1b2e48SStefano Zampini if (pressures) { 3204f1b2e48SStefano Zampini IS t_pressure_subs; 3214f1b2e48SStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(l2g,IS_GTOLM_DROP,pressures,&t_pressure_subs);CHKERRQ(ierr); 3224f1b2e48SStefano Zampini ierr = ISEqual(t_pressure_subs,t_zerodiag_subs,&valid);CHKERRQ(ierr); 3234f1b2e48SStefano Zampini ierr = ISDestroy(&t_pressure_subs);CHKERRQ(ierr); 3244f1b2e48SStefano Zampini } 3254f1b2e48SStefano Zampini if (valid) { 3264f1b2e48SStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(l2g,t_zerodiag_subs,&zerodiag_subs[pcbddc->benign_n]);CHKERRQ(ierr); 3274f1b2e48SStefano Zampini pcbddc->benign_n++; 3284f1b2e48SStefano Zampini } else { 3294f1b2e48SStefano Zampini recompute_zerodiag = PETSC_TRUE; 3304f1b2e48SStefano Zampini } 3314f1b2e48SStefano Zampini } 3324f1b2e48SStefano Zampini ierr = ISDestroy(&t_zerodiag_subs);CHKERRQ(ierr); 3334f1b2e48SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&l2g);CHKERRQ(ierr); 3344f1b2e48SStefano Zampini } 3354f1b2e48SStefano Zampini } else { /* there's just one subdomain (or zero if they have not been detected */ 3364f1b2e48SStefano Zampini PetscBool valid = PETSC_TRUE; 3374f1b2e48SStefano Zampini if (pressures) { 3384f1b2e48SStefano Zampini ierr = ISEqual(pressures,zerodiag,&valid);CHKERRQ(ierr); 3394f1b2e48SStefano Zampini } 3404f1b2e48SStefano Zampini if (valid) { 3414f1b2e48SStefano Zampini pcbddc->benign_n = 1; 3424f1b2e48SStefano Zampini ierr = PetscCalloc1(1,&zerodiag_subs);CHKERRQ(ierr); 3434f1b2e48SStefano Zampini ierr = PetscObjectReference((PetscObject)zerodiag);CHKERRQ(ierr); 3444f1b2e48SStefano Zampini zerodiag_subs[0] = zerodiag; 3454f1b2e48SStefano Zampini } 3464f1b2e48SStefano Zampini } 3474f1b2e48SStefano Zampini } 3484f1b2e48SStefano Zampini 3494f1b2e48SStefano Zampini if (!pcbddc->benign_n) { 3504f1b2e48SStefano Zampini ierr = ISDestroy(&zerodiag);CHKERRQ(ierr); 3514f1b2e48SStefano Zampini recompute_zerodiag = PETSC_FALSE; 3524f1b2e48SStefano Zampini has_null_pressures = PETSC_FALSE; 3534f1b2e48SStefano Zampini have_null = PETSC_FALSE; 3544f1b2e48SStefano Zampini } 3554f1b2e48SStefano Zampini 3564f1b2e48SStefano Zampini /* final check for null pressures */ 3574f1b2e48SStefano Zampini if (zerodiag && pressures) { 3584f1b2e48SStefano Zampini PetscInt nz,np; 3594f1b2e48SStefano Zampini ierr = ISGetLocalSize(zerodiag,&nz);CHKERRQ(ierr); 3604f1b2e48SStefano Zampini ierr = ISGetLocalSize(pressures,&np);CHKERRQ(ierr); 3614f1b2e48SStefano Zampini if (nz != np) have_null = PETSC_FALSE; 3624f1b2e48SStefano Zampini } 3634f1b2e48SStefano Zampini 3644f1b2e48SStefano Zampini if (recompute_zerodiag) { 3654f1b2e48SStefano Zampini ierr = ISDestroy(&zerodiag);CHKERRQ(ierr); 3664f1b2e48SStefano Zampini if (pcbddc->benign_n == 1) { 3674f1b2e48SStefano Zampini ierr = PetscObjectReference((PetscObject)zerodiag_subs[0]);CHKERRQ(ierr); 3684f1b2e48SStefano Zampini zerodiag = zerodiag_subs[0]; 3694f1b2e48SStefano Zampini } else { 3704f1b2e48SStefano Zampini PetscInt i,nzn,*new_idxs; 3714f1b2e48SStefano Zampini 3724f1b2e48SStefano Zampini nzn = 0; 3734f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 3744f1b2e48SStefano Zampini PetscInt ns; 3754f1b2e48SStefano Zampini ierr = ISGetLocalSize(zerodiag_subs[i],&ns);CHKERRQ(ierr); 3764f1b2e48SStefano Zampini nzn += ns; 3774f1b2e48SStefano Zampini } 3784f1b2e48SStefano Zampini ierr = PetscMalloc1(nzn,&new_idxs);CHKERRQ(ierr); 3794f1b2e48SStefano Zampini nzn = 0; 3804f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 3814f1b2e48SStefano Zampini PetscInt ns,*idxs; 3824f1b2e48SStefano Zampini ierr = ISGetLocalSize(zerodiag_subs[i],&ns);CHKERRQ(ierr); 3834f1b2e48SStefano Zampini ierr = ISGetIndices(zerodiag_subs[i],(const PetscInt**)&idxs);CHKERRQ(ierr); 3844f1b2e48SStefano Zampini ierr = PetscMemcpy(new_idxs+nzn,idxs,ns*sizeof(PetscInt));CHKERRQ(ierr); 3854f1b2e48SStefano Zampini ierr = ISRestoreIndices(zerodiag_subs[i],(const PetscInt**)&idxs);CHKERRQ(ierr); 3864f1b2e48SStefano Zampini nzn += ns; 3874f1b2e48SStefano Zampini } 3884f1b2e48SStefano Zampini ierr = PetscSortInt(nzn,new_idxs);CHKERRQ(ierr); 3894f1b2e48SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,nzn,new_idxs,PETSC_OWN_POINTER,&zerodiag);CHKERRQ(ierr); 3904f1b2e48SStefano Zampini } 3914f1b2e48SStefano Zampini have_null = PETSC_FALSE; 3924f1b2e48SStefano Zampini } 3934f1b2e48SStefano Zampini 3944f1b2e48SStefano Zampini if (has_null_pressures) { 3954f1b2e48SStefano Zampini IS zerodiagc; 3964f1b2e48SStefano Zampini const PetscInt *idxs,*idxsc; 3974f1b2e48SStefano Zampini PetscInt i,s,*nnz; 3981cf9b237SStefano Zampini Mat M; 3994f1b2e48SStefano Zampini 4004f1b2e48SStefano Zampini ierr = ISGetLocalSize(zerodiag,&nz);CHKERRQ(ierr); 401339f8db1SStefano Zampini ierr = ISComplement(zerodiag,0,n,&zerodiagc);CHKERRQ(ierr); 402339f8db1SStefano Zampini ierr = ISGetIndices(zerodiagc,&idxsc);CHKERRQ(ierr); 403339f8db1SStefano Zampini /* local change of basis for pressures */ 404339f8db1SStefano Zampini ierr = MatDestroy(&pcbddc->benign_change);CHKERRQ(ierr); 405339f8db1SStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)pcbddc->benign_original_mat),&pcbddc->benign_change);CHKERRQ(ierr); 406339f8db1SStefano Zampini ierr = MatSetType(pcbddc->benign_change,MATAIJ);CHKERRQ(ierr); 407339f8db1SStefano Zampini ierr = MatSetSizes(pcbddc->benign_change,n,n,PETSC_DECIDE,PETSC_DECIDE);CHKERRQ(ierr); 408339f8db1SStefano Zampini ierr = PetscMalloc1(n,&nnz);CHKERRQ(ierr); 4094f1b2e48SStefano Zampini for (i=0;i<n-nz;i++) nnz[idxsc[i]] = 1; /* identity on velocities plus pressure dofs for non-singular subdomains */ 4104f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 4114f1b2e48SStefano Zampini PetscInt nzs,j; 4124f1b2e48SStefano Zampini 4134f1b2e48SStefano Zampini ierr = ISGetLocalSize(zerodiag_subs[i],&nzs);CHKERRQ(ierr); 4144f1b2e48SStefano Zampini ierr = ISGetIndices(zerodiag_subs[i],&idxs);CHKERRQ(ierr); 4154f1b2e48SStefano Zampini for (j=0;j<nzs-1;j++) nnz[idxs[j]] = 2; /* change on pressures */ 4164f1b2e48SStefano Zampini nnz[idxs[nzs-1]] = nzs; /* last local pressure dof in subdomain */ 4174f1b2e48SStefano Zampini ierr = ISRestoreIndices(zerodiag_subs[i],&idxs);CHKERRQ(ierr); 4184f1b2e48SStefano Zampini } 419339f8db1SStefano Zampini ierr = MatSeqAIJSetPreallocation(pcbddc->benign_change,0,nnz);CHKERRQ(ierr); 420339f8db1SStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 421339f8db1SStefano Zampini /* set identity on velocities */ 422339f8db1SStefano Zampini for (i=0;i<n-nz;i++) { 423339f8db1SStefano Zampini ierr = MatSetValue(pcbddc->benign_change,idxsc[i],idxsc[i],1.,INSERT_VALUES);CHKERRQ(ierr); 424339f8db1SStefano Zampini } 4254f1b2e48SStefano Zampini ierr = ISRestoreIndices(zerodiagc,&idxsc);CHKERRQ(ierr); 4264f1b2e48SStefano Zampini ierr = ISDestroy(&zerodiagc);CHKERRQ(ierr); 4279f47a83aSStefano Zampini ierr = PetscFree3(pcbddc->benign_p0_lidx,pcbddc->benign_p0_gidx,pcbddc->benign_p0);CHKERRQ(ierr); 4284f1b2e48SStefano 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); 429339f8db1SStefano Zampini /* set change on pressures */ 4304f1b2e48SStefano Zampini for (s=0;s<pcbddc->benign_n;s++) { 4314f1b2e48SStefano Zampini PetscScalar *array; 4324f1b2e48SStefano Zampini PetscInt nzs; 4334f1b2e48SStefano Zampini 4344f1b2e48SStefano Zampini ierr = ISGetLocalSize(zerodiag_subs[s],&nzs);CHKERRQ(ierr); 4354f1b2e48SStefano Zampini ierr = ISGetIndices(zerodiag_subs[s],&idxs);CHKERRQ(ierr); 4364f1b2e48SStefano Zampini for (i=0;i<nzs-1;i++) { 437339f8db1SStefano Zampini PetscScalar vals[2]; 438339f8db1SStefano Zampini PetscInt cols[2]; 439339f8db1SStefano Zampini 440339f8db1SStefano Zampini cols[0] = idxs[i]; 4414f1b2e48SStefano Zampini cols[1] = idxs[nzs-1]; 442339f8db1SStefano Zampini vals[0] = 1.; 443b0f5fe93SStefano Zampini vals[1] = 1.; 4444f1b2e48SStefano Zampini ierr = MatSetValues(pcbddc->benign_change,1,cols,2,cols,vals,INSERT_VALUES);CHKERRQ(ierr); 445339f8db1SStefano Zampini } 4464f1b2e48SStefano Zampini ierr = PetscMalloc1(nzs,&array);CHKERRQ(ierr); 4474f1b2e48SStefano Zampini for (i=0;i<nzs-1;i++) array[i] = -1.; 4484f1b2e48SStefano Zampini array[nzs-1] = 1.; 4494f1b2e48SStefano Zampini ierr = MatSetValues(pcbddc->benign_change,1,idxs+nzs-1,nzs,idxs,array,INSERT_VALUES);CHKERRQ(ierr); 4504f1b2e48SStefano Zampini /* store local idxs for p0 */ 4514f1b2e48SStefano Zampini pcbddc->benign_p0_lidx[s] = idxs[nzs-1]; 4524f1b2e48SStefano Zampini ierr = ISRestoreIndices(zerodiag_subs[s],&idxs);CHKERRQ(ierr); 4534f1b2e48SStefano Zampini ierr = ISDestroy(&zerodiag_subs[s]);CHKERRQ(ierr); 454339f8db1SStefano Zampini ierr = PetscFree(array);CHKERRQ(ierr); 4554f1b2e48SStefano Zampini } 456339f8db1SStefano Zampini ierr = MatAssemblyBegin(pcbddc->benign_change,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 457339f8db1SStefano Zampini ierr = MatAssemblyEnd(pcbddc->benign_change,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 458339f8db1SStefano Zampini /* TODO: need optimization? */ 459339f8db1SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 4601cf9b237SStefano Zampini ierr = MatPtAP(pcbddc->benign_original_mat,pcbddc->benign_change,MAT_INITIAL_MATRIX,2.0,&M);CHKERRQ(ierr); 4611cf9b237SStefano Zampini ierr = MatSeqAIJCompress(M,&pcbddc->local_mat);CHKERRQ(ierr); 4621cf9b237SStefano Zampini ierr = MatDestroy(&M);CHKERRQ(ierr); 4634f1b2e48SStefano Zampini /* store global idxs for p0 */ 4644f1b2e48SStefano Zampini ierr = ISLocalToGlobalMappingApply(pc->pmat->rmap->mapping,pcbddc->benign_n,pcbddc->benign_p0_lidx,pcbddc->benign_p0_gidx);CHKERRQ(ierr); 465339f8db1SStefano Zampini } 4664f1b2e48SStefano Zampini ierr = PetscFree(zerodiag_subs);CHKERRQ(ierr); 4674f1b2e48SStefano Zampini ierr = ISDestroy(&pressures);CHKERRQ(ierr); 468b0f5fe93SStefano Zampini 469b0f5fe93SStefano Zampini /* determines if the coarse solver will be singular or not */ 470b0f5fe93SStefano Zampini ierr = MPI_Allreduce(&have_null,&pcbddc->benign_null,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 471339f8db1SStefano Zampini *zerodiaglocal = zerodiag; 472339f8db1SStefano Zampini PetscFunctionReturn(0); 473339f8db1SStefano Zampini } 474339f8db1SStefano Zampini 475339f8db1SStefano Zampini #undef __FUNCT__ 476015636ebSStefano Zampini #define __FUNCT__ "PCBDDCBenignGetOrSetP0" 477015636ebSStefano Zampini PetscErrorCode PCBDDCBenignGetOrSetP0(PC pc, Vec v, PetscBool get) 478efc2fbd9SStefano Zampini { 479efc2fbd9SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 480efc2fbd9SStefano Zampini PetscErrorCode ierr; 481efc2fbd9SStefano Zampini 482efc2fbd9SStefano Zampini PetscFunctionBegin; 483efc2fbd9SStefano Zampini if (!pcbddc->benign_sf) { 484efc2fbd9SStefano Zampini ierr = PetscSFCreate(PetscObjectComm((PetscObject)pc),&pcbddc->benign_sf);CHKERRQ(ierr); 4854f1b2e48SStefano Zampini ierr = PetscSFSetGraphLayout(pcbddc->benign_sf,pc->pmat->rmap,pcbddc->benign_n,NULL,PETSC_OWN_POINTER,pcbddc->benign_p0_gidx);CHKERRQ(ierr); 486efc2fbd9SStefano Zampini } 487015636ebSStefano Zampini if (get) { /* use SF to get values */ 488efc2fbd9SStefano Zampini PetscScalar *array; 489efc2fbd9SStefano Zampini 490efc2fbd9SStefano Zampini ierr = VecGetArrayRead(v,(const PetscScalar**)&array);CHKERRQ(ierr); 4914f1b2e48SStefano Zampini ierr = PetscSFBcastBegin(pcbddc->benign_sf,MPIU_SCALAR,array,pcbddc->benign_p0);CHKERRQ(ierr); 4924f1b2e48SStefano Zampini ierr = PetscSFBcastEnd(pcbddc->benign_sf,MPIU_SCALAR,array,pcbddc->benign_p0);CHKERRQ(ierr); 493efc2fbd9SStefano Zampini ierr = VecRestoreArrayRead(v,(const PetscScalar**)&array);CHKERRQ(ierr); 4944f1b2e48SStefano Zampini } else { /* use VecSetValues (not scalable, I should try to find a better solution (defining a new MPI_OP for reduction) */ 4954f1b2e48SStefano Zampini ierr = VecSetValues(v,pcbddc->benign_n,pcbddc->benign_p0_gidx,pcbddc->benign_p0,INSERT_VALUES);CHKERRQ(ierr); 496efc2fbd9SStefano Zampini ierr = VecAssemblyBegin(v);CHKERRQ(ierr); 497efc2fbd9SStefano Zampini ierr = VecAssemblyEnd(v);CHKERRQ(ierr); 498efc2fbd9SStefano Zampini } 499efc2fbd9SStefano Zampini PetscFunctionReturn(0); 500efc2fbd9SStefano Zampini } 501efc2fbd9SStefano Zampini 502efc2fbd9SStefano Zampini #undef __FUNCT__ 503c263805aSStefano Zampini #define __FUNCT__ "PCBDDCBenignPopOrPushB0" 504c263805aSStefano Zampini PetscErrorCode PCBDDCBenignPopOrPushB0(PC pc, PetscBool pop) 505c263805aSStefano Zampini { 506c263805aSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 507c263805aSStefano Zampini PetscErrorCode ierr; 508c263805aSStefano Zampini 509c263805aSStefano Zampini PetscFunctionBegin; 510c263805aSStefano Zampini /* TODO: add error checking 511c263805aSStefano Zampini - avoid nested pop (or push) calls. 512c263805aSStefano Zampini - cannot push before pop. 5131c604dc7SStefano Zampini - cannot call this if pcbddc->local_mat is NULL 514c263805aSStefano Zampini */ 5154f1b2e48SStefano Zampini if (!pcbddc->benign_n) { 516efc2fbd9SStefano Zampini PetscFunctionReturn(0); 517efc2fbd9SStefano Zampini } 518c263805aSStefano Zampini if (pop) { 5194f1b2e48SStefano Zampini IS is_p0; 5204f1b2e48SStefano Zampini MatReuse reuse; 521c263805aSStefano Zampini 522c263805aSStefano Zampini /* extract B_0 */ 5234f1b2e48SStefano Zampini reuse = MAT_INITIAL_MATRIX; 5244f1b2e48SStefano Zampini if (pcbddc->benign_B0) { 5254f1b2e48SStefano Zampini reuse = MAT_REUSE_MATRIX; 5264f1b2e48SStefano Zampini } 5274f1b2e48SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,pcbddc->benign_n,pcbddc->benign_p0_lidx,PETSC_COPY_VALUES,&is_p0);CHKERRQ(ierr); 5284f1b2e48SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,is_p0,NULL,reuse,&pcbddc->benign_B0);CHKERRQ(ierr); 529c263805aSStefano Zampini /* remove rows and cols from local problem */ 530c263805aSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_KEEP_NONZERO_PATTERN,PETSC_TRUE);CHKERRQ(ierr); 5314f1b2e48SStefano Zampini ierr = MatZeroRowsColumnsIS(pcbddc->local_mat,is_p0,1.0,NULL,NULL);CHKERRQ(ierr); 5324f1b2e48SStefano Zampini ierr = ISDestroy(&is_p0);CHKERRQ(ierr); 533c263805aSStefano Zampini } else { /* push */ 5344f1b2e48SStefano Zampini PetscInt i; 5354f1b2e48SStefano Zampini 5364f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 5374f1b2e48SStefano Zampini PetscScalar *B0_vals; 5384f1b2e48SStefano Zampini PetscInt *B0_cols,B0_ncol; 5394f1b2e48SStefano Zampini 5404f1b2e48SStefano Zampini ierr = MatGetRow(pcbddc->benign_B0,i,&B0_ncol,(const PetscInt**)&B0_cols,(const PetscScalar**)&B0_vals);CHKERRQ(ierr); 5414f1b2e48SStefano Zampini ierr = MatSetValues(pcbddc->local_mat,1,pcbddc->benign_p0_lidx+i,B0_ncol,B0_cols,B0_vals,INSERT_VALUES);CHKERRQ(ierr); 5424f1b2e48SStefano Zampini ierr = MatSetValues(pcbddc->local_mat,B0_ncol,B0_cols,1,pcbddc->benign_p0_lidx+1,B0_vals,INSERT_VALUES);CHKERRQ(ierr); 5434f1b2e48SStefano Zampini ierr = MatSetValue(pcbddc->local_mat,pcbddc->benign_p0_lidx[i],pcbddc->benign_p0_lidx[i],0.0,INSERT_VALUES);CHKERRQ(ierr); 5444f1b2e48SStefano Zampini ierr = MatRestoreRow(pcbddc->benign_B0,i,&B0_ncol,(const PetscInt**)&B0_cols,(const PetscScalar**)&B0_vals);CHKERRQ(ierr); 5454f1b2e48SStefano Zampini } 546c263805aSStefano Zampini ierr = MatAssemblyBegin(pcbddc->local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 547c263805aSStefano Zampini ierr = MatAssemblyEnd(pcbddc->local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 548c263805aSStefano Zampini } 549c263805aSStefano Zampini PetscFunctionReturn(0); 550c263805aSStefano Zampini } 551c263805aSStefano Zampini 552c263805aSStefano Zampini #undef __FUNCT__ 553b1b3d7a2SStefano Zampini #define __FUNCT__ "PCBDDCAdaptiveSelection" 55408122e43SStefano Zampini PetscErrorCode PCBDDCAdaptiveSelection(PC pc) 555b1b3d7a2SStefano Zampini { 556b1b3d7a2SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 55708122e43SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 55808122e43SStefano Zampini PetscBLASInt B_dummyint,B_neigs,B_ierr,B_lwork; 55908122e43SStefano Zampini PetscBLASInt *B_iwork,*B_ifail; 56008122e43SStefano Zampini PetscScalar *work,lwork; 56108122e43SStefano Zampini PetscScalar *St,*S,*eigv; 56208122e43SStefano Zampini PetscScalar *Sarray,*Starray; 56308122e43SStefano Zampini PetscReal *eigs,thresh; 5641b968477SStefano Zampini PetscInt i,nmax,nmin,nv,cum,mss,cum2,cumarray,maxneigs; 565f6f667cfSStefano Zampini PetscBool allocated_S_St; 56608122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 56708122e43SStefano Zampini PetscReal *rwork; 56808122e43SStefano Zampini #endif 569b1b3d7a2SStefano Zampini PetscErrorCode ierr; 570b1b3d7a2SStefano Zampini 571b1b3d7a2SStefano Zampini PetscFunctionBegin; 57208122e43SStefano Zampini if (!sub_schurs->use_mumps) { 57308122e43SStefano Zampini SETERRQ(PetscObjectComm((PetscObject)pc),PETSC_ERR_SUP,"Adaptive selection of constraints requires MUMPS"); 57408122e43SStefano Zampini } 57508122e43SStefano Zampini 57606a4e24aSStefano Zampini if (sub_schurs->n_subs && (!sub_schurs->is_hermitian || !sub_schurs->is_posdef)) { 57706a4e24aSStefano 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); 57806a4e24aSStefano Zampini } 57906a4e24aSStefano Zampini 580fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 581fd14bc51SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 582fd14bc51SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 583fd14bc51SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Check adaptive selection of constraints\n");CHKERRQ(ierr); 584fd14bc51SStefano Zampini ierr = PetscViewerASCIISynchronizedAllow(pcbddc->dbg_viewer,PETSC_TRUE);CHKERRQ(ierr); 585fd14bc51SStefano Zampini } 586fd14bc51SStefano Zampini 587e496cd5dSStefano Zampini if (pcbddc->dbg_flag) { 588e496cd5dSStefano 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); 589e496cd5dSStefano Zampini } 590e496cd5dSStefano Zampini 59108122e43SStefano Zampini /* max size of subsets */ 59208122e43SStefano Zampini mss = 0; 59308122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 59408122e43SStefano Zampini PetscInt subset_size; 595862806e4SStefano Zampini 59608122e43SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 59708122e43SStefano Zampini mss = PetscMax(mss,subset_size); 59808122e43SStefano Zampini } 59908122e43SStefano Zampini 60008122e43SStefano Zampini /* min/max and threshold */ 60108122e43SStefano Zampini nmax = pcbddc->adaptive_nmax > 0 ? pcbddc->adaptive_nmax : mss; 602f6f667cfSStefano Zampini nmin = pcbddc->adaptive_nmin > 0 ? pcbddc->adaptive_nmin : 0; 60308122e43SStefano Zampini nmax = PetscMax(nmin,nmax); 604f6f667cfSStefano Zampini allocated_S_St = PETSC_FALSE; 605f6f667cfSStefano Zampini if (nmin) { 606f6f667cfSStefano Zampini allocated_S_St = PETSC_TRUE; 607f6f667cfSStefano Zampini } 60808122e43SStefano Zampini 60908122e43SStefano Zampini /* allocate lapack workspace */ 61008122e43SStefano Zampini cum = cum2 = 0; 61108122e43SStefano Zampini maxneigs = 0; 61208122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 61308122e43SStefano Zampini PetscInt n,subset_size; 614f6f667cfSStefano Zampini 61508122e43SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 61608122e43SStefano Zampini n = PetscMin(subset_size,nmax); 6179162d606SStefano Zampini cum += subset_size; 6189162d606SStefano Zampini cum2 += subset_size*n; 61908122e43SStefano Zampini maxneigs = PetscMax(maxneigs,n); 62008122e43SStefano Zampini } 62108122e43SStefano Zampini if (mss) { 6229ab7bb16SStefano Zampini if (sub_schurs->is_hermitian && sub_schurs->is_posdef) { 62308122e43SStefano Zampini PetscBLASInt B_itype = 1; 62408122e43SStefano Zampini PetscBLASInt B_N = mss; 6254c6709b3SStefano Zampini PetscReal zero = 0.0; 6264c6709b3SStefano Zampini PetscReal eps = 0.0; /* dlamch? */ 62708122e43SStefano Zampini 62808122e43SStefano Zampini B_lwork = -1; 62908122e43SStefano Zampini S = NULL; 63008122e43SStefano Zampini St = NULL; 631a58a30b4SStefano Zampini eigs = NULL; 632a58a30b4SStefano Zampini eigv = NULL; 633a58a30b4SStefano Zampini B_iwork = NULL; 634a58a30b4SStefano Zampini B_ifail = NULL; 635d1710679SStefano Zampini #if defined(PETSC_USE_COMPLEX) 636d1710679SStefano Zampini rwork = NULL; 637d1710679SStefano Zampini #endif 6388bec7fa6SStefano Zampini thresh = 1.0; 63908122e43SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 64008122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 64108122e43SStefano Zampini PetscStackCallBLAS("LAPACKsygvx",LAPACKsygvx_(&B_itype,"V","V","L",&B_N,St,&B_N,S,&B_N,&zero,&thresh,&B_dummyint,&B_dummyint,&eps,&B_neigs,eigs,eigv,&B_N,&lwork,&B_lwork,rwork,B_iwork,B_ifail,&B_ierr)); 64208122e43SStefano Zampini #else 64308122e43SStefano 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)); 64408122e43SStefano Zampini #endif 64508122e43SStefano Zampini if (B_ierr != 0) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to SYGVX Lapack routine %d",(int)B_ierr); 64608122e43SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 64708122e43SStefano Zampini } else { 64808122e43SStefano Zampini /* TODO */ 64908122e43SStefano Zampini } 65008122e43SStefano Zampini } else { 65108122e43SStefano Zampini lwork = 0; 65208122e43SStefano Zampini } 65308122e43SStefano Zampini 65408122e43SStefano Zampini nv = 0; 655d62866d3SStefano 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) */ 656d62866d3SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_vertices,&nv);CHKERRQ(ierr); 65708122e43SStefano Zampini } 6584c6709b3SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lwork),&B_lwork);CHKERRQ(ierr); 659f6f667cfSStefano Zampini if (allocated_S_St) { 660f6f667cfSStefano Zampini ierr = PetscMalloc2(mss*mss,&S,mss*mss,&St);CHKERRQ(ierr); 661f6f667cfSStefano Zampini } 662f6f667cfSStefano Zampini ierr = PetscMalloc5(mss*mss,&eigv,mss,&eigs,B_lwork,&work,5*mss,&B_iwork,mss,&B_ifail);CHKERRQ(ierr); 66308122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 66408122e43SStefano Zampini ierr = PetscMalloc1(7*mss,&rwork);CHKERRQ(ierr); 66508122e43SStefano Zampini #endif 6669162d606SStefano Zampini ierr = PetscMalloc5(nv+sub_schurs->n_subs,&pcbddc->adaptive_constraints_n, 6679162d606SStefano Zampini nv+sub_schurs->n_subs+1,&pcbddc->adaptive_constraints_idxs_ptr, 6689162d606SStefano Zampini nv+sub_schurs->n_subs+1,&pcbddc->adaptive_constraints_data_ptr, 66908122e43SStefano Zampini nv+cum,&pcbddc->adaptive_constraints_idxs, 6709162d606SStefano Zampini nv+cum2,&pcbddc->adaptive_constraints_data);CHKERRQ(ierr); 67108122e43SStefano Zampini ierr = PetscMemzero(pcbddc->adaptive_constraints_n,(nv+sub_schurs->n_subs)*sizeof(PetscInt));CHKERRQ(ierr); 67208122e43SStefano Zampini 67308122e43SStefano Zampini maxneigs = 0; 67408122e43SStefano Zampini cum = cum2 = cumarray = 0; 6759162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr[0] = 0; 6769162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr[0] = 0; 677d62866d3SStefano Zampini if (sub_schurs->is_vertices && pcbddc->use_vertices) { 67808122e43SStefano Zampini const PetscInt *idxs; 67908122e43SStefano Zampini 680d62866d3SStefano Zampini ierr = ISGetIndices(sub_schurs->is_vertices,&idxs);CHKERRQ(ierr); 68108122e43SStefano Zampini for (cum=0;cum<nv;cum++) { 68208122e43SStefano Zampini pcbddc->adaptive_constraints_n[cum] = 1; 68308122e43SStefano Zampini pcbddc->adaptive_constraints_idxs[cum] = idxs[cum]; 68408122e43SStefano Zampini pcbddc->adaptive_constraints_data[cum] = 1.0; 6859162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr[cum+1] = pcbddc->adaptive_constraints_idxs_ptr[cum]+1; 6869162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr[cum+1] = pcbddc->adaptive_constraints_data_ptr[cum]+1; 68708122e43SStefano Zampini } 68808122e43SStefano Zampini cum2 = cum; 689d62866d3SStefano Zampini ierr = ISRestoreIndices(sub_schurs->is_vertices,&idxs);CHKERRQ(ierr); 69008122e43SStefano Zampini } 69108122e43SStefano Zampini 69208122e43SStefano Zampini if (mss) { /* multilevel */ 69308122e43SStefano Zampini ierr = MatSeqAIJGetArray(sub_schurs->sum_S_Ej_inv_all,&Sarray);CHKERRQ(ierr); 69408122e43SStefano Zampini ierr = MatSeqAIJGetArray(sub_schurs->sum_S_Ej_tilda_all,&Starray);CHKERRQ(ierr); 69508122e43SStefano Zampini } 69608122e43SStefano Zampini 697ffd830a3SStefano Zampini thresh = pcbddc->adaptive_threshold; 69808122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 69908122e43SStefano Zampini const PetscInt *idxs; 700f6f667cfSStefano Zampini PetscReal infty = PETSC_MAX_REAL; 701862806e4SStefano Zampini PetscInt j,subset_size,eigs_start = 0; 70208122e43SStefano Zampini PetscBLASInt B_N; 703aff50787SStefano Zampini PetscBool same_data = PETSC_FALSE; 70408122e43SStefano Zampini 705862806e4SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 706ffd830a3SStefano Zampini ierr = ISGetIndices(sub_schurs->is_subs[i],&idxs);CHKERRQ(ierr); 707f6f667cfSStefano Zampini ierr = PetscBLASIntCast(subset_size,&B_N);CHKERRQ(ierr); 708f6f667cfSStefano Zampini if (allocated_S_St) { /* S and S_t should be copied since we could need them later */ 7099ab7bb16SStefano Zampini if (sub_schurs->is_hermitian) { 710aff50787SStefano Zampini PetscInt j,k; 711aff50787SStefano Zampini if (sub_schurs->n_subs == 1) { /* zeroing memory to use PetscMemcmp later */ 712aff50787SStefano Zampini ierr = PetscMemzero(S,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 713aff50787SStefano Zampini ierr = PetscMemzero(St,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 71408122e43SStefano Zampini } 71508122e43SStefano Zampini for (j=0;j<subset_size;j++) { 716aff50787SStefano Zampini for (k=j;k<subset_size;k++) { 717aff50787SStefano Zampini S [j*subset_size+k] = Sarray [cumarray+j*subset_size+k]; 718aff50787SStefano Zampini St[j*subset_size+k] = Starray[cumarray+j*subset_size+k]; 719aff50787SStefano Zampini } 72008122e43SStefano Zampini } 72108122e43SStefano Zampini } else { 72208122e43SStefano Zampini ierr = PetscMemcpy(S,Sarray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 72308122e43SStefano Zampini ierr = PetscMemcpy(St,Starray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 72408122e43SStefano Zampini } 7258bec7fa6SStefano Zampini } else { 726f6f667cfSStefano Zampini S = Sarray + cumarray; 727f6f667cfSStefano Zampini St = Starray + cumarray; 7288bec7fa6SStefano Zampini } 72908122e43SStefano Zampini 730aff50787SStefano Zampini /* see if we can save some work */ 731aff50787SStefano Zampini if (sub_schurs->n_subs == 1) { 732aff50787SStefano Zampini ierr = PetscMemcmp(S,St,subset_size*subset_size*sizeof(PetscScalar),&same_data);CHKERRQ(ierr); 733aff50787SStefano Zampini } 734aff50787SStefano Zampini 735aff50787SStefano Zampini if (same_data) { /* there's no need of constraints here, deluxe scaling is enough */ 736aff50787SStefano Zampini B_neigs = 0; 737aff50787SStefano Zampini } else { 7389ab7bb16SStefano Zampini if (sub_schurs->is_hermitian && sub_schurs->is_posdef) { 73908122e43SStefano Zampini PetscBLASInt B_itype = 1; 740f6f667cfSStefano Zampini PetscBLASInt B_IL, B_IU; 7414c6709b3SStefano Zampini PetscReal eps = -1.0; /* dlamch? */ 7429552c7c7SStefano Zampini PetscInt nmin_s; 74308122e43SStefano Zampini 744fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 7458bec7fa6SStefano 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]]); 746fd14bc51SStefano Zampini } 747d16cbb6bSStefano Zampini 74808122e43SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 749d16cbb6bSStefano Zampini if (thresh > 1.+PETSC_SMALL) { 750d16cbb6bSStefano Zampini 751d16cbb6bSStefano Zampini /* ask for eigenvalues larger than thresh */ 75208122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 753f6f667cfSStefano Zampini PetscStackCallBLAS("LAPACKsygvx",LAPACKsygvx_(&B_itype,"V","V","L",&B_N,St,&B_N,S,&B_N,&thresh,&infty,&B_IL,&B_IU,&eps,&B_neigs,eigs,eigv,&B_N,work,&B_lwork,rwork,B_iwork,B_ifail,&B_ierr)); 75408122e43SStefano Zampini #else 755f6f667cfSStefano 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)); 75608122e43SStefano Zampini #endif 757d16cbb6bSStefano Zampini } else { 758d16cbb6bSStefano Zampini B_IU = PetscMax(1,PetscMin(B_N,nmax)); 759d16cbb6bSStefano Zampini B_IL = 1; 760d16cbb6bSStefano Zampini #if defined(PETSC_USE_COMPLEX) 761d16cbb6bSStefano Zampini PetscStackCallBLAS("LAPACKsygvx",LAPACKsygvx_(&B_itype,"V","I","L",&B_N,St,&B_N,S,&B_N,&thresh,&infty,&B_IL,&B_IU,&eps,&B_neigs,eigs,eigv,&B_N,work,&B_lwork,rwork,B_iwork,B_ifail,&B_ierr)); 762d16cbb6bSStefano Zampini #else 763d16cbb6bSStefano 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)); 764d16cbb6bSStefano Zampini #endif 765d16cbb6bSStefano Zampini } 76608122e43SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 76708122e43SStefano Zampini if (B_ierr) { 76808122e43SStefano Zampini if (B_ierr < 0 ) { 76908122e43SStefano Zampini SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYGVX Lapack routine: illegal value for argument %d",-(int)B_ierr); 77008122e43SStefano Zampini } else if (B_ierr <= B_N) { 77108122e43SStefano Zampini SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYGVX Lapack routine: %d eigenvalues failed to converge",(int)B_ierr); 77208122e43SStefano Zampini } else { 7739552c7c7SStefano 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); 77408122e43SStefano Zampini } 77508122e43SStefano Zampini } 77608122e43SStefano Zampini 77708122e43SStefano Zampini if (B_neigs > nmax) { 778fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 779fd14bc51SStefano Zampini PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," found %d eigs, more than maximum required %d.\n",B_neigs,nmax); 780fd14bc51SStefano Zampini } 781f6f667cfSStefano Zampini eigs_start = B_neigs -nmax; 78208122e43SStefano Zampini B_neigs = nmax; 78308122e43SStefano Zampini } 78408122e43SStefano Zampini 7859552c7c7SStefano Zampini nmin_s = PetscMin(nmin,B_N); 7869552c7c7SStefano Zampini if (B_neigs < nmin_s) { 78708122e43SStefano Zampini PetscBLASInt B_neigs2; 78808122e43SStefano Zampini 789f6f667cfSStefano Zampini B_IU = B_N - B_neigs; 790f6f667cfSStefano Zampini B_IL = B_N - nmin_s + 1; 791fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 792fd14bc51SStefano 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); 793fd14bc51SStefano Zampini } 7949ab7bb16SStefano Zampini if (sub_schurs->is_hermitian) { 79508122e43SStefano Zampini PetscInt j; 79608122e43SStefano Zampini for (j=0;j<subset_size;j++) { 79708122e43SStefano Zampini ierr = PetscMemcpy(S+j*(subset_size+1),Sarray+cumarray+j*(subset_size+1),(subset_size-j)*sizeof(PetscScalar));CHKERRQ(ierr); 79808122e43SStefano Zampini } 79908122e43SStefano Zampini for (j=0;j<subset_size;j++) { 80008122e43SStefano Zampini ierr = PetscMemcpy(St+j*(subset_size+1),Starray+cumarray+j*(subset_size+1),(subset_size-j)*sizeof(PetscScalar));CHKERRQ(ierr); 80108122e43SStefano Zampini } 80208122e43SStefano Zampini } else { 80308122e43SStefano Zampini ierr = PetscMemcpy(S,Sarray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 80408122e43SStefano Zampini ierr = PetscMemcpy(St,Starray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 80508122e43SStefano Zampini } 80608122e43SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 80708122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 808f6f667cfSStefano Zampini PetscStackCallBLAS("LAPACKsygvx",LAPACKsygvx_(&B_itype,"V","I","L",&B_N,St,&B_N,S,&B_N,&thresh,&infty,&B_IL,&B_IU,&eps,&B_neigs2,eigs+B_neigs,eigv+B_neigs*subset_size,&B_N,work,&B_lwork,rwork,B_iwork,B_ifail,&B_ierr)); 80908122e43SStefano Zampini #else 810f6f667cfSStefano 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)); 81108122e43SStefano Zampini #endif 81208122e43SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 81308122e43SStefano Zampini B_neigs += B_neigs2; 81408122e43SStefano Zampini } 81508122e43SStefano Zampini if (B_ierr) { 81608122e43SStefano Zampini if (B_ierr < 0 ) { 81708122e43SStefano Zampini SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYGVX Lapack routine: illegal value for argument %d",-(int)B_ierr); 81808122e43SStefano Zampini } else if (B_ierr <= B_N) { 81908122e43SStefano Zampini SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYGVX Lapack routine: %d eigenvalues failed to converge",(int)B_ierr); 82008122e43SStefano Zampini } else { 8219552c7c7SStefano 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); 82208122e43SStefano Zampini } 82308122e43SStefano Zampini } 824fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 825ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," -> Got %d eigs\n",B_neigs);CHKERRQ(ierr); 82608122e43SStefano Zampini for (j=0;j<B_neigs;j++) { 82708122e43SStefano Zampini if (eigs[j] == 0.0) { 828ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," Inf\n");CHKERRQ(ierr); 82908122e43SStefano Zampini } else { 830ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," %1.6e\n",eigs[j+eigs_start]);CHKERRQ(ierr); 831fd14bc51SStefano Zampini } 83208122e43SStefano Zampini } 83308122e43SStefano Zampini } 83408122e43SStefano Zampini } else { 83508122e43SStefano Zampini /* TODO */ 83608122e43SStefano Zampini } 837aff50787SStefano Zampini } 8388bec7fa6SStefano Zampini maxneigs = PetscMax(B_neigs,maxneigs); 8398bec7fa6SStefano Zampini pcbddc->adaptive_constraints_n[i+nv] = B_neigs; 8409162d606SStefano Zampini if (B_neigs) { 8419162d606SStefano 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); 842fd14bc51SStefano Zampini 843fd14bc51SStefano Zampini if (pcbddc->dbg_flag > 1) { 8449552c7c7SStefano Zampini PetscInt ii; 8459552c7c7SStefano Zampini for (ii=0;ii<B_neigs;ii++) { 846ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," -> Eigenvector %d/%d (%d)\n",ii,B_neigs,B_N);CHKERRQ(ierr); 8479552c7c7SStefano Zampini for (j=0;j<B_N;j++) { 848ac47001eSStefano Zampini #if defined(PETSC_USE_COMPLEX) 849ac47001eSStefano Zampini PetscReal r = PetscRealPart(pcbddc->adaptive_constraints_data[ii*subset_size+j+pcbddc->adaptive_constraints_data_ptr[cum]]); 850ac47001eSStefano Zampini PetscReal c = PetscImaginaryPart(pcbddc->adaptive_constraints_data[ii*subset_size+j+pcbddc->adaptive_constraints_data_ptr[cum]]); 851ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," %1.4e + %1.4e i\n",r,c);CHKERRQ(ierr); 852ac47001eSStefano Zampini #else 853ac47001eSStefano 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); 854ac47001eSStefano Zampini #endif 8559552c7c7SStefano Zampini } 8569552c7c7SStefano Zampini } 857fd14bc51SStefano Zampini } 85808122e43SStefano Zampini #if 0 8599162d606SStefano Zampini for (j=0;j<B_neigs;j++) { 86008122e43SStefano Zampini PetscBLASInt Blas_N,Blas_one = 1.0; 86108122e43SStefano Zampini PetscScalar norm; 86208122e43SStefano Zampini ierr = PetscBLASIntCast(subset_size,&Blas_N);CHKERRQ(ierr); 8639162d606SStefano Zampini PetscStackCallBLAS("BLASdot",norm = BLASdot_(&Blas_N,pcbddc->adaptive_constraints_data+pcbddc->adaptive_constraints_data_ptr[cum]+j*subset_size, 8649162d606SStefano Zampini &Blas_one,pcbddc->adaptive_constraints_data+pcbddc->adaptive_constraints_data_ptr[cum]+j*subset_size,&Blas_one)); 86508122e43SStefano Zampini if (pcbddc->adaptive_constraints_data[cum2] > 0.0) { 86608122e43SStefano Zampini norm = 1.0/PetscSqrtReal(PetscRealPart(norm)); 86708122e43SStefano Zampini } else { 86808122e43SStefano Zampini norm = -1.0/PetscSqrtReal(PetscRealPart(norm)); 86908122e43SStefano Zampini } 8709162d606SStefano Zampini PetscStackCallBLAS("BLASscal",BLASscal_(&Blas_N,&norm,pcbddc->adaptive_constraints_data+pcbddc->adaptive_constraints_data_ptr[cum]+j*subset_size,&Blas_one)); 871b1b3d7a2SStefano Zampini } 872b1b3d7a2SStefano Zampini #endif 8739162d606SStefano Zampini ierr = PetscMemcpy(pcbddc->adaptive_constraints_idxs+pcbddc->adaptive_constraints_idxs_ptr[cum],idxs,subset_size*sizeof(PetscInt));CHKERRQ(ierr); 8749162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr[cum+1] = pcbddc->adaptive_constraints_idxs_ptr[cum] + subset_size; 8759162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr[cum+1] = pcbddc->adaptive_constraints_data_ptr[cum] + subset_size*B_neigs; 8769162d606SStefano Zampini cum++; 87708122e43SStefano Zampini } 87808122e43SStefano Zampini ierr = ISRestoreIndices(sub_schurs->is_subs[i],&idxs);CHKERRQ(ierr); 87908122e43SStefano Zampini /* shift for next computation */ 88008122e43SStefano Zampini cumarray += subset_size*subset_size; 88108122e43SStefano Zampini } 882fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 883fd14bc51SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 884fd14bc51SStefano Zampini } 88508122e43SStefano Zampini 88608122e43SStefano Zampini if (mss) { 88708122e43SStefano Zampini ierr = MatSeqAIJRestoreArray(sub_schurs->sum_S_Ej_inv_all,&Sarray);CHKERRQ(ierr); 88808122e43SStefano Zampini ierr = MatSeqAIJRestoreArray(sub_schurs->sum_S_Ej_tilda_all,&Starray);CHKERRQ(ierr); 889f6f667cfSStefano Zampini /* destroy matrices (junk) */ 890f6f667cfSStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_inv_all);CHKERRQ(ierr); 891f6f667cfSStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_tilda_all);CHKERRQ(ierr); 89208122e43SStefano Zampini } 893f6f667cfSStefano Zampini if (allocated_S_St) { 894f6f667cfSStefano Zampini ierr = PetscFree2(S,St);CHKERRQ(ierr); 895f6f667cfSStefano Zampini } 896f6f667cfSStefano Zampini ierr = PetscFree5(eigv,eigs,work,B_iwork,B_ifail);CHKERRQ(ierr); 89708122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 89808122e43SStefano Zampini ierr = PetscFree(rwork);CHKERRQ(ierr); 89908122e43SStefano Zampini #endif 90008122e43SStefano Zampini if (pcbddc->dbg_flag) { 9011b968477SStefano Zampini PetscInt maxneigs_r; 90208122e43SStefano Zampini ierr = MPI_Allreduce(&maxneigs,&maxneigs_r,1,MPIU_INT,MPI_MAX,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 9039b28b941SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Maximum number of constraints per cc %d\n",maxneigs_r);CHKERRQ(ierr); 90408122e43SStefano Zampini } 90508122e43SStefano Zampini PetscFunctionReturn(0); 90608122e43SStefano Zampini } 907b1b3d7a2SStefano Zampini 908674ae819SStefano Zampini #undef __FUNCT__ 909c8587f34SStefano Zampini #define __FUNCT__ "PCBDDCSetUpSolvers" 910c8587f34SStefano Zampini PetscErrorCode PCBDDCSetUpSolvers(PC pc) 911c8587f34SStefano Zampini { 912c8587f34SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 9138629588bSStefano Zampini PetscScalar *coarse_submat_vals; 914c8587f34SStefano Zampini PetscErrorCode ierr; 915c8587f34SStefano Zampini 916c8587f34SStefano Zampini PetscFunctionBegin; 917f4ddd8eeSStefano Zampini /* Setup local scatters R_to_B and (optionally) R_to_D */ 9185e8657edSStefano Zampini /* PCBDDCSetUpLocalWorkVectors should be called first! */ 919c8587f34SStefano Zampini ierr = PCBDDCSetUpLocalScatters(pc);CHKERRQ(ierr); 920c8587f34SStefano Zampini 921684f6988SStefano Zampini /* Setup local neumann solver ksp_R */ 9220fccc4e9SStefano Zampini /* PCBDDCSetUpLocalScatters should be called first! */ 923684f6988SStefano Zampini ierr = PCBDDCSetUpLocalSolvers(pc,PETSC_FALSE,PETSC_TRUE);CHKERRQ(ierr); 924c8587f34SStefano Zampini 925c8587f34SStefano Zampini /* Change global null space passed in by the user if change of basis has been requested */ 926b9b85e73SStefano Zampini if (pcbddc->NullSpace && pcbddc->ChangeOfBasisMatrix) { 927c8587f34SStefano Zampini ierr = PCBDDCNullSpaceAdaptGlobal(pc);CHKERRQ(ierr); 928c8587f34SStefano Zampini } 929c8587f34SStefano Zampini 9308629588bSStefano Zampini /* 9318629588bSStefano Zampini Setup local correction and local part of coarse basis. 9328629588bSStefano Zampini Gives back the dense local part of the coarse matrix in column major ordering 9338629588bSStefano Zampini */ 93447f4ddc3SStefano Zampini ierr = PCBDDCSetUpCorrection(pc,&coarse_submat_vals);CHKERRQ(ierr); 9358629588bSStefano Zampini 9368629588bSStefano Zampini /* Compute total number of coarse nodes and setup coarse solver */ 9378629588bSStefano Zampini ierr = PCBDDCSetUpCoarseSolver(pc,coarse_submat_vals);CHKERRQ(ierr); 9388629588bSStefano Zampini 9398629588bSStefano Zampini /* free */ 9408629588bSStefano Zampini ierr = PetscFree(coarse_submat_vals);CHKERRQ(ierr); 941c8587f34SStefano Zampini PetscFunctionReturn(0); 942c8587f34SStefano Zampini } 943c8587f34SStefano Zampini 944c8587f34SStefano Zampini #undef __FUNCT__ 945674ae819SStefano Zampini #define __FUNCT__ "PCBDDCResetCustomization" 946674ae819SStefano Zampini PetscErrorCode PCBDDCResetCustomization(PC pc) 947674ae819SStefano Zampini { 948674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 949674ae819SStefano Zampini PetscErrorCode ierr; 950674ae819SStefano Zampini 951674ae819SStefano Zampini PetscFunctionBegin; 952674ae819SStefano Zampini ierr = PCBDDCGraphResetCSR(pcbddc->mat_graph);CHKERRQ(ierr); 953674ae819SStefano Zampini ierr = ISDestroy(&pcbddc->user_primal_vertices);CHKERRQ(ierr); 95430368db7SStefano Zampini ierr = ISDestroy(&pcbddc->user_primal_vertices_local);CHKERRQ(ierr); 955674ae819SStefano Zampini ierr = MatNullSpaceDestroy(&pcbddc->NullSpace);CHKERRQ(ierr); 956674ae819SStefano Zampini ierr = ISDestroy(&pcbddc->NeumannBoundaries);CHKERRQ(ierr); 957785d1243SStefano Zampini ierr = ISDestroy(&pcbddc->NeumannBoundariesLocal);CHKERRQ(ierr); 958674ae819SStefano Zampini ierr = ISDestroy(&pcbddc->DirichletBoundaries);CHKERRQ(ierr); 959f4ddd8eeSStefano Zampini ierr = MatNullSpaceDestroy(&pcbddc->onearnullspace);CHKERRQ(ierr); 960f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->onearnullvecs_state);CHKERRQ(ierr); 961785d1243SStefano Zampini ierr = ISDestroy(&pcbddc->DirichletBoundariesLocal);CHKERRQ(ierr); 96263602bcaSStefano Zampini ierr = PCBDDCSetDofsSplitting(pc,0,NULL);CHKERRQ(ierr); 96363602bcaSStefano Zampini ierr = PCBDDCSetDofsSplittingLocal(pc,0,NULL);CHKERRQ(ierr); 964674ae819SStefano Zampini PetscFunctionReturn(0); 965674ae819SStefano Zampini } 966674ae819SStefano Zampini 967674ae819SStefano Zampini #undef __FUNCT__ 968674ae819SStefano Zampini #define __FUNCT__ "PCBDDCResetTopography" 969674ae819SStefano Zampini PetscErrorCode PCBDDCResetTopography(PC pc) 970674ae819SStefano Zampini { 971674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 9724f1b2e48SStefano Zampini PetscInt i; 973674ae819SStefano Zampini PetscErrorCode ierr; 974674ae819SStefano Zampini 975674ae819SStefano Zampini PetscFunctionBegin; 976b9b85e73SStefano Zampini ierr = MatDestroy(&pcbddc->user_ChangeOfBasisMatrix);CHKERRQ(ierr); 977674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 978674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 979674ae819SStefano Zampini ierr = PCBDDCGraphReset(pcbddc->mat_graph);CHKERRQ(ierr); 9804f1b2e48SStefano Zampini for (i=0;i<pcbddc->n_local_subs;i++) { 9814f1b2e48SStefano Zampini ierr = ISDestroy(&pcbddc->local_subs[i]);CHKERRQ(ierr); 9824f1b2e48SStefano Zampini } 9834f1b2e48SStefano Zampini ierr = PetscFree(pcbddc->local_subs);CHKERRQ(ierr); 984b96c3477SStefano Zampini ierr = PCBDDCSubSchursReset(pcbddc->sub_schurs);CHKERRQ(ierr); 985674ae819SStefano Zampini PetscFunctionReturn(0); 986674ae819SStefano Zampini } 987674ae819SStefano Zampini 988674ae819SStefano Zampini #undef __FUNCT__ 989674ae819SStefano Zampini #define __FUNCT__ "PCBDDCResetSolvers" 990674ae819SStefano Zampini PetscErrorCode PCBDDCResetSolvers(PC pc) 991674ae819SStefano Zampini { 992674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 99306656605SStefano Zampini PetscScalar *array; 994674ae819SStefano Zampini PetscErrorCode ierr; 995674ae819SStefano Zampini 996674ae819SStefano Zampini PetscFunctionBegin; 997674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->coarse_vec);CHKERRQ(ierr); 99858da7f69SStefano Zampini if (pcbddc->coarse_phi_B) { 99906656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&array);CHKERRQ(ierr); 100006656605SStefano Zampini ierr = PetscFree(array);CHKERRQ(ierr); 100158da7f69SStefano Zampini } 1002674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_B);CHKERRQ(ierr); 1003674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_D);CHKERRQ(ierr); 100415aaf578SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_B);CHKERRQ(ierr); 100515aaf578SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_D);CHKERRQ(ierr); 1006674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec1_P);CHKERRQ(ierr); 1007674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec1_C);CHKERRQ(ierr); 1008674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat2);CHKERRQ(ierr); 100906656605SStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat1);CHKERRQ(ierr); 1010674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec1_R);CHKERRQ(ierr); 1011674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec2_R);CHKERRQ(ierr); 10128ce42a96SStefano Zampini ierr = ISDestroy(&pcbddc->is_R_local);CHKERRQ(ierr); 1013674ae819SStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_B);CHKERRQ(ierr); 1014674ae819SStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_D);CHKERRQ(ierr); 1015674ae819SStefano Zampini ierr = VecScatterDestroy(&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 1016f4ddd8eeSStefano Zampini ierr = KSPDestroy(&pcbddc->ksp_D);CHKERRQ(ierr); 1017f4ddd8eeSStefano Zampini ierr = KSPDestroy(&pcbddc->ksp_R);CHKERRQ(ierr); 1018f4ddd8eeSStefano Zampini ierr = KSPDestroy(&pcbddc->coarse_ksp);CHKERRQ(ierr); 1019f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 1020727cdba6SStefano Zampini ierr = PetscFree(pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 10210e6343abSStefano Zampini ierr = PetscFree2(pcbddc->local_primal_ref_node,pcbddc->local_primal_ref_mult);CHKERRQ(ierr); 1022f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->global_primal_indices);CHKERRQ(ierr); 102370cf5478SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling);CHKERRQ(ierr); 10246e683305SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 102581d14e9dSStefano Zampini ierr = MatDestroy(&pcbddc->benign_change);CHKERRQ(ierr); 10260369aaf7SStefano Zampini ierr = VecDestroy(&pcbddc->benign_vec);CHKERRQ(ierr); 10278b9f24d4SStefano Zampini ierr = MatDestroy(&pcbddc->benign_original_mat);CHKERRQ(ierr); 10284f1b2e48SStefano Zampini ierr = MatDestroy(&pcbddc->benign_original_mat);CHKERRQ(ierr); 10294f1b2e48SStefano Zampini ierr = MatDestroy(&pcbddc->benign_B0);CHKERRQ(ierr); 10308b9f24d4SStefano Zampini ierr = PetscSFDestroy(&pcbddc->benign_sf);CHKERRQ(ierr); 10314f1b2e48SStefano Zampini ierr = PetscFree3(pcbddc->benign_p0_lidx,pcbddc->benign_p0_gidx,pcbddc->benign_p0);CHKERRQ(ierr); 1032674ae819SStefano Zampini PetscFunctionReturn(0); 1033674ae819SStefano Zampini } 1034674ae819SStefano Zampini 1035674ae819SStefano Zampini #undef __FUNCT__ 1036f4ddd8eeSStefano Zampini #define __FUNCT__ "PCBDDCSetUpLocalWorkVectors" 1037f4ddd8eeSStefano Zampini PetscErrorCode PCBDDCSetUpLocalWorkVectors(PC pc) 10386bfb1811SStefano Zampini { 10396bfb1811SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 10406bfb1811SStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 10416bfb1811SStefano Zampini VecType impVecType; 10424f1b2e48SStefano Zampini PetscInt n_constraints,n_R,old_size; 10436bfb1811SStefano Zampini PetscErrorCode ierr; 10446bfb1811SStefano Zampini 10456bfb1811SStefano Zampini PetscFunctionBegin; 1046f4ddd8eeSStefano Zampini if (!pcbddc->ConstraintMatrix) { 1047019a44ceSStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"BDDC Constraint matrix has not been created"); 1048f4ddd8eeSStefano Zampini } 1049e7b262bdSStefano Zampini /* get sizes */ 10504f1b2e48SStefano Zampini n_constraints = pcbddc->local_primal_size - pcbddc->benign_n - pcbddc->n_vertices; 1051b371cd4fSStefano Zampini n_R = pcis->n - pcbddc->n_vertices; 10526bfb1811SStefano Zampini ierr = VecGetType(pcis->vec1_N,&impVecType);CHKERRQ(ierr); 1053e7b262bdSStefano Zampini /* local work vectors (try to avoid unneeded work)*/ 1054e7b262bdSStefano Zampini /* R nodes */ 1055e7b262bdSStefano Zampini old_size = -1; 1056e7b262bdSStefano Zampini if (pcbddc->vec1_R) { 1057e7b262bdSStefano Zampini ierr = VecGetSize(pcbddc->vec1_R,&old_size);CHKERRQ(ierr); 1058e7b262bdSStefano Zampini } 1059e7b262bdSStefano Zampini if (n_R != old_size) { 1060e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec1_R);CHKERRQ(ierr); 1061e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec2_R);CHKERRQ(ierr); 10626bfb1811SStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_R);CHKERRQ(ierr); 10636bfb1811SStefano Zampini ierr = VecSetSizes(pcbddc->vec1_R,PETSC_DECIDE,n_R);CHKERRQ(ierr); 10646bfb1811SStefano Zampini ierr = VecSetType(pcbddc->vec1_R,impVecType);CHKERRQ(ierr); 10656bfb1811SStefano Zampini ierr = VecDuplicate(pcbddc->vec1_R,&pcbddc->vec2_R);CHKERRQ(ierr); 1066e7b262bdSStefano Zampini } 1067e7b262bdSStefano Zampini /* local primal dofs */ 1068e7b262bdSStefano Zampini old_size = -1; 1069e7b262bdSStefano Zampini if (pcbddc->vec1_P) { 1070e7b262bdSStefano Zampini ierr = VecGetSize(pcbddc->vec1_P,&old_size);CHKERRQ(ierr); 1071e7b262bdSStefano Zampini } 1072e9189074SStefano Zampini if (pcbddc->local_primal_size != old_size) { 1073e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec1_P);CHKERRQ(ierr); 107483b7ccabSStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_P);CHKERRQ(ierr); 1075e9189074SStefano Zampini ierr = VecSetSizes(pcbddc->vec1_P,PETSC_DECIDE,pcbddc->local_primal_size);CHKERRQ(ierr); 10766bfb1811SStefano Zampini ierr = VecSetType(pcbddc->vec1_P,impVecType);CHKERRQ(ierr); 1077e7b262bdSStefano Zampini } 1078e7b262bdSStefano Zampini /* local explicit constraints */ 1079e7b262bdSStefano Zampini old_size = -1; 1080e7b262bdSStefano Zampini if (pcbddc->vec1_C) { 1081e7b262bdSStefano Zampini ierr = VecGetSize(pcbddc->vec1_C,&old_size);CHKERRQ(ierr); 1082e7b262bdSStefano Zampini } 1083e7b262bdSStefano Zampini if (n_constraints && n_constraints != old_size) { 1084e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec1_C);CHKERRQ(ierr); 108583b7ccabSStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_C);CHKERRQ(ierr); 108683b7ccabSStefano Zampini ierr = VecSetSizes(pcbddc->vec1_C,PETSC_DECIDE,n_constraints);CHKERRQ(ierr); 108783b7ccabSStefano Zampini ierr = VecSetType(pcbddc->vec1_C,impVecType);CHKERRQ(ierr); 108883b7ccabSStefano Zampini } 10896bfb1811SStefano Zampini PetscFunctionReturn(0); 10906bfb1811SStefano Zampini } 10916bfb1811SStefano Zampini 10926bfb1811SStefano Zampini #undef __FUNCT__ 109347f4ddc3SStefano Zampini #define __FUNCT__ "PCBDDCSetUpCorrection" 109447f4ddc3SStefano Zampini PetscErrorCode PCBDDCSetUpCorrection(PC pc, PetscScalar **coarse_submat_vals_n) 109588ebb749SStefano Zampini { 109625084f0cSStefano Zampini PetscErrorCode ierr; 109725084f0cSStefano Zampini /* pointers to pcis and pcbddc */ 109888ebb749SStefano Zampini PC_IS* pcis = (PC_IS*)pc->data; 109988ebb749SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 1100d62866d3SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 110125084f0cSStefano Zampini /* submatrices of local problem */ 110280677318SStefano Zampini Mat A_RV,A_VR,A_VV,local_auxmat2_R; 1103019a44ceSStefano Zampini /* submatrices of benign trick */ 1104d16cbb6bSStefano Zampini Mat B0_V = NULL; 110506656605SStefano Zampini /* submatrices of local coarse problem */ 110606656605SStefano Zampini Mat S_VV,S_CV,S_VC,S_CC; 110725084f0cSStefano Zampini /* working matrices */ 110806656605SStefano Zampini Mat C_CR; 110925084f0cSStefano Zampini /* additional working stuff */ 111006656605SStefano Zampini PC pc_R; 11114f1b2e48SStefano Zampini Mat F; 111206656605SStefano Zampini PetscBool isLU,isCHOL,isILU; 111306656605SStefano Zampini 111425084f0cSStefano Zampini PetscScalar *coarse_submat_vals; /* TODO: use a PETSc matrix */ 111506656605SStefano Zampini PetscScalar *work; 111606656605SStefano Zampini PetscInt *idx_V_B; 1117ffd830a3SStefano Zampini PetscInt lda_rhs,n,n_vertices,n_constraints,*p0_lidx_I; 111806656605SStefano Zampini PetscInt i,n_R,n_D,n_B; 1119ffd830a3SStefano Zampini 112025084f0cSStefano Zampini /* some shortcuts to scalars */ 112106656605SStefano Zampini PetscScalar one=1.0,m_one=-1.0; 112288ebb749SStefano Zampini 112388ebb749SStefano Zampini PetscFunctionBegin; 1124ffd830a3SStefano Zampini if (!pcbddc->symmetric_primal && pcbddc->benign_n) { 1125ffd830a3SStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Non-symmetric primal basis computation with benign trick not yet implemented"); 1126ffd830a3SStefano Zampini } 1127ffd830a3SStefano Zampini 1128ffd830a3SStefano Zampini /* Set Non-overlapping dimensions */ 1129b371cd4fSStefano Zampini n_vertices = pcbddc->n_vertices; 11304f1b2e48SStefano Zampini n_constraints = pcbddc->local_primal_size - pcbddc->benign_n - n_vertices; 1131b371cd4fSStefano Zampini n_B = pcis->n_B; 1132b371cd4fSStefano Zampini n_D = pcis->n - n_B; 113388ebb749SStefano Zampini n_R = pcis->n - n_vertices; 113488ebb749SStefano Zampini 113588ebb749SStefano Zampini /* vertices in boundary numbering */ 1136785e854fSJed Brown ierr = PetscMalloc1(n_vertices,&idx_V_B);CHKERRQ(ierr); 11370e6343abSStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,n_vertices,pcbddc->local_primal_ref_node,&i,idx_V_B);CHKERRQ(ierr); 113888ebb749SStefano Zampini if (i != n_vertices) { 113922d5777bSStefano Zampini SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Error in boundary numbering for BDDC vertices! %d != %d\n",n_vertices,i); 114088ebb749SStefano Zampini } 114188ebb749SStefano Zampini 114206656605SStefano Zampini /* Subdomain contribution (Non-overlapping) to coarse matrix */ 1143019a44ceSStefano Zampini ierr = PetscCalloc1(pcbddc->local_primal_size*pcbddc->local_primal_size,&coarse_submat_vals);CHKERRQ(ierr); 114406656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_vertices,n_vertices,coarse_submat_vals,&S_VV);CHKERRQ(ierr); 114506656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_VV,pcbddc->local_primal_size);CHKERRQ(ierr); 114606656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_constraints,n_vertices,coarse_submat_vals+n_vertices,&S_CV);CHKERRQ(ierr); 114706656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_CV,pcbddc->local_primal_size);CHKERRQ(ierr); 114806656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_vertices,n_constraints,coarse_submat_vals+pcbddc->local_primal_size*n_vertices,&S_VC);CHKERRQ(ierr); 114906656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_VC,pcbddc->local_primal_size);CHKERRQ(ierr); 115006656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_constraints,n_constraints,coarse_submat_vals+(pcbddc->local_primal_size+1)*n_vertices,&S_CC);CHKERRQ(ierr); 115106656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_CC,pcbddc->local_primal_size);CHKERRQ(ierr); 115206656605SStefano Zampini 115306656605SStefano Zampini /* determine if can use MatSolve routines instead of calling KSPSolve on ksp_R */ 115406656605SStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_R);CHKERRQ(ierr); 115506656605SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_R,PCLU,&isLU);CHKERRQ(ierr); 115606656605SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_R,PCILU,&isILU);CHKERRQ(ierr); 115706656605SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_R,PCCHOLESKY,&isCHOL);CHKERRQ(ierr); 1158ffd830a3SStefano Zampini lda_rhs = n_R; 115906656605SStefano Zampini if (isLU || isILU || isCHOL) { 116006656605SStefano Zampini ierr = PCFactorGetMatrix(pc_R,&F);CHKERRQ(ierr); 1161d62866d3SStefano Zampini } else if (sub_schurs->reuse_mumps) { 1162d62866d3SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 1163d62866d3SStefano Zampini MatFactorType type; 1164d62866d3SStefano Zampini 11656816873aSStefano Zampini F = reuse_mumps->F; 11666816873aSStefano Zampini ierr = MatGetFactorType(F,&type);CHKERRQ(ierr); 1167d62866d3SStefano Zampini if (type == MAT_FACTOR_CHOLESKY) isCHOL = PETSC_TRUE; 1168ffd830a3SStefano Zampini ierr = MatGetSize(F,&lda_rhs,NULL);CHKERRQ(ierr); 116906656605SStefano Zampini } else { 117006656605SStefano Zampini F = NULL; 117106656605SStefano Zampini } 117206656605SStefano Zampini 1173ffd830a3SStefano Zampini /* allocate workspace */ 1174ffd830a3SStefano Zampini n = 0; 1175ffd830a3SStefano Zampini if (n_constraints) { 1176ffd830a3SStefano Zampini n += lda_rhs*n_constraints; 1177ffd830a3SStefano Zampini } 1178ffd830a3SStefano Zampini if (n_vertices) { 1179ffd830a3SStefano Zampini n = PetscMax(2*lda_rhs*n_vertices,n); 1180ffd830a3SStefano Zampini n = PetscMax((lda_rhs+n_B)*n_vertices,n); 1181ffd830a3SStefano Zampini } 1182ffd830a3SStefano Zampini ierr = PetscMalloc1(n,&work);CHKERRQ(ierr); 1183ffd830a3SStefano Zampini 118488ebb749SStefano Zampini /* Precompute stuffs needed for preprocessing and application of BDDC*/ 118588ebb749SStefano Zampini if (n_constraints) { 118606656605SStefano Zampini Mat M1,M2,M3; 118780677318SStefano Zampini Mat auxmat; 118806656605SStefano Zampini IS is_aux; 118980677318SStefano Zampini PetscScalar *array,*array2; 119006656605SStefano Zampini 1191f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat1);CHKERRQ(ierr); 119280677318SStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat2);CHKERRQ(ierr); 119388ebb749SStefano Zampini 119425084f0cSStefano Zampini /* Extract constraints on R nodes: C_{CR} */ 119525084f0cSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,n_constraints,n_vertices,1,&is_aux);CHKERRQ(ierr); 11968ce42a96SStefano Zampini ierr = MatGetSubMatrix(pcbddc->ConstraintMatrix,is_aux,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&C_CR);CHKERRQ(ierr); 119780677318SStefano Zampini ierr = MatGetSubMatrix(pcbddc->ConstraintMatrix,is_aux,pcis->is_B_local,MAT_INITIAL_MATRIX,&auxmat);CHKERRQ(ierr); 119888ebb749SStefano Zampini 119980677318SStefano Zampini /* Assemble local_auxmat2_R = (- A_{RR}^{-1} C^T_{CR}) needed by BDDC setup */ 120080677318SStefano Zampini /* Assemble pcbddc->local_auxmat2 = R_to_B (- A_{RR}^{-1} C^T_{CR}) needed by BDDC application */ 1201ffd830a3SStefano Zampini ierr = PetscMemzero(work,lda_rhs*n_constraints*sizeof(PetscScalar));CHKERRQ(ierr); 120288ebb749SStefano Zampini for (i=0;i<n_constraints;i++) { 120306656605SStefano Zampini const PetscScalar *row_cmat_values; 120406656605SStefano Zampini const PetscInt *row_cmat_indices; 120506656605SStefano Zampini PetscInt size_of_constraint,j; 120688ebb749SStefano Zampini 120706656605SStefano Zampini ierr = MatGetRow(C_CR,i,&size_of_constraint,&row_cmat_indices,&row_cmat_values);CHKERRQ(ierr); 120806656605SStefano Zampini for (j=0;j<size_of_constraint;j++) { 1209ffd830a3SStefano Zampini work[row_cmat_indices[j]+i*lda_rhs] = -row_cmat_values[j]; 121006656605SStefano Zampini } 121106656605SStefano Zampini ierr = MatRestoreRow(C_CR,i,&size_of_constraint,&row_cmat_indices,&row_cmat_values);CHKERRQ(ierr); 121206656605SStefano Zampini } 1213ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_constraints,NULL,&local_auxmat2_R);CHKERRQ(ierr); 121406656605SStefano Zampini if (F) { 121506656605SStefano Zampini Mat B; 121606656605SStefano Zampini 1217ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_constraints,work,&B);CHKERRQ(ierr); 121880677318SStefano Zampini ierr = MatMatSolve(F,B,local_auxmat2_R);CHKERRQ(ierr); 121906656605SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 122006656605SStefano Zampini } else { 122180677318SStefano Zampini PetscScalar *marr; 122280677318SStefano Zampini 122380677318SStefano Zampini ierr = MatDenseGetArray(local_auxmat2_R,&marr);CHKERRQ(ierr); 122406656605SStefano Zampini for (i=0;i<n_constraints;i++) { 1225ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*lda_rhs);CHKERRQ(ierr); 1226ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,marr+i*lda_rhs);CHKERRQ(ierr); 122706656605SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 122806656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 122906656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 123006656605SStefano Zampini } 123180677318SStefano Zampini ierr = MatDenseRestoreArray(local_auxmat2_R,&marr);CHKERRQ(ierr); 123206656605SStefano Zampini } 123380677318SStefano Zampini if (!pcbddc->switch_static) { 123480677318SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,n_constraints,NULL,&pcbddc->local_auxmat2);CHKERRQ(ierr); 123580677318SStefano Zampini ierr = MatDenseGetArray(pcbddc->local_auxmat2,&array);CHKERRQ(ierr); 123680677318SStefano Zampini ierr = MatDenseGetArray(local_auxmat2_R,&array2);CHKERRQ(ierr); 123780677318SStefano Zampini for (i=0;i<n_constraints;i++) { 1238ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,array2+i*lda_rhs);CHKERRQ(ierr); 123980677318SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,array+i*n_B);CHKERRQ(ierr); 124080677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 124180677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 124280677318SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 124380677318SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 124480677318SStefano Zampini } 124580677318SStefano Zampini ierr = MatDenseRestoreArray(local_auxmat2_R,&array2);CHKERRQ(ierr); 124680677318SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->local_auxmat2,&array);CHKERRQ(ierr); 124780677318SStefano Zampini ierr = MatMatMult(auxmat,pcbddc->local_auxmat2,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&M3);CHKERRQ(ierr); 124880677318SStefano Zampini } else { 1249ffd830a3SStefano Zampini if (lda_rhs != n_R) { 1250ffd830a3SStefano Zampini IS dummy; 1251ffd830a3SStefano Zampini 1252ffd830a3SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,n_R,0,1,&dummy);CHKERRQ(ierr); 1253ffd830a3SStefano Zampini ierr = MatGetSubMatrix(local_auxmat2_R,dummy,dummy,MAT_INITIAL_MATRIX,&pcbddc->local_auxmat2);CHKERRQ(ierr); 1254ffd830a3SStefano Zampini ierr = ISDestroy(&dummy);CHKERRQ(ierr); 1255ffd830a3SStefano Zampini } else { 125680677318SStefano Zampini ierr = PetscObjectReference((PetscObject)local_auxmat2_R);CHKERRQ(ierr); 125780677318SStefano Zampini pcbddc->local_auxmat2 = local_auxmat2_R; 1258ffd830a3SStefano Zampini } 125925084f0cSStefano Zampini ierr = MatMatMult(C_CR,pcbddc->local_auxmat2,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&M3);CHKERRQ(ierr); 126080677318SStefano Zampini } 126180677318SStefano Zampini ierr = ISDestroy(&is_aux);CHKERRQ(ierr); 126280677318SStefano Zampini /* Assemble explicitly S_CC = ( C_{CR} A_{RR}^{-1} C^T_{CR} )^{-1} */ 126380677318SStefano Zampini ierr = MatScale(M3,m_one);CHKERRQ(ierr); 126406656605SStefano Zampini ierr = MatDuplicate(M3,MAT_DO_NOT_COPY_VALUES,&M1);CHKERRQ(ierr); 126506656605SStefano Zampini ierr = MatDuplicate(M3,MAT_DO_NOT_COPY_VALUES,&M2);CHKERRQ(ierr); 126680677318SStefano Zampini if (isCHOL) { 126780677318SStefano Zampini ierr = MatCholeskyFactor(M3,NULL,NULL);CHKERRQ(ierr); 126880677318SStefano Zampini } else { 126925084f0cSStefano Zampini ierr = MatLUFactor(M3,NULL,NULL,NULL);CHKERRQ(ierr); 127080677318SStefano Zampini } 127180677318SStefano Zampini ierr = VecSet(pcbddc->vec1_C,one);CHKERRQ(ierr); 127206656605SStefano Zampini ierr = MatDiagonalSet(M2,pcbddc->vec1_C,INSERT_VALUES);CHKERRQ(ierr); 127325084f0cSStefano Zampini ierr = MatMatSolve(M3,M2,M1);CHKERRQ(ierr); 127425084f0cSStefano Zampini ierr = MatDestroy(&M2);CHKERRQ(ierr); 127525084f0cSStefano Zampini ierr = MatDestroy(&M3);CHKERRQ(ierr); 127680677318SStefano Zampini /* Assemble local_auxmat1 = S_CC*C_{CB} needed by BDDC application in KSP and in preproc */ 127780677318SStefano Zampini ierr = MatMatMult(M1,auxmat,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&pcbddc->local_auxmat1);CHKERRQ(ierr); 127880677318SStefano Zampini ierr = MatDestroy(&auxmat);CHKERRQ(ierr); 127906656605SStefano Zampini ierr = MatCopy(M1,S_CC,SAME_NONZERO_PATTERN);CHKERRQ(ierr); /* S_CC can have a different LDA, MatMatSolve doesn't support it */ 128006656605SStefano Zampini ierr = MatDestroy(&M1);CHKERRQ(ierr); 1281f4ddd8eeSStefano Zampini } 128288ebb749SStefano Zampini /* Get submatrices from subdomain matrix */ 12834f1b2e48SStefano Zampini if (pcbddc->benign_n) { 1284d16cbb6bSStefano Zampini IS dummy; 1285d16cbb6bSStefano Zampini Mat B0_R; 1286d16cbb6bSStefano Zampini PetscReal norm; 1287d16cbb6bSStefano Zampini 12884f1b2e48SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcbddc->benign_n,0,1,&dummy);CHKERRQ(ierr); 12894f1b2e48SStefano Zampini ierr = MatGetSubMatrix(pcbddc->benign_B0,dummy,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&B0_R);CHKERRQ(ierr); 1290d16cbb6bSStefano Zampini ierr = MatNorm(B0_R,NORM_INFINITY,&norm);CHKERRQ(ierr); 1291d16cbb6bSStefano Zampini if (norm > PETSC_SMALL) { 1292d16cbb6bSStefano Zampini SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_SUP,"Benign trick can not be applied! ||B0_R|| = %f (should be numerically 0.)",norm); 1293d16cbb6bSStefano Zampini } 1294d16cbb6bSStefano Zampini ierr = MatDestroy(&B0_R);CHKERRQ(ierr); 1295d16cbb6bSStefano Zampini ierr = ISDestroy(&dummy);CHKERRQ(ierr); 1296d16cbb6bSStefano Zampini } 1297d16cbb6bSStefano Zampini 129888ebb749SStefano Zampini if (n_vertices) { 129906656605SStefano Zampini IS is_aux; 13003a50541eSStefano Zampini 13016816873aSStefano Zampini if (sub_schurs->reuse_mumps) { /* is_R_local is not sorted, ISComplement doesn't like it */ 13026816873aSStefano Zampini IS tis; 13036816873aSStefano Zampini 13046816873aSStefano Zampini ierr = ISDuplicate(pcbddc->is_R_local,&tis);CHKERRQ(ierr); 13056816873aSStefano Zampini ierr = ISSort(tis);CHKERRQ(ierr); 13066816873aSStefano Zampini ierr = ISComplement(tis,0,pcis->n,&is_aux);CHKERRQ(ierr); 13076816873aSStefano Zampini ierr = ISDestroy(&tis);CHKERRQ(ierr); 13086816873aSStefano Zampini } else { 13093a50541eSStefano Zampini ierr = ISComplement(pcbddc->is_R_local,0,pcis->n,&is_aux);CHKERRQ(ierr); 13106816873aSStefano Zampini } 13119577ea80SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,pcbddc->is_R_local,is_aux,MAT_INITIAL_MATRIX,&A_RV);CHKERRQ(ierr); 13129577ea80SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,is_aux,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&A_VR);CHKERRQ(ierr); 131304708bb6SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,is_aux,is_aux,MAT_INITIAL_MATRIX,&A_VV);CHKERRQ(ierr); 13144f1b2e48SStefano Zampini if (pcbddc->benign_n) { 1315019a44ceSStefano Zampini IS dummy; 1316019a44ceSStefano Zampini 13174f1b2e48SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcbddc->benign_n,0,1,&dummy);CHKERRQ(ierr); 13184f1b2e48SStefano Zampini ierr = MatGetSubMatrix(pcbddc->benign_B0,dummy,is_aux,MAT_INITIAL_MATRIX,&B0_V);CHKERRQ(ierr); 1319019a44ceSStefano Zampini ierr = ISDestroy(&dummy);CHKERRQ(ierr); 1320019a44ceSStefano Zampini } 132125084f0cSStefano Zampini ierr = ISDestroy(&is_aux);CHKERRQ(ierr); 132288ebb749SStefano Zampini } 132388ebb749SStefano Zampini 132488ebb749SStefano Zampini /* Matrix of coarse basis functions (local) */ 1325f4ddd8eeSStefano Zampini if (pcbddc->coarse_phi_B) { 132606656605SStefano Zampini PetscInt on_B,on_primal,on_D=n_D; 132706656605SStefano Zampini if (pcbddc->coarse_phi_D) { 132806656605SStefano Zampini ierr = MatGetSize(pcbddc->coarse_phi_D,&on_D,NULL);CHKERRQ(ierr); 132906656605SStefano Zampini } 1330f4ddd8eeSStefano Zampini ierr = MatGetSize(pcbddc->coarse_phi_B,&on_B,&on_primal);CHKERRQ(ierr); 133106656605SStefano Zampini if (on_B != n_B || on_primal != pcbddc->local_primal_size || on_D != n_D) { 133206656605SStefano Zampini PetscScalar *marray; 133306656605SStefano Zampini 133406656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&marray);CHKERRQ(ierr); 133506656605SStefano Zampini ierr = PetscFree(marray);CHKERRQ(ierr); 1336f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_B);CHKERRQ(ierr); 1337f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_B);CHKERRQ(ierr); 1338f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_D);CHKERRQ(ierr); 1339f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_D);CHKERRQ(ierr); 1340f4ddd8eeSStefano Zampini } 1341f4ddd8eeSStefano Zampini } 134206656605SStefano Zampini 1343f4ddd8eeSStefano Zampini if (!pcbddc->coarse_phi_B) { 134406656605SStefano Zampini PetscScalar *marray; 134588ebb749SStefano Zampini 134606656605SStefano Zampini n = n_B*pcbddc->local_primal_size; 13478eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 134806656605SStefano Zampini n += n_D*pcbddc->local_primal_size; 134988ebb749SStefano Zampini } 13503301b35fSStefano Zampini if (!pcbddc->symmetric_primal) { 135106656605SStefano Zampini n *= 2; 135288ebb749SStefano Zampini } 135306656605SStefano Zampini ierr = PetscCalloc1(n,&marray);CHKERRQ(ierr); 135406656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,pcbddc->local_primal_size,marray,&pcbddc->coarse_phi_B);CHKERRQ(ierr); 135506656605SStefano Zampini n = n_B*pcbddc->local_primal_size; 13568eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 135706656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_D,pcbddc->local_primal_size,marray+n,&pcbddc->coarse_phi_D);CHKERRQ(ierr); 135806656605SStefano Zampini n += n_D*pcbddc->local_primal_size; 135988ebb749SStefano Zampini } 13603301b35fSStefano Zampini if (!pcbddc->symmetric_primal) { 136106656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,pcbddc->local_primal_size,marray+n,&pcbddc->coarse_psi_B);CHKERRQ(ierr); 13628eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 136306656605SStefano Zampini n = n_B*pcbddc->local_primal_size; 136406656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_D,pcbddc->local_primal_size,marray+n,&pcbddc->coarse_psi_D);CHKERRQ(ierr); 136588ebb749SStefano Zampini } 136688ebb749SStefano Zampini } else { 1367c0553b1fSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->coarse_phi_B);CHKERRQ(ierr); 1368c0553b1fSStefano Zampini pcbddc->coarse_psi_B = pcbddc->coarse_phi_B; 13691b968477SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 1370c0553b1fSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->coarse_phi_D);CHKERRQ(ierr); 1371c0553b1fSStefano Zampini pcbddc->coarse_psi_D = pcbddc->coarse_phi_D; 1372c0553b1fSStefano Zampini } 137388ebb749SStefano Zampini } 137406656605SStefano Zampini } 1375019a44ceSStefano Zampini 137606656605SStefano Zampini /* We are now ready to evaluate coarse basis functions and subdomain contribution to coarse problem */ 13774f1b2e48SStefano Zampini p0_lidx_I = NULL; 13784f1b2e48SStefano Zampini if (pcbddc->benign_n && (pcbddc->switch_static || pcbddc->dbg_flag)) { 1379d12edf2fSStefano Zampini const PetscInt *idxs; 1380d12edf2fSStefano Zampini 1381d12edf2fSStefano Zampini ierr = ISGetIndices(pcis->is_I_local,&idxs);CHKERRQ(ierr); 13824f1b2e48SStefano Zampini ierr = PetscMalloc1(pcbddc->benign_n,&p0_lidx_I);CHKERRQ(ierr); 13834f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 13844f1b2e48SStefano Zampini ierr = PetscFindInt(pcbddc->benign_p0_lidx[i],pcis->n-pcis->n_B,idxs,&p0_lidx_I[i]);CHKERRQ(ierr); 13854f1b2e48SStefano Zampini } 1386d12edf2fSStefano Zampini ierr = ISRestoreIndices(pcis->is_I_local,&idxs);CHKERRQ(ierr); 1387d12edf2fSStefano Zampini } 1388d16cbb6bSStefano Zampini 138906656605SStefano Zampini /* vertices */ 139006656605SStefano Zampini if (n_vertices) { 139116f15bc4SStefano Zampini 1392ffd830a3SStefano Zampini ierr = MatConvert(A_VV,MATDENSE,MAT_REUSE_MATRIX,&A_VV);CHKERRQ(ierr); 139304708bb6SStefano Zampini 139416f15bc4SStefano Zampini if (n_R) { 139506656605SStefano Zampini Mat A_RRmA_RV,S_VVt; /* S_VVt with LDA=N */ 139606656605SStefano Zampini PetscBLASInt B_N,B_one = 1; 139716f15bc4SStefano Zampini PetscScalar *x,*y; 139804708bb6SStefano Zampini PetscBool isseqaij; 139906656605SStefano Zampini 140021eccb56SStefano Zampini ierr = MatScale(A_RV,m_one);CHKERRQ(ierr); 1401ffd830a3SStefano Zampini if (lda_rhs == n_R) { 1402ffd830a3SStefano Zampini ierr = MatConvert(A_RV,MATDENSE,MAT_REUSE_MATRIX,&A_RV);CHKERRQ(ierr); 1403ffd830a3SStefano Zampini } else { 1404ffd830a3SStefano Zampini Mat A_RVt; 1405ffd830a3SStefano Zampini PetscScalar *array; 1406ffd830a3SStefano Zampini 1407ffd830a3SStefano Zampini /* TODO optimize */ 1408ffd830a3SStefano Zampini ierr = MatConvert(A_RV,MATDENSE,MAT_INITIAL_MATRIX,&A_RVt);CHKERRQ(ierr); 1409ffd830a3SStefano Zampini ierr = MatDestroy(&A_RV);CHKERRQ(ierr); 1410ffd830a3SStefano Zampini ierr = PetscMemzero(work+lda_rhs*n_vertices,lda_rhs*n_vertices*sizeof(PetscScalar));CHKERRQ(ierr); 1411ffd830a3SStefano Zampini ierr = MatDenseGetArray(A_RVt,&array);CHKERRQ(ierr); 1412ffd830a3SStefano Zampini for (i=0;i<n_vertices;i++) { 1413ffd830a3SStefano Zampini ierr = PetscMemcpy(work+(n_vertices+i)*lda_rhs,array+i*n_R,n_R*sizeof(PetscScalar));CHKERRQ(ierr); 1414ffd830a3SStefano Zampini } 1415ffd830a3SStefano Zampini ierr = MatDenseRestoreArray(A_RVt,&array);CHKERRQ(ierr); 1416ffd830a3SStefano Zampini ierr = MatDestroy(&A_RVt);CHKERRQ(ierr); 1417ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_vertices,work+lda_rhs*n_vertices,&A_RV);CHKERRQ(ierr); 1418ffd830a3SStefano Zampini } 1419ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_vertices,work,&A_RRmA_RV);CHKERRQ(ierr); 14206816873aSStefano Zampini if (F) { /* TODO could be optimized for symmetric problems */ 142106656605SStefano Zampini ierr = MatMatSolve(F,A_RV,A_RRmA_RV);CHKERRQ(ierr); 142206656605SStefano Zampini } else { 142306656605SStefano Zampini ierr = MatDenseGetArray(A_RV,&y);CHKERRQ(ierr); 142406656605SStefano Zampini for (i=0;i<n_vertices;i++) { 1425ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,y+i*lda_rhs);CHKERRQ(ierr); 1426ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,work+i*lda_rhs);CHKERRQ(ierr); 142706656605SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 142806656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 142906656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 143006656605SStefano Zampini } 143106656605SStefano Zampini ierr = MatDenseRestoreArray(A_RV,&y);CHKERRQ(ierr); 143206656605SStefano Zampini } 143380677318SStefano Zampini ierr = MatDestroy(&A_RV);CHKERRQ(ierr); 1434ffd830a3SStefano Zampini /* S_VV and S_CV */ 143506656605SStefano Zampini if (n_constraints) { 143606656605SStefano Zampini Mat B; 143780677318SStefano Zampini 1438ffd830a3SStefano Zampini ierr = PetscMemzero(work+lda_rhs*n_vertices,n_B*n_vertices*sizeof(PetscScalar));CHKERRQ(ierr); 143980677318SStefano Zampini for (i=0;i<n_vertices;i++) { 1440ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*lda_rhs);CHKERRQ(ierr); 1441ffd830a3SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,work+lda_rhs*n_vertices+i*n_B);CHKERRQ(ierr); 144280677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 144380677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 144480677318SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 144580677318SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 144680677318SStefano Zampini } 1447ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,n_vertices,work+lda_rhs*n_vertices,&B);CHKERRQ(ierr); 144880677318SStefano Zampini ierr = MatMatMult(pcbddc->local_auxmat1,B,MAT_REUSE_MATRIX,PETSC_DEFAULT,&S_CV);CHKERRQ(ierr); 144980677318SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 1450ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_vertices,work+lda_rhs*n_vertices,&B);CHKERRQ(ierr); 145180677318SStefano Zampini ierr = MatMatMult(local_auxmat2_R,S_CV,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B);CHKERRQ(ierr); 145206656605SStefano Zampini ierr = MatScale(S_CV,m_one);CHKERRQ(ierr); 1453ffd830a3SStefano Zampini ierr = PetscBLASIntCast(lda_rhs*n_vertices,&B_N);CHKERRQ(ierr); 1454ffd830a3SStefano Zampini PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&B_N,&one,work+lda_rhs*n_vertices,&B_one,work,&B_one)); 145506656605SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 145606656605SStefano Zampini } 145704708bb6SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)A_VR,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 145804708bb6SStefano Zampini if (!isseqaij) { /* MatMatMult with SEQ(S)BAIJ below will raise an error */ 145904708bb6SStefano Zampini ierr = MatConvert(A_VR,MATSEQAIJ,MAT_REUSE_MATRIX,&A_VR);CHKERRQ(ierr); 146004708bb6SStefano Zampini } 1461ffd830a3SStefano Zampini if (lda_rhs != n_R) { 1462ffd830a3SStefano Zampini ierr = MatDestroy(&A_RRmA_RV);CHKERRQ(ierr); 1463ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_vertices,work,&A_RRmA_RV);CHKERRQ(ierr); 1464ffd830a3SStefano Zampini ierr = MatSeqDenseSetLDA(A_RRmA_RV,lda_rhs);CHKERRQ(ierr); 1465ffd830a3SStefano Zampini } 146606656605SStefano Zampini ierr = MatMatMult(A_VR,A_RRmA_RV,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&S_VVt);CHKERRQ(ierr); 146780677318SStefano Zampini ierr = MatDestroy(&A_RRmA_RV);CHKERRQ(ierr); 146806656605SStefano Zampini ierr = PetscBLASIntCast(n_vertices*n_vertices,&B_N);CHKERRQ(ierr); 146906656605SStefano Zampini ierr = MatDenseGetArray(A_VV,&x);CHKERRQ(ierr); 147006656605SStefano Zampini ierr = MatDenseGetArray(S_VVt,&y);CHKERRQ(ierr); 147106656605SStefano Zampini PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&B_N,&one,x,&B_one,y,&B_one)); 147206656605SStefano Zampini ierr = MatDenseRestoreArray(A_VV,&x);CHKERRQ(ierr); 147306656605SStefano Zampini ierr = MatDenseRestoreArray(S_VVt,&y);CHKERRQ(ierr); 147406656605SStefano Zampini ierr = MatCopy(S_VVt,S_VV,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 1475d16cbb6bSStefano Zampini ierr = MatDestroy(&S_VVt);CHKERRQ(ierr); 1476019a44ceSStefano Zampini } else { 1477d16cbb6bSStefano Zampini ierr = MatCopy(A_VV,S_VV,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 1478d16cbb6bSStefano Zampini } 14794f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 1480019a44ceSStefano Zampini const PetscScalar *vals; 1481019a44ceSStefano Zampini const PetscInt *idxs; 14824f1b2e48SStefano Zampini PetscInt n,j,primal_idx; 1483019a44ceSStefano Zampini 14844f1b2e48SStefano Zampini ierr = MatGetRow(B0_V,i,&n,&idxs,&vals);CHKERRQ(ierr); 14854f1b2e48SStefano Zampini primal_idx = pcbddc->local_primal_size - pcbddc->benign_n + i; 1486d16cbb6bSStefano Zampini for (j=0;j<n;j++) { 14874f1b2e48SStefano Zampini coarse_submat_vals[primal_idx*pcbddc->local_primal_size+idxs[j]] = vals[j]; 14884f1b2e48SStefano Zampini coarse_submat_vals[idxs[j]*pcbddc->local_primal_size+primal_idx] = vals[j]; 1489019a44ceSStefano Zampini } 14904f1b2e48SStefano Zampini ierr = MatRestoreRow(B0_V,i,&n,&idxs,&vals);CHKERRQ(ierr); 149116f15bc4SStefano Zampini } 149221eccb56SStefano Zampini ierr = MatDestroy(&A_VV);CHKERRQ(ierr); 1493d16cbb6bSStefano Zampini 149406656605SStefano Zampini /* coarse basis functions */ 149506656605SStefano Zampini for (i=0;i<n_vertices;i++) { 149616f15bc4SStefano Zampini PetscScalar *y; 149716f15bc4SStefano Zampini 1498ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+lda_rhs*i);CHKERRQ(ierr); 149906656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 150006656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*i);CHKERRQ(ierr); 150106656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 150206656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 150306656605SStefano Zampini y[n_B*i+idx_V_B[i]] = 1.0; 150406656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 150506656605SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 150606656605SStefano Zampini 150706656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 15084f1b2e48SStefano Zampini PetscInt j; 15094f1b2e48SStefano Zampini 151006656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 151106656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*i);CHKERRQ(ierr); 151206656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 151306656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 151406656605SStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 15154f1b2e48SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) y[n_D*i+p0_lidx_I[j]] = 0.0; 151606656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 151706656605SStefano Zampini } 151806656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 151906656605SStefano Zampini } 152004708bb6SStefano Zampini /* if n_R == 0 the object is not destroyed */ 152104708bb6SStefano Zampini ierr = MatDestroy(&A_RV);CHKERRQ(ierr); 152206656605SStefano Zampini } 152306656605SStefano Zampini 152406656605SStefano Zampini if (n_constraints) { 152506656605SStefano Zampini Mat B; 152606656605SStefano Zampini 1527ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_constraints,work,&B);CHKERRQ(ierr); 152806656605SStefano Zampini ierr = MatScale(S_CC,m_one);CHKERRQ(ierr); 152980677318SStefano Zampini ierr = MatMatMult(local_auxmat2_R,S_CC,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B);CHKERRQ(ierr); 153006656605SStefano Zampini ierr = MatScale(S_CC,m_one);CHKERRQ(ierr); 153106656605SStefano Zampini if (n_vertices) { 153280677318SStefano Zampini if (isCHOL) { /* if we can solve the interior problem with cholesky, we should also be fine with transposing here */ 153380677318SStefano Zampini ierr = MatTranspose(S_CV,MAT_REUSE_MATRIX,&S_VC);CHKERRQ(ierr); 153480677318SStefano Zampini } else { 153580677318SStefano Zampini Mat S_VCt; 153680677318SStefano Zampini 1537ffd830a3SStefano Zampini if (lda_rhs != n_R) { 1538ffd830a3SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 1539ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_vertices,work,&B);CHKERRQ(ierr); 1540ffd830a3SStefano Zampini ierr = MatSeqDenseSetLDA(B,lda_rhs);CHKERRQ(ierr); 1541ffd830a3SStefano Zampini } 154280677318SStefano Zampini ierr = MatMatMult(A_VR,B,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&S_VCt);CHKERRQ(ierr); 154380677318SStefano Zampini ierr = MatCopy(S_VCt,S_VC,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 154480677318SStefano Zampini ierr = MatDestroy(&S_VCt);CHKERRQ(ierr); 154580677318SStefano Zampini } 154606656605SStefano Zampini } 154706656605SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 154806656605SStefano Zampini /* coarse basis functions */ 154906656605SStefano Zampini for (i=0;i<n_constraints;i++) { 155006656605SStefano Zampini PetscScalar *y; 155106656605SStefano Zampini 1552ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+lda_rhs*i);CHKERRQ(ierr); 155306656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 155406656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*(i+n_vertices));CHKERRQ(ierr); 155506656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 155606656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 155706656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 155806656605SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 155906656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 15604f1b2e48SStefano Zampini PetscInt j; 15614f1b2e48SStefano Zampini 156206656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 156306656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*(i+n_vertices));CHKERRQ(ierr); 156406656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 156506656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 156606656605SStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 15674f1b2e48SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) y[n_D*i+p0_lidx_I[j]] = 0.0; 156806656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 156906656605SStefano Zampini } 157006656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 157106656605SStefano Zampini } 157206656605SStefano Zampini } 157380677318SStefano Zampini if (n_constraints) { 157480677318SStefano Zampini ierr = MatDestroy(&local_auxmat2_R);CHKERRQ(ierr); 157580677318SStefano Zampini } 15764f1b2e48SStefano Zampini ierr = PetscFree(p0_lidx_I);CHKERRQ(ierr); 1577019a44ceSStefano Zampini ierr = MatDestroy(&B0_V);CHKERRQ(ierr); 1578019a44ceSStefano Zampini 157906656605SStefano Zampini /* compute other basis functions for non-symmetric problems */ 15803301b35fSStefano Zampini if (!pcbddc->symmetric_primal) { 1581ffd830a3SStefano Zampini Mat B_V=NULL,B_C=NULL; 1582ffd830a3SStefano Zampini PetscScalar *marray; 158306656605SStefano Zampini 158406656605SStefano Zampini if (n_constraints) { 1585ffd830a3SStefano Zampini Mat S_CCT,C_CRT; 158606656605SStefano Zampini 1587ffd830a3SStefano Zampini ierr = MatTranspose(C_CR,MAT_INITIAL_MATRIX,&C_CRT);CHKERRQ(ierr); 158806656605SStefano Zampini ierr = MatTranspose(S_CC,MAT_INITIAL_MATRIX,&S_CCT);CHKERRQ(ierr); 1589ffd830a3SStefano Zampini ierr = MatMatMult(C_CRT,S_CCT,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&B_C);CHKERRQ(ierr); 159016f15bc4SStefano Zampini ierr = MatDestroy(&S_CCT);CHKERRQ(ierr); 159106656605SStefano Zampini if (n_vertices) { 1592ffd830a3SStefano Zampini Mat S_VCT; 159306656605SStefano Zampini 159406656605SStefano Zampini ierr = MatTranspose(S_VC,MAT_INITIAL_MATRIX,&S_VCT);CHKERRQ(ierr); 1595ffd830a3SStefano Zampini ierr = MatMatMult(C_CRT,S_VCT,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&B_V);CHKERRQ(ierr); 159616f15bc4SStefano Zampini ierr = MatDestroy(&S_VCT);CHKERRQ(ierr); 159706656605SStefano Zampini } 1598ffd830a3SStefano Zampini ierr = MatDestroy(&C_CRT);CHKERRQ(ierr); 159906656605SStefano Zampini } 160016f15bc4SStefano Zampini if (n_vertices && n_R) { 1601ffd830a3SStefano Zampini PetscScalar *av,*marray; 1602ffd830a3SStefano Zampini const PetscInt *xadj,*adjncy; 1603ffd830a3SStefano Zampini PetscInt n; 1604ffd830a3SStefano Zampini PetscBool flg_row; 160506656605SStefano Zampini 1606ffd830a3SStefano Zampini /* B_V = B_V - A_VR^T */ 1607ffd830a3SStefano Zampini ierr = MatConvert(A_VR,MATSEQAIJ,MAT_REUSE_MATRIX,&A_VR);CHKERRQ(ierr); 1608ffd830a3SStefano Zampini ierr = MatGetRowIJ(A_VR,0,PETSC_FALSE,PETSC_FALSE,&n,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 1609ffd830a3SStefano Zampini ierr = MatSeqAIJGetArray(A_VR,&av);CHKERRQ(ierr); 1610ffd830a3SStefano Zampini ierr = MatDenseGetArray(B_V,&marray);CHKERRQ(ierr); 1611ffd830a3SStefano Zampini for (i=0;i<n;i++) { 1612ffd830a3SStefano Zampini PetscInt j; 1613ffd830a3SStefano Zampini for (j=xadj[i];j<xadj[i+1];j++) marray[i*n_R + adjncy[j]] -= av[j]; 1614ffd830a3SStefano Zampini } 1615ffd830a3SStefano Zampini ierr = MatDenseRestoreArray(B_V,&marray);CHKERRQ(ierr); 1616ffd830a3SStefano Zampini ierr = MatRestoreRowIJ(A_VR,0,PETSC_FALSE,PETSC_FALSE,&n,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 1617ffd830a3SStefano Zampini ierr = MatDestroy(&A_VR);CHKERRQ(ierr); 161806656605SStefano Zampini } 161906656605SStefano Zampini 1620ffd830a3SStefano Zampini /* currently there's no support for MatTransposeMatSolve(F,B,X) */ 1621ffd830a3SStefano Zampini ierr = MatDenseGetArray(B_V,&marray);CHKERRQ(ierr); 1622ffd830a3SStefano Zampini for (i=0;i<n_vertices;i++) { 1623ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,marray+i*n_R);CHKERRQ(ierr); 1624ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,work+i*n_R);CHKERRQ(ierr); 162506656605SStefano Zampini ierr = KSPSolveTranspose(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 162606656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 162706656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 162806656605SStefano Zampini } 1629ffd830a3SStefano Zampini ierr = MatDenseRestoreArray(B_V,&marray);CHKERRQ(ierr); 1630ffd830a3SStefano Zampini ierr = MatDenseGetArray(B_C,&marray);CHKERRQ(ierr); 1631ffd830a3SStefano Zampini for (i=n_vertices;i<n_constraints+n_vertices;i++) { 1632ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,marray+(i-n_vertices)*n_R);CHKERRQ(ierr); 1633ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,work+i*n_R);CHKERRQ(ierr); 1634ffd830a3SStefano Zampini ierr = KSPSolveTranspose(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 1635ffd830a3SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 1636ffd830a3SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 163706656605SStefano Zampini } 1638ffd830a3SStefano Zampini ierr = MatDenseRestoreArray(B_C,&marray);CHKERRQ(ierr); 163906656605SStefano Zampini /* coarse basis functions */ 164006656605SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 164106656605SStefano Zampini PetscScalar *y; 164206656605SStefano Zampini 1643ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*n_R);CHKERRQ(ierr); 164406656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_psi_B,&y);CHKERRQ(ierr); 164506656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*i);CHKERRQ(ierr); 164606656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 164706656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 164806656605SStefano Zampini if (i<n_vertices) { 164906656605SStefano Zampini y[n_B*i+idx_V_B[i]] = 1.0; 165006656605SStefano Zampini } 165106656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_psi_B,&y);CHKERRQ(ierr); 165206656605SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 165306656605SStefano Zampini 165406656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 165506656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_psi_D,&y);CHKERRQ(ierr); 165606656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*i);CHKERRQ(ierr); 165706656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 165806656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 165906656605SStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 166006656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_psi_D,&y);CHKERRQ(ierr); 166106656605SStefano Zampini } 166206656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 166306656605SStefano Zampini } 1664ffd830a3SStefano Zampini ierr = MatDestroy(&B_V);CHKERRQ(ierr); 1665ffd830a3SStefano Zampini ierr = MatDestroy(&B_C);CHKERRQ(ierr); 166606656605SStefano Zampini } 1667d62866d3SStefano Zampini /* free memory */ 166888ebb749SStefano Zampini ierr = PetscFree(idx_V_B);CHKERRQ(ierr); 166906656605SStefano Zampini ierr = MatDestroy(&S_VV);CHKERRQ(ierr); 167006656605SStefano Zampini ierr = MatDestroy(&S_CV);CHKERRQ(ierr); 167106656605SStefano Zampini ierr = MatDestroy(&S_VC);CHKERRQ(ierr); 167206656605SStefano Zampini ierr = MatDestroy(&S_CC);CHKERRQ(ierr); 1673d62866d3SStefano Zampini ierr = PetscFree(work);CHKERRQ(ierr); 1674d62866d3SStefano Zampini if (n_vertices) { 1675d62866d3SStefano Zampini ierr = MatDestroy(&A_VR);CHKERRQ(ierr); 1676d62866d3SStefano Zampini } 1677d62866d3SStefano Zampini if (n_constraints) { 1678d62866d3SStefano Zampini ierr = MatDestroy(&C_CR);CHKERRQ(ierr); 1679d62866d3SStefano Zampini } 168088ebb749SStefano Zampini /* Checking coarse_sub_mat and coarse basis functios */ 168188ebb749SStefano Zampini /* Symmetric case : It should be \Phi^{(j)^T} A^{(j)} \Phi^{(j)}=coarse_sub_mat */ 168288ebb749SStefano Zampini /* Non-symmetric case : It should be \Psi^{(j)^T} A^{(j)} \Phi^{(j)}=coarse_sub_mat */ 1683d12edf2fSStefano Zampini if (pcbddc->dbg_flag) { 168488ebb749SStefano Zampini Mat coarse_sub_mat; 168525084f0cSStefano Zampini Mat AUXMAT,TM1,TM2,TM3,TM4; 168688ebb749SStefano Zampini Mat coarse_phi_D,coarse_phi_B; 168788ebb749SStefano Zampini Mat coarse_psi_D,coarse_psi_B; 168888ebb749SStefano Zampini Mat A_II,A_BB,A_IB,A_BI; 16898bec7fa6SStefano Zampini Mat C_B,CPHI; 16908bec7fa6SStefano Zampini IS is_dummy; 16918bec7fa6SStefano Zampini Vec mones; 169288ebb749SStefano Zampini MatType checkmattype=MATSEQAIJ; 169388ebb749SStefano Zampini PetscReal real_value; 169488ebb749SStefano Zampini 169588ebb749SStefano Zampini ierr = MatConvert(pcis->A_II,checkmattype,MAT_INITIAL_MATRIX,&A_II);CHKERRQ(ierr); 169688ebb749SStefano Zampini ierr = MatConvert(pcis->A_IB,checkmattype,MAT_INITIAL_MATRIX,&A_IB);CHKERRQ(ierr); 169788ebb749SStefano Zampini ierr = MatConvert(pcis->A_BI,checkmattype,MAT_INITIAL_MATRIX,&A_BI);CHKERRQ(ierr); 169888ebb749SStefano Zampini ierr = MatConvert(pcis->A_BB,checkmattype,MAT_INITIAL_MATRIX,&A_BB);CHKERRQ(ierr); 169988ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_phi_D,checkmattype,MAT_INITIAL_MATRIX,&coarse_phi_D);CHKERRQ(ierr); 170088ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_phi_B,checkmattype,MAT_INITIAL_MATRIX,&coarse_phi_B);CHKERRQ(ierr); 1701ffd830a3SStefano Zampini if (!pcbddc->symmetric_primal) { 170288ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_psi_D,checkmattype,MAT_INITIAL_MATRIX,&coarse_psi_D);CHKERRQ(ierr); 170388ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_psi_B,checkmattype,MAT_INITIAL_MATRIX,&coarse_psi_B);CHKERRQ(ierr); 170488ebb749SStefano Zampini } 170588ebb749SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,pcbddc->local_primal_size,pcbddc->local_primal_size,coarse_submat_vals,&coarse_sub_mat);CHKERRQ(ierr); 170688ebb749SStefano Zampini 170725084f0cSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 17083301b35fSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Check coarse sub mat computation (symmetric %d)\n",pcbddc->symmetric_primal);CHKERRQ(ierr); 170925084f0cSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1710ffd830a3SStefano Zampini if (!pcbddc->symmetric_primal) { 171188ebb749SStefano Zampini ierr = MatMatMult(A_II,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 171288ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM1);CHKERRQ(ierr); 171388ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 171488ebb749SStefano Zampini ierr = MatMatMult(A_BB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 171588ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM2);CHKERRQ(ierr); 171688ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 171788ebb749SStefano Zampini ierr = MatMatMult(A_IB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 171888ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM3);CHKERRQ(ierr); 171988ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 172088ebb749SStefano Zampini ierr = MatMatMult(A_BI,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 172188ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM4);CHKERRQ(ierr); 172288ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 172388ebb749SStefano Zampini } else { 172488ebb749SStefano Zampini ierr = MatPtAP(A_II,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&TM1);CHKERRQ(ierr); 172588ebb749SStefano Zampini ierr = MatPtAP(A_BB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&TM2);CHKERRQ(ierr); 172688ebb749SStefano Zampini ierr = MatMatMult(A_IB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 172788ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_phi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM3);CHKERRQ(ierr); 172888ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 172988ebb749SStefano Zampini ierr = MatMatMult(A_BI,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 173088ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_phi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM4);CHKERRQ(ierr); 173188ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 173288ebb749SStefano Zampini } 173388ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM2,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 173488ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM3,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 173588ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM4,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 173688ebb749SStefano Zampini ierr = MatConvert(TM1,MATSEQDENSE,MAT_REUSE_MATRIX,&TM1);CHKERRQ(ierr); 17374f1b2e48SStefano Zampini if (pcbddc->benign_n) { 1738d12edf2fSStefano Zampini Mat B0_I,B0_B,B0_BPHI,B0_IPHI; 1739d12edf2fSStefano Zampini PetscScalar *data,*data2; 17404f1b2e48SStefano Zampini PetscInt j; 1741d12edf2fSStefano Zampini 17424f1b2e48SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcbddc->benign_n,0,1,&is_dummy);CHKERRQ(ierr); 17434f1b2e48SStefano Zampini ierr = MatGetSubMatrix(pcbddc->benign_B0,is_dummy,pcis->is_B_local,MAT_INITIAL_MATRIX,&B0_B); 17444f1b2e48SStefano Zampini ierr = MatGetSubMatrix(pcbddc->benign_B0,is_dummy,pcis->is_I_local,MAT_INITIAL_MATRIX,&B0_I); 1745d12edf2fSStefano Zampini ierr = MatMatMult(B0_B,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&B0_BPHI);CHKERRQ(ierr); 1746d12edf2fSStefano Zampini ierr = MatConvert(B0_BPHI,MATSEQDENSE,MAT_REUSE_MATRIX,&B0_BPHI);CHKERRQ(ierr); 1747d12edf2fSStefano Zampini ierr = MatDenseGetArray(TM1,&data);CHKERRQ(ierr); 1748d12edf2fSStefano Zampini ierr = MatDenseGetArray(B0_BPHI,&data2);CHKERRQ(ierr); 17494f1b2e48SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) { 17504f1b2e48SStefano Zampini PetscInt primal_idx = pcbddc->local_primal_size - pcbddc->benign_n + j; 1751d12edf2fSStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 17524f1b2e48SStefano Zampini data[primal_idx*pcbddc->local_primal_size+i] += data2[i*pcbddc->benign_n+j]; 17534f1b2e48SStefano Zampini data[i*pcbddc->local_primal_size+primal_idx] += data2[i*pcbddc->benign_n+j]; 17544f1b2e48SStefano Zampini } 1755d12edf2fSStefano Zampini } 1756d12edf2fSStefano Zampini ierr = MatDenseRestoreArray(TM1,&data);CHKERRQ(ierr); 1757d12edf2fSStefano Zampini ierr = MatDenseRestoreArray(B0_BPHI,&data2);CHKERRQ(ierr); 1758d12edf2fSStefano Zampini ierr = MatDestroy(&B0_B);CHKERRQ(ierr); 1759d12edf2fSStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 1760d12edf2fSStefano Zampini ierr = MatDestroy(&B0_BPHI);CHKERRQ(ierr); 1761d12edf2fSStefano Zampini ierr = MatMatMult(B0_I,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&B0_IPHI);CHKERRQ(ierr); 1762d12edf2fSStefano Zampini ierr = MatDestroy(&B0_I);CHKERRQ(ierr); 1763d12edf2fSStefano Zampini ierr = MatNorm(B0_IPHI,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 1764d12edf2fSStefano Zampini ierr = MatDestroy(&B0_IPHI);CHKERRQ(ierr); 1765d12edf2fSStefano Zampini } 1766d12edf2fSStefano Zampini #if 0 1767d12edf2fSStefano Zampini { 1768d12edf2fSStefano Zampini PetscViewer viewer; 1769d12edf2fSStefano Zampini char filename[256]; 1770ffd830a3SStefano Zampini sprintf(filename,"details_local_coarse_mat%d_level%d.m",PetscGlobalRank,pcbddc->current_level); 1771d12edf2fSStefano Zampini ierr = PetscViewerASCIIOpen(PETSC_COMM_SELF,filename,&viewer);CHKERRQ(ierr); 1772d12edf2fSStefano Zampini ierr = PetscViewerSetFormat(viewer,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 1773ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)coarse_sub_mat,"computed");CHKERRQ(ierr); 1774ffd830a3SStefano Zampini ierr = MatView(coarse_sub_mat,viewer);CHKERRQ(ierr); 1775ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)TM1,"projected");CHKERRQ(ierr); 1776d12edf2fSStefano Zampini ierr = MatView(TM1,viewer);CHKERRQ(ierr); 1777ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)pcbddc->coarse_phi_B,"phi_B");CHKERRQ(ierr); 1778ffd830a3SStefano Zampini ierr = MatView(pcbddc->coarse_phi_B,viewer);CHKERRQ(ierr); 1779ffd830a3SStefano Zampini if (pcbddc->coarse_phi_D) { 1780ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)pcbddc->coarse_phi_D,"phi_D");CHKERRQ(ierr); 1781ffd830a3SStefano Zampini ierr = MatView(pcbddc->coarse_phi_D,viewer);CHKERRQ(ierr); 1782ffd830a3SStefano Zampini } 1783ffd830a3SStefano Zampini if (pcbddc->coarse_psi_B) { 1784ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)pcbddc->coarse_psi_B,"psi_B");CHKERRQ(ierr); 1785ffd830a3SStefano Zampini ierr = MatView(pcbddc->coarse_psi_B,viewer);CHKERRQ(ierr); 1786ffd830a3SStefano Zampini } 1787ffd830a3SStefano Zampini if (pcbddc->coarse_phi_B) { 1788ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)pcbddc->coarse_psi_D,"psi_D");CHKERRQ(ierr); 1789ffd830a3SStefano Zampini ierr = MatView(pcbddc->coarse_psi_D,viewer);CHKERRQ(ierr); 1790ffd830a3SStefano Zampini } 1791d12edf2fSStefano Zampini ierr = PetscViewerDestroy(&viewer);CHKERRQ(ierr); 1792d12edf2fSStefano Zampini } 1793d12edf2fSStefano Zampini #endif 179481d9aea3SBarry Smith ierr = MatAXPY(TM1,m_one,coarse_sub_mat,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 17958bec7fa6SStefano Zampini ierr = MatNorm(TM1,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 17960fccc4e9SStefano Zampini ierr = PetscViewerASCIISynchronizedAllow(pcbddc->dbg_viewer,PETSC_TRUE);CHKERRQ(ierr); 179706656605SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d matrix error % 1.14e\n",PetscGlobalRank,real_value);CHKERRQ(ierr); 17988bec7fa6SStefano Zampini 17998bec7fa6SStefano Zampini /* check constraints */ 18004f1b2e48SStefano Zampini if (!pcbddc->benign_n) { /* TODO: add benign case */ 18018bec7fa6SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcbddc->local_primal_size,0,1,&is_dummy);CHKERRQ(ierr); 18028bec7fa6SStefano Zampini ierr = MatGetSubMatrix(pcbddc->ConstraintMatrix,is_dummy,pcis->is_B_local,MAT_INITIAL_MATRIX,&C_B); 18038bec7fa6SStefano Zampini ierr = MatMatMult(C_B,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&CPHI);CHKERRQ(ierr); 18048bec7fa6SStefano Zampini ierr = MatCreateVecs(CPHI,&mones,NULL);CHKERRQ(ierr); 18058bec7fa6SStefano Zampini ierr = VecSet(mones,-1.0);CHKERRQ(ierr); 18068bec7fa6SStefano Zampini ierr = MatDiagonalSet(CPHI,mones,ADD_VALUES);CHKERRQ(ierr); 18078bec7fa6SStefano Zampini ierr = MatNorm(CPHI,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 1808bdae7319SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d phi constraints error % 1.14e\n",PetscGlobalRank,real_value);CHKERRQ(ierr); 1809ffd830a3SStefano Zampini if (!pcbddc->symmetric_primal) { 1810bdae7319SStefano Zampini ierr = MatMatMult(C_B,coarse_psi_B,MAT_REUSE_MATRIX,1.0,&CPHI);CHKERRQ(ierr); 1811bdae7319SStefano Zampini ierr = VecSet(mones,-1.0);CHKERRQ(ierr); 1812bdae7319SStefano Zampini ierr = MatDiagonalSet(CPHI,mones,ADD_VALUES);CHKERRQ(ierr); 1813bdae7319SStefano Zampini ierr = MatNorm(CPHI,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 1814bdae7319SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d psi constraints error % 1.14e\n",PetscGlobalRank,real_value);CHKERRQ(ierr); 181588ebb749SStefano Zampini } 18168bec7fa6SStefano Zampini ierr = MatDestroy(&C_B);CHKERRQ(ierr); 18178bec7fa6SStefano Zampini ierr = MatDestroy(&CPHI);CHKERRQ(ierr); 18188bec7fa6SStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 18198bec7fa6SStefano Zampini ierr = VecDestroy(&mones);CHKERRQ(ierr); 1820d12edf2fSStefano Zampini } 182125084f0cSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 182288ebb749SStefano Zampini ierr = MatDestroy(&A_II);CHKERRQ(ierr); 182388ebb749SStefano Zampini ierr = MatDestroy(&A_BB);CHKERRQ(ierr); 182488ebb749SStefano Zampini ierr = MatDestroy(&A_IB);CHKERRQ(ierr); 182588ebb749SStefano Zampini ierr = MatDestroy(&A_BI);CHKERRQ(ierr); 182688ebb749SStefano Zampini ierr = MatDestroy(&TM1);CHKERRQ(ierr); 182788ebb749SStefano Zampini ierr = MatDestroy(&TM2);CHKERRQ(ierr); 182888ebb749SStefano Zampini ierr = MatDestroy(&TM3);CHKERRQ(ierr); 182988ebb749SStefano Zampini ierr = MatDestroy(&TM4);CHKERRQ(ierr); 183088ebb749SStefano Zampini ierr = MatDestroy(&coarse_phi_D);CHKERRQ(ierr); 183188ebb749SStefano Zampini ierr = MatDestroy(&coarse_phi_B);CHKERRQ(ierr); 1832ffd830a3SStefano Zampini if (!pcbddc->symmetric_primal) { 183388ebb749SStefano Zampini ierr = MatDestroy(&coarse_psi_D);CHKERRQ(ierr); 183488ebb749SStefano Zampini ierr = MatDestroy(&coarse_psi_B);CHKERRQ(ierr); 183588ebb749SStefano Zampini } 183688ebb749SStefano Zampini ierr = MatDestroy(&coarse_sub_mat);CHKERRQ(ierr); 183788ebb749SStefano Zampini } 18388629588bSStefano Zampini /* get back data */ 18398629588bSStefano Zampini *coarse_submat_vals_n = coarse_submat_vals; 184088ebb749SStefano Zampini PetscFunctionReturn(0); 184188ebb749SStefano Zampini } 184288ebb749SStefano Zampini 184388ebb749SStefano Zampini #undef __FUNCT__ 1844d65f70fdSStefano Zampini #define __FUNCT__ "MatGetSubMatrixUnsorted" 1845d648f858SStefano Zampini PetscErrorCode MatGetSubMatrixUnsorted(Mat A, IS isrow, IS iscol, Mat* B) 1846aa0d41d4SStefano Zampini { 1847d65f70fdSStefano Zampini Mat *work_mat; 1848d65f70fdSStefano Zampini IS isrow_s,iscol_s; 1849d65f70fdSStefano Zampini PetscBool rsorted,csorted; 1850d65f70fdSStefano Zampini PetscInt rsize,*idxs_perm_r,csize,*idxs_perm_c; 1851aa0d41d4SStefano Zampini PetscErrorCode ierr; 1852aa0d41d4SStefano Zampini 1853aa0d41d4SStefano Zampini PetscFunctionBegin; 1854d65f70fdSStefano Zampini ierr = ISSorted(isrow,&rsorted);CHKERRQ(ierr); 1855d65f70fdSStefano Zampini ierr = ISSorted(iscol,&csorted);CHKERRQ(ierr); 1856d65f70fdSStefano Zampini ierr = ISGetLocalSize(isrow,&rsize);CHKERRQ(ierr); 1857d65f70fdSStefano Zampini ierr = ISGetLocalSize(iscol,&csize);CHKERRQ(ierr); 1858aa0d41d4SStefano Zampini 1859d65f70fdSStefano Zampini if (!rsorted) { 1860906d46d4SStefano Zampini const PetscInt *idxs; 1861906d46d4SStefano Zampini PetscInt *idxs_sorted,i; 1862aa0d41d4SStefano Zampini 1863d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_perm_r);CHKERRQ(ierr); 1864d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_sorted);CHKERRQ(ierr); 1865d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 1866d65f70fdSStefano Zampini idxs_perm_r[i] = i; 1867aa0d41d4SStefano Zampini } 1868d65f70fdSStefano Zampini ierr = ISGetIndices(isrow,&idxs);CHKERRQ(ierr); 1869d65f70fdSStefano Zampini ierr = PetscSortIntWithPermutation(rsize,idxs,idxs_perm_r);CHKERRQ(ierr); 1870d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 1871d65f70fdSStefano Zampini idxs_sorted[i] = idxs[idxs_perm_r[i]]; 1872aa0d41d4SStefano Zampini } 1873d65f70fdSStefano Zampini ierr = ISRestoreIndices(isrow,&idxs);CHKERRQ(ierr); 1874d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,rsize,idxs_sorted,PETSC_OWN_POINTER,&isrow_s);CHKERRQ(ierr); 1875d65f70fdSStefano Zampini } else { 1876d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)isrow);CHKERRQ(ierr); 1877d65f70fdSStefano Zampini isrow_s = isrow; 1878aa0d41d4SStefano Zampini } 1879906d46d4SStefano Zampini 1880d65f70fdSStefano Zampini if (!csorted) { 1881d65f70fdSStefano Zampini if (isrow == iscol) { 1882d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)isrow_s);CHKERRQ(ierr); 1883d65f70fdSStefano Zampini iscol_s = isrow_s; 1884d65f70fdSStefano Zampini } else { 1885d65f70fdSStefano Zampini const PetscInt *idxs; 1886d65f70fdSStefano Zampini PetscInt *idxs_sorted,i; 1887906d46d4SStefano Zampini 1888d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_perm_c);CHKERRQ(ierr); 1889d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_sorted);CHKERRQ(ierr); 1890d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 1891d65f70fdSStefano Zampini idxs_perm_c[i] = i; 1892d65f70fdSStefano Zampini } 1893d65f70fdSStefano Zampini ierr = ISGetIndices(iscol,&idxs);CHKERRQ(ierr); 1894d65f70fdSStefano Zampini ierr = PetscSortIntWithPermutation(csize,idxs,idxs_perm_c);CHKERRQ(ierr); 1895d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 1896d65f70fdSStefano Zampini idxs_sorted[i] = idxs[idxs_perm_c[i]]; 1897d65f70fdSStefano Zampini } 1898d65f70fdSStefano Zampini ierr = ISRestoreIndices(iscol,&idxs);CHKERRQ(ierr); 1899d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,csize,idxs_sorted,PETSC_OWN_POINTER,&iscol_s);CHKERRQ(ierr); 1900d65f70fdSStefano Zampini } 1901d65f70fdSStefano Zampini } else { 1902d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)iscol);CHKERRQ(ierr); 1903d65f70fdSStefano Zampini iscol_s = iscol; 1904d65f70fdSStefano Zampini } 1905d65f70fdSStefano Zampini 1906d648f858SStefano Zampini ierr = MatGetSubMatrices(A,1,&isrow_s,&iscol_s,MAT_INITIAL_MATRIX,&work_mat);CHKERRQ(ierr); 1907d65f70fdSStefano Zampini 1908d65f70fdSStefano Zampini if (!rsorted || !csorted) { 1909906d46d4SStefano Zampini Mat new_mat; 1910d65f70fdSStefano Zampini IS is_perm_r,is_perm_c; 1911906d46d4SStefano Zampini 1912d65f70fdSStefano Zampini if (!rsorted) { 1913d65f70fdSStefano Zampini PetscInt *idxs_r,i; 1914d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_r);CHKERRQ(ierr); 1915d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 1916d65f70fdSStefano Zampini idxs_r[idxs_perm_r[i]] = i; 1917906d46d4SStefano Zampini } 1918d65f70fdSStefano Zampini ierr = PetscFree(idxs_perm_r);CHKERRQ(ierr); 1919d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,rsize,idxs_r,PETSC_OWN_POINTER,&is_perm_r);CHKERRQ(ierr); 1920d65f70fdSStefano Zampini } else { 1921d65f70fdSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,rsize,0,1,&is_perm_r);CHKERRQ(ierr); 1922906d46d4SStefano Zampini } 1923d65f70fdSStefano Zampini ierr = ISSetPermutation(is_perm_r);CHKERRQ(ierr); 1924d65f70fdSStefano Zampini 1925d65f70fdSStefano Zampini if (!csorted) { 1926d65f70fdSStefano Zampini if (isrow_s == iscol_s) { 1927d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)is_perm_r);CHKERRQ(ierr); 1928d65f70fdSStefano Zampini is_perm_c = is_perm_r; 1929d65f70fdSStefano Zampini } else { 1930d65f70fdSStefano Zampini PetscInt *idxs_c,i; 1931d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_c);CHKERRQ(ierr); 1932d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 1933d65f70fdSStefano Zampini idxs_c[idxs_perm_c[i]] = i; 1934d65f70fdSStefano Zampini } 1935d65f70fdSStefano Zampini ierr = PetscFree(idxs_perm_c);CHKERRQ(ierr); 1936d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,csize,idxs_c,PETSC_OWN_POINTER,&is_perm_c);CHKERRQ(ierr); 1937d65f70fdSStefano Zampini } 1938d65f70fdSStefano Zampini } else { 1939d65f70fdSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,csize,0,1,&is_perm_c);CHKERRQ(ierr); 1940d65f70fdSStefano Zampini } 1941d65f70fdSStefano Zampini ierr = ISSetPermutation(is_perm_c);CHKERRQ(ierr); 1942d65f70fdSStefano Zampini 1943d65f70fdSStefano Zampini ierr = MatPermute(work_mat[0],is_perm_r,is_perm_c,&new_mat);CHKERRQ(ierr); 1944d65f70fdSStefano Zampini ierr = MatDestroy(&work_mat[0]);CHKERRQ(ierr); 1945d65f70fdSStefano Zampini work_mat[0] = new_mat; 1946d65f70fdSStefano Zampini ierr = ISDestroy(&is_perm_r);CHKERRQ(ierr); 1947d65f70fdSStefano Zampini ierr = ISDestroy(&is_perm_c);CHKERRQ(ierr); 1948d65f70fdSStefano Zampini } 1949d65f70fdSStefano Zampini 1950d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)work_mat[0]);CHKERRQ(ierr); 1951d65f70fdSStefano Zampini *B = work_mat[0]; 1952d65f70fdSStefano Zampini ierr = MatDestroyMatrices(1,&work_mat);CHKERRQ(ierr); 1953d65f70fdSStefano Zampini ierr = ISDestroy(&isrow_s);CHKERRQ(ierr); 1954d65f70fdSStefano Zampini ierr = ISDestroy(&iscol_s);CHKERRQ(ierr); 1955d65f70fdSStefano Zampini PetscFunctionReturn(0); 1956d65f70fdSStefano Zampini } 1957d65f70fdSStefano Zampini 1958d65f70fdSStefano Zampini #undef __FUNCT__ 19595e8657edSStefano Zampini #define __FUNCT__ "PCBDDCComputeLocalMatrix" 19605e8657edSStefano Zampini PetscErrorCode PCBDDCComputeLocalMatrix(PC pc, Mat ChangeOfBasisMatrix) 1961aa0d41d4SStefano Zampini { 1962aa0d41d4SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 19635e8657edSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 1964d65f70fdSStefano Zampini Mat new_mat; 19655e8657edSStefano Zampini IS is_local,is_global; 1966d65f70fdSStefano Zampini PetscInt local_size; 1967d65f70fdSStefano Zampini PetscBool isseqaij; 1968aa0d41d4SStefano Zampini PetscErrorCode ierr; 1969aa0d41d4SStefano Zampini 1970aa0d41d4SStefano Zampini PetscFunctionBegin; 1971aa0d41d4SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 19725e8657edSStefano Zampini ierr = MatGetSize(matis->A,&local_size,NULL);CHKERRQ(ierr); 19735e8657edSStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)matis->A),local_size,0,1,&is_local);CHKERRQ(ierr); 1974b087196eSStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(pc->pmat->rmap->mapping,is_local,&is_global);CHKERRQ(ierr); 1975aa0d41d4SStefano Zampini ierr = ISDestroy(&is_local);CHKERRQ(ierr); 1976d648f858SStefano Zampini ierr = MatGetSubMatrixUnsorted(ChangeOfBasisMatrix,is_global,is_global,&new_mat);CHKERRQ(ierr); 1977aa0d41d4SStefano Zampini ierr = ISDestroy(&is_global);CHKERRQ(ierr); 1978906d46d4SStefano Zampini 1979906d46d4SStefano Zampini /* check */ 1980906d46d4SStefano Zampini if (pcbddc->dbg_flag) { 1981906d46d4SStefano Zampini Vec x,x_change; 1982906d46d4SStefano Zampini PetscReal error; 1983906d46d4SStefano Zampini 19845e8657edSStefano Zampini ierr = MatCreateVecs(ChangeOfBasisMatrix,&x,&x_change);CHKERRQ(ierr); 1985906d46d4SStefano Zampini ierr = VecSetRandom(x,NULL);CHKERRQ(ierr); 19865e8657edSStefano Zampini ierr = MatMult(ChangeOfBasisMatrix,x,x_change);CHKERRQ(ierr); 1987e176bc59SStefano Zampini ierr = VecScatterBegin(matis->cctx,x,matis->x,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1988e176bc59SStefano Zampini ierr = VecScatterEnd(matis->cctx,x,matis->x,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1989d65f70fdSStefano Zampini ierr = MatMult(new_mat,matis->x,matis->y);CHKERRQ(ierr); 1990e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,matis->y,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1991e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,matis->y,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1992906d46d4SStefano Zampini ierr = VecAXPY(x,-1.0,x_change);CHKERRQ(ierr); 1993906d46d4SStefano Zampini ierr = VecNorm(x,NORM_INFINITY,&error);CHKERRQ(ierr); 1994906d46d4SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1995906d46d4SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Error global vs local change on N: %1.6e\n",error);CHKERRQ(ierr); 1996906d46d4SStefano Zampini ierr = VecDestroy(&x);CHKERRQ(ierr); 1997906d46d4SStefano Zampini ierr = VecDestroy(&x_change);CHKERRQ(ierr); 1998906d46d4SStefano Zampini } 1999906d46d4SStefano Zampini 200022d5777bSStefano Zampini /* TODO: HOW TO WORK WITH BAIJ and SBAIJ and SEQDENSE? */ 20019b28b3ffSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)matis->A,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 200222d5777bSStefano Zampini if (isseqaij) { 20031cf9b237SStefano Zampini Mat M; 20041cf9b237SStefano Zampini 20051cf9b237SStefano Zampini ierr = MatPtAP(matis->A,new_mat,MAT_INITIAL_MATRIX,2.0,&M);CHKERRQ(ierr); 20061cf9b237SStefano Zampini ierr = MatSeqAIJCompress(M,&pcbddc->local_mat);CHKERRQ(ierr); 20071cf9b237SStefano Zampini ierr = MatDestroy(&M);CHKERRQ(ierr); 2008aa0d41d4SStefano Zampini } else { 20091cf9b237SStefano Zampini Mat work_mat,M; 20101cf9b237SStefano Zampini 2011aa0d41d4SStefano Zampini ierr = MatConvert(matis->A,MATSEQAIJ,MAT_INITIAL_MATRIX,&work_mat);CHKERRQ(ierr); 20121cf9b237SStefano Zampini ierr = MatPtAP(work_mat,new_mat,MAT_INITIAL_MATRIX,2.0,&M);CHKERRQ(ierr); 20131cf9b237SStefano Zampini ierr = MatSeqAIJCompress(M,&pcbddc->local_mat);CHKERRQ(ierr); 20141cf9b237SStefano Zampini ierr = MatDestroy(&M);CHKERRQ(ierr); 2015aa0d41d4SStefano Zampini ierr = MatDestroy(&work_mat);CHKERRQ(ierr); 2016aa0d41d4SStefano Zampini } 20173301b35fSStefano Zampini if (matis->A->symmetric_set) { 20183301b35fSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_SYMMETRIC,matis->A->symmetric);CHKERRQ(ierr); 2019e496cd5dSStefano Zampini #if !defined(PETSC_USE_COMPLEX) 20203301b35fSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_HERMITIAN,matis->A->symmetric);CHKERRQ(ierr); 2021e496cd5dSStefano Zampini #endif 20223301b35fSStefano Zampini } 202345a1bb75SStefano Zampini /* 202445a1bb75SStefano Zampini ierr = PetscViewerSetFormat(PETSC_VIEWER_STDOUT_SELF,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 2025d65f70fdSStefano Zampini ierr = MatView(new_mat,(PetscViewer)0);CHKERRQ(ierr); 202645a1bb75SStefano Zampini */ 2027d65f70fdSStefano Zampini ierr = MatDestroy(&new_mat);CHKERRQ(ierr); 2028aa0d41d4SStefano Zampini PetscFunctionReturn(0); 2029aa0d41d4SStefano Zampini } 2030aa0d41d4SStefano Zampini 2031aa0d41d4SStefano Zampini #undef __FUNCT__ 2032a64d13efSStefano Zampini #define __FUNCT__ "PCBDDCSetUpLocalScatters" 20338ce42a96SStefano Zampini PetscErrorCode PCBDDCSetUpLocalScatters(PC pc) 2034a64d13efSStefano Zampini { 2035a64d13efSStefano Zampini PC_IS* pcis = (PC_IS*)(pc->data); 2036a64d13efSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 2037d62866d3SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 203853892102SStefano Zampini PetscInt *idx_R_local=NULL; 20393a50541eSStefano Zampini PetscInt n_vertices,i,j,n_R,n_D,n_B; 20403a50541eSStefano Zampini PetscInt vbs,bs; 20416816873aSStefano Zampini PetscBT bitmask=NULL; 2042a64d13efSStefano Zampini PetscErrorCode ierr; 2043a64d13efSStefano Zampini 2044a64d13efSStefano Zampini PetscFunctionBegin; 2045b23d619eSStefano Zampini /* 2046b23d619eSStefano Zampini No need to setup local scatters if 2047b23d619eSStefano Zampini - primal space is unchanged 2048b23d619eSStefano Zampini AND 2049b23d619eSStefano Zampini - we actually have locally some primal dofs (could not be true in multilevel or for isolated subdomains) 2050b23d619eSStefano Zampini AND 2051b23d619eSStefano Zampini - we are not in debugging mode (this is needed since there are Synchronized prints at the end of the subroutine 2052b23d619eSStefano Zampini */ 2053b23d619eSStefano Zampini if (!pcbddc->new_primal_space_local && pcbddc->local_primal_size && !pcbddc->dbg_flag) { 2054f4ddd8eeSStefano Zampini PetscFunctionReturn(0); 2055f4ddd8eeSStefano Zampini } 2056f4ddd8eeSStefano Zampini /* destroy old objects */ 2057f4ddd8eeSStefano Zampini ierr = ISDestroy(&pcbddc->is_R_local);CHKERRQ(ierr); 2058f4ddd8eeSStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_B);CHKERRQ(ierr); 2059f4ddd8eeSStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_D);CHKERRQ(ierr); 2060a64d13efSStefano Zampini /* Set Non-overlapping dimensions */ 2061b371cd4fSStefano Zampini n_B = pcis->n_B; 2062b371cd4fSStefano Zampini n_D = pcis->n - n_B; 2063b371cd4fSStefano Zampini n_vertices = pcbddc->n_vertices; 20643a50541eSStefano Zampini 2065a64d13efSStefano Zampini /* Dohrmann's notation: dofs splitted in R (Remaining: all dofs but the vertices) and V (Vertices) */ 20666816873aSStefano Zampini 206753892102SStefano Zampini /* create auxiliary bitmask and allocate workspace */ 20686816873aSStefano Zampini if (!sub_schurs->reuse_mumps) { 2069854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n-n_vertices,&idx_R_local);CHKERRQ(ierr); 2070a64d13efSStefano Zampini ierr = PetscBTCreate(pcis->n,&bitmask);CHKERRQ(ierr); 2071a64d13efSStefano Zampini for (i=0;i<n_vertices;i++) { 20720e6343abSStefano Zampini ierr = PetscBTSet(bitmask,pcbddc->local_primal_ref_node[i]);CHKERRQ(ierr); 2073a64d13efSStefano Zampini } 2074a64d13efSStefano Zampini 2075a64d13efSStefano Zampini for (i=0, n_R=0; i<pcis->n; i++) { 20764641a718SStefano Zampini if (!PetscBTLookup(bitmask,i)) { 20776816873aSStefano Zampini idx_R_local[n_R++] = i; 2078a64d13efSStefano Zampini } 2079a64d13efSStefano Zampini } 208053892102SStefano Zampini } else { /* A different ordering (already computed) is present if we are reusing MUMPS Schur solver */ 20816816873aSStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 20826816873aSStefano Zampini 208353892102SStefano Zampini ierr = ISGetIndices(reuse_mumps->is_R,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 208453892102SStefano Zampini ierr = ISGetLocalSize(reuse_mumps->is_R,&n_R);CHKERRQ(ierr); 20856816873aSStefano Zampini } 20863a50541eSStefano Zampini 20873a50541eSStefano Zampini /* Block code */ 20883a50541eSStefano Zampini vbs = 1; 20893a50541eSStefano Zampini ierr = MatGetBlockSize(pcbddc->local_mat,&bs);CHKERRQ(ierr); 20903a50541eSStefano Zampini if (bs>1 && !(n_vertices%bs)) { 20913a50541eSStefano Zampini PetscBool is_blocked = PETSC_TRUE; 20923a50541eSStefano Zampini PetscInt *vary; 2093d3df7717SStefano Zampini if (!sub_schurs->reuse_mumps) { 2094785e854fSJed Brown ierr = PetscMalloc1(pcis->n/bs,&vary);CHKERRQ(ierr); 20953a50541eSStefano Zampini ierr = PetscMemzero(vary,pcis->n/bs*sizeof(PetscInt));CHKERRQ(ierr); 2096d3df7717SStefano Zampini /* Verify that the vertex indices correspond to each element in a block (code taken from sbaij2.c) */ 2097d3df7717SStefano 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 */ 20980e6343abSStefano Zampini for (i=0; i<n_vertices; i++) vary[pcbddc->local_primal_ref_node[i]/bs]++; 2099d3df7717SStefano Zampini for (i=0; i<pcis->n/bs; i++) { 21003a50541eSStefano Zampini if (vary[i]!=0 && vary[i]!=bs) { 21013a50541eSStefano Zampini is_blocked = PETSC_FALSE; 21023a50541eSStefano Zampini break; 21033a50541eSStefano Zampini } 21043a50541eSStefano Zampini } 2105d3df7717SStefano Zampini ierr = PetscFree(vary);CHKERRQ(ierr); 2106d3df7717SStefano Zampini } else { 2107d3df7717SStefano Zampini /* Verify directly the R set */ 2108d3df7717SStefano Zampini for (i=0; i<n_R/bs; i++) { 2109d3df7717SStefano Zampini PetscInt j,node=idx_R_local[bs*i]; 2110d3df7717SStefano Zampini for (j=1; j<bs; j++) { 2111d3df7717SStefano Zampini if (node != idx_R_local[bs*i+j]-j) { 2112d3df7717SStefano Zampini is_blocked = PETSC_FALSE; 2113d3df7717SStefano Zampini break; 2114d3df7717SStefano Zampini } 2115d3df7717SStefano Zampini } 2116d3df7717SStefano Zampini } 2117d3df7717SStefano Zampini } 21183a50541eSStefano Zampini if (is_blocked) { /* build compressed IS for R nodes (complement of vertices) */ 21193a50541eSStefano Zampini vbs = bs; 21203a50541eSStefano Zampini for (i=0;i<n_R/vbs;i++) { 21213a50541eSStefano Zampini idx_R_local[i] = idx_R_local[vbs*i]/vbs; 21223a50541eSStefano Zampini } 21233a50541eSStefano Zampini } 21243a50541eSStefano Zampini } 21253a50541eSStefano Zampini ierr = ISCreateBlock(PETSC_COMM_SELF,vbs,n_R/vbs,idx_R_local,PETSC_COPY_VALUES,&pcbddc->is_R_local);CHKERRQ(ierr); 212653892102SStefano Zampini if (sub_schurs->reuse_mumps) { 212753892102SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 212853892102SStefano Zampini 212953892102SStefano Zampini ierr = ISRestoreIndices(reuse_mumps->is_R,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 213053892102SStefano Zampini ierr = ISDestroy(&reuse_mumps->is_R);CHKERRQ(ierr); 213153892102SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->is_R_local);CHKERRQ(ierr); 213253892102SStefano Zampini reuse_mumps->is_R = pcbddc->is_R_local; 213353892102SStefano Zampini } else { 21343a50541eSStefano Zampini ierr = PetscFree(idx_R_local);CHKERRQ(ierr); 213553892102SStefano Zampini } 2136a64d13efSStefano Zampini 2137a64d13efSStefano Zampini /* print some info if requested */ 2138a64d13efSStefano Zampini if (pcbddc->dbg_flag) { 2139a64d13efSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 2140a64d13efSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 21410fccc4e9SStefano Zampini ierr = PetscViewerASCIISynchronizedAllow(pcbddc->dbg_viewer,PETSC_TRUE);CHKERRQ(ierr); 2142a64d13efSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d local dimensions\n",PetscGlobalRank);CHKERRQ(ierr); 2143a64d13efSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"local_size = %d, dirichlet_size = %d, boundary_size = %d\n",pcis->n,n_D,n_B);CHKERRQ(ierr); 21444f1b2e48SStefano 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); 2145a64d13efSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 2146a64d13efSStefano Zampini } 2147a64d13efSStefano Zampini 2148a64d13efSStefano Zampini /* VecScatters pcbddc->R_to_B and (optionally) pcbddc->R_to_D */ 21496816873aSStefano Zampini if (!sub_schurs->reuse_mumps) { 21506816873aSStefano Zampini IS is_aux1,is_aux2; 21516816873aSStefano Zampini PetscInt *aux_array1,*aux_array2,*is_indices,*idx_R_local; 21526816873aSStefano Zampini 21533a50541eSStefano Zampini ierr = ISGetIndices(pcbddc->is_R_local,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 2154854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n_B-n_vertices,&aux_array1);CHKERRQ(ierr); 2155854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n_B-n_vertices,&aux_array2);CHKERRQ(ierr); 2156a64d13efSStefano Zampini ierr = ISGetIndices(pcis->is_I_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 21574641a718SStefano Zampini for (i=0; i<n_D; i++) { 21584641a718SStefano Zampini ierr = PetscBTSet(bitmask,is_indices[i]);CHKERRQ(ierr); 21594641a718SStefano Zampini } 2160a64d13efSStefano Zampini ierr = ISRestoreIndices(pcis->is_I_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 2161a64d13efSStefano Zampini for (i=0, j=0; i<n_R; i++) { 21624641a718SStefano Zampini if (!PetscBTLookup(bitmask,idx_R_local[i])) { 21634641a718SStefano Zampini aux_array1[j++] = i; 2164a64d13efSStefano Zampini } 2165a64d13efSStefano Zampini } 2166a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array1,PETSC_OWN_POINTER,&is_aux1);CHKERRQ(ierr); 2167a64d13efSStefano Zampini ierr = ISGetIndices(pcis->is_B_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 2168a64d13efSStefano Zampini for (i=0, j=0; i<n_B; i++) { 21694641a718SStefano Zampini if (!PetscBTLookup(bitmask,is_indices[i])) { 21704641a718SStefano Zampini aux_array2[j++] = i; 2171a64d13efSStefano Zampini } 2172a64d13efSStefano Zampini } 2173a64d13efSStefano Zampini ierr = ISRestoreIndices(pcis->is_B_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 2174a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array2,PETSC_OWN_POINTER,&is_aux2);CHKERRQ(ierr); 2175a64d13efSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,is_aux1,pcis->vec1_B,is_aux2,&pcbddc->R_to_B);CHKERRQ(ierr); 2176a64d13efSStefano Zampini ierr = ISDestroy(&is_aux1);CHKERRQ(ierr); 2177a64d13efSStefano Zampini ierr = ISDestroy(&is_aux2);CHKERRQ(ierr); 2178a64d13efSStefano Zampini 21798eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 2180785e854fSJed Brown ierr = PetscMalloc1(n_D,&aux_array1);CHKERRQ(ierr); 2181a64d13efSStefano Zampini for (i=0, j=0; i<n_R; i++) { 21824641a718SStefano Zampini if (PetscBTLookup(bitmask,idx_R_local[i])) { 21834641a718SStefano Zampini aux_array1[j++] = i; 2184a64d13efSStefano Zampini } 2185a64d13efSStefano Zampini } 2186a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array1,PETSC_OWN_POINTER,&is_aux1);CHKERRQ(ierr); 2187a64d13efSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,is_aux1,pcis->vec1_D,(IS)0,&pcbddc->R_to_D);CHKERRQ(ierr); 2188a64d13efSStefano Zampini ierr = ISDestroy(&is_aux1);CHKERRQ(ierr); 2189a64d13efSStefano Zampini } 21904641a718SStefano Zampini ierr = PetscBTDestroy(&bitmask);CHKERRQ(ierr); 21913a50541eSStefano Zampini ierr = ISRestoreIndices(pcbddc->is_R_local,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 2192d62866d3SStefano Zampini } else { 219353892102SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 21946816873aSStefano Zampini IS tis; 21956816873aSStefano Zampini PetscInt schur_size; 21966816873aSStefano Zampini 219753892102SStefano Zampini ierr = ISGetLocalSize(reuse_mumps->is_B,&schur_size);CHKERRQ(ierr); 21986816873aSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,schur_size,n_D,1,&tis);CHKERRQ(ierr); 219953892102SStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,tis,pcis->vec1_B,reuse_mumps->is_B,&pcbddc->R_to_B);CHKERRQ(ierr); 22006816873aSStefano Zampini ierr = ISDestroy(&tis);CHKERRQ(ierr); 22016816873aSStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 22026816873aSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,n_D,0,1,&tis);CHKERRQ(ierr); 22036816873aSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,tis,pcis->vec1_D,(IS)0,&pcbddc->R_to_D);CHKERRQ(ierr); 22046816873aSStefano Zampini ierr = ISDestroy(&tis);CHKERRQ(ierr); 2205d62866d3SStefano Zampini } 2206d62866d3SStefano Zampini } 2207a64d13efSStefano Zampini PetscFunctionReturn(0); 2208a64d13efSStefano Zampini } 2209a64d13efSStefano Zampini 2210304d26faSStefano Zampini 2211304d26faSStefano Zampini #undef __FUNCT__ 2212304d26faSStefano Zampini #define __FUNCT__ "PCBDDCSetUpLocalSolvers" 2213684f6988SStefano Zampini PetscErrorCode PCBDDCSetUpLocalSolvers(PC pc, PetscBool dirichlet, PetscBool neumann) 2214304d26faSStefano Zampini { 2215304d26faSStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 2216304d26faSStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 2217304d26faSStefano Zampini PC pc_temp; 2218304d26faSStefano Zampini Mat A_RR; 2219f4ddd8eeSStefano Zampini MatReuse reuse; 2220304d26faSStefano Zampini PetscScalar m_one = -1.0; 2221304d26faSStefano Zampini PetscReal value; 222204708bb6SStefano Zampini PetscInt n_D,n_R; 22239577ea80SStefano Zampini PetscBool use_exact,use_exact_reduced,issbaij; 2224304d26faSStefano Zampini PetscErrorCode ierr; 2225e604994aSStefano Zampini /* prefixes stuff */ 2226312be037SStefano Zampini char dir_prefix[256],neu_prefix[256],str_level[16]; 2227e604994aSStefano Zampini size_t len; 2228304d26faSStefano Zampini 2229304d26faSStefano Zampini PetscFunctionBegin; 2230304d26faSStefano Zampini 2231e604994aSStefano Zampini /* compute prefixes */ 2232e604994aSStefano Zampini ierr = PetscStrcpy(dir_prefix,"");CHKERRQ(ierr); 2233e604994aSStefano Zampini ierr = PetscStrcpy(neu_prefix,"");CHKERRQ(ierr); 2234e604994aSStefano Zampini if (!pcbddc->current_level) { 2235e604994aSStefano Zampini ierr = PetscStrcpy(dir_prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 2236e604994aSStefano Zampini ierr = PetscStrcpy(neu_prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 2237e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,"pc_bddc_dirichlet_");CHKERRQ(ierr); 2238e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,"pc_bddc_neumann_");CHKERRQ(ierr); 2239e604994aSStefano Zampini } else { 2240e604994aSStefano Zampini ierr = PetscStrcpy(str_level,"");CHKERRQ(ierr); 2241312be037SStefano Zampini sprintf(str_level,"l%d_",(int)(pcbddc->current_level)); 2242e604994aSStefano Zampini ierr = PetscStrlen(((PetscObject)pc)->prefix,&len);CHKERRQ(ierr); 2243e604994aSStefano Zampini len -= 15; /* remove "pc_bddc_coarse_" */ 2244312be037SStefano Zampini if (pcbddc->current_level>1) len -= 3; /* remove "lX_" with X level number */ 2245312be037SStefano Zampini if (pcbddc->current_level>10) len -= 1; /* remove another char from level number */ 224634d6797cSStefano Zampini ierr = PetscStrncpy(dir_prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 224734d6797cSStefano Zampini ierr = PetscStrncpy(neu_prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 2248e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,"pc_bddc_dirichlet_");CHKERRQ(ierr); 2249e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,"pc_bddc_neumann_");CHKERRQ(ierr); 2250e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,str_level);CHKERRQ(ierr); 2251e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,str_level);CHKERRQ(ierr); 2252e604994aSStefano Zampini } 2253e604994aSStefano Zampini 2254304d26faSStefano Zampini /* DIRICHLET PROBLEM */ 2255684f6988SStefano Zampini if (dirichlet) { 2256d5574798SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 22573301b35fSStefano Zampini if (pcbddc->local_mat->symmetric_set) { 22583301b35fSStefano Zampini ierr = MatSetOption(pcis->A_II,MAT_SYMMETRIC,pcbddc->local_mat->symmetric_set);CHKERRQ(ierr); 2259964fefecSStefano Zampini } 2260ac78edfcSStefano Zampini /* Matrix for Dirichlet problem is pcis->A_II */ 2261964fefecSStefano Zampini n_D = pcis->n - pcis->n_B; 2262304d26faSStefano Zampini if (!pcbddc->ksp_D) { /* create object if not yet build */ 2263304d26faSStefano Zampini ierr = KSPCreate(PETSC_COMM_SELF,&pcbddc->ksp_D);CHKERRQ(ierr); 2264304d26faSStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->ksp_D,(PetscObject)pc,1);CHKERRQ(ierr); 2265304d26faSStefano Zampini /* default */ 2266304d26faSStefano Zampini ierr = KSPSetType(pcbddc->ksp_D,KSPPREONLY);CHKERRQ(ierr); 2267e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->ksp_D,dir_prefix);CHKERRQ(ierr); 22689577ea80SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcis->A_II,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 2269304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_D,&pc_temp);CHKERRQ(ierr); 22709577ea80SStefano Zampini if (issbaij) { 22719577ea80SStefano Zampini ierr = PCSetType(pc_temp,PCCHOLESKY);CHKERRQ(ierr); 22729577ea80SStefano Zampini } else { 2273304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCLU);CHKERRQ(ierr); 22749577ea80SStefano Zampini } 2275304d26faSStefano Zampini /* Allow user's customization */ 2276304d26faSStefano Zampini ierr = KSPSetFromOptions(pcbddc->ksp_D);CHKERRQ(ierr); 2277304d26faSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 2278304d26faSStefano Zampini } 2279d1e9a80fSBarry Smith ierr = KSPSetOperators(pcbddc->ksp_D,pcis->A_II,pcis->A_II);CHKERRQ(ierr); 2280d62866d3SStefano Zampini if (sub_schurs->reuse_mumps) { 2281d62866d3SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 2282d62866d3SStefano Zampini 2283d62866d3SStefano Zampini ierr = KSPSetPC(pcbddc->ksp_D,reuse_mumps->interior_solver);CHKERRQ(ierr); 2284d5574798SStefano Zampini } 2285304d26faSStefano Zampini /* umfpack interface has a bug when matrix dimension is zero. TODO solve from umfpack interface */ 2286304d26faSStefano Zampini if (!n_D) { 2287304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_D,&pc_temp);CHKERRQ(ierr); 2288304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCNONE);CHKERRQ(ierr); 2289304d26faSStefano Zampini } 2290304d26faSStefano Zampini /* Set Up KSP for Dirichlet problem of BDDC */ 2291304d26faSStefano Zampini ierr = KSPSetUp(pcbddc->ksp_D);CHKERRQ(ierr); 2292304d26faSStefano Zampini /* set ksp_D into pcis data */ 2293304d26faSStefano Zampini ierr = KSPDestroy(&pcis->ksp_D);CHKERRQ(ierr); 2294304d26faSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->ksp_D);CHKERRQ(ierr); 2295304d26faSStefano Zampini pcis->ksp_D = pcbddc->ksp_D; 2296684f6988SStefano Zampini } 2297304d26faSStefano Zampini 2298304d26faSStefano Zampini /* NEUMANN PROBLEM */ 2299684f6988SStefano Zampini A_RR = 0; 2300684f6988SStefano Zampini if (neumann) { 2301d62866d3SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 230204708bb6SStefano Zampini PetscInt ibs,mbs; 230304708bb6SStefano Zampini PetscBool issbaij; 230404708bb6SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 2305f4ddd8eeSStefano Zampini /* Matrix for Neumann problem is A_RR -> we need to create/reuse it at this point */ 23068ce42a96SStefano Zampini ierr = ISGetSize(pcbddc->is_R_local,&n_R);CHKERRQ(ierr); 2307f4ddd8eeSStefano Zampini if (pcbddc->ksp_R) { /* already created ksp */ 2308f4ddd8eeSStefano Zampini PetscInt nn_R; 230981d9aea3SBarry Smith ierr = KSPGetOperators(pcbddc->ksp_R,NULL,&A_RR);CHKERRQ(ierr); 2310f4ddd8eeSStefano Zampini ierr = PetscObjectReference((PetscObject)A_RR);CHKERRQ(ierr); 2311f4ddd8eeSStefano Zampini ierr = MatGetSize(A_RR,&nn_R,NULL);CHKERRQ(ierr); 2312f4ddd8eeSStefano Zampini if (nn_R != n_R) { /* old ksp is not reusable, so reset it */ 2313f4ddd8eeSStefano Zampini ierr = KSPReset(pcbddc->ksp_R);CHKERRQ(ierr); 2314f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 2315f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 2316f4ddd8eeSStefano Zampini } else { /* same sizes, but nonzero pattern depend on primal vertices so it can be changed */ 2317727cdba6SStefano Zampini if (pcbddc->new_primal_space_local) { /* we are not sure the matrix will have the same nonzero pattern */ 2318f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 2319f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 2320f4ddd8eeSStefano Zampini } else { /* safe to reuse the matrix */ 2321f4ddd8eeSStefano Zampini reuse = MAT_REUSE_MATRIX; 2322f4ddd8eeSStefano Zampini } 2323f4ddd8eeSStefano Zampini } 2324f4ddd8eeSStefano Zampini /* last check */ 2325d1e9a80fSBarry Smith if (pc->flag == DIFFERENT_NONZERO_PATTERN) { 2326f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 2327f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 2328f4ddd8eeSStefano Zampini } 2329f4ddd8eeSStefano Zampini } else { /* first time, so we need to create the matrix */ 2330f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 2331f4ddd8eeSStefano Zampini } 2332f4ddd8eeSStefano Zampini /* extract A_RR */ 2333af732b37SStefano Zampini ierr = MatGetBlockSize(pcbddc->local_mat,&mbs);CHKERRQ(ierr); 2334af732b37SStefano Zampini ierr = ISGetBlockSize(pcbddc->is_R_local,&ibs);CHKERRQ(ierr); 233504708bb6SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcbddc->local_mat,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 233604708bb6SStefano Zampini if (ibs != mbs) { /* need to convert to SEQAIJ to extract any submatrix with is_R_local */ 233704708bb6SStefano Zampini if (matis->A == pcbddc->local_mat) { 233804708bb6SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 233904708bb6SStefano Zampini ierr = MatConvert(matis->A,MATSEQAIJ,MAT_INITIAL_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 2340af732b37SStefano Zampini } else { 234104708bb6SStefano Zampini ierr = MatConvert(pcbddc->local_mat,MATSEQAIJ,MAT_REUSE_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 23426816873aSStefano Zampini } 234304708bb6SStefano Zampini } else if (issbaij) { /* need to convert to BAIJ to get offdiagonal blocks */ 234404708bb6SStefano Zampini if (matis->A == pcbddc->local_mat) { 234504708bb6SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 234604708bb6SStefano Zampini ierr = MatConvert(matis->A,MATSEQBAIJ,MAT_INITIAL_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 234704708bb6SStefano Zampini } else { 234804708bb6SStefano Zampini ierr = MatConvert(pcbddc->local_mat,MATSEQBAIJ,MAT_REUSE_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 234904708bb6SStefano Zampini } 235004708bb6SStefano Zampini } 2351f4ddd8eeSStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,pcbddc->is_R_local,pcbddc->is_R_local,reuse,&A_RR);CHKERRQ(ierr); 23523301b35fSStefano Zampini if (pcbddc->local_mat->symmetric_set) { 23533301b35fSStefano Zampini ierr = MatSetOption(A_RR,MAT_SYMMETRIC,pcbddc->local_mat->symmetric_set);CHKERRQ(ierr); 23546816873aSStefano Zampini } 2355f4ddd8eeSStefano Zampini if (!pcbddc->ksp_R) { /* create object if not present */ 2356304d26faSStefano Zampini ierr = KSPCreate(PETSC_COMM_SELF,&pcbddc->ksp_R);CHKERRQ(ierr); 2357304d26faSStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->ksp_R,(PetscObject)pc,1);CHKERRQ(ierr); 2358304d26faSStefano Zampini /* default */ 2359304d26faSStefano Zampini ierr = KSPSetType(pcbddc->ksp_R,KSPPREONLY);CHKERRQ(ierr); 2360e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->ksp_R,neu_prefix);CHKERRQ(ierr); 2361304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_temp);CHKERRQ(ierr); 23629577ea80SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)A_RR,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 23639577ea80SStefano Zampini if (issbaij) { 23649577ea80SStefano Zampini ierr = PCSetType(pc_temp,PCCHOLESKY);CHKERRQ(ierr); 23659577ea80SStefano Zampini } else { 2366304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCLU);CHKERRQ(ierr); 23679577ea80SStefano Zampini } 2368304d26faSStefano Zampini /* Allow user's customization */ 2369304d26faSStefano Zampini ierr = KSPSetFromOptions(pcbddc->ksp_R);CHKERRQ(ierr); 2370304d26faSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 2371304d26faSStefano Zampini } 2372304d26faSStefano Zampini /* umfpack interface has a bug when matrix dimension is zero. TODO solve from umfpack interface */ 2373304d26faSStefano Zampini if (!n_R) { 2374304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_temp);CHKERRQ(ierr); 2375304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCNONE);CHKERRQ(ierr); 2376304d26faSStefano Zampini } 2377d62866d3SStefano Zampini /* Reuse MUMPS solver if it is present */ 2378d62866d3SStefano Zampini if (sub_schurs->reuse_mumps) { 2379d62866d3SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 2380d62866d3SStefano Zampini 2381d62866d3SStefano Zampini ierr = KSPSetPC(pcbddc->ksp_R,reuse_mumps->correction_solver);CHKERRQ(ierr); 2382d62866d3SStefano Zampini } 2383ffd830a3SStefano Zampini ierr = KSPSetOperators(pcbddc->ksp_R,A_RR,A_RR);CHKERRQ(ierr); 2384304d26faSStefano Zampini /* Set Up KSP for Neumann problem of BDDC */ 2385304d26faSStefano Zampini ierr = KSPSetUp(pcbddc->ksp_R);CHKERRQ(ierr); 2386684f6988SStefano Zampini } 23876816873aSStefano Zampini /* free Neumann problem's matrix */ 23886816873aSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 2389304d26faSStefano Zampini 2390304d26faSStefano Zampini /* check Dirichlet and Neumann solvers and adapt them if a nullspace correction is needed */ 23910fccc4e9SStefano Zampini if (pcbddc->NullSpace || pcbddc->dbg_flag) { 2392684f6988SStefano Zampini if (pcbddc->dbg_flag) { 2393684f6988SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 2394684f6988SStefano Zampini ierr = PetscViewerASCIISynchronizedAllow(pcbddc->dbg_viewer,PETSC_TRUE);CHKERRQ(ierr); 2395684f6988SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 2396684f6988SStefano Zampini } 2397684f6988SStefano Zampini if (dirichlet) { /* Dirichlet */ 23980fccc4e9SStefano Zampini ierr = VecSetRandom(pcis->vec1_D,NULL);CHKERRQ(ierr); 23990fccc4e9SStefano Zampini ierr = MatMult(pcis->A_II,pcis->vec1_D,pcis->vec2_D);CHKERRQ(ierr); 24000fccc4e9SStefano Zampini ierr = KSPSolve(pcbddc->ksp_D,pcis->vec2_D,pcis->vec2_D);CHKERRQ(ierr); 24010fccc4e9SStefano Zampini ierr = VecAXPY(pcis->vec1_D,m_one,pcis->vec2_D);CHKERRQ(ierr); 24020fccc4e9SStefano Zampini ierr = VecNorm(pcis->vec1_D,NORM_INFINITY,&value);CHKERRQ(ierr); 2403304d26faSStefano Zampini /* need to be adapted? */ 2404b8ffe317SStefano Zampini use_exact = (PetscAbsReal(value) > 1.e-4 ? PETSC_FALSE : PETSC_TRUE); 2405b8ffe317SStefano Zampini ierr = MPI_Allreduce(&use_exact,&use_exact_reduced,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 2406b8ffe317SStefano Zampini ierr = PCBDDCSetUseExactDirichlet(pc,use_exact_reduced);CHKERRQ(ierr); 2407304d26faSStefano Zampini /* print info */ 2408304d26faSStefano Zampini if (pcbddc->dbg_flag) { 2409e604994aSStefano 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); 2410304d26faSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 2411304d26faSStefano Zampini } 2412b8ffe317SStefano Zampini if (pcbddc->NullSpace && !use_exact_reduced && !pcbddc->switch_static) { 2413298c0119SStefano Zampini ierr = PCBDDCNullSpaceAssembleCorrection(pc,PETSC_TRUE,pcis->is_I_local);CHKERRQ(ierr); 2414304d26faSStefano Zampini } 2415684f6988SStefano Zampini } 2416684f6988SStefano Zampini if (neumann) { /* Neumann */ 24176816873aSStefano Zampini ierr = KSPGetOperators(pcbddc->ksp_R,&A_RR,NULL);CHKERRQ(ierr); 24180fccc4e9SStefano Zampini ierr = VecSetRandom(pcbddc->vec1_R,NULL);CHKERRQ(ierr); 24190fccc4e9SStefano Zampini ierr = MatMult(A_RR,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 24200fccc4e9SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec2_R,pcbddc->vec2_R);CHKERRQ(ierr); 24210fccc4e9SStefano Zampini ierr = VecAXPY(pcbddc->vec1_R,m_one,pcbddc->vec2_R);CHKERRQ(ierr); 24220fccc4e9SStefano Zampini ierr = VecNorm(pcbddc->vec1_R,NORM_INFINITY,&value);CHKERRQ(ierr); 2423304d26faSStefano Zampini /* need to be adapted? */ 2424b8ffe317SStefano Zampini use_exact = (PetscAbsReal(value) > 1.e-4 ? PETSC_FALSE : PETSC_TRUE); 2425b8ffe317SStefano Zampini ierr = MPI_Allreduce(&use_exact,&use_exact_reduced,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 2426304d26faSStefano Zampini /* print info */ 2427304d26faSStefano Zampini if (pcbddc->dbg_flag) { 2428e604994aSStefano 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); 2429304d26faSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 2430304d26faSStefano Zampini } 2431b8ffe317SStefano Zampini if (pcbddc->NullSpace && !use_exact_reduced) { /* is it the right logic? */ 2432298c0119SStefano Zampini ierr = PCBDDCNullSpaceAssembleCorrection(pc,PETSC_FALSE,pcbddc->is_R_local);CHKERRQ(ierr); 2433304d26faSStefano Zampini } 24340fccc4e9SStefano Zampini } 2435684f6988SStefano Zampini } 2436304d26faSStefano Zampini PetscFunctionReturn(0); 2437304d26faSStefano Zampini } 2438304d26faSStefano Zampini 2439304d26faSStefano Zampini #undef __FUNCT__ 2440ba15a52eSStefano Zampini #define __FUNCT__ "PCBDDCSolveSubstructureCorrection" 244180677318SStefano Zampini static PetscErrorCode PCBDDCSolveSubstructureCorrection(PC pc, Vec inout_B, Vec inout_D, PetscBool applytranspose) 2442674ae819SStefano Zampini { 2443674ae819SStefano Zampini PetscErrorCode ierr; 2444674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 2445be83ff47SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 2446674ae819SStefano Zampini 2447674ae819SStefano Zampini PetscFunctionBegin; 2448be83ff47SStefano Zampini if (!sub_schurs->reuse_mumps) { 244980677318SStefano Zampini ierr = VecSet(pcbddc->vec1_R,0.);CHKERRQ(ierr); 245020c7b377SStefano Zampini } 245180677318SStefano Zampini if (!pcbddc->switch_static) { 245280677318SStefano Zampini if (applytranspose && pcbddc->local_auxmat1) { 245380677318SStefano Zampini ierr = MatMultTranspose(pcbddc->local_auxmat2,inout_B,pcbddc->vec1_C);CHKERRQ(ierr); 245480677318SStefano Zampini ierr = MatMultTransposeAdd(pcbddc->local_auxmat1,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr); 245520c7b377SStefano Zampini } 2456be83ff47SStefano Zampini if (!sub_schurs->reuse_mumps) { 245780677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 245880677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 245920c7b377SStefano Zampini } else { 2460be83ff47SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 2461be83ff47SStefano Zampini 246253892102SStefano Zampini ierr = VecScatterBegin(reuse_mumps->correction_scatter_B,inout_B,reuse_mumps->rhs_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 246353892102SStefano Zampini ierr = VecScatterEnd(reuse_mumps->correction_scatter_B,inout_B,reuse_mumps->rhs_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 246420c7b377SStefano Zampini } 2465be83ff47SStefano Zampini } else { 246680677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 246780677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 246880677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,inout_D,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 246980677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,inout_D,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 247080677318SStefano Zampini if (applytranspose && pcbddc->local_auxmat1) { 247180677318SStefano Zampini ierr = MatMultTranspose(pcbddc->local_auxmat2,pcbddc->vec1_R,pcbddc->vec1_C);CHKERRQ(ierr); 247280677318SStefano Zampini ierr = MatMultTransposeAdd(pcbddc->local_auxmat1,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr); 247380677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 247480677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2475674ae819SStefano Zampini } 2476674ae819SStefano Zampini } 2477be83ff47SStefano Zampini if (!sub_schurs->reuse_mumps) { 247880677318SStefano Zampini if (applytranspose) { 247980677318SStefano Zampini ierr = KSPSolveTranspose(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 248080677318SStefano Zampini } else { 248180677318SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 248280677318SStefano Zampini } 248353892102SStefano Zampini #if defined(PETSC_HAVE_MUMPS) 2484be83ff47SStefano Zampini } else { 2485be83ff47SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 2486be83ff47SStefano Zampini 2487be83ff47SStefano Zampini if (applytranspose) { 248853892102SStefano Zampini ierr = MatMumpsSolveSchurComplementTranspose(reuse_mumps->F,reuse_mumps->rhs_B,reuse_mumps->sol_B);CHKERRQ(ierr); 2489be83ff47SStefano Zampini } else { 249053892102SStefano Zampini ierr = MatMumpsSolveSchurComplement(reuse_mumps->F,reuse_mumps->rhs_B,reuse_mumps->sol_B);CHKERRQ(ierr); 2491be83ff47SStefano Zampini } 249253892102SStefano Zampini #endif 2493be83ff47SStefano Zampini } 249480677318SStefano Zampini ierr = VecSet(inout_B,0.);CHKERRQ(ierr); 249580677318SStefano Zampini if (!pcbddc->switch_static) { 2496be83ff47SStefano Zampini if (!sub_schurs->reuse_mumps) { 249780677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 249880677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2499be83ff47SStefano Zampini } else { 2500be83ff47SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 2501be83ff47SStefano Zampini 250253892102SStefano Zampini ierr = VecScatterBegin(reuse_mumps->correction_scatter_B,reuse_mumps->sol_B,inout_B,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 250353892102SStefano Zampini ierr = VecScatterEnd(reuse_mumps->correction_scatter_B,reuse_mumps->sol_B,inout_B,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2504be83ff47SStefano Zampini } 250580677318SStefano Zampini if (!applytranspose && pcbddc->local_auxmat1) { 250680677318SStefano Zampini ierr = MatMult(pcbddc->local_auxmat1,inout_B,pcbddc->vec1_C);CHKERRQ(ierr); 250780677318SStefano Zampini ierr = MatMultAdd(pcbddc->local_auxmat2,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr); 250880677318SStefano Zampini } 250980677318SStefano Zampini } else { 251080677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 251180677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 251280677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 251380677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 251480677318SStefano Zampini if (!applytranspose && pcbddc->local_auxmat1) { 251580677318SStefano Zampini ierr = MatMult(pcbddc->local_auxmat1,inout_B,pcbddc->vec1_C);CHKERRQ(ierr); 251680677318SStefano Zampini ierr = MatMultAdd(pcbddc->local_auxmat2,pcbddc->vec1_C,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 251780677318SStefano Zampini } 251880677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 251980677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 252080677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 252180677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2522674ae819SStefano Zampini } 2523674ae819SStefano Zampini PetscFunctionReturn(0); 2524674ae819SStefano Zampini } 2525674ae819SStefano Zampini 2526dc359a40SStefano Zampini /* parameter apply transpose determines if the interface preconditioner should be applied transposed or not */ 2527674ae819SStefano Zampini #undef __FUNCT__ 2528674ae819SStefano Zampini #define __FUNCT__ "PCBDDCApplyInterfacePreconditioner" 2529dc359a40SStefano Zampini PetscErrorCode PCBDDCApplyInterfacePreconditioner(PC pc, PetscBool applytranspose) 2530674ae819SStefano Zampini { 2531674ae819SStefano Zampini PetscErrorCode ierr; 2532674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 2533674ae819SStefano Zampini PC_IS* pcis = (PC_IS*) (pc->data); 2534674ae819SStefano Zampini const PetscScalar zero = 0.0; 2535674ae819SStefano Zampini 2536674ae819SStefano Zampini PetscFunctionBegin; 2537dc359a40SStefano Zampini /* Application of PSI^T or PHI^T (depending on applytranspose, see comment above) */ 2538dc359a40SStefano Zampini if (applytranspose) { 2539674ae819SStefano Zampini ierr = MatMultTranspose(pcbddc->coarse_phi_B,pcis->vec1_B,pcbddc->vec1_P);CHKERRQ(ierr); 25408eeda7d8SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultTransposeAdd(pcbddc->coarse_phi_D,pcis->vec1_D,pcbddc->vec1_P,pcbddc->vec1_P);CHKERRQ(ierr); } 2541dc359a40SStefano Zampini } else { 2542674ae819SStefano Zampini ierr = MatMultTranspose(pcbddc->coarse_psi_B,pcis->vec1_B,pcbddc->vec1_P);CHKERRQ(ierr); 2543674ae819SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultTransposeAdd(pcbddc->coarse_psi_D,pcis->vec1_D,pcbddc->vec1_P,pcbddc->vec1_P);CHKERRQ(ierr); } 254415aaf578SStefano Zampini } 2545efc2fbd9SStefano Zampini 2546efc2fbd9SStefano Zampini /* add p0 to the last value of vec1_P holding the coarse dof relative to p0 */ 25474f1b2e48SStefano Zampini if (pcbddc->benign_n) { 2548efc2fbd9SStefano Zampini PetscScalar *array; 25494f1b2e48SStefano Zampini PetscInt j; 2550efc2fbd9SStefano Zampini 2551efc2fbd9SStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 25524f1b2e48SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) array[pcbddc->local_primal_size-pcbddc->benign_n+j] += pcbddc->benign_p0[j]; 2553efc2fbd9SStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 2554efc2fbd9SStefano Zampini } 2555efc2fbd9SStefano Zampini 255612edc857SStefano Zampini /* start communications from local primal nodes to rhs of coarse solver */ 255712edc857SStefano Zampini ierr = VecSet(pcbddc->coarse_vec,zero);CHKERRQ(ierr); 255812edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataBegin(pc,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 255912edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataEnd(pc,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 256012edc857SStefano Zampini 25619f00e9b4SStefano Zampini /* Coarse solution -> rhs and sol updated inside PCBDDCScattarCoarseDataBegin/End */ 256212edc857SStefano Zampini /* TODO remove null space when doing multilevel */ 256312edc857SStefano Zampini if (pcbddc->coarse_ksp) { 256451694757SStefano Zampini Mat coarse_mat; 2565964fefecSStefano Zampini Vec rhs,sol; 256651694757SStefano Zampini MatNullSpace nullsp; 2567964fefecSStefano Zampini 2568964fefecSStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&rhs);CHKERRQ(ierr); 2569964fefecSStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&sol);CHKERRQ(ierr); 257051694757SStefano Zampini ierr = KSPGetOperators(pcbddc->coarse_ksp,&coarse_mat,NULL);CHKERRQ(ierr); 257151694757SStefano Zampini ierr = MatGetNullSpace(coarse_mat,&nullsp);CHKERRQ(ierr); 257251694757SStefano Zampini if (nullsp) { 257351694757SStefano Zampini ierr = MatNullSpaceRemove(nullsp,rhs);CHKERRQ(ierr); 257451694757SStefano Zampini } 257512edc857SStefano Zampini if (applytranspose) { 2576964fefecSStefano Zampini ierr = KSPSolveTranspose(pcbddc->coarse_ksp,rhs,sol);CHKERRQ(ierr); 257712edc857SStefano Zampini } else { 2578964fefecSStefano Zampini ierr = KSPSolve(pcbddc->coarse_ksp,rhs,sol);CHKERRQ(ierr); 257912edc857SStefano Zampini } 258051694757SStefano Zampini if (nullsp) { 258151694757SStefano Zampini ierr = MatNullSpaceRemove(nullsp,sol);CHKERRQ(ierr); 258251694757SStefano Zampini } 258312edc857SStefano Zampini } 2584674ae819SStefano Zampini 2585674ae819SStefano Zampini /* Local solution on R nodes */ 258680677318SStefano Zampini if (pcis->n) { /* in/out pcbddc->vec1_B,pcbddc->vec1_D */ 258780677318SStefano Zampini ierr = PCBDDCSolveSubstructureCorrection(pc,pcis->vec1_B,pcis->vec1_D,applytranspose);CHKERRQ(ierr); 25889f00e9b4SStefano Zampini } 2589674ae819SStefano Zampini 25909f00e9b4SStefano Zampini /* communications from coarse sol to local primal nodes */ 25919f00e9b4SStefano Zampini ierr = PCBDDCScatterCoarseDataBegin(pc,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 259212edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataEnd(pc,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2593674ae819SStefano Zampini 2594674ae819SStefano Zampini /* Sum contributions from two levels */ 2595dc359a40SStefano Zampini if (applytranspose) { 2596dc359a40SStefano Zampini ierr = MatMultAdd(pcbddc->coarse_psi_B,pcbddc->vec1_P,pcis->vec1_B,pcis->vec1_B);CHKERRQ(ierr); 2597dc359a40SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultAdd(pcbddc->coarse_psi_D,pcbddc->vec1_P,pcis->vec1_D,pcis->vec1_D);CHKERRQ(ierr); } 2598dc359a40SStefano Zampini } else { 2599674ae819SStefano Zampini ierr = MatMultAdd(pcbddc->coarse_phi_B,pcbddc->vec1_P,pcis->vec1_B,pcis->vec1_B);CHKERRQ(ierr); 26008eeda7d8SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultAdd(pcbddc->coarse_phi_D,pcbddc->vec1_P,pcis->vec1_D,pcis->vec1_D);CHKERRQ(ierr); } 2601dc359a40SStefano Zampini } 2602efc2fbd9SStefano Zampini /* store p0 */ 26034f1b2e48SStefano Zampini if (pcbddc->benign_n) { 2604efc2fbd9SStefano Zampini PetscScalar *array; 26054f1b2e48SStefano Zampini PetscInt j; 2606efc2fbd9SStefano Zampini 2607efc2fbd9SStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 26084f1b2e48SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) pcbddc->benign_p0[j] = array[pcbddc->local_primal_size-pcbddc->benign_n+j]; 2609efc2fbd9SStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 2610efc2fbd9SStefano Zampini } 2611674ae819SStefano Zampini PetscFunctionReturn(0); 2612674ae819SStefano Zampini } 2613674ae819SStefano Zampini 2614674ae819SStefano Zampini #undef __FUNCT__ 2615674ae819SStefano Zampini #define __FUNCT__ "PCBDDCScatterCoarseDataBegin" 261612edc857SStefano Zampini PetscErrorCode PCBDDCScatterCoarseDataBegin(PC pc,InsertMode imode, ScatterMode smode) 2617674ae819SStefano Zampini { 2618674ae819SStefano Zampini PetscErrorCode ierr; 2619674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 262058da7f69SStefano Zampini PetscScalar *array; 262112edc857SStefano Zampini Vec from,to; 2622674ae819SStefano Zampini 2623674ae819SStefano Zampini PetscFunctionBegin; 262412edc857SStefano Zampini if (smode == SCATTER_REVERSE) { /* from global to local -> get data from coarse solution */ 262512edc857SStefano Zampini from = pcbddc->coarse_vec; 262612edc857SStefano Zampini to = pcbddc->vec1_P; 262712edc857SStefano Zampini if (pcbddc->coarse_ksp) { /* get array from coarse processes */ 262812edc857SStefano Zampini Vec tvec; 262958da7f69SStefano Zampini 263058da7f69SStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 263158da7f69SStefano Zampini ierr = VecResetArray(tvec);CHKERRQ(ierr); 263212edc857SStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 263358da7f69SStefano Zampini ierr = VecGetArray(tvec,&array);CHKERRQ(ierr); 263458da7f69SStefano Zampini ierr = VecPlaceArray(from,array);CHKERRQ(ierr); 263558da7f69SStefano Zampini ierr = VecRestoreArray(tvec,&array);CHKERRQ(ierr); 263612edc857SStefano Zampini } 263712edc857SStefano Zampini } else { /* from local to global -> put data in coarse right hand side */ 263812edc857SStefano Zampini from = pcbddc->vec1_P; 263912edc857SStefano Zampini to = pcbddc->coarse_vec; 264012edc857SStefano Zampini } 264112edc857SStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,from,to,imode,smode);CHKERRQ(ierr); 2642674ae819SStefano Zampini PetscFunctionReturn(0); 2643674ae819SStefano Zampini } 2644674ae819SStefano Zampini 2645674ae819SStefano Zampini #undef __FUNCT__ 2646674ae819SStefano Zampini #define __FUNCT__ "PCBDDCScatterCoarseDataEnd" 264712edc857SStefano Zampini PetscErrorCode PCBDDCScatterCoarseDataEnd(PC pc, InsertMode imode, ScatterMode smode) 2648674ae819SStefano Zampini { 2649674ae819SStefano Zampini PetscErrorCode ierr; 2650674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 265158da7f69SStefano Zampini PetscScalar *array; 265212edc857SStefano Zampini Vec from,to; 2653674ae819SStefano Zampini 2654674ae819SStefano Zampini PetscFunctionBegin; 265512edc857SStefano Zampini if (smode == SCATTER_REVERSE) { /* from global to local -> get data from coarse solution */ 265612edc857SStefano Zampini from = pcbddc->coarse_vec; 265712edc857SStefano Zampini to = pcbddc->vec1_P; 265812edc857SStefano Zampini } else { /* from local to global -> put data in coarse right hand side */ 265912edc857SStefano Zampini from = pcbddc->vec1_P; 266012edc857SStefano Zampini to = pcbddc->coarse_vec; 266112edc857SStefano Zampini } 266212edc857SStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,from,to,imode,smode);CHKERRQ(ierr); 266312edc857SStefano Zampini if (smode == SCATTER_FORWARD) { 266412edc857SStefano Zampini if (pcbddc->coarse_ksp) { /* get array from coarse processes */ 266512edc857SStefano Zampini Vec tvec; 266658da7f69SStefano Zampini 266712edc857SStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 266858da7f69SStefano Zampini ierr = VecGetArray(to,&array);CHKERRQ(ierr); 266958da7f69SStefano Zampini ierr = VecPlaceArray(tvec,array);CHKERRQ(ierr); 267058da7f69SStefano Zampini ierr = VecRestoreArray(to,&array);CHKERRQ(ierr); 267158da7f69SStefano Zampini } 267258da7f69SStefano Zampini } else { 267358da7f69SStefano Zampini if (pcbddc->coarse_ksp) { /* restore array of pcbddc->coarse_vec */ 267458da7f69SStefano Zampini ierr = VecResetArray(from);CHKERRQ(ierr); 267512edc857SStefano Zampini } 267612edc857SStefano Zampini } 2677674ae819SStefano Zampini PetscFunctionReturn(0); 2678674ae819SStefano Zampini } 2679674ae819SStefano Zampini 2680984c4197SStefano Zampini /* uncomment for testing purposes */ 2681984c4197SStefano Zampini /* #define PETSC_MISSING_LAPACK_GESVD 1 */ 2682674ae819SStefano Zampini #undef __FUNCT__ 2683674ae819SStefano Zampini #define __FUNCT__ "PCBDDCConstraintsSetUp" 2684674ae819SStefano Zampini PetscErrorCode PCBDDCConstraintsSetUp(PC pc) 2685674ae819SStefano Zampini { 2686674ae819SStefano Zampini PetscErrorCode ierr; 2687674ae819SStefano Zampini PC_IS* pcis = (PC_IS*)(pc->data); 2688674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 2689674ae819SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 2690984c4197SStefano Zampini /* one and zero */ 2691984c4197SStefano Zampini PetscScalar one=1.0,zero=0.0; 2692984c4197SStefano Zampini /* space to store constraints and their local indices */ 26939162d606SStefano Zampini PetscScalar *constraints_data; 26949162d606SStefano Zampini PetscInt *constraints_idxs,*constraints_idxs_B; 26959162d606SStefano Zampini PetscInt *constraints_idxs_ptr,*constraints_data_ptr; 26969162d606SStefano Zampini PetscInt *constraints_n; 2697984c4197SStefano Zampini /* iterators */ 2698b3d85658SStefano Zampini PetscInt i,j,k,total_counts,total_counts_cc,cum; 2699984c4197SStefano Zampini /* BLAS integers */ 2700e310c8b4SStefano Zampini PetscBLASInt lwork,lierr; 2701e310c8b4SStefano Zampini PetscBLASInt Blas_N,Blas_M,Blas_K,Blas_one=1; 2702c4303822SStefano Zampini PetscBLASInt Blas_LDA,Blas_LDB,Blas_LDC; 2703727cdba6SStefano Zampini /* reuse */ 27040e6343abSStefano Zampini PetscInt olocal_primal_size,olocal_primal_size_cc; 27050e6343abSStefano Zampini PetscInt *olocal_primal_ref_node,*olocal_primal_ref_mult; 2706984c4197SStefano Zampini /* change of basis */ 2707b3d85658SStefano Zampini PetscBool qr_needed; 27089162d606SStefano Zampini PetscBT change_basis,qr_needed_idx; 2709984c4197SStefano Zampini /* auxiliary stuff */ 271064efe560SStefano Zampini PetscInt *nnz,*is_indices; 27118a0068c3SStefano Zampini PetscInt ncc; 2712984c4197SStefano Zampini /* some quantities */ 271345a1bb75SStefano Zampini PetscInt n_vertices,total_primal_vertices,valid_constraints; 2714a58a30b4SStefano Zampini PetscInt size_of_constraint,max_size_of_constraint=0,max_constraints,temp_constraints; 2715984c4197SStefano Zampini 2716674ae819SStefano Zampini PetscFunctionBegin; 27178e61c736SStefano Zampini /* Destroy Mat objects computed previously */ 27188e61c736SStefano Zampini ierr = MatDestroy(&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 27198e61c736SStefano Zampini ierr = MatDestroy(&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 2720088faed8SStefano Zampini /* save info on constraints from previous setup (if any) */ 2721088faed8SStefano Zampini olocal_primal_size = pcbddc->local_primal_size; 27220e6343abSStefano Zampini olocal_primal_size_cc = pcbddc->local_primal_size_cc; 27230e6343abSStefano Zampini ierr = PetscMalloc2(olocal_primal_size_cc,&olocal_primal_ref_node,olocal_primal_size_cc,&olocal_primal_ref_mult);CHKERRQ(ierr); 27240e6343abSStefano Zampini ierr = PetscMemcpy(olocal_primal_ref_node,pcbddc->local_primal_ref_node,olocal_primal_size_cc*sizeof(PetscInt));CHKERRQ(ierr); 27250e6343abSStefano Zampini ierr = PetscMemcpy(olocal_primal_ref_mult,pcbddc->local_primal_ref_mult,olocal_primal_size_cc*sizeof(PetscInt));CHKERRQ(ierr); 27260e6343abSStefano Zampini ierr = PetscFree2(pcbddc->local_primal_ref_node,pcbddc->local_primal_ref_mult);CHKERRQ(ierr); 2727088faed8SStefano Zampini ierr = PetscFree(pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 2728cf5a6209SStefano Zampini 2729cf5a6209SStefano Zampini /* print some info */ 2730cf5a6209SStefano Zampini if (pcbddc->dbg_flag) { 2731cf5a6209SStefano Zampini IS vertices; 2732cf5a6209SStefano Zampini PetscInt nv,nedges,nfaces; 2733cf5a6209SStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,&nfaces,NULL,&nedges,NULL,&vertices);CHKERRQ(ierr); 2734cf5a6209SStefano Zampini ierr = ISGetSize(vertices,&nv);CHKERRQ(ierr); 2735cf5a6209SStefano Zampini ierr = ISDestroy(&vertices);CHKERRQ(ierr); 2736cf5a6209SStefano Zampini ierr = PetscViewerASCIISynchronizedAllow(pcbddc->dbg_viewer,PETSC_TRUE);CHKERRQ(ierr); 2737cf5a6209SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"--------------------------------------------------------------\n");CHKERRQ(ierr); 2738cf5a6209SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate vertices (%d)\n",PetscGlobalRank,nv,pcbddc->use_vertices);CHKERRQ(ierr); 2739fd14bc51SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate edges (%d)\n",PetscGlobalRank,nedges,pcbddc->use_edges);CHKERRQ(ierr); 2740fd14bc51SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate faces (%d)\n",PetscGlobalRank,nfaces,pcbddc->use_faces);CHKERRQ(ierr); 2741cf5a6209SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 2742cf5a6209SStefano Zampini } 2743cf5a6209SStefano Zampini 2744cf5a6209SStefano Zampini if (!pcbddc->adaptive_selection) { 27459162d606SStefano Zampini IS ISForVertices,*ISForFaces,*ISForEdges; 2746cf5a6209SStefano Zampini MatNullSpace nearnullsp; 2747cf5a6209SStefano Zampini const Vec *nearnullvecs; 2748cf5a6209SStefano Zampini Vec *localnearnullsp; 2749cf5a6209SStefano Zampini PetscScalar *array; 2750cf5a6209SStefano Zampini PetscInt n_ISForFaces,n_ISForEdges,nnsp_size; 2751cf5a6209SStefano Zampini PetscBool nnsp_has_cnst; 2752674ae819SStefano Zampini /* LAPACK working arrays for SVD or POD */ 2753b3d85658SStefano Zampini PetscBool skip_lapack,boolforchange; 2754674ae819SStefano Zampini PetscScalar *work; 2755674ae819SStefano Zampini PetscReal *singular_vals; 2756674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX) 2757674ae819SStefano Zampini PetscReal *rwork; 2758674ae819SStefano Zampini #endif 2759674ae819SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 2760674ae819SStefano Zampini PetscScalar *temp_basis,*correlation_mat; 2761674ae819SStefano Zampini #else 2762964fefecSStefano Zampini PetscBLASInt dummy_int=1; 2763964fefecSStefano Zampini PetscScalar dummy_scalar=1.; 2764674ae819SStefano Zampini #endif 2765674ae819SStefano Zampini 2766674ae819SStefano Zampini /* Get index sets for faces, edges and vertices from graph */ 2767d06fc5fdSStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,&n_ISForFaces,&ISForFaces,&n_ISForEdges,&ISForEdges,&ISForVertices);CHKERRQ(ierr); 2768d06fc5fdSStefano Zampini /* free unneeded index sets */ 2769d06fc5fdSStefano Zampini if (!pcbddc->use_vertices) { 2770d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForVertices);CHKERRQ(ierr); 2771674ae819SStefano Zampini } 2772d06fc5fdSStefano Zampini if (!pcbddc->use_edges) { 2773d06fc5fdSStefano Zampini for (i=0;i<n_ISForEdges;i++) { 2774d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForEdges[i]);CHKERRQ(ierr); 2775d06fc5fdSStefano Zampini } 2776d06fc5fdSStefano Zampini ierr = PetscFree(ISForEdges);CHKERRQ(ierr); 2777d06fc5fdSStefano Zampini n_ISForEdges = 0; 2778d06fc5fdSStefano Zampini } 2779d06fc5fdSStefano Zampini if (!pcbddc->use_faces) { 2780d06fc5fdSStefano Zampini for (i=0;i<n_ISForFaces;i++) { 2781d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForFaces[i]);CHKERRQ(ierr); 2782d06fc5fdSStefano Zampini } 2783d06fc5fdSStefano Zampini ierr = PetscFree(ISForFaces);CHKERRQ(ierr); 2784d06fc5fdSStefano Zampini n_ISForFaces = 0; 2785d06fc5fdSStefano Zampini } 278670022509SStefano Zampini 278770022509SStefano Zampini #if defined(PETSC_USE_DEBUG) 278870022509SStefano Zampini /* HACK: when solving singular problems not using vertices, a change of basis is mandatory. 278970022509SStefano Zampini Also use_change_of_basis should be consistent among processors */ 279070022509SStefano Zampini if (pcbddc->NullSpace) { 279170022509SStefano Zampini PetscBool tbool[2],gbool[2]; 279270022509SStefano Zampini 279370022509SStefano Zampini if (!ISForVertices && !pcbddc->user_ChangeOfBasisMatrix) { 2794b8ffe317SStefano Zampini pcbddc->use_change_of_basis = PETSC_TRUE; 2795d06fc5fdSStefano Zampini if (!ISForEdges) { 2796d06fc5fdSStefano Zampini pcbddc->use_change_on_faces = PETSC_TRUE; 2797d06fc5fdSStefano Zampini } 2798b8ffe317SStefano Zampini } 2799d06fc5fdSStefano Zampini tbool[0] = pcbddc->use_change_of_basis; 2800d06fc5fdSStefano Zampini tbool[1] = pcbddc->use_change_on_faces; 2801d06fc5fdSStefano Zampini ierr = MPI_Allreduce(tbool,gbool,2,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 2802d06fc5fdSStefano Zampini pcbddc->use_change_of_basis = gbool[0]; 2803d06fc5fdSStefano Zampini pcbddc->use_change_on_faces = gbool[1]; 280498a51de6SStefano Zampini } 280570022509SStefano Zampini #endif 280608122e43SStefano Zampini 2807674ae819SStefano Zampini /* check if near null space is attached to global mat */ 2808674ae819SStefano Zampini ierr = MatGetNearNullSpace(pc->pmat,&nearnullsp);CHKERRQ(ierr); 2809674ae819SStefano Zampini if (nearnullsp) { 2810674ae819SStefano Zampini ierr = MatNullSpaceGetVecs(nearnullsp,&nnsp_has_cnst,&nnsp_size,&nearnullvecs);CHKERRQ(ierr); 2811f4ddd8eeSStefano Zampini /* remove any stored info */ 2812f4ddd8eeSStefano Zampini ierr = MatNullSpaceDestroy(&pcbddc->onearnullspace);CHKERRQ(ierr); 2813f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->onearnullvecs_state);CHKERRQ(ierr); 2814f4ddd8eeSStefano Zampini /* store information for BDDC solver reuse */ 2815f4ddd8eeSStefano Zampini ierr = PetscObjectReference((PetscObject)nearnullsp);CHKERRQ(ierr); 2816f4ddd8eeSStefano Zampini pcbddc->onearnullspace = nearnullsp; 2817473ba861SJed Brown ierr = PetscMalloc1(nnsp_size,&pcbddc->onearnullvecs_state);CHKERRQ(ierr); 2818f4ddd8eeSStefano Zampini for (i=0;i<nnsp_size;i++) { 2819f4ddd8eeSStefano Zampini ierr = PetscObjectStateGet((PetscObject)nearnullvecs[i],&pcbddc->onearnullvecs_state[i]);CHKERRQ(ierr); 2820f4ddd8eeSStefano Zampini } 2821984c4197SStefano Zampini } else { /* if near null space is not provided BDDC uses constants by default */ 2822984c4197SStefano Zampini nnsp_size = 0; 2823674ae819SStefano Zampini nnsp_has_cnst = PETSC_TRUE; 2824674ae819SStefano Zampini } 2825984c4197SStefano Zampini /* get max number of constraints on a single cc */ 2826984c4197SStefano Zampini max_constraints = nnsp_size; 2827984c4197SStefano Zampini if (nnsp_has_cnst) max_constraints++; 2828984c4197SStefano Zampini 2829674ae819SStefano Zampini /* 2830674ae819SStefano Zampini Evaluate maximum storage size needed by the procedure 28319162d606SStefano Zampini - Indices for connected component i stored at "constraints_idxs + constraints_idxs_ptr[i]" 28329162d606SStefano Zampini - Values for constraints on connected component i stored at "constraints_data + constraints_data_ptr[i]" 28339162d606SStefano Zampini There can be multiple constraints per connected component 2834674ae819SStefano Zampini */ 2835674ae819SStefano Zampini n_vertices = 0; 2836674ae819SStefano Zampini if (ISForVertices) { 2837674ae819SStefano Zampini ierr = ISGetSize(ISForVertices,&n_vertices);CHKERRQ(ierr); 2838674ae819SStefano Zampini } 28399162d606SStefano Zampini ncc = n_vertices+n_ISForFaces+n_ISForEdges; 28409162d606SStefano Zampini ierr = PetscMalloc3(ncc+1,&constraints_idxs_ptr,ncc+1,&constraints_data_ptr,ncc,&constraints_n);CHKERRQ(ierr); 28419162d606SStefano Zampini 28429162d606SStefano Zampini total_counts = n_ISForFaces+n_ISForEdges; 28439162d606SStefano Zampini total_counts *= max_constraints; 2844674ae819SStefano Zampini total_counts += n_vertices; 28454641a718SStefano Zampini ierr = PetscBTCreate(total_counts,&change_basis);CHKERRQ(ierr); 28469162d606SStefano Zampini 2847674ae819SStefano Zampini total_counts = 0; 2848674ae819SStefano Zampini max_size_of_constraint = 0; 2849674ae819SStefano Zampini for (i=0;i<n_ISForEdges+n_ISForFaces;i++) { 28509162d606SStefano Zampini IS used_is; 2851674ae819SStefano Zampini if (i<n_ISForEdges) { 28529162d606SStefano Zampini used_is = ISForEdges[i]; 2853674ae819SStefano Zampini } else { 28549162d606SStefano Zampini used_is = ISForFaces[i-n_ISForEdges]; 2855674ae819SStefano Zampini } 28569162d606SStefano Zampini ierr = ISGetSize(used_is,&j);CHKERRQ(ierr); 2857674ae819SStefano Zampini total_counts += j; 2858674ae819SStefano Zampini max_size_of_constraint = PetscMax(j,max_size_of_constraint); 2859674ae819SStefano Zampini } 28609162d606SStefano 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); 28619162d606SStefano Zampini 2862984c4197SStefano Zampini /* get local part of global near null space vectors */ 2863785e854fSJed Brown ierr = PetscMalloc1(nnsp_size,&localnearnullsp);CHKERRQ(ierr); 2864984c4197SStefano Zampini for (k=0;k<nnsp_size;k++) { 2865984c4197SStefano Zampini ierr = VecDuplicate(pcis->vec1_N,&localnearnullsp[k]);CHKERRQ(ierr); 2866e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,nearnullvecs[k],localnearnullsp[k],INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2867e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,nearnullvecs[k],localnearnullsp[k],INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2868984c4197SStefano Zampini } 2869674ae819SStefano Zampini 2870242a89d7SStefano Zampini /* whether or not to skip lapack calls */ 2871242a89d7SStefano Zampini skip_lapack = PETSC_TRUE; 2872a773dcb8SStefano Zampini if (n_ISForFaces+n_ISForEdges && max_constraints > 1 && !pcbddc->use_nnsp_true) skip_lapack = PETSC_FALSE; 2873242a89d7SStefano Zampini 2874984c4197SStefano Zampini /* First we issue queries to allocate optimal workspace for LAPACKgesvd (or LAPACKsyev if SVD is missing) */ 2875a773dcb8SStefano Zampini if (!skip_lapack) { 2876674ae819SStefano Zampini PetscScalar temp_work; 2877911cabfeSStefano Zampini 2878674ae819SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 2879984c4197SStefano Zampini /* Proper Orthogonal Decomposition (POD) using the snapshot method */ 2880785e854fSJed Brown ierr = PetscMalloc1(max_constraints*max_constraints,&correlation_mat);CHKERRQ(ierr); 2881785e854fSJed Brown ierr = PetscMalloc1(max_constraints,&singular_vals);CHKERRQ(ierr); 2882785e854fSJed Brown ierr = PetscMalloc1(max_size_of_constraint*max_constraints,&temp_basis);CHKERRQ(ierr); 2883674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX) 2884785e854fSJed Brown ierr = PetscMalloc1(3*max_constraints,&rwork);CHKERRQ(ierr); 2885674ae819SStefano Zampini #endif 2886674ae819SStefano Zampini /* now we evaluate the optimal workspace using query with lwork=-1 */ 2887c8244a33SStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_N);CHKERRQ(ierr); 2888c8244a33SStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_LDA);CHKERRQ(ierr); 2889674ae819SStefano Zampini lwork = -1; 2890674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 2891674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 2892c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,&temp_work,&lwork,&lierr)); 2893674ae819SStefano Zampini #else 2894c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,&temp_work,&lwork,rwork,&lierr)); 2895674ae819SStefano Zampini #endif 2896674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2897984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to SYEV Lapack routine %d",(int)lierr); 2898674ae819SStefano Zampini #else /* on missing GESVD */ 2899674ae819SStefano Zampini /* SVD */ 2900674ae819SStefano Zampini PetscInt max_n,min_n; 2901674ae819SStefano Zampini max_n = max_size_of_constraint; 2902984c4197SStefano Zampini min_n = max_constraints; 2903984c4197SStefano Zampini if (max_size_of_constraint < max_constraints) { 2904674ae819SStefano Zampini min_n = max_size_of_constraint; 2905984c4197SStefano Zampini max_n = max_constraints; 2906674ae819SStefano Zampini } 2907785e854fSJed Brown ierr = PetscMalloc1(min_n,&singular_vals);CHKERRQ(ierr); 2908674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX) 2909785e854fSJed Brown ierr = PetscMalloc1(5*min_n,&rwork);CHKERRQ(ierr); 2910674ae819SStefano Zampini #endif 2911674ae819SStefano Zampini /* now we evaluate the optimal workspace using query with lwork=-1 */ 2912674ae819SStefano Zampini lwork = -1; 2913e310c8b4SStefano Zampini ierr = PetscBLASIntCast(max_n,&Blas_M);CHKERRQ(ierr); 2914e310c8b4SStefano Zampini ierr = PetscBLASIntCast(min_n,&Blas_N);CHKERRQ(ierr); 2915b7d8b9f8SStefano Zampini ierr = PetscBLASIntCast(max_n,&Blas_LDA);CHKERRQ(ierr); 2916674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 2917674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 29189162d606SStefano 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)); 2919674ae819SStefano Zampini #else 29209162d606SStefano 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)); 2921674ae819SStefano Zampini #endif 2922674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2923984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to GESVD Lapack routine %d",(int)lierr); 2924984c4197SStefano Zampini #endif /* on missing GESVD */ 2925674ae819SStefano Zampini /* Allocate optimal workspace */ 2926674ae819SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(temp_work),&lwork);CHKERRQ(ierr); 2927854ce69bSBarry Smith ierr = PetscMalloc1(lwork,&work);CHKERRQ(ierr); 2928674ae819SStefano Zampini } 2929674ae819SStefano Zampini /* Now we can loop on constraining sets */ 2930674ae819SStefano Zampini total_counts = 0; 29319162d606SStefano Zampini constraints_idxs_ptr[0] = 0; 29329162d606SStefano Zampini constraints_data_ptr[0] = 0; 2933674ae819SStefano Zampini /* vertices */ 29349162d606SStefano Zampini if (n_vertices) { 2935674ae819SStefano Zampini ierr = ISGetIndices(ISForVertices,(const PetscInt**)&is_indices);CHKERRQ(ierr); 29369162d606SStefano Zampini ierr = PetscMemcpy(constraints_idxs,is_indices,n_vertices*sizeof(PetscInt));CHKERRQ(ierr); 2937674ae819SStefano Zampini for (i=0;i<n_vertices;i++) { 29389162d606SStefano Zampini constraints_n[total_counts] = 1; 29399162d606SStefano Zampini constraints_data[total_counts] = 1.0; 29409162d606SStefano Zampini constraints_idxs_ptr[total_counts+1] = constraints_idxs_ptr[total_counts]+1; 29419162d606SStefano Zampini constraints_data_ptr[total_counts+1] = constraints_data_ptr[total_counts]+1; 2942674ae819SStefano Zampini total_counts++; 2943674ae819SStefano Zampini } 2944674ae819SStefano Zampini ierr = ISRestoreIndices(ISForVertices,(const PetscInt**)&is_indices);CHKERRQ(ierr); 2945674ae819SStefano Zampini n_vertices = total_counts; 2946674ae819SStefano Zampini } 2947984c4197SStefano Zampini 2948674ae819SStefano Zampini /* edges and faces */ 29499162d606SStefano Zampini total_counts_cc = total_counts; 2950911cabfeSStefano Zampini for (ncc=0;ncc<n_ISForEdges+n_ISForFaces;ncc++) { 29519162d606SStefano Zampini IS used_is; 29529162d606SStefano Zampini PetscBool idxs_copied = PETSC_FALSE; 29539162d606SStefano Zampini 2954911cabfeSStefano Zampini if (ncc<n_ISForEdges) { 29559162d606SStefano Zampini used_is = ISForEdges[ncc]; 2956984c4197SStefano Zampini boolforchange = pcbddc->use_change_of_basis; /* change or not the basis on the edge */ 2957674ae819SStefano Zampini } else { 29589162d606SStefano Zampini used_is = ISForFaces[ncc-n_ISForEdges]; 2959984c4197SStefano Zampini boolforchange = (PetscBool)(pcbddc->use_change_of_basis && pcbddc->use_change_on_faces); /* change or not the basis on the face */ 2960674ae819SStefano Zampini } 2961674ae819SStefano Zampini temp_constraints = 0; /* zero the number of constraints I have on this conn comp */ 29629162d606SStefano Zampini 29639162d606SStefano Zampini ierr = ISGetSize(used_is,&size_of_constraint);CHKERRQ(ierr); 29649162d606SStefano Zampini ierr = ISGetIndices(used_is,(const PetscInt**)&is_indices);CHKERRQ(ierr); 2965984c4197SStefano Zampini /* change of basis should not be performed on local periodic nodes */ 2966984c4197SStefano Zampini if (pcbddc->mat_graph->mirrors && pcbddc->mat_graph->mirrors[is_indices[0]]) boolforchange = PETSC_FALSE; 2967674ae819SStefano Zampini if (nnsp_has_cnst) { 29685b08dc53SStefano Zampini PetscScalar quad_value; 29699162d606SStefano Zampini 29709162d606SStefano Zampini ierr = PetscMemcpy(constraints_idxs + constraints_idxs_ptr[total_counts_cc],is_indices,size_of_constraint*sizeof(PetscInt));CHKERRQ(ierr); 29719162d606SStefano Zampini idxs_copied = PETSC_TRUE; 29729162d606SStefano Zampini 2973a773dcb8SStefano Zampini if (!pcbddc->use_nnsp_true) { 2974674ae819SStefano Zampini quad_value = (PetscScalar)(1.0/PetscSqrtReal((PetscReal)size_of_constraint)); 2975a773dcb8SStefano Zampini } else { 2976a773dcb8SStefano Zampini quad_value = 1.0; 2977a773dcb8SStefano Zampini } 2978674ae819SStefano Zampini for (j=0;j<size_of_constraint;j++) { 29799162d606SStefano Zampini constraints_data[constraints_data_ptr[total_counts_cc]+j] = quad_value; 2980674ae819SStefano Zampini } 29819162d606SStefano Zampini temp_constraints++; 2982674ae819SStefano Zampini total_counts++; 2983674ae819SStefano Zampini } 2984674ae819SStefano Zampini for (k=0;k<nnsp_size;k++) { 2985984c4197SStefano Zampini PetscReal real_value; 29869162d606SStefano Zampini PetscScalar *ptr_to_data; 29879162d606SStefano Zampini 2988984c4197SStefano Zampini ierr = VecGetArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 29899162d606SStefano Zampini ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]+temp_constraints*size_of_constraint]; 2990674ae819SStefano Zampini for (j=0;j<size_of_constraint;j++) { 29919162d606SStefano Zampini ptr_to_data[j] = array[is_indices[j]]; 2992674ae819SStefano Zampini } 2993984c4197SStefano Zampini ierr = VecRestoreArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 2994984c4197SStefano Zampini /* check if array is null on the connected component */ 2995e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 29969162d606SStefano Zampini PetscStackCallBLAS("BLASasum",real_value = BLASasum_(&Blas_N,ptr_to_data,&Blas_one)); 29975b08dc53SStefano Zampini if (real_value > 0.0) { /* keep indices and values */ 2998674ae819SStefano Zampini temp_constraints++; 2999674ae819SStefano Zampini total_counts++; 30009162d606SStefano Zampini if (!idxs_copied) { 30019162d606SStefano Zampini ierr = PetscMemcpy(constraints_idxs + constraints_idxs_ptr[total_counts_cc],is_indices,size_of_constraint*sizeof(PetscInt));CHKERRQ(ierr); 30029162d606SStefano Zampini idxs_copied = PETSC_TRUE; 3003674ae819SStefano Zampini } 3004674ae819SStefano Zampini } 30059162d606SStefano Zampini } 30069162d606SStefano Zampini ierr = ISRestoreIndices(used_is,(const PetscInt**)&is_indices);CHKERRQ(ierr); 300745a1bb75SStefano Zampini valid_constraints = temp_constraints; 3008eb97c9d2SStefano Zampini if (!pcbddc->use_nnsp_true && temp_constraints) { 3009a773dcb8SStefano Zampini if (temp_constraints == 1) { /* just normalize the constraint */ 30109162d606SStefano Zampini PetscScalar norm,*ptr_to_data; 30119162d606SStefano Zampini 30129162d606SStefano Zampini ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]]; 3013a773dcb8SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 30149162d606SStefano Zampini PetscStackCallBLAS("BLASdot",norm = BLASdot_(&Blas_N,ptr_to_data,&Blas_one,ptr_to_data,&Blas_one)); 3015a773dcb8SStefano Zampini norm = 1.0/PetscSqrtReal(PetscRealPart(norm)); 30169162d606SStefano Zampini PetscStackCallBLAS("BLASscal",BLASscal_(&Blas_N,&norm,ptr_to_data,&Blas_one)); 3017a773dcb8SStefano Zampini } else { /* perform SVD */ 3018984c4197SStefano Zampini PetscReal tol = 1.0e-8; /* tolerance for retaining eigenmodes */ 30199162d606SStefano Zampini PetscScalar *ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]]; 3020674ae819SStefano Zampini 3021674ae819SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 3022984c4197SStefano Zampini /* SVD: Y = U*S*V^H -> U (eigenvectors of Y*Y^H) = Y*V*(S)^\dag 3023984c4197SStefano Zampini POD: Y^H*Y = V*D*V^H, D = S^H*S -> U = Y*V*D^(-1/2) 3024984c4197SStefano Zampini -> When PETSC_USE_COMPLEX and PETSC_MISSING_LAPACK_GESVD are defined 3025984c4197SStefano Zampini the constraints basis will differ (by a complex factor with absolute value equal to 1) 3026984c4197SStefano Zampini from that computed using LAPACKgesvd 3027984c4197SStefano Zampini -> This is due to a different computation of eigenvectors in LAPACKheev 3028984c4197SStefano Zampini -> The quality of the POD-computed basis will be the same */ 3029984c4197SStefano Zampini ierr = PetscMemzero(correlation_mat,temp_constraints*temp_constraints*sizeof(PetscScalar));CHKERRQ(ierr); 3030674ae819SStefano Zampini /* Store upper triangular part of correlation matrix */ 3031e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 3032984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 3033674ae819SStefano Zampini for (j=0;j<temp_constraints;j++) { 3034674ae819SStefano Zampini for (k=0;k<j+1;k++) { 30359162d606SStefano 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)); 3036674ae819SStefano Zampini } 3037674ae819SStefano Zampini } 3038e310c8b4SStefano Zampini /* compute eigenvalues and eigenvectors of correlation matrix */ 3039e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 3040e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_LDA);CHKERRQ(ierr); 3041674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 3042c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,work,&lwork,&lierr)); 3043674ae819SStefano Zampini #else 3044c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,work,&lwork,rwork,&lierr)); 3045674ae819SStefano Zampini #endif 3046674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 3047984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYEV Lapack routine %d",(int)lierr); 3048984c4197SStefano Zampini /* retain eigenvalues greater than tol: note that LAPACKsyev gives eigs in ascending order */ 3049674ae819SStefano Zampini j = 0; 3050984c4197SStefano Zampini while (j < temp_constraints && singular_vals[j] < tol) j++; 3051674ae819SStefano Zampini total_counts = total_counts-j; 305245a1bb75SStefano Zampini valid_constraints = temp_constraints-j; 3053e310c8b4SStefano Zampini /* scale and copy POD basis into used quadrature memory */ 3054c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 3055c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 3056c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_K);CHKERRQ(ierr); 3057c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 3058c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_LDB);CHKERRQ(ierr); 3059c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDC);CHKERRQ(ierr); 3060674ae819SStefano Zampini if (j<temp_constraints) { 3061984c4197SStefano Zampini PetscInt ii; 3062984c4197SStefano Zampini for (k=j;k<temp_constraints;k++) singular_vals[k] = 1.0/PetscSqrtReal(singular_vals[k]); 3063674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 30649162d606SStefano 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)); 3065674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 3066984c4197SStefano Zampini for (k=0;k<temp_constraints-j;k++) { 3067674ae819SStefano Zampini for (ii=0;ii<size_of_constraint;ii++) { 30689162d606SStefano 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]; 3069674ae819SStefano Zampini } 3070674ae819SStefano Zampini } 3071674ae819SStefano Zampini } 3072674ae819SStefano Zampini #else /* on missing GESVD */ 3073e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 3074e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 3075b7d8b9f8SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 3076674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 3077674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 30789162d606SStefano 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)); 3079674ae819SStefano Zampini #else 30809162d606SStefano 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)); 3081674ae819SStefano Zampini #endif 3082984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GESVD Lapack routine %d",(int)lierr); 3083674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 3084984c4197SStefano Zampini /* retain eigenvalues greater than tol: note that LAPACKgesvd gives eigs in descending order */ 3085e310c8b4SStefano Zampini k = temp_constraints; 3086e310c8b4SStefano Zampini if (k > size_of_constraint) k = size_of_constraint; 3087674ae819SStefano Zampini j = 0; 3088e310c8b4SStefano Zampini while (j < k && singular_vals[k-j-1] < tol) j++; 308945a1bb75SStefano Zampini valid_constraints = k-j; 3090911cabfeSStefano Zampini total_counts = total_counts-temp_constraints+valid_constraints; 3091984c4197SStefano Zampini #endif /* on missing GESVD */ 3092674ae819SStefano Zampini } 3093a773dcb8SStefano Zampini } 30949162d606SStefano Zampini /* update pointers information */ 30959162d606SStefano Zampini if (valid_constraints) { 30969162d606SStefano Zampini constraints_n[total_counts_cc] = valid_constraints; 30979162d606SStefano Zampini constraints_idxs_ptr[total_counts_cc+1] = constraints_idxs_ptr[total_counts_cc]+size_of_constraint; 30989162d606SStefano Zampini constraints_data_ptr[total_counts_cc+1] = constraints_data_ptr[total_counts_cc]+size_of_constraint*valid_constraints; 30999162d606SStefano Zampini /* set change_of_basis flag */ 310045a1bb75SStefano Zampini if (boolforchange) { 3101b3d85658SStefano Zampini PetscBTSet(change_basis,total_counts_cc); 31029162d606SStefano Zampini } 3103b3d85658SStefano Zampini total_counts_cc++; 310445a1bb75SStefano Zampini } 310545a1bb75SStefano Zampini } 3106984c4197SStefano Zampini /* free workspace */ 31078f1c130eSStefano Zampini if (!skip_lapack) { 3108984c4197SStefano Zampini ierr = PetscFree(work);CHKERRQ(ierr); 3109984c4197SStefano Zampini #if defined(PETSC_USE_COMPLEX) 3110984c4197SStefano Zampini ierr = PetscFree(rwork);CHKERRQ(ierr); 3111984c4197SStefano Zampini #endif 3112984c4197SStefano Zampini ierr = PetscFree(singular_vals);CHKERRQ(ierr); 3113984c4197SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 3114984c4197SStefano Zampini ierr = PetscFree(correlation_mat);CHKERRQ(ierr); 3115984c4197SStefano Zampini ierr = PetscFree(temp_basis);CHKERRQ(ierr); 3116984c4197SStefano Zampini #endif 3117984c4197SStefano Zampini } 3118984c4197SStefano Zampini for (k=0;k<nnsp_size;k++) { 3119984c4197SStefano Zampini ierr = VecDestroy(&localnearnullsp[k]);CHKERRQ(ierr); 3120984c4197SStefano Zampini } 3121984c4197SStefano Zampini ierr = PetscFree(localnearnullsp);CHKERRQ(ierr); 3122cf5a6209SStefano Zampini /* free index sets of faces, edges and vertices */ 3123cf5a6209SStefano Zampini for (i=0;i<n_ISForFaces;i++) { 3124cf5a6209SStefano Zampini ierr = ISDestroy(&ISForFaces[i]);CHKERRQ(ierr); 3125cf5a6209SStefano Zampini } 3126cf5a6209SStefano Zampini if (n_ISForFaces) { 3127cf5a6209SStefano Zampini ierr = PetscFree(ISForFaces);CHKERRQ(ierr); 3128cf5a6209SStefano Zampini } 3129cf5a6209SStefano Zampini for (i=0;i<n_ISForEdges;i++) { 3130cf5a6209SStefano Zampini ierr = ISDestroy(&ISForEdges[i]);CHKERRQ(ierr); 3131cf5a6209SStefano Zampini } 3132cf5a6209SStefano Zampini if (n_ISForEdges) { 3133cf5a6209SStefano Zampini ierr = PetscFree(ISForEdges);CHKERRQ(ierr); 3134cf5a6209SStefano Zampini } 3135cf5a6209SStefano Zampini ierr = ISDestroy(&ISForVertices);CHKERRQ(ierr); 313608122e43SStefano Zampini } else { 313708122e43SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 3138984c4197SStefano Zampini 313908122e43SStefano Zampini total_counts = 0; 314008122e43SStefano Zampini n_vertices = 0; 3141d62866d3SStefano Zampini if (sub_schurs->is_vertices && pcbddc->use_vertices) { 3142d62866d3SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_vertices,&n_vertices);CHKERRQ(ierr); 314308122e43SStefano Zampini } 314408122e43SStefano Zampini max_constraints = 0; 31459162d606SStefano Zampini total_counts_cc = 0; 314608122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs+n_vertices;i++) { 314708122e43SStefano Zampini total_counts += pcbddc->adaptive_constraints_n[i]; 31489162d606SStefano Zampini if (pcbddc->adaptive_constraints_n[i]) total_counts_cc++; 314908122e43SStefano Zampini max_constraints = PetscMax(max_constraints,pcbddc->adaptive_constraints_n[i]); 315008122e43SStefano Zampini } 31519162d606SStefano Zampini constraints_idxs_ptr = pcbddc->adaptive_constraints_idxs_ptr; 31529162d606SStefano Zampini constraints_data_ptr = pcbddc->adaptive_constraints_data_ptr; 31539162d606SStefano Zampini constraints_idxs = pcbddc->adaptive_constraints_idxs; 31549162d606SStefano Zampini constraints_data = pcbddc->adaptive_constraints_data; 315574d5cdf7SStefano Zampini /* constraints_n differs from pcbddc->adaptive_constraints_n */ 31569162d606SStefano Zampini ierr = PetscMalloc1(total_counts_cc,&constraints_n);CHKERRQ(ierr); 31579162d606SStefano Zampini total_counts_cc = 0; 31589162d606SStefano Zampini for (i=0;i<sub_schurs->n_subs+n_vertices;i++) { 31599162d606SStefano Zampini if (pcbddc->adaptive_constraints_n[i]) { 31609162d606SStefano Zampini constraints_n[total_counts_cc++] = pcbddc->adaptive_constraints_n[i]; 316108122e43SStefano Zampini } 316208122e43SStefano Zampini } 31639162d606SStefano Zampini #if 0 31649162d606SStefano Zampini printf("Found %d totals (%d)\n",total_counts_cc,total_counts); 31659162d606SStefano Zampini for (i=0;i<total_counts_cc;i++) { 31669162d606SStefano Zampini printf("const %d, start %d",i,constraints_idxs_ptr[i]); 31679162d606SStefano Zampini printf(" end %d:\n",constraints_idxs_ptr[i+1]); 31689162d606SStefano Zampini for (j=constraints_idxs_ptr[i];j<constraints_idxs_ptr[i+1];j++) { 31699162d606SStefano Zampini printf(" %d",constraints_idxs[j]); 31709162d606SStefano Zampini } 31719162d606SStefano Zampini printf("\n"); 31729162d606SStefano Zampini printf("number of cc: %d\n",constraints_n[i]); 31739162d606SStefano Zampini } 31741b968477SStefano Zampini for (i=0;i<n_vertices;i++) { 31758bec7fa6SStefano Zampini PetscPrintf(PETSC_COMM_SELF,"[%d] vertex %d, n %d\n",PetscGlobalRank,i,pcbddc->adaptive_constraints_n[i]); 31761b968477SStefano Zampini } 31771b968477SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 31788bec7fa6SStefano 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]); 31791b968477SStefano Zampini } 318008122e43SStefano Zampini #endif 318108122e43SStefano Zampini 31828bec7fa6SStefano Zampini max_size_of_constraint = 0; 31839162d606SStefano 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]); 31849162d606SStefano Zampini ierr = PetscMalloc1(constraints_idxs_ptr[total_counts_cc],&constraints_idxs_B);CHKERRQ(ierr); 318508122e43SStefano Zampini /* Change of basis */ 3186b3d85658SStefano Zampini ierr = PetscBTCreate(total_counts_cc,&change_basis);CHKERRQ(ierr); 318708122e43SStefano Zampini if (pcbddc->use_change_of_basis) { 318808122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 318908122e43SStefano Zampini if (PetscBTLookup(sub_schurs->is_edge,i) || pcbddc->use_change_on_faces) { 3190b3d85658SStefano Zampini ierr = PetscBTSet(change_basis,i+n_vertices);CHKERRQ(ierr); 319108122e43SStefano Zampini } 319208122e43SStefano Zampini } 319308122e43SStefano Zampini } 319408122e43SStefano Zampini } 3195984c4197SStefano Zampini pcbddc->local_primal_size = total_counts; 31964f1b2e48SStefano Zampini ierr = PetscMalloc1(pcbddc->local_primal_size+pcbddc->benign_n,&pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 319708122e43SStefano Zampini 31989162d606SStefano Zampini /* map constraints_idxs in boundary numbering */ 31999162d606SStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,constraints_idxs_ptr[total_counts_cc],constraints_idxs,&i,constraints_idxs_B);CHKERRQ(ierr); 32009162d606SStefano Zampini if (i != constraints_idxs_ptr[total_counts_cc]) { 32019162d606SStefano 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); 320208122e43SStefano Zampini } 3203674ae819SStefano Zampini 3204674ae819SStefano Zampini /* Create constraint matrix */ 3205674ae819SStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 320616f15bc4SStefano Zampini ierr = MatSetType(pcbddc->ConstraintMatrix,MATAIJ);CHKERRQ(ierr); 3207984c4197SStefano Zampini ierr = MatSetSizes(pcbddc->ConstraintMatrix,pcbddc->local_primal_size,pcis->n,pcbddc->local_primal_size,pcis->n);CHKERRQ(ierr); 3208984c4197SStefano Zampini 3209984c4197SStefano Zampini /* find primal_dofs: subdomain corners plus dofs selected as primal after change of basis */ 3210a717540cSStefano Zampini /* determine if a QR strategy is needed for change of basis */ 3211a717540cSStefano Zampini qr_needed = PETSC_FALSE; 321274d5cdf7SStefano Zampini ierr = PetscBTCreate(total_counts_cc,&qr_needed_idx);CHKERRQ(ierr); 3213984c4197SStefano Zampini total_primal_vertices=0; 3214b3d85658SStefano Zampini pcbddc->local_primal_size_cc = 0; 32159162d606SStefano Zampini for (i=0;i<total_counts_cc;i++) { 32169162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 3217984c4197SStefano Zampini if (size_of_constraint == 1) { 32189162d606SStefano Zampini pcbddc->primal_indices_local_idxs[total_primal_vertices++] = constraints_idxs[constraints_idxs_ptr[i]]; 3219b3d85658SStefano Zampini pcbddc->local_primal_size_cc += 1; 322064efe560SStefano Zampini } else if (PetscBTLookup(change_basis,i)) { 32219162d606SStefano Zampini for (k=0;k<constraints_n[i];k++) { 32229162d606SStefano Zampini pcbddc->primal_indices_local_idxs[total_primal_vertices++] = constraints_idxs[constraints_idxs_ptr[i]+k]; 3223a717540cSStefano Zampini } 3224b3d85658SStefano Zampini pcbddc->local_primal_size_cc += constraints_n[i]; 322574d5cdf7SStefano Zampini if (constraints_n[i] > 1 || pcbddc->use_qr_single || pcbddc->faster_deluxe) { 3226a717540cSStefano Zampini PetscBTSet(qr_needed_idx,i); 3227a717540cSStefano Zampini qr_needed = PETSC_TRUE; 3228a717540cSStefano Zampini } 3229fa434743SStefano Zampini } else { 3230b3d85658SStefano Zampini pcbddc->local_primal_size_cc += 1; 3231fa434743SStefano Zampini } 3232a717540cSStefano Zampini } 3233b371cd4fSStefano Zampini /* note that the local variable n_vertices used below stores the number of pointwise constraints */ 3234b371cd4fSStefano Zampini pcbddc->n_vertices = total_primal_vertices; 3235674ae819SStefano Zampini /* permute indices in order to have a sorted set of vertices */ 323670022509SStefano Zampini ierr = PetscSortInt(total_primal_vertices,pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 3237b3d85658SStefano Zampini 32384f1b2e48SStefano 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); 32390e6343abSStefano Zampini ierr = PetscMemcpy(pcbddc->local_primal_ref_node,pcbddc->primal_indices_local_idxs,total_primal_vertices*sizeof(PetscInt));CHKERRQ(ierr); 32400e6343abSStefano Zampini for (i=0;i<total_primal_vertices;i++) pcbddc->local_primal_ref_mult[i] = 1; 3241984c4197SStefano Zampini 3242984c4197SStefano Zampini /* nonzero structure of constraint matrix */ 324374d5cdf7SStefano Zampini /* and get reference dof for local constraints */ 3244785e854fSJed Brown ierr = PetscMalloc1(pcbddc->local_primal_size,&nnz);CHKERRQ(ierr); 3245984c4197SStefano Zampini for (i=0;i<total_primal_vertices;i++) nnz[i] = 1; 324674d5cdf7SStefano Zampini 3247984c4197SStefano Zampini j = total_primal_vertices; 324874d5cdf7SStefano Zampini total_counts = total_primal_vertices; 3249b3d85658SStefano Zampini cum = total_primal_vertices; 32509162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 32514641a718SStefano Zampini if (!PetscBTLookup(change_basis,i)) { 3252b3d85658SStefano Zampini pcbddc->local_primal_ref_node[cum] = constraints_idxs[constraints_idxs_ptr[i]]; 3253b3d85658SStefano Zampini pcbddc->local_primal_ref_mult[cum] = constraints_n[i]; 3254b3d85658SStefano Zampini cum++; 32559162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 325674d5cdf7SStefano Zampini for (k=0;k<constraints_n[i];k++) { 325774d5cdf7SStefano Zampini pcbddc->primal_indices_local_idxs[total_counts++] = constraints_idxs[constraints_idxs_ptr[i]+k]; 325874d5cdf7SStefano Zampini nnz[j+k] = size_of_constraint; 325974d5cdf7SStefano Zampini } 32609162d606SStefano Zampini j += constraints_n[i]; 3261674ae819SStefano Zampini } 3262674ae819SStefano Zampini } 3263674ae819SStefano Zampini ierr = MatSeqAIJSetPreallocation(pcbddc->ConstraintMatrix,0,nnz);CHKERRQ(ierr); 3264674ae819SStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 3265088faed8SStefano Zampini 3266674ae819SStefano Zampini /* set values in constraint matrix */ 3267984c4197SStefano Zampini for (i=0;i<total_primal_vertices;i++) { 32680e6343abSStefano Zampini ierr = MatSetValue(pcbddc->ConstraintMatrix,i,pcbddc->local_primal_ref_node[i],1.0,INSERT_VALUES);CHKERRQ(ierr); 3269674ae819SStefano Zampini } 3270984c4197SStefano Zampini total_counts = total_primal_vertices; 32719162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 32724641a718SStefano Zampini if (!PetscBTLookup(change_basis,i)) { 32739162d606SStefano Zampini PetscInt *cols; 32749162d606SStefano Zampini 32759162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 32769162d606SStefano Zampini cols = constraints_idxs+constraints_idxs_ptr[i]; 32779162d606SStefano Zampini for (k=0;k<constraints_n[i];k++) { 32789162d606SStefano Zampini PetscInt row = total_counts+k; 32799162d606SStefano Zampini PetscScalar *vals; 32809162d606SStefano Zampini 32819162d606SStefano Zampini vals = constraints_data+constraints_data_ptr[i]+k*size_of_constraint; 32829162d606SStefano Zampini ierr = MatSetValues(pcbddc->ConstraintMatrix,1,&row,size_of_constraint,cols,vals,INSERT_VALUES);CHKERRQ(ierr); 32839162d606SStefano Zampini } 32849162d606SStefano Zampini total_counts += constraints_n[i]; 3285674ae819SStefano Zampini } 3286674ae819SStefano Zampini } 3287674ae819SStefano Zampini /* assembling */ 3288674ae819SStefano Zampini ierr = MatAssemblyBegin(pcbddc->ConstraintMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3289674ae819SStefano Zampini ierr = MatAssemblyEnd(pcbddc->ConstraintMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3290088faed8SStefano Zampini 3291984c4197SStefano Zampini /* 329245a1bb75SStefano Zampini ierr = PetscViewerSetFormat(PETSC_VIEWER_STDOUT_SELF,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 3293984c4197SStefano Zampini ierr = MatView(pcbddc->ConstraintMatrix,(PetscViewer)0);CHKERRQ(ierr); 3294984c4197SStefano Zampini */ 3295674ae819SStefano Zampini /* Create matrix for change of basis. We don't need it in case pcbddc->use_change_of_basis is FALSE */ 3296674ae819SStefano Zampini if (pcbddc->use_change_of_basis) { 3297026de310SStefano Zampini /* dual and primal dofs on a single cc */ 3298984c4197SStefano Zampini PetscInt dual_dofs,primal_dofs; 3299984c4197SStefano Zampini /* working stuff for GEQRF */ 330081d9aea3SBarry Smith PetscScalar *qr_basis,*qr_tau = NULL,*qr_work,lqr_work_t; 3301984c4197SStefano Zampini PetscBLASInt lqr_work; 3302984c4197SStefano Zampini /* working stuff for UNGQR */ 3303984c4197SStefano Zampini PetscScalar *gqr_work,lgqr_work_t; 3304984c4197SStefano Zampini PetscBLASInt lgqr_work; 3305984c4197SStefano Zampini /* working stuff for TRTRS */ 3306984c4197SStefano Zampini PetscScalar *trs_rhs; 33073f08241aSStefano Zampini PetscBLASInt Blas_NRHS; 3308984c4197SStefano Zampini /* pointers for values insertion into change of basis matrix */ 3309984c4197SStefano Zampini PetscInt *start_rows,*start_cols; 3310984c4197SStefano Zampini PetscScalar *start_vals; 3311984c4197SStefano Zampini /* working stuff for values insertion */ 33124641a718SStefano Zampini PetscBT is_primal; 331364efe560SStefano Zampini PetscInt *aux_primal_numbering_B; 3314906d46d4SStefano Zampini /* matrix sizes */ 3315906d46d4SStefano Zampini PetscInt global_size,local_size; 3316906d46d4SStefano Zampini /* temporary change of basis */ 3317906d46d4SStefano Zampini Mat localChangeOfBasisMatrix; 3318cf5a6209SStefano Zampini /* extra space for debugging */ 3319cf5a6209SStefano Zampini PetscScalar *dbg_work; 3320984c4197SStefano Zampini 3321906d46d4SStefano Zampini /* local temporary change of basis acts on local interfaces -> dimension is n_B x n_B */ 3322906d46d4SStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&localChangeOfBasisMatrix);CHKERRQ(ierr); 332316f15bc4SStefano Zampini ierr = MatSetType(localChangeOfBasisMatrix,MATAIJ);CHKERRQ(ierr); 3324bbb9e6c6SStefano Zampini ierr = MatSetSizes(localChangeOfBasisMatrix,pcis->n,pcis->n,pcis->n,pcis->n);CHKERRQ(ierr); 3325906d46d4SStefano Zampini /* nonzeros for local mat */ 3326bbb9e6c6SStefano Zampini ierr = PetscMalloc1(pcis->n,&nnz);CHKERRQ(ierr); 3327bbb9e6c6SStefano Zampini for (i=0;i<pcis->n;i++) nnz[i]=1; 33289162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 3329a717540cSStefano Zampini if (PetscBTLookup(change_basis,i)) { 33309162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 3331a717540cSStefano Zampini if (PetscBTLookup(qr_needed_idx,i)) { 33329162d606SStefano Zampini for (j=0;j<size_of_constraint;j++) nnz[constraints_idxs[constraints_idxs_ptr[i]+j]] = size_of_constraint; 3333a717540cSStefano Zampini } else { 33349162d606SStefano Zampini nnz[constraints_idxs[constraints_idxs_ptr[i]]] = size_of_constraint; 33359162d606SStefano Zampini for (j=1;j<size_of_constraint;j++) nnz[constraints_idxs[constraints_idxs_ptr[i]+j]] = 2; 3336a717540cSStefano Zampini } 3337a717540cSStefano Zampini } 3338a717540cSStefano Zampini } 3339906d46d4SStefano Zampini ierr = MatSeqAIJSetPreallocation(localChangeOfBasisMatrix,0,nnz);CHKERRQ(ierr); 3340bbb9e6c6SStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 3341a717540cSStefano Zampini /* Set initial identity in the matrix */ 3342bbb9e6c6SStefano Zampini for (i=0;i<pcis->n;i++) { 3343906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,i,i,1.0,INSERT_VALUES);CHKERRQ(ierr); 3344a717540cSStefano Zampini } 3345a717540cSStefano Zampini 3346a717540cSStefano Zampini if (pcbddc->dbg_flag) { 3347a717540cSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"--------------------------------------------------------------\n");CHKERRQ(ierr); 3348a717540cSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Checking change of basis computation for subdomain %04d\n",PetscGlobalRank);CHKERRQ(ierr); 3349a717540cSStefano Zampini } 3350a717540cSStefano Zampini 3351a717540cSStefano Zampini 3352a717540cSStefano Zampini /* Now we loop on the constraints which need a change of basis */ 3353a717540cSStefano Zampini /* 3354a717540cSStefano Zampini Change of basis matrix is evaluated similarly to the FIRST APPROACH in 3355a717540cSStefano Zampini Klawonn and Widlund, Dual-primal FETI-DP methods for linear elasticity, (see Sect 6.2.1) 3356a717540cSStefano Zampini 3357a6b551f4SStefano Zampini Basic blocks of change of basis matrix T computed by 3358a717540cSStefano Zampini 3359a6b551f4SStefano 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) 3360a6b551f4SStefano Zampini 3361a6b551f4SStefano Zampini | 1 0 ... 0 s_1/S | 3362a6b551f4SStefano Zampini | 0 1 ... 0 s_2/S | 3363a717540cSStefano Zampini | ... | 3364a6b551f4SStefano Zampini | 0 ... 1 s_{n-1}/S | 3365a6b551f4SStefano Zampini | -s_1/s_n ... -s_{n-1}/s_n s_n/S | 3366a717540cSStefano Zampini 3367a6b551f4SStefano Zampini with S = \sum_{i=1}^n s_i^2 3368a6b551f4SStefano Zampini NOTE: in the above example, the primal dof is the last one of the edge in LOCAL ordering 3369a6b551f4SStefano Zampini in the current implementation, the primal dof is the first one of the edge in GLOBAL ordering 3370a6b551f4SStefano Zampini 3371a6b551f4SStefano Zampini - QR decomposition of constraints otherwise 3372a717540cSStefano Zampini */ 3373a717540cSStefano Zampini if (qr_needed) { 3374984c4197SStefano Zampini /* space to store Q */ 3375854ce69bSBarry Smith ierr = PetscMalloc1(max_size_of_constraint*max_size_of_constraint,&qr_basis);CHKERRQ(ierr); 3376984c4197SStefano Zampini /* first we issue queries for optimal work */ 33773f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_M);CHKERRQ(ierr); 33783f08241aSStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_N);CHKERRQ(ierr); 33793f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 3380984c4197SStefano Zampini lqr_work = -1; 33813f08241aSStefano Zampini PetscStackCallBLAS("LAPACKgeqrf",LAPACKgeqrf_(&Blas_M,&Blas_N,qr_basis,&Blas_LDA,qr_tau,&lqr_work_t,&lqr_work,&lierr)); 3382984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to GEQRF Lapack routine %d",(int)lierr); 3383984c4197SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lqr_work_t),&lqr_work);CHKERRQ(ierr); 3384785e854fSJed Brown ierr = PetscMalloc1((PetscInt)PetscRealPart(lqr_work_t),&qr_work);CHKERRQ(ierr); 3385984c4197SStefano Zampini lgqr_work = -1; 33863f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_M);CHKERRQ(ierr); 33873f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_N);CHKERRQ(ierr); 33883f08241aSStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_K);CHKERRQ(ierr); 33893f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 33903f08241aSStefano Zampini if (Blas_K>Blas_M) Blas_K=Blas_M; /* adjust just for computing optimal work */ 33913f08241aSStefano Zampini PetscStackCallBLAS("LAPACKungqr",LAPACKungqr_(&Blas_M,&Blas_N,&Blas_K,qr_basis,&Blas_LDA,qr_tau,&lgqr_work_t,&lgqr_work,&lierr)); 3392984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to UNGQR Lapack routine %d",(int)lierr); 3393984c4197SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lgqr_work_t),&lgqr_work);CHKERRQ(ierr); 3394785e854fSJed Brown ierr = PetscMalloc1((PetscInt)PetscRealPart(lgqr_work_t),&gqr_work);CHKERRQ(ierr); 3395984c4197SStefano Zampini /* array to store scaling factors for reflectors */ 3396785e854fSJed Brown ierr = PetscMalloc1(max_constraints,&qr_tau);CHKERRQ(ierr); 3397984c4197SStefano Zampini /* array to store rhs and solution of triangular solver */ 3398785e854fSJed Brown ierr = PetscMalloc1(max_constraints*max_constraints,&trs_rhs);CHKERRQ(ierr); 3399a717540cSStefano Zampini /* allocating workspace for check */ 3400a717540cSStefano Zampini if (pcbddc->dbg_flag) { 3401cf5a6209SStefano Zampini ierr = PetscMalloc1(max_size_of_constraint*(max_constraints+max_size_of_constraint),&dbg_work);CHKERRQ(ierr); 3402a717540cSStefano Zampini } 3403a717540cSStefano Zampini } 3404984c4197SStefano Zampini /* array to store whether a node is primal or not */ 34054641a718SStefano Zampini ierr = PetscBTCreate(pcis->n_B,&is_primal);CHKERRQ(ierr); 3406473ba861SJed Brown ierr = PetscMalloc1(total_primal_vertices,&aux_primal_numbering_B);CHKERRQ(ierr); 34070e6343abSStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,total_primal_vertices,pcbddc->local_primal_ref_node,&i,aux_primal_numbering_B);CHKERRQ(ierr); 340839e2fb2aSStefano Zampini if (i != total_primal_vertices) { 340939e2fb2aSStefano Zampini SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_SUP,"Error in boundary numbering for BDDC vertices! %d != %d\n",total_primal_vertices,i); 34104641a718SStefano Zampini } 341139e2fb2aSStefano Zampini for (i=0;i<total_primal_vertices;i++) { 341239e2fb2aSStefano Zampini ierr = PetscBTSet(is_primal,aux_primal_numbering_B[i]);CHKERRQ(ierr); 341339e2fb2aSStefano Zampini } 341439e2fb2aSStefano Zampini ierr = PetscFree(aux_primal_numbering_B);CHKERRQ(ierr); 3415984c4197SStefano Zampini 3416a717540cSStefano Zampini /* loop on constraints and see whether or not they need a change of basis and compute it */ 34179162d606SStefano Zampini for (total_counts=n_vertices;total_counts<total_counts_cc;total_counts++) { 34189162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[total_counts+1]-constraints_idxs_ptr[total_counts]; 34194641a718SStefano Zampini if (PetscBTLookup(change_basis,total_counts)) { 3420984c4197SStefano Zampini /* get constraint info */ 34219162d606SStefano Zampini primal_dofs = constraints_n[total_counts]; 3422984c4197SStefano Zampini dual_dofs = size_of_constraint-primal_dofs; 3423984c4197SStefano Zampini 3424984c4197SStefano Zampini if (pcbddc->dbg_flag) { 34259162d606SStefano 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); 3426674ae819SStefano Zampini } 3427984c4197SStefano Zampini 3428fa434743SStefano Zampini if (PetscBTLookup(qr_needed_idx,total_counts)) { /* QR */ 3429a717540cSStefano Zampini 3430a717540cSStefano Zampini /* copy quadrature constraints for change of basis check */ 3431a717540cSStefano Zampini if (pcbddc->dbg_flag) { 34329162d606SStefano Zampini ierr = PetscMemcpy(dbg_work,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 3433a717540cSStefano Zampini } 3434984c4197SStefano Zampini /* copy temporary constraints into larger work vector (in order to store all columns of Q) */ 34359162d606SStefano Zampini ierr = PetscMemcpy(qr_basis,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 3436984c4197SStefano Zampini 3437984c4197SStefano Zampini /* compute QR decomposition of constraints */ 34383f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 34393f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 34403f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 3441674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 34423f08241aSStefano Zampini PetscStackCallBLAS("LAPACKgeqrf",LAPACKgeqrf_(&Blas_M,&Blas_N,qr_basis,&Blas_LDA,qr_tau,qr_work,&lqr_work,&lierr)); 3443984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GEQRF Lapack routine %d",(int)lierr); 3444674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 3445984c4197SStefano Zampini 3446984c4197SStefano Zampini /* explictly compute R^-T */ 3447984c4197SStefano Zampini ierr = PetscMemzero(trs_rhs,primal_dofs*primal_dofs*sizeof(*trs_rhs));CHKERRQ(ierr); 3448984c4197SStefano Zampini for (j=0;j<primal_dofs;j++) trs_rhs[j*(primal_dofs+1)] = 1.0; 34493f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 34503f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_NRHS);CHKERRQ(ierr); 34513f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 34523f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDB);CHKERRQ(ierr); 3453984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 34543f08241aSStefano Zampini PetscStackCallBLAS("LAPACKtrtrs",LAPACKtrtrs_("U","T","N",&Blas_N,&Blas_NRHS,qr_basis,&Blas_LDA,trs_rhs,&Blas_LDB,&lierr)); 3455984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in TRTRS Lapack routine %d",(int)lierr); 3456984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 3457984c4197SStefano Zampini 3458a717540cSStefano Zampini /* explicitly compute all columns of Q (Q = [Q1 | Q2] ) overwriting QR factorization in qr_basis */ 34593f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 34603f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 34613f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_K);CHKERRQ(ierr); 34623f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 3463984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 34643f08241aSStefano Zampini PetscStackCallBLAS("LAPACKungqr",LAPACKungqr_(&Blas_M,&Blas_N,&Blas_K,qr_basis,&Blas_LDA,qr_tau,gqr_work,&lgqr_work,&lierr)); 3465984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in UNGQR Lapack routine %d",(int)lierr); 3466984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 3467984c4197SStefano Zampini 3468984c4197SStefano Zampini /* first primal_dofs columns of Q need to be re-scaled in order to be unitary w.r.t constraints 3469984c4197SStefano Zampini i.e. C_{pxn}*Q_{nxn} should be equal to [I_pxp | 0_pxd] (see check below) 3470984c4197SStefano Zampini where n=size_of_constraint, p=primal_dofs, d=dual_dofs (n=p+d), I and 0 identity and null matrix resp. */ 34713f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 34723f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 34733f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_K);CHKERRQ(ierr); 34743f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 34753f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDB);CHKERRQ(ierr); 34763f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDC);CHKERRQ(ierr); 3477984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 34789162d606SStefano 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)); 3479984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 34809162d606SStefano Zampini ierr = PetscMemcpy(qr_basis,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 3481984c4197SStefano Zampini 3482984c4197SStefano Zampini /* insert values in change of basis matrix respecting global ordering of new primal dofs */ 34839162d606SStefano Zampini start_rows = &constraints_idxs[constraints_idxs_ptr[total_counts]]; 3484984c4197SStefano Zampini /* insert cols for primal dofs */ 3485984c4197SStefano Zampini for (j=0;j<primal_dofs;j++) { 3486984c4197SStefano Zampini start_vals = &qr_basis[j*size_of_constraint]; 34879162d606SStefano Zampini start_cols = &constraints_idxs[constraints_idxs_ptr[total_counts]+j]; 3488906d46d4SStefano Zampini ierr = MatSetValues(localChangeOfBasisMatrix,size_of_constraint,start_rows,1,start_cols,start_vals,INSERT_VALUES);CHKERRQ(ierr); 3489984c4197SStefano Zampini } 3490984c4197SStefano Zampini /* insert cols for dual dofs */ 3491984c4197SStefano Zampini for (j=0,k=0;j<dual_dofs;k++) { 34929162d606SStefano Zampini if (!PetscBTLookup(is_primal,constraints_idxs_B[constraints_idxs_ptr[total_counts]+k])) { 3493984c4197SStefano Zampini start_vals = &qr_basis[(primal_dofs+j)*size_of_constraint]; 34949162d606SStefano Zampini start_cols = &constraints_idxs[constraints_idxs_ptr[total_counts]+k]; 3495906d46d4SStefano Zampini ierr = MatSetValues(localChangeOfBasisMatrix,size_of_constraint,start_rows,1,start_cols,start_vals,INSERT_VALUES);CHKERRQ(ierr); 3496984c4197SStefano Zampini j++; 3497674ae819SStefano Zampini } 3498674ae819SStefano Zampini } 3499984c4197SStefano Zampini 3500984c4197SStefano Zampini /* check change of basis */ 3501984c4197SStefano Zampini if (pcbddc->dbg_flag) { 3502984c4197SStefano Zampini PetscInt ii,jj; 3503984c4197SStefano Zampini PetscBool valid_qr=PETSC_TRUE; 3504c4303822SStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_M);CHKERRQ(ierr); 3505c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 3506c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_K);CHKERRQ(ierr); 3507c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 3508c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDB);CHKERRQ(ierr); 3509c4303822SStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDC);CHKERRQ(ierr); 3510984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 3511cf5a6209SStefano 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)); 3512984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 3513984c4197SStefano Zampini for (jj=0;jj<size_of_constraint;jj++) { 3514984c4197SStefano Zampini for (ii=0;ii<primal_dofs;ii++) { 3515cf5a6209SStefano Zampini if (ii != jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]) > 1.e-12) valid_qr = PETSC_FALSE; 3516cf5a6209SStefano 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; 3517674ae819SStefano Zampini } 3518674ae819SStefano Zampini } 3519984c4197SStefano Zampini if (!valid_qr) { 352022d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> wrong change of basis!\n");CHKERRQ(ierr); 3521984c4197SStefano Zampini for (jj=0;jj<size_of_constraint;jj++) { 3522984c4197SStefano Zampini for (ii=0;ii<primal_dofs;ii++) { 3523cf5a6209SStefano Zampini if (ii != jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]) > 1.e-12) { 3524cf5a6209SStefano 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])); 3525674ae819SStefano Zampini } 3526cf5a6209SStefano Zampini if (ii == jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]-1.0) > 1.e-12) { 3527cf5a6209SStefano 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])); 3528984c4197SStefano Zampini } 3529984c4197SStefano Zampini } 3530984c4197SStefano Zampini } 3531674ae819SStefano Zampini } else { 353222d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> right change of basis!\n");CHKERRQ(ierr); 3533674ae819SStefano Zampini } 3534674ae819SStefano Zampini } 3535a717540cSStefano Zampini } else { /* simple transformation block */ 3536a717540cSStefano Zampini PetscInt row,col; 3537a6b551f4SStefano Zampini PetscScalar val,norm; 3538a6b551f4SStefano Zampini 3539a6b551f4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 35409162d606SStefano 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)); 3541a717540cSStefano Zampini for (j=0;j<size_of_constraint;j++) { 35429162d606SStefano Zampini PetscInt row_B = constraints_idxs_B[constraints_idxs_ptr[total_counts]+j]; 35439162d606SStefano Zampini row = constraints_idxs[constraints_idxs_ptr[total_counts]+j]; 3544bbb9e6c6SStefano Zampini if (!PetscBTLookup(is_primal,row_B)) { 35459162d606SStefano Zampini col = constraints_idxs[constraints_idxs_ptr[total_counts]]; 3546906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,row,1.0,INSERT_VALUES);CHKERRQ(ierr); 35479162d606SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,col,constraints_data[constraints_data_ptr[total_counts]+j]/norm,INSERT_VALUES);CHKERRQ(ierr); 3548a717540cSStefano Zampini } else { 3549a717540cSStefano Zampini for (k=0;k<size_of_constraint;k++) { 35509162d606SStefano Zampini col = constraints_idxs[constraints_idxs_ptr[total_counts]+k]; 3551a717540cSStefano Zampini if (row != col) { 35529162d606SStefano Zampini val = -constraints_data[constraints_data_ptr[total_counts]+k]/constraints_data[constraints_data_ptr[total_counts]]; 3553a717540cSStefano Zampini } else { 35549162d606SStefano Zampini val = constraints_data[constraints_data_ptr[total_counts]]/norm; 3555a717540cSStefano Zampini } 3556906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,col,val,INSERT_VALUES);CHKERRQ(ierr); 3557a717540cSStefano Zampini } 3558a717540cSStefano Zampini } 3559a717540cSStefano Zampini } 356098a51de6SStefano Zampini if (pcbddc->dbg_flag) { 356122d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> using standard change of basis\n");CHKERRQ(ierr); 3562a717540cSStefano Zampini } 3563674ae819SStefano Zampini } 3564984c4197SStefano Zampini } else { 3565984c4197SStefano Zampini if (pcbddc->dbg_flag) { 35669162d606SStefano 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); 3567674ae819SStefano Zampini } 3568674ae819SStefano Zampini } 3569674ae819SStefano Zampini } 3570a717540cSStefano Zampini 3571a717540cSStefano Zampini /* free workspace */ 3572a717540cSStefano Zampini if (qr_needed) { 3573984c4197SStefano Zampini if (pcbddc->dbg_flag) { 3574cf5a6209SStefano Zampini ierr = PetscFree(dbg_work);CHKERRQ(ierr); 3575984c4197SStefano Zampini } 3576984c4197SStefano Zampini ierr = PetscFree(trs_rhs);CHKERRQ(ierr); 3577984c4197SStefano Zampini ierr = PetscFree(qr_tau);CHKERRQ(ierr); 3578984c4197SStefano Zampini ierr = PetscFree(qr_work);CHKERRQ(ierr); 3579984c4197SStefano Zampini ierr = PetscFree(gqr_work);CHKERRQ(ierr); 3580984c4197SStefano Zampini ierr = PetscFree(qr_basis);CHKERRQ(ierr); 3581674ae819SStefano Zampini } 3582a717540cSStefano Zampini ierr = PetscBTDestroy(&is_primal);CHKERRQ(ierr); 3583906d46d4SStefano Zampini ierr = MatAssemblyBegin(localChangeOfBasisMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3584906d46d4SStefano Zampini ierr = MatAssemblyEnd(localChangeOfBasisMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3585906d46d4SStefano Zampini 3586906d46d4SStefano Zampini /* assembling of global change of variable */ 3587bbb9e6c6SStefano Zampini { 3588bbb9e6c6SStefano Zampini Mat tmat; 358916f15bc4SStefano Zampini PetscInt bs; 359016f15bc4SStefano Zampini 3591906d46d4SStefano Zampini ierr = VecGetSize(pcis->vec1_global,&global_size);CHKERRQ(ierr); 3592906d46d4SStefano Zampini ierr = VecGetLocalSize(pcis->vec1_global,&local_size);CHKERRQ(ierr); 3593bbb9e6c6SStefano Zampini ierr = MatDuplicate(pc->pmat,MAT_DO_NOT_COPY_VALUES,&tmat);CHKERRQ(ierr); 3594bbb9e6c6SStefano Zampini ierr = MatISSetLocalMat(tmat,localChangeOfBasisMatrix);CHKERRQ(ierr); 3595bbb9e6c6SStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)pc),&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 3596bbb9e6c6SStefano Zampini ierr = MatSetType(pcbddc->ChangeOfBasisMatrix,MATAIJ);CHKERRQ(ierr); 359716f15bc4SStefano Zampini ierr = MatGetBlockSize(pc->pmat,&bs);CHKERRQ(ierr); 359816f15bc4SStefano Zampini ierr = MatSetBlockSize(pcbddc->ChangeOfBasisMatrix,bs);CHKERRQ(ierr); 3599906d46d4SStefano Zampini ierr = MatSetSizes(pcbddc->ChangeOfBasisMatrix,local_size,local_size,global_size,global_size);CHKERRQ(ierr); 3600bbb9e6c6SStefano Zampini ierr = MatISSetMPIXAIJPreallocation_Private(tmat,pcbddc->ChangeOfBasisMatrix,PETSC_TRUE);CHKERRQ(ierr); 3601bbb9e6c6SStefano Zampini ierr = MatISGetMPIXAIJ(tmat,MAT_REUSE_MATRIX,&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 3602bbb9e6c6SStefano Zampini ierr = MatDestroy(&tmat);CHKERRQ(ierr); 3603bbb9e6c6SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 3604bbb9e6c6SStefano Zampini ierr = VecSet(pcis->vec1_N,1.0);CHKERRQ(ierr); 3605e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 3606e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 3607bbb9e6c6SStefano Zampini ierr = VecReciprocal(pcis->vec1_global);CHKERRQ(ierr); 3608bbb9e6c6SStefano Zampini ierr = MatDiagonalScale(pcbddc->ChangeOfBasisMatrix,pcis->vec1_global,NULL);CHKERRQ(ierr); 3609906d46d4SStefano Zampini } 3610906d46d4SStefano Zampini /* check */ 3611906d46d4SStefano Zampini if (pcbddc->dbg_flag) { 3612906d46d4SStefano Zampini PetscReal error; 3613906d46d4SStefano Zampini Vec x,x_change; 3614906d46d4SStefano Zampini 3615906d46d4SStefano Zampini ierr = VecDuplicate(pcis->vec1_global,&x);CHKERRQ(ierr); 3616906d46d4SStefano Zampini ierr = VecDuplicate(pcis->vec1_global,&x_change);CHKERRQ(ierr); 3617906d46d4SStefano Zampini ierr = VecSetRandom(x,NULL);CHKERRQ(ierr); 3618906d46d4SStefano Zampini ierr = VecCopy(x,pcis->vec1_global);CHKERRQ(ierr); 3619e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,x,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 3620e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,x,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 3621bbb9e6c6SStefano Zampini ierr = MatMult(localChangeOfBasisMatrix,pcis->vec1_N,pcis->vec2_N);CHKERRQ(ierr); 3622e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec2_N,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 3623e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec2_N,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 3624906d46d4SStefano Zampini ierr = MatMult(pcbddc->ChangeOfBasisMatrix,pcis->vec1_global,x_change);CHKERRQ(ierr); 3625906d46d4SStefano Zampini ierr = VecAXPY(x,-1.0,x_change);CHKERRQ(ierr); 3626906d46d4SStefano Zampini ierr = VecNorm(x,NORM_INFINITY,&error);CHKERRQ(ierr); 3627906d46d4SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 3628bbb9e6c6SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Error global vs local change: %1.6e\n",error);CHKERRQ(ierr); 3629906d46d4SStefano Zampini ierr = VecDestroy(&x);CHKERRQ(ierr); 3630906d46d4SStefano Zampini ierr = VecDestroy(&x_change);CHKERRQ(ierr); 3631906d46d4SStefano Zampini } 3632b96c3477SStefano Zampini 3633b96c3477SStefano Zampini /* adapt sub_schurs computed (if any) */ 3634b96c3477SStefano Zampini if (pcbddc->use_deluxe_scaling) { 3635b96c3477SStefano Zampini PCBDDCSubSchurs sub_schurs=pcbddc->sub_schurs; 3636b96c3477SStefano Zampini if (sub_schurs->S_Ej_all) { 3637ac632422SStefano Zampini Mat S_new,tmat; 3638b087196eSStefano Zampini ISLocalToGlobalMapping NtoSall; 3639b087196eSStefano Zampini IS is_all_N,is_V,is_V_Sall; 3640b087196eSStefano Zampini const PetscScalar *array; 3641b087196eSStefano Zampini const PetscInt *idxs_V,*idxs_all; 3642b087196eSStefano Zampini PetscInt i,n_V; 3643bbb9e6c6SStefano Zampini 3644bbb9e6c6SStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(pcis->BtoNmap,sub_schurs->is_Ej_all,&is_all_N);CHKERRQ(ierr); 36456816873aSStefano Zampini ierr = MatGetSubMatrix(localChangeOfBasisMatrix,is_all_N,is_all_N,MAT_INITIAL_MATRIX,&tmat);CHKERRQ(ierr); 3646b087196eSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,pcbddc->n_vertices,pcbddc->local_primal_ref_node,PETSC_COPY_VALUES,&is_V);CHKERRQ(ierr); 3647b087196eSStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(is_all_N,&NtoSall);CHKERRQ(ierr); 3648b087196eSStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(NtoSall,IS_GTOLM_DROP,is_V,&is_V_Sall);CHKERRQ(ierr); 3649b087196eSStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&NtoSall);CHKERRQ(ierr); 3650bbb9e6c6SStefano Zampini ierr = ISDestroy(&is_all_N);CHKERRQ(ierr); 3651b087196eSStefano Zampini ierr = ISDestroy(&is_V);CHKERRQ(ierr); 3652ac632422SStefano Zampini ierr = MatPtAP(sub_schurs->S_Ej_all,tmat,MAT_INITIAL_MATRIX,1.0,&S_new);CHKERRQ(ierr); 3653b96c3477SStefano Zampini ierr = MatDestroy(&sub_schurs->S_Ej_all);CHKERRQ(ierr); 3654ac632422SStefano Zampini ierr = PetscObjectReference((PetscObject)S_new);CHKERRQ(ierr); 3655b087196eSStefano Zampini ierr = MatZeroRowsColumnsIS(S_new,is_V_Sall,1.,NULL,NULL);CHKERRQ(ierr); 3656b087196eSStefano Zampini ierr = ISGetLocalSize(is_V_Sall,&n_V);CHKERRQ(ierr); 3657b087196eSStefano Zampini ierr = ISGetIndices(is_V_Sall,&idxs_V);CHKERRQ(ierr); 3658b087196eSStefano Zampini ierr = ISGetIndices(sub_schurs->is_Ej_all,&idxs_all);CHKERRQ(ierr); 3659b087196eSStefano Zampini ierr = VecGetArrayRead(pcis->D,&array);CHKERRQ(ierr); 3660b087196eSStefano Zampini for (i=0;i<n_V;i++) { 3661b087196eSStefano Zampini PetscScalar val; 3662b087196eSStefano Zampini PetscInt idx; 3663b087196eSStefano Zampini 3664b087196eSStefano Zampini idx = idxs_V[i]; 3665b087196eSStefano Zampini val = array[idxs_all[idxs_V[i]]]; 3666b087196eSStefano Zampini ierr = MatSetValue(S_new,idx,idx,val,INSERT_VALUES);CHKERRQ(ierr); 3667b087196eSStefano Zampini } 3668b087196eSStefano Zampini ierr = MatAssemblyBegin(S_new,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3669b087196eSStefano Zampini ierr = MatAssemblyEnd(S_new,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3670ac632422SStefano Zampini sub_schurs->S_Ej_all = S_new; 3671ac632422SStefano Zampini ierr = MatDestroy(&S_new);CHKERRQ(ierr); 3672ac632422SStefano Zampini if (sub_schurs->sum_S_Ej_all) { 3673ac632422SStefano Zampini ierr = MatPtAP(sub_schurs->sum_S_Ej_all,tmat,MAT_INITIAL_MATRIX,1.0,&S_new);CHKERRQ(ierr); 3674b96c3477SStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_all);CHKERRQ(ierr); 3675ac632422SStefano Zampini ierr = PetscObjectReference((PetscObject)S_new);CHKERRQ(ierr); 3676b087196eSStefano Zampini ierr = MatZeroRowsColumnsIS(S_new,is_V_Sall,1.,NULL,NULL);CHKERRQ(ierr); 3677ac632422SStefano Zampini sub_schurs->sum_S_Ej_all = S_new; 3678ac632422SStefano Zampini ierr = MatDestroy(&S_new);CHKERRQ(ierr); 3679ac632422SStefano Zampini } 3680b087196eSStefano Zampini ierr = VecRestoreArrayRead(pcis->D,&array);CHKERRQ(ierr); 3681b087196eSStefano Zampini ierr = ISRestoreIndices(sub_schurs->is_Ej_all,&idxs_all);CHKERRQ(ierr); 3682b087196eSStefano Zampini ierr = ISRestoreIndices(is_V_Sall,&idxs_V);CHKERRQ(ierr); 3683b96c3477SStefano Zampini ierr = MatDestroy(&tmat);CHKERRQ(ierr); 3684b087196eSStefano Zampini ierr = ISDestroy(&is_V_Sall);CHKERRQ(ierr); 3685b96c3477SStefano Zampini } 3686b96c3477SStefano Zampini } 3687906d46d4SStefano Zampini ierr = MatDestroy(&localChangeOfBasisMatrix);CHKERRQ(ierr); 3688906d46d4SStefano Zampini } else if (pcbddc->user_ChangeOfBasisMatrix) { 3689b9b85e73SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->user_ChangeOfBasisMatrix);CHKERRQ(ierr); 3690b9b85e73SStefano Zampini pcbddc->ChangeOfBasisMatrix = pcbddc->user_ChangeOfBasisMatrix; 3691b9b85e73SStefano Zampini } 3692906d46d4SStefano Zampini 3693906d46d4SStefano Zampini /* set up change of basis context */ 3694906d46d4SStefano Zampini if (pcbddc->ChangeOfBasisMatrix) { 3695906d46d4SStefano Zampini PCBDDCChange_ctx change_ctx; 3696906d46d4SStefano Zampini 3697906d46d4SStefano Zampini if (!pcbddc->new_global_mat) { 3698906d46d4SStefano Zampini PetscInt global_size,local_size; 3699906d46d4SStefano Zampini 3700906d46d4SStefano Zampini ierr = VecGetSize(pcis->vec1_global,&global_size);CHKERRQ(ierr); 3701906d46d4SStefano Zampini ierr = VecGetLocalSize(pcis->vec1_global,&local_size);CHKERRQ(ierr); 3702906d46d4SStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)pc),&pcbddc->new_global_mat);CHKERRQ(ierr); 3703906d46d4SStefano Zampini ierr = MatSetSizes(pcbddc->new_global_mat,local_size,local_size,global_size,global_size);CHKERRQ(ierr); 3704906d46d4SStefano Zampini ierr = MatSetType(pcbddc->new_global_mat,MATSHELL);CHKERRQ(ierr); 3705906d46d4SStefano Zampini ierr = MatShellSetOperation(pcbddc->new_global_mat,MATOP_MULT,(void (*)(void))PCBDDCMatMult_Private);CHKERRQ(ierr); 3706906d46d4SStefano Zampini ierr = MatShellSetOperation(pcbddc->new_global_mat,MATOP_MULT_TRANSPOSE,(void (*)(void))PCBDDCMatMultTranspose_Private);CHKERRQ(ierr); 3707906d46d4SStefano Zampini ierr = PetscNew(&change_ctx);CHKERRQ(ierr); 3708906d46d4SStefano Zampini ierr = MatShellSetContext(pcbddc->new_global_mat,change_ctx);CHKERRQ(ierr); 3709906d46d4SStefano Zampini } else { 3710906d46d4SStefano Zampini ierr = MatShellGetContext(pcbddc->new_global_mat,&change_ctx);CHKERRQ(ierr); 3711906d46d4SStefano Zampini ierr = MatDestroy(&change_ctx->global_change);CHKERRQ(ierr); 3712906d46d4SStefano Zampini ierr = VecDestroyVecs(2,&change_ctx->work);CHKERRQ(ierr); 3713906d46d4SStefano Zampini } 3714906d46d4SStefano Zampini if (!pcbddc->user_ChangeOfBasisMatrix) { 3715906d46d4SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 3716906d46d4SStefano Zampini change_ctx->global_change = pcbddc->ChangeOfBasisMatrix; 3717906d46d4SStefano Zampini } else { 3718906d46d4SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->user_ChangeOfBasisMatrix);CHKERRQ(ierr); 3719906d46d4SStefano Zampini change_ctx->global_change = pcbddc->user_ChangeOfBasisMatrix; 3720906d46d4SStefano Zampini } 3721906d46d4SStefano Zampini ierr = VecDuplicateVecs(pcis->vec1_global,2,&change_ctx->work);CHKERRQ(ierr); 3722906d46d4SStefano Zampini ierr = MatSetUp(pcbddc->new_global_mat);CHKERRQ(ierr); 3723906d46d4SStefano Zampini ierr = MatAssemblyBegin(pcbddc->new_global_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3724906d46d4SStefano Zampini ierr = MatAssemblyEnd(pcbddc->new_global_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3725b9b85e73SStefano Zampini } 3726a717540cSStefano Zampini 37274f1b2e48SStefano Zampini /* add pressure dofs to set of primal nodes for numbering purposes */ 37284f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 37294f1b2e48SStefano Zampini pcbddc->local_primal_ref_node[pcbddc->local_primal_size_cc] = pcbddc->benign_p0_lidx[i]; 37304f1b2e48SStefano Zampini pcbddc->primal_indices_local_idxs[pcbddc->local_primal_size] = pcbddc->benign_p0_lidx[i]; 3731019a44ceSStefano Zampini pcbddc->local_primal_ref_mult[pcbddc->local_primal_size_cc] = 1; 3732019a44ceSStefano Zampini pcbddc->local_primal_size_cc++; 3733019a44ceSStefano Zampini pcbddc->local_primal_size++; 3734019a44ceSStefano Zampini } 3735019a44ceSStefano Zampini 3736019a44ceSStefano Zampini /* check if a new primal space has been introduced (also take into account benign trick) */ 3737727cdba6SStefano Zampini pcbddc->new_primal_space_local = PETSC_TRUE; 3738727cdba6SStefano Zampini if (olocal_primal_size == pcbddc->local_primal_size) { 37399f47a83aSStefano 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); 3740c1c8e736SStefano Zampini pcbddc->new_primal_space_local = (PetscBool)(!pcbddc->new_primal_space_local); 37410e6343abSStefano Zampini if (!pcbddc->new_primal_space_local) { 37429f47a83aSStefano 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); 3743727cdba6SStefano Zampini pcbddc->new_primal_space_local = (PetscBool)(!pcbddc->new_primal_space_local); 3744727cdba6SStefano Zampini } 37450e6343abSStefano Zampini } 37460e6343abSStefano Zampini ierr = PetscFree2(olocal_primal_ref_node,olocal_primal_ref_mult);CHKERRQ(ierr); 3747727cdba6SStefano Zampini /* new_primal_space will be used for numbering of coarse dofs, so it should be the same across all subdomains */ 3748727cdba6SStefano Zampini ierr = MPI_Allreduce(&pcbddc->new_primal_space_local,&pcbddc->new_primal_space,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 3749727cdba6SStefano Zampini 3750a717540cSStefano Zampini /* flush dbg viewer */ 3751b8ffe317SStefano Zampini if (pcbddc->dbg_flag) { 3752b8ffe317SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 3753b8ffe317SStefano Zampini } 3754a717540cSStefano Zampini 3755e310c8b4SStefano Zampini /* free workspace */ 3756a717540cSStefano Zampini ierr = PetscBTDestroy(&qr_needed_idx);CHKERRQ(ierr); 37574641a718SStefano Zampini ierr = PetscBTDestroy(&change_basis);CHKERRQ(ierr); 375808122e43SStefano Zampini if (!pcbddc->adaptive_selection) { 37599162d606SStefano Zampini ierr = PetscFree3(constraints_idxs_ptr,constraints_data_ptr,constraints_n);CHKERRQ(ierr); 37609162d606SStefano Zampini ierr = PetscFree3(constraints_data,constraints_idxs,constraints_idxs_B);CHKERRQ(ierr); 376108122e43SStefano Zampini } else { 37629162d606SStefano Zampini ierr = PetscFree5(pcbddc->adaptive_constraints_n, 37639162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr, 37649162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr, 376508122e43SStefano Zampini pcbddc->adaptive_constraints_idxs, 376608122e43SStefano Zampini pcbddc->adaptive_constraints_data);CHKERRQ(ierr); 37679162d606SStefano Zampini ierr = PetscFree(constraints_n);CHKERRQ(ierr); 37689162d606SStefano Zampini ierr = PetscFree(constraints_idxs_B);CHKERRQ(ierr); 376908122e43SStefano Zampini } 3770674ae819SStefano Zampini PetscFunctionReturn(0); 3771674ae819SStefano Zampini } 3772674ae819SStefano Zampini 3773674ae819SStefano Zampini #undef __FUNCT__ 3774674ae819SStefano Zampini #define __FUNCT__ "PCBDDCAnalyzeInterface" 3775674ae819SStefano Zampini PetscErrorCode PCBDDCAnalyzeInterface(PC pc) 3776674ae819SStefano Zampini { 3777674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 3778674ae819SStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 3779674ae819SStefano Zampini Mat_IS *matis = (Mat_IS*)pc->pmat->data; 37807fb0e2dbSStefano Zampini PetscInt ierr,i,vertex_size,N; 3781674ae819SStefano Zampini PetscViewer viewer=pcbddc->dbg_viewer; 3782674ae819SStefano Zampini 3783674ae819SStefano Zampini PetscFunctionBegin; 37848e61c736SStefano Zampini /* Reset previously computed graph */ 37858e61c736SStefano Zampini ierr = PCBDDCGraphReset(pcbddc->mat_graph);CHKERRQ(ierr); 3786674ae819SStefano Zampini /* Init local Graph struct */ 37877fb0e2dbSStefano Zampini ierr = MatGetSize(pc->pmat,&N,NULL);CHKERRQ(ierr); 37883bbff08aSStefano Zampini ierr = PCBDDCGraphInit(pcbddc->mat_graph,pcis->mapping,N);CHKERRQ(ierr); 3789674ae819SStefano Zampini 3790575ad6abSStefano Zampini /* Check validity of the csr graph passed in by the user */ 37915099eff2SStefano Zampini if (pcbddc->mat_graph->nvtxs_csr && pcbddc->mat_graph->nvtxs_csr != pcbddc->mat_graph->nvtxs) { 37925099eff2SStefano 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); 3793575ad6abSStefano Zampini } 37949577ea80SStefano Zampini 3795674ae819SStefano Zampini /* Set default CSR adjacency of local dofs if not provided by the user with PCBDDCSetLocalAdjacencyGraph */ 37964d379d7bSStefano Zampini if (!pcbddc->mat_graph->xadj || !pcbddc->mat_graph->adjncy) { 37974d379d7bSStefano Zampini PetscInt *xadj,*adjncy; 37984d379d7bSStefano Zampini PetscInt nvtxs; 3799e496cd5dSStefano Zampini PetscBool flg_row=PETSC_FALSE; 3800674ae819SStefano Zampini 38014d379d7bSStefano Zampini if (pcbddc->use_local_adj) { 38022fffb893SStefano Zampini 38032fffb893SStefano Zampini ierr = MatGetRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 38042fffb893SStefano Zampini if (flg_row) { 38054d379d7bSStefano Zampini ierr = PCBDDCSetLocalAdjacencyGraph(pc,nvtxs,xadj,adjncy,PETSC_COPY_VALUES);CHKERRQ(ierr); 3806b96c3477SStefano Zampini pcbddc->computed_rowadj = PETSC_TRUE; 38072fffb893SStefano Zampini } 38082fffb893SStefano Zampini ierr = MatRestoreRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 38099b28b941SStefano 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 */ 38104d379d7bSStefano Zampini IS is_dummy; 38114d379d7bSStefano Zampini ISLocalToGlobalMapping l2gmap_dummy; 38124d379d7bSStefano Zampini PetscInt j,sum; 38134d379d7bSStefano Zampini PetscInt *cxadj,*cadjncy; 38144d379d7bSStefano Zampini const PetscInt *idxs; 38154d379d7bSStefano Zampini PCBDDCGraph graph; 38164d379d7bSStefano Zampini PetscBT is_on_boundary; 38174d379d7bSStefano Zampini 38184d379d7bSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcis->n,0,1,&is_dummy);CHKERRQ(ierr); 38194d379d7bSStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(is_dummy,&l2gmap_dummy);CHKERRQ(ierr); 38204d379d7bSStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 38214d379d7bSStefano Zampini ierr = PCBDDCGraphCreate(&graph);CHKERRQ(ierr); 38227fb0e2dbSStefano Zampini ierr = PCBDDCGraphInit(graph,l2gmap_dummy,pcis->n);CHKERRQ(ierr); 38234d379d7bSStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&l2gmap_dummy);CHKERRQ(ierr); 3824e496cd5dSStefano Zampini ierr = MatGetRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 3825e496cd5dSStefano Zampini if (flg_row) { 38264d379d7bSStefano Zampini graph->xadj = xadj; 38274d379d7bSStefano Zampini graph->adjncy = adjncy; 3828e496cd5dSStefano Zampini } 38294d379d7bSStefano Zampini ierr = PCBDDCGraphSetUp(graph,1,NULL,NULL,0,NULL,NULL);CHKERRQ(ierr); 38304d379d7bSStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(graph);CHKERRQ(ierr); 3831e496cd5dSStefano Zampini ierr = MatRestoreRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 38324d379d7bSStefano Zampini 38334d379d7bSStefano Zampini if (pcbddc->dbg_flag) { 38349b28b941SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"[%d] Found %d subdomains (local size %d)\n",PetscGlobalRank,graph->ncc,pcis->n);CHKERRQ(ierr); 38354d379d7bSStefano Zampini for (i=0;i<graph->ncc;i++) { 38364d379d7bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"[%d] %d cc size %d\n",PetscGlobalRank,i,graph->cptr[i+1]-graph->cptr[i]);CHKERRQ(ierr); 38374d379d7bSStefano Zampini } 38384d379d7bSStefano Zampini } 38394d379d7bSStefano Zampini 3840e496cd5dSStefano Zampini ierr = PetscBTCreate(pcis->n,&is_on_boundary);CHKERRQ(ierr); 38414d379d7bSStefano Zampini ierr = ISGetIndices(pcis->is_B_local,&idxs);CHKERRQ(ierr); 38424d379d7bSStefano Zampini for (i=0;i<pcis->n_B;i++) { 38434d379d7bSStefano Zampini ierr = PetscBTSet(is_on_boundary,idxs[i]);CHKERRQ(ierr); 38444d379d7bSStefano Zampini } 38454d379d7bSStefano Zampini ierr = ISRestoreIndices(pcis->is_B_local,&idxs);CHKERRQ(ierr); 38464d379d7bSStefano Zampini 3847e496cd5dSStefano Zampini ierr = PetscCalloc1(pcis->n+1,&cxadj);CHKERRQ(ierr); 38484d379d7bSStefano Zampini sum = 0; 38494d379d7bSStefano Zampini for (i=0;i<graph->ncc;i++) { 38504d379d7bSStefano Zampini PetscInt sizecc = 0; 38514d379d7bSStefano Zampini for (j=graph->cptr[i];j<graph->cptr[i+1];j++) { 38524d379d7bSStefano Zampini if (PetscBTLookup(is_on_boundary,graph->queue[j])) { 38534d379d7bSStefano Zampini sizecc++; 38544d379d7bSStefano Zampini } 38554d379d7bSStefano Zampini } 38564d379d7bSStefano Zampini for (j=graph->cptr[i];j<graph->cptr[i+1];j++) { 38574d379d7bSStefano Zampini if (PetscBTLookup(is_on_boundary,graph->queue[j])) { 38584d379d7bSStefano Zampini cxadj[graph->queue[j]] = sizecc; 38594d379d7bSStefano Zampini } 38604d379d7bSStefano Zampini } 38614d379d7bSStefano Zampini sum += sizecc*sizecc; 38624d379d7bSStefano Zampini } 38634d379d7bSStefano Zampini ierr = PetscMalloc1(sum,&cadjncy);CHKERRQ(ierr); 38644d379d7bSStefano Zampini sum = 0; 3865e496cd5dSStefano Zampini for (i=0;i<pcis->n;i++) { 38664d379d7bSStefano Zampini PetscInt temp = cxadj[i]; 38674d379d7bSStefano Zampini cxadj[i] = sum; 38684d379d7bSStefano Zampini sum += temp; 38694d379d7bSStefano Zampini } 3870e496cd5dSStefano Zampini cxadj[pcis->n] = sum; 38714d379d7bSStefano Zampini for (i=0;i<graph->ncc;i++) { 38724d379d7bSStefano Zampini for (j=graph->cptr[i];j<graph->cptr[i+1];j++) { 38734d379d7bSStefano Zampini if (PetscBTLookup(is_on_boundary,graph->queue[j])) { 38744d379d7bSStefano Zampini PetscInt k,sizecc = 0; 38754d379d7bSStefano Zampini for (k=graph->cptr[i];k<graph->cptr[i+1];k++) { 38764d379d7bSStefano Zampini if (PetscBTLookup(is_on_boundary,graph->queue[k])) { 38774d379d7bSStefano Zampini cadjncy[cxadj[graph->queue[j]]+sizecc]=graph->queue[k]; 38784d379d7bSStefano Zampini sizecc++; 38794d379d7bSStefano Zampini } 38804d379d7bSStefano Zampini } 38814d379d7bSStefano Zampini } 38824d379d7bSStefano Zampini } 38834d379d7bSStefano Zampini } 38849b28b941SStefano Zampini if (sum) { 3885e496cd5dSStefano Zampini ierr = PCBDDCSetLocalAdjacencyGraph(pc,pcis->n,cxadj,cadjncy,PETSC_OWN_POINTER);CHKERRQ(ierr); 38864d379d7bSStefano Zampini } else { 38874d379d7bSStefano Zampini ierr = PetscFree(cxadj);CHKERRQ(ierr); 38884d379d7bSStefano Zampini ierr = PetscFree(cadjncy);CHKERRQ(ierr); 38894d379d7bSStefano Zampini } 38904d379d7bSStefano Zampini graph->xadj = 0; 38914d379d7bSStefano Zampini graph->adjncy = 0; 38924d379d7bSStefano Zampini ierr = PCBDDCGraphDestroy(&graph);CHKERRQ(ierr); 38934d379d7bSStefano Zampini ierr = PetscBTDestroy(&is_on_boundary);CHKERRQ(ierr); 38944d379d7bSStefano Zampini } 3895674ae819SStefano Zampini } 38969b28b941SStefano Zampini if (pcbddc->dbg_flag) { 38979b28b941SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 3898674ae819SStefano Zampini } 3899674ae819SStefano Zampini 390063602bcaSStefano Zampini /* Set default dofs' splitting if no information has been provided by the user with PCBDDCSetDofsSplitting or PCBDDCSetDofsSplittingLocal */ 3901674ae819SStefano Zampini vertex_size = 1; 390263602bcaSStefano Zampini if (pcbddc->user_provided_isfordofs) { 390363602bcaSStefano Zampini if (pcbddc->n_ISForDofs) { /* need to convert from global to local and remove references to global dofs splitting */ 390495ecbf38SStefano Zampini ierr = PetscMalloc1(pcbddc->n_ISForDofs,&pcbddc->ISForDofsLocal);CHKERRQ(ierr); 390563602bcaSStefano Zampini for (i=0;i<pcbddc->n_ISForDofs;i++) { 3906e176bc59SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->rctx,pcis->vec1_global,pcis->vec1_N,pcbddc->ISForDofs[i],&pcbddc->ISForDofsLocal[i]);CHKERRQ(ierr); 390763602bcaSStefano Zampini ierr = ISDestroy(&pcbddc->ISForDofs[i]);CHKERRQ(ierr); 3908674ae819SStefano Zampini } 390963602bcaSStefano Zampini pcbddc->n_ISForDofsLocal = pcbddc->n_ISForDofs; 391063602bcaSStefano Zampini pcbddc->n_ISForDofs = 0; 391163602bcaSStefano Zampini ierr = PetscFree(pcbddc->ISForDofs);CHKERRQ(ierr); 3912674ae819SStefano Zampini } 391363602bcaSStefano Zampini /* mat block size as vertex size (used for elasticity with rigid body modes as nearnullspace) */ 3914674ae819SStefano Zampini ierr = MatGetBlockSize(matis->A,&vertex_size);CHKERRQ(ierr); 391563602bcaSStefano Zampini } else { 391663602bcaSStefano Zampini if (!pcbddc->n_ISForDofsLocal) { /* field split not present, create it in local ordering */ 391763602bcaSStefano Zampini ierr = MatGetBlockSize(pc->pmat,&pcbddc->n_ISForDofsLocal);CHKERRQ(ierr); 3918854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->n_ISForDofsLocal,&pcbddc->ISForDofsLocal);CHKERRQ(ierr); 391963602bcaSStefano Zampini for (i=0;i<pcbddc->n_ISForDofsLocal;i++) { 392063602bcaSStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)pc),pcis->n/pcbddc->n_ISForDofsLocal,i,pcbddc->n_ISForDofsLocal,&pcbddc->ISForDofsLocal[i]);CHKERRQ(ierr); 392163602bcaSStefano Zampini } 392263602bcaSStefano Zampini } 3923674ae819SStefano Zampini } 3924674ae819SStefano Zampini 3925674ae819SStefano Zampini /* Setup of Graph */ 3926785d1243SStefano Zampini if (!pcbddc->DirichletBoundariesLocal && pcbddc->DirichletBoundaries) { /* need to convert from global to local */ 3927e176bc59SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->rctx,pcis->vec1_global,pcis->vec1_N,pcbddc->DirichletBoundaries,&pcbddc->DirichletBoundariesLocal);CHKERRQ(ierr); 3928785d1243SStefano Zampini } 3929785d1243SStefano Zampini if (!pcbddc->NeumannBoundariesLocal && pcbddc->NeumannBoundaries) { /* need to convert from global to local */ 3930e176bc59SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->rctx,pcis->vec1_global,pcis->vec1_N,pcbddc->NeumannBoundaries,&pcbddc->NeumannBoundariesLocal);CHKERRQ(ierr); 3931785d1243SStefano Zampini } 393230368db7SStefano Zampini if (!pcbddc->user_primal_vertices_local && pcbddc->user_primal_vertices) { /* need to convert from global to local */ 393330368db7SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->rctx,pcis->vec1_global,pcis->vec1_N,pcbddc->user_primal_vertices,&pcbddc->user_primal_vertices_local);CHKERRQ(ierr); 393430368db7SStefano Zampini } 393530368db7SStefano Zampini ierr = PCBDDCGraphSetUp(pcbddc->mat_graph,vertex_size,pcbddc->NeumannBoundariesLocal,pcbddc->DirichletBoundariesLocal,pcbddc->n_ISForDofsLocal,pcbddc->ISForDofsLocal,pcbddc->user_primal_vertices_local);CHKERRQ(ierr); 3936674ae819SStefano Zampini 39374f1b2e48SStefano Zampini /* attach info on disconnected subdomains if present */ 39384f1b2e48SStefano Zampini if (pcbddc->n_local_subs) { 39394f1b2e48SStefano Zampini PetscInt *local_subs; 39404f1b2e48SStefano Zampini 39414f1b2e48SStefano Zampini ierr = PetscMalloc1(N,&local_subs);CHKERRQ(ierr); 39424f1b2e48SStefano Zampini for (i=0;i<pcbddc->n_local_subs;i++) { 39434f1b2e48SStefano Zampini const PetscInt *idxs; 39444f1b2e48SStefano Zampini PetscInt nl,j; 39454f1b2e48SStefano Zampini 39464f1b2e48SStefano Zampini ierr = ISGetLocalSize(pcbddc->local_subs[i],&nl);CHKERRQ(ierr); 39474f1b2e48SStefano Zampini ierr = ISGetIndices(pcbddc->local_subs[i],&idxs);CHKERRQ(ierr); 39484f1b2e48SStefano Zampini for (j=0;j<nl;j++) { 39494f1b2e48SStefano Zampini local_subs[idxs[j]] = i; 39504f1b2e48SStefano Zampini } 39514f1b2e48SStefano Zampini ierr = ISRestoreIndices(pcbddc->local_subs[i],&idxs);CHKERRQ(ierr); 39524f1b2e48SStefano Zampini } 39534f1b2e48SStefano Zampini pcbddc->mat_graph->n_local_subs = pcbddc->n_local_subs; 39544f1b2e48SStefano Zampini pcbddc->mat_graph->local_subs = local_subs; 39554f1b2e48SStefano Zampini } 39564f1b2e48SStefano Zampini 3957674ae819SStefano Zampini /* Graph's connected components analysis */ 3958674ae819SStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(pcbddc->mat_graph);CHKERRQ(ierr); 3959674ae819SStefano Zampini 3960674ae819SStefano Zampini /* print some info to stdout */ 3961674ae819SStefano Zampini if (pcbddc->dbg_flag) { 3962302440fdSBarry Smith ierr = PCBDDCGraphASCIIView(pcbddc->mat_graph,pcbddc->dbg_flag,viewer);CHKERRQ(ierr); 3963674ae819SStefano Zampini } 3964fb180af4SStefano Zampini 3965fb180af4SStefano Zampini /* mark topography has done */ 3966fb180af4SStefano Zampini pcbddc->recompute_topography = PETSC_FALSE; 3967674ae819SStefano Zampini PetscFunctionReturn(0); 3968674ae819SStefano Zampini } 3969674ae819SStefano Zampini 3970dc456d91SStefano Zampini /* given an index sets possibly with holes, renumbers the indexes removing the holes */ 3971674ae819SStefano Zampini #undef __FUNCT__ 3972674ae819SStefano Zampini #define __FUNCT__ "PCBDDCSubsetNumbering" 3973dc456d91SStefano Zampini PetscErrorCode PCBDDCSubsetNumbering(IS subset, IS subset_mult, PetscInt *N_n, IS *subset_n) 3974674ae819SStefano Zampini { 3975dc456d91SStefano Zampini PetscSF sf; 3976dc456d91SStefano Zampini PetscLayout map; 3977dc456d91SStefano Zampini const PetscInt *idxs; 3978dc456d91SStefano Zampini PetscInt *leaf_data,*root_data,*gidxs; 3979dc456d91SStefano Zampini PetscInt N,n,i,lbounds[2],gbounds[2],Nl; 3980dc456d91SStefano Zampini PetscInt n_n,nlocals,start,first_index; 3981dc456d91SStefano Zampini PetscMPIInt commsize; 3982674ae819SStefano Zampini PetscBool first_found; 3983674ae819SStefano Zampini PetscErrorCode ierr; 3984674ae819SStefano Zampini 3985674ae819SStefano Zampini PetscFunctionBegin; 3986dc456d91SStefano Zampini ierr = ISGetLocalSize(subset,&n);CHKERRQ(ierr); 3987dc456d91SStefano Zampini if (subset_mult) { 3988dc456d91SStefano Zampini PetscCheckSameComm(subset,1,subset_mult,2); 3989dc456d91SStefano Zampini ierr = ISGetLocalSize(subset,&i);CHKERRQ(ierr); 3990dc456d91SStefano Zampini if (i != n) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Local subset and multiplicity sizes don't match! %d != %d",n,i); 3991674ae819SStefano Zampini } 3992dc456d91SStefano Zampini /* create workspace layout for computing global indices of subset */ 3993dc456d91SStefano Zampini ierr = ISGetIndices(subset,&idxs);CHKERRQ(ierr); 3994dc456d91SStefano Zampini lbounds[0] = lbounds[1] = 0; 3995dc456d91SStefano Zampini for (i=0;i<n;i++) { 3996dc456d91SStefano Zampini if (idxs[i] < lbounds[0]) lbounds[0] = idxs[i]; 3997dc456d91SStefano Zampini else if (idxs[i] > lbounds[1]) lbounds[1] = idxs[i]; 3998674ae819SStefano Zampini } 3999dc456d91SStefano Zampini lbounds[0] = -lbounds[0]; 4000dc456d91SStefano Zampini ierr = MPI_Allreduce(lbounds,gbounds,2,MPIU_INT,MPI_MAX,PetscObjectComm((PetscObject)subset));CHKERRQ(ierr); 4001dc456d91SStefano Zampini gbounds[0] = -gbounds[0]; 4002dc456d91SStefano Zampini N = gbounds[1] - gbounds[0] + 1; 4003dc456d91SStefano Zampini ierr = PetscLayoutCreate(PetscObjectComm((PetscObject)subset),&map);CHKERRQ(ierr); 4004dc456d91SStefano Zampini ierr = PetscLayoutSetBlockSize(map,1);CHKERRQ(ierr); 4005dc456d91SStefano Zampini ierr = PetscLayoutSetSize(map,N);CHKERRQ(ierr); 4006dc456d91SStefano Zampini ierr = PetscLayoutSetUp(map);CHKERRQ(ierr); 4007dc456d91SStefano Zampini ierr = PetscLayoutGetLocalSize(map,&Nl);CHKERRQ(ierr); 4008dc456d91SStefano Zampini 4009dc456d91SStefano Zampini /* create sf : leaf_data == multiplicity of indexes, root data == global index in layout */ 4010dc456d91SStefano Zampini ierr = PetscMalloc2(n,&leaf_data,Nl,&root_data);CHKERRQ(ierr); 4011dc456d91SStefano Zampini if (subset_mult) { 4012dc456d91SStefano Zampini const PetscInt* idxs_mult; 4013dc456d91SStefano Zampini 4014dc456d91SStefano Zampini ierr = ISGetIndices(subset_mult,&idxs_mult);CHKERRQ(ierr); 4015dc456d91SStefano Zampini ierr = PetscMemcpy(leaf_data,idxs_mult,n*sizeof(PetscInt));CHKERRQ(ierr); 4016dc456d91SStefano Zampini ierr = ISRestoreIndices(subset_mult,&idxs_mult);CHKERRQ(ierr); 4017674ae819SStefano Zampini } else { 4018dc456d91SStefano Zampini for (i=0;i<n;i++) leaf_data[i] = 1; 4019674ae819SStefano Zampini } 4020dc456d91SStefano Zampini /* local size of new subset */ 4021dc456d91SStefano Zampini n_n = 0; 4022dc456d91SStefano Zampini for (i=0;i<n;i++) n_n += leaf_data[i]; 4023dc456d91SStefano Zampini 4024dc456d91SStefano Zampini /* global indexes in layout */ 4025dc456d91SStefano Zampini ierr = PetscMalloc1(n_n,&gidxs);CHKERRQ(ierr); /* allocating possibly extra space in gidxs which will be used later */ 4026dc456d91SStefano Zampini for (i=0;i<n;i++) gidxs[i] = idxs[i] - gbounds[0]; 4027dc456d91SStefano Zampini ierr = ISRestoreIndices(subset,&idxs);CHKERRQ(ierr); 4028dc456d91SStefano Zampini ierr = PetscSFCreate(PetscObjectComm((PetscObject)subset),&sf);CHKERRQ(ierr); 4029dc456d91SStefano Zampini ierr = PetscSFSetGraphLayout(sf,map,n,NULL,PETSC_COPY_VALUES,gidxs);CHKERRQ(ierr); 4030dc456d91SStefano Zampini ierr = PetscLayoutDestroy(&map);CHKERRQ(ierr); 4031dc456d91SStefano Zampini 4032dc456d91SStefano Zampini /* reduce from leaves to roots */ 4033dc456d91SStefano Zampini ierr = PetscMemzero(root_data,Nl*sizeof(PetscInt));CHKERRQ(ierr); 403464a8e5bcSStefano Zampini ierr = PetscSFReduceBegin(sf,MPIU_INT,leaf_data,root_data,MPI_MAX);CHKERRQ(ierr); 403564a8e5bcSStefano Zampini ierr = PetscSFReduceEnd(sf,MPIU_INT,leaf_data,root_data,MPI_MAX);CHKERRQ(ierr); 4036dc456d91SStefano Zampini 4037dc456d91SStefano Zampini /* count indexes in local part of layout */ 4038674ae819SStefano Zampini nlocals = 0; 4039674ae819SStefano Zampini first_index = -1; 4040674ae819SStefano Zampini first_found = PETSC_FALSE; 4041dc456d91SStefano Zampini for (i=0;i<Nl;i++) { 4042dc456d91SStefano Zampini if (!first_found && root_data[i]) { 4043674ae819SStefano Zampini first_found = PETSC_TRUE; 4044674ae819SStefano Zampini first_index = i; 4045674ae819SStefano Zampini } 4046dc456d91SStefano Zampini nlocals += root_data[i]; 4047674ae819SStefano Zampini } 4048dc456d91SStefano Zampini 4049dc456d91SStefano Zampini /* cumulative of number of indexes and size of subset without holes */ 40505fa240b1SStefano Zampini #if defined(PETSC_HAVE_MPI_EXSCAN) 4051dc456d91SStefano Zampini start = 0; 405264a8e5bcSStefano Zampini ierr = MPI_Exscan(&nlocals,&start,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)subset));CHKERRQ(ierr); 40535fa240b1SStefano Zampini #else 405464a8e5bcSStefano Zampini ierr = MPI_Scan(&nlocals,&start,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)subset));CHKERRQ(ierr); 40555fa240b1SStefano Zampini start = start-nlocals; 40565fa240b1SStefano Zampini #endif 40575fa240b1SStefano Zampini 4058dc456d91SStefano Zampini if (N_n) { /* compute total size of new subset if requested */ 4059dc456d91SStefano Zampini *N_n = start + nlocals; 4060dc456d91SStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)subset),&commsize);CHKERRQ(ierr); 4061dc456d91SStefano Zampini ierr = MPI_Bcast(N_n,1,MPIU_INT,commsize-1,PetscObjectComm((PetscObject)subset));CHKERRQ(ierr); 4062674ae819SStefano Zampini } 40635fa240b1SStefano Zampini 40645fa240b1SStefano Zampini /* adapt root data with cumulative */ 4065674ae819SStefano Zampini if (first_found) { 4066dc456d91SStefano Zampini PetscInt old_index; 4067dc456d91SStefano Zampini 4068dc456d91SStefano Zampini root_data[first_index] += start; 4069674ae819SStefano Zampini old_index = first_index; 4070dc456d91SStefano Zampini for (i=first_index+1;i<Nl;i++) { 4071dc456d91SStefano Zampini if (root_data[i]) { 4072dc456d91SStefano Zampini root_data[i] += root_data[old_index]; 4073674ae819SStefano Zampini old_index = i; 4074674ae819SStefano Zampini } 4075674ae819SStefano Zampini } 4076674ae819SStefano Zampini } 4077dc456d91SStefano Zampini 4078dc456d91SStefano Zampini /* from roots to leaves */ 4079dc456d91SStefano Zampini ierr = PetscSFBcastBegin(sf,MPIU_INT,root_data,leaf_data);CHKERRQ(ierr); 4080dc456d91SStefano Zampini ierr = PetscSFBcastEnd(sf,MPIU_INT,root_data,leaf_data);CHKERRQ(ierr); 4081dc456d91SStefano Zampini ierr = PetscSFDestroy(&sf);CHKERRQ(ierr); 4082dc456d91SStefano Zampini 4083dc456d91SStefano Zampini /* create new IS with global indexes without holes */ 4084dc456d91SStefano Zampini if (subset_mult) { 4085dc456d91SStefano Zampini const PetscInt* idxs_mult; 4086dc456d91SStefano Zampini PetscInt cum; 4087dc456d91SStefano Zampini 4088dc456d91SStefano Zampini cum = 0; 4089dc456d91SStefano Zampini ierr = ISGetIndices(subset_mult,&idxs_mult);CHKERRQ(ierr); 4090dc456d91SStefano Zampini for (i=0;i<n;i++) { 4091dc456d91SStefano Zampini PetscInt j; 4092dc456d91SStefano Zampini for (j=0;j<idxs_mult[i];j++) gidxs[cum++] = leaf_data[i] - idxs_mult[i] + j; 4093674ae819SStefano Zampini } 4094dc456d91SStefano Zampini ierr = ISRestoreIndices(subset_mult,&idxs_mult);CHKERRQ(ierr); 4095674ae819SStefano Zampini } else { 4096dc456d91SStefano Zampini for (i=0;i<n;i++) { 4097dc456d91SStefano Zampini gidxs[i] = leaf_data[i]-1; 4098674ae819SStefano Zampini } 4099674ae819SStefano Zampini } 4100dc456d91SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)subset),n_n,gidxs,PETSC_OWN_POINTER,subset_n);CHKERRQ(ierr); 4101dc456d91SStefano Zampini ierr = PetscFree2(leaf_data,root_data);CHKERRQ(ierr); 4102674ae819SStefano Zampini PetscFunctionReturn(0); 4103674ae819SStefano Zampini } 41049a7d3425SStefano Zampini 41059a7d3425SStefano Zampini #undef __FUNCT__ 41069a7d3425SStefano Zampini #define __FUNCT__ "PCBDDCOrthonormalizeVecs" 41079a7d3425SStefano Zampini PetscErrorCode PCBDDCOrthonormalizeVecs(PetscInt n, Vec vecs[]) 41089a7d3425SStefano Zampini { 41099a7d3425SStefano Zampini PetscInt i,j; 41109a7d3425SStefano Zampini PetscScalar *alphas; 41119a7d3425SStefano Zampini PetscErrorCode ierr; 41129a7d3425SStefano Zampini 41139a7d3425SStefano Zampini PetscFunctionBegin; 41149a7d3425SStefano Zampini /* this implements stabilized Gram-Schmidt */ 4115785e854fSJed Brown ierr = PetscMalloc1(n,&alphas);CHKERRQ(ierr); 41169a7d3425SStefano Zampini for (i=0;i<n;i++) { 41179a7d3425SStefano Zampini ierr = VecNormalize(vecs[i],NULL);CHKERRQ(ierr); 41189a7d3425SStefano Zampini if (i<n) { ierr = VecMDot(vecs[i],n-i-1,&vecs[i+1],&alphas[i+1]);CHKERRQ(ierr); } 41199a7d3425SStefano Zampini for (j=i+1;j<n;j++) { ierr = VecAXPY(vecs[j],PetscConj(-alphas[j]),vecs[i]);CHKERRQ(ierr); } 41209a7d3425SStefano Zampini } 41219a7d3425SStefano Zampini ierr = PetscFree(alphas);CHKERRQ(ierr); 41229a7d3425SStefano Zampini PetscFunctionReturn(0); 41239a7d3425SStefano Zampini } 41249a7d3425SStefano Zampini 4125e7931f94SStefano Zampini #undef __FUNCT__ 412670cf5478SStefano Zampini #define __FUNCT__ "MatISGetSubassemblingPattern" 4127b0c7d250SStefano Zampini PetscErrorCode MatISGetSubassemblingPattern(Mat mat, PetscInt n_subdomains, PetscInt redprocs, IS* is_sends) 4128e7931f94SStefano Zampini { 412952e5ac9dSStefano Zampini IS ranks_send_to; 4130e7931f94SStefano Zampini PetscInt n_neighs,*neighs,*n_shared,**shared; 4131e7931f94SStefano Zampini PetscMPIInt size,rank,color; 413252e5ac9dSStefano Zampini PetscInt *xadj,*adjncy; 413352e5ac9dSStefano Zampini PetscInt *adjncy_wgt,*v_wgt,*ranks_send_to_idx; 41343837a79fSStefano Zampini PetscInt i,local_size,threshold=0; 41352b510759SStefano Zampini PetscBool use_vwgt=PETSC_FALSE,use_square=PETSC_FALSE; 4136e7931f94SStefano Zampini PetscSubcomm subcomm; 413752e5ac9dSStefano Zampini PetscErrorCode ierr; 4138a57a6d2fSStefano Zampini 4139e7931f94SStefano Zampini PetscFunctionBegin; 41402b510759SStefano Zampini ierr = PetscOptionsGetBool(NULL,"-matis_partitioning_use_square",&use_square,NULL);CHKERRQ(ierr); 41412b510759SStefano Zampini ierr = PetscOptionsGetBool(NULL,"-matis_partitioning_use_vwgt",&use_vwgt,NULL);CHKERRQ(ierr); 41422b510759SStefano Zampini ierr = PetscOptionsGetInt(NULL,"-matis_partitioning_threshold",&threshold,NULL);CHKERRQ(ierr); 4143e7931f94SStefano Zampini 4144e7931f94SStefano Zampini /* Get info on mapping */ 41453bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetSize(mat->rmap->mapping,&local_size);CHKERRQ(ierr); 41463bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetInfo(mat->rmap->mapping,&n_neighs,&neighs,&n_shared,&shared);CHKERRQ(ierr); 4147e7931f94SStefano Zampini 4148e7931f94SStefano Zampini /* build local CSR graph of subdomains' connectivity */ 4149785e854fSJed Brown ierr = PetscMalloc1(2,&xadj);CHKERRQ(ierr); 4150e7931f94SStefano Zampini xadj[0] = 0; 4151e7931f94SStefano Zampini xadj[1] = PetscMax(n_neighs-1,0); 4152785e854fSJed Brown ierr = PetscMalloc1(xadj[1],&adjncy);CHKERRQ(ierr); 4153785e854fSJed Brown ierr = PetscMalloc1(xadj[1],&adjncy_wgt);CHKERRQ(ierr); 4154e7931f94SStefano Zampini 41552b510759SStefano Zampini if (threshold) { 4156d023bfaeSStefano Zampini PetscInt xadj_count = 0; 41572b510759SStefano Zampini for (i=1;i<n_neighs;i++) { 4158d023bfaeSStefano Zampini if (n_shared[i] > threshold) { 4159d023bfaeSStefano Zampini adjncy[xadj_count] = neighs[i]; 4160d023bfaeSStefano Zampini adjncy_wgt[xadj_count] = n_shared[i]; 4161d023bfaeSStefano Zampini xadj_count++; 4162e7931f94SStefano Zampini } 4163e7931f94SStefano Zampini } 4164d023bfaeSStefano Zampini xadj[1] = xadj_count; 4165c8587f34SStefano Zampini } else { 4166e7931f94SStefano Zampini if (xadj[1]) { 4167e7931f94SStefano Zampini ierr = PetscMemcpy(adjncy,&neighs[1],xadj[1]*sizeof(*adjncy));CHKERRQ(ierr); 4168e7931f94SStefano Zampini ierr = PetscMemcpy(adjncy_wgt,&n_shared[1],xadj[1]*sizeof(*adjncy_wgt));CHKERRQ(ierr); 4169c8587f34SStefano Zampini } 4170e7931f94SStefano Zampini } 41713bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingRestoreInfo(mat->rmap->mapping,&n_neighs,&neighs,&n_shared,&shared);CHKERRQ(ierr); 4172e7931f94SStefano Zampini if (use_square) { 4173e7931f94SStefano Zampini for (i=0;i<xadj[1];i++) { 4174e7931f94SStefano Zampini adjncy_wgt[i] = adjncy_wgt[i]*adjncy_wgt[i]; 4175e7931f94SStefano Zampini } 4176e7931f94SStefano Zampini } 4177e7931f94SStefano Zampini ierr = PetscSortIntWithArray(xadj[1],adjncy,adjncy_wgt);CHKERRQ(ierr); 4178e7931f94SStefano Zampini 41793837a79fSStefano Zampini ierr = PetscMalloc1(1,&ranks_send_to_idx);CHKERRQ(ierr); 4180e7931f94SStefano Zampini 4181e7931f94SStefano Zampini /* 4182e7931f94SStefano Zampini Restrict work on active processes only. 4183e7931f94SStefano Zampini */ 4184e7931f94SStefano Zampini ierr = PetscSubcommCreate(PetscObjectComm((PetscObject)mat),&subcomm);CHKERRQ(ierr); 4185e7931f94SStefano Zampini ierr = PetscSubcommSetNumber(subcomm,2);CHKERRQ(ierr); /* 2 groups, active process and not active processes */ 4186e7931f94SStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)mat),&rank);CHKERRQ(ierr); 41872b510759SStefano Zampini ierr = PetscMPIIntCast(!local_size,&color);CHKERRQ(ierr); 4188d3531aaaSJed Brown ierr = PetscSubcommSetTypeGeneral(subcomm,color,rank);CHKERRQ(ierr); 4189e7931f94SStefano Zampini if (color) { 4190e7931f94SStefano Zampini ierr = PetscFree(xadj);CHKERRQ(ierr); 4191e7931f94SStefano Zampini ierr = PetscFree(adjncy);CHKERRQ(ierr); 4192e7931f94SStefano Zampini ierr = PetscFree(adjncy_wgt);CHKERRQ(ierr); 4193c8587f34SStefano Zampini } else { 419452e5ac9dSStefano Zampini Mat subdomain_adj; 419552e5ac9dSStefano Zampini IS new_ranks,new_ranks_contig; 419652e5ac9dSStefano Zampini MatPartitioning partitioner; 419752e5ac9dSStefano Zampini PetscInt prank,rstart=0,rend=0; 419852e5ac9dSStefano Zampini PetscInt *is_indices,*oldranks; 4199b0c7d250SStefano Zampini PetscBool aggregate; 4200b0c7d250SStefano Zampini 4201306c2d5bSBarry Smith ierr = MPI_Comm_size(PetscSubcommChild(subcomm),&size);CHKERRQ(ierr); 4202785e854fSJed Brown ierr = PetscMalloc1(size,&oldranks);CHKERRQ(ierr); 4203e7931f94SStefano Zampini prank = rank; 4204306c2d5bSBarry Smith ierr = MPI_Allgather(&prank,1,MPIU_INT,oldranks,1,MPIU_INT,PetscSubcommChild(subcomm));CHKERRQ(ierr); 42058002ef2cSStefano Zampini /* 4206e7931f94SStefano Zampini for (i=0;i<size;i++) { 4207e7931f94SStefano Zampini PetscPrintf(subcomm->comm,"oldranks[%d] = %d\n",i,oldranks[i]); 4208c8587f34SStefano Zampini } 42098002ef2cSStefano Zampini */ 4210e7931f94SStefano Zampini for (i=0;i<xadj[1];i++) { 4211e7931f94SStefano Zampini ierr = PetscFindInt(adjncy[i],size,oldranks,&adjncy[i]);CHKERRQ(ierr); 4212c8587f34SStefano Zampini } 4213e7931f94SStefano Zampini ierr = PetscSortIntWithArray(xadj[1],adjncy,adjncy_wgt);CHKERRQ(ierr); 4214b0c7d250SStefano Zampini aggregate = ((redprocs > 0 && redprocs < size) ? PETSC_TRUE : PETSC_FALSE); 4215b0c7d250SStefano Zampini if (aggregate) { 4216b0c7d250SStefano Zampini PetscInt lrows,row,ncols,*cols; 4217b0c7d250SStefano Zampini PetscMPIInt nrank; 4218b0c7d250SStefano Zampini PetscScalar *vals; 4219b0c7d250SStefano Zampini 4220b0c7d250SStefano Zampini ierr = MPI_Comm_rank(PetscSubcommChild(subcomm),&nrank);CHKERRQ(ierr); 4221b0c7d250SStefano Zampini lrows = 0; 4222b0c7d250SStefano Zampini if (nrank<redprocs) { 4223b0c7d250SStefano Zampini lrows = size/redprocs; 4224b0c7d250SStefano Zampini if (nrank<size%redprocs) lrows++; 4225b0c7d250SStefano Zampini } 42265fa240b1SStefano Zampini ierr = MatCreateAIJ(PetscSubcommChild(subcomm),lrows,lrows,size,size,50,NULL,50,NULL,&subdomain_adj);CHKERRQ(ierr); 4227b0c7d250SStefano Zampini ierr = MatGetOwnershipRange(subdomain_adj,&rstart,&rend);CHKERRQ(ierr); 4228b0c7d250SStefano Zampini ierr = MatSetOption(subdomain_adj,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_FALSE);CHKERRQ(ierr); 4229b0c7d250SStefano Zampini ierr = MatSetOption(subdomain_adj,MAT_NEW_NONZERO_ALLOCATION_ERR,PETSC_FALSE);CHKERRQ(ierr); 4230b0c7d250SStefano Zampini row = nrank; 4231b0c7d250SStefano Zampini ncols = xadj[1]-xadj[0]; 4232b0c7d250SStefano Zampini cols = adjncy; 4233b0c7d250SStefano Zampini ierr = PetscMalloc1(ncols,&vals);CHKERRQ(ierr); 4234b0c7d250SStefano Zampini for (i=0;i<ncols;i++) vals[i] = adjncy_wgt[i]; 4235b0c7d250SStefano Zampini ierr = MatSetValues(subdomain_adj,1,&row,ncols,cols,vals,INSERT_VALUES);CHKERRQ(ierr); 4236b0c7d250SStefano Zampini ierr = MatAssemblyBegin(subdomain_adj,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4237b0c7d250SStefano Zampini ierr = MatAssemblyEnd(subdomain_adj,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 423852e5ac9dSStefano Zampini ierr = PetscFree(xadj);CHKERRQ(ierr); 423952e5ac9dSStefano Zampini ierr = PetscFree(adjncy);CHKERRQ(ierr); 424052e5ac9dSStefano Zampini ierr = PetscFree(adjncy_wgt);CHKERRQ(ierr); 4241b0c7d250SStefano Zampini ierr = PetscFree(vals);CHKERRQ(ierr); 4242b0c7d250SStefano Zampini } else { 4243306c2d5bSBarry Smith ierr = MatCreateMPIAdj(PetscSubcommChild(subcomm),1,(PetscInt)size,xadj,adjncy,adjncy_wgt,&subdomain_adj);CHKERRQ(ierr); 4244b0c7d250SStefano Zampini } 424522b6e8a2SStefano Zampini /* ierr = MatView(subdomain_adj,0);CHKERRQ(ierr); */ 4246e7931f94SStefano Zampini 4247e7931f94SStefano Zampini /* Partition */ 4248306c2d5bSBarry Smith ierr = MatPartitioningCreate(PetscSubcommChild(subcomm),&partitioner);CHKERRQ(ierr); 4249e7931f94SStefano Zampini ierr = MatPartitioningSetAdjacency(partitioner,subdomain_adj);CHKERRQ(ierr); 4250e7931f94SStefano Zampini if (use_vwgt) { 42513837a79fSStefano Zampini ierr = PetscMalloc1(1,&v_wgt);CHKERRQ(ierr); 4252e7931f94SStefano Zampini v_wgt[0] = local_size; 4253e7931f94SStefano Zampini ierr = MatPartitioningSetVertexWeights(partitioner,v_wgt);CHKERRQ(ierr); 4254c8587f34SStefano Zampini } 425528143c3dSStefano Zampini n_subdomains = PetscMin((PetscInt)size,n_subdomains); 425628143c3dSStefano Zampini ierr = MatPartitioningSetNParts(partitioner,n_subdomains);CHKERRQ(ierr); 4257e7931f94SStefano Zampini ierr = MatPartitioningSetFromOptions(partitioner);CHKERRQ(ierr); 4258e7931f94SStefano Zampini ierr = MatPartitioningApply(partitioner,&new_ranks);CHKERRQ(ierr); 425922b6e8a2SStefano Zampini /* ierr = MatPartitioningView(partitioner,0);CHKERRQ(ierr); */ 4260e7931f94SStefano Zampini 426152e5ac9dSStefano Zampini /* renumber new_ranks to avoid "holes" in new set of processors */ 426252e5ac9dSStefano Zampini ierr = PCBDDCSubsetNumbering(new_ranks,NULL,NULL,&new_ranks_contig);CHKERRQ(ierr); 426352e5ac9dSStefano Zampini ierr = ISDestroy(&new_ranks);CHKERRQ(ierr); 426452e5ac9dSStefano Zampini ierr = ISGetIndices(new_ranks_contig,(const PetscInt**)&is_indices);CHKERRQ(ierr); 4265b0c7d250SStefano Zampini if (!redprocs) { 4266b0c7d250SStefano Zampini ranks_send_to_idx[0] = oldranks[is_indices[0]]; 426728143c3dSStefano Zampini } else { 4268b0c7d250SStefano Zampini PetscInt idxs[1]; 4269b0c7d250SStefano Zampini PetscMPIInt tag; 4270b0c7d250SStefano Zampini MPI_Request *reqs; 4271b0c7d250SStefano Zampini 4272b0c7d250SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)subdomain_adj,&tag);CHKERRQ(ierr); 4273b0c7d250SStefano Zampini ierr = PetscMalloc1(rend-rstart,&reqs);CHKERRQ(ierr); 4274b0c7d250SStefano Zampini for (i=rstart;i<rend;i++) { 4275b0c7d250SStefano Zampini ierr = MPI_Isend(is_indices+i-rstart,1,MPIU_INT,i,tag,PetscSubcommChild(subcomm),&reqs[i-rstart]);CHKERRQ(ierr); 427628143c3dSStefano Zampini } 4277b0c7d250SStefano Zampini ierr = MPI_Recv(idxs,1,MPIU_INT,MPI_ANY_SOURCE,tag,PetscSubcommChild(subcomm),MPI_STATUS_IGNORE);CHKERRQ(ierr); 4278b0c7d250SStefano Zampini ierr = MPI_Waitall(rend-rstart,reqs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 4279b0c7d250SStefano Zampini ierr = PetscFree(reqs);CHKERRQ(ierr); 4280b0c7d250SStefano Zampini ranks_send_to_idx[0] = oldranks[idxs[0]]; 4281e7931f94SStefano Zampini } 428252e5ac9dSStefano Zampini ierr = ISRestoreIndices(new_ranks_contig,(const PetscInt**)&is_indices);CHKERRQ(ierr); 4283e7931f94SStefano Zampini /* clean up */ 4284e7931f94SStefano Zampini ierr = PetscFree(oldranks);CHKERRQ(ierr); 428552e5ac9dSStefano Zampini ierr = ISDestroy(&new_ranks_contig);CHKERRQ(ierr); 4286e7931f94SStefano Zampini ierr = MatDestroy(&subdomain_adj);CHKERRQ(ierr); 4287e7931f94SStefano Zampini ierr = MatPartitioningDestroy(&partitioner);CHKERRQ(ierr); 4288e7931f94SStefano Zampini } 4289e7931f94SStefano Zampini ierr = PetscSubcommDestroy(&subcomm);CHKERRQ(ierr); 4290e7931f94SStefano Zampini 4291e7931f94SStefano Zampini /* assemble parallel IS for sends */ 4292e7931f94SStefano Zampini i = 1; 4293e7931f94SStefano Zampini if (color) i=0; 4294e7931f94SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)mat),i,ranks_send_to_idx,PETSC_OWN_POINTER,&ranks_send_to);CHKERRQ(ierr); 4295e7931f94SStefano Zampini /* get back IS */ 4296e7931f94SStefano Zampini *is_sends = ranks_send_to; 4297e7931f94SStefano Zampini PetscFunctionReturn(0); 4298e7931f94SStefano Zampini } 4299e7931f94SStefano Zampini 4300e7931f94SStefano Zampini typedef enum {MATDENSE_PRIVATE=0,MATAIJ_PRIVATE,MATBAIJ_PRIVATE,MATSBAIJ_PRIVATE}MatTypePrivate; 4301e7931f94SStefano Zampini 4302e7931f94SStefano Zampini #undef __FUNCT__ 4303e7931f94SStefano Zampini #define __FUNCT__ "MatISSubassemble" 430453a05cb3SStefano 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[]) 4305e7931f94SStefano Zampini { 430670cf5478SStefano Zampini Mat local_mat; 4307e7931f94SStefano Zampini IS is_sends_internal; 43089d30be91SStefano Zampini PetscInt rows,cols,new_local_rows; 430928143c3dSStefano Zampini PetscInt i,bs,buf_size_idxs,buf_size_idxs_is,buf_size_vals; 43109d30be91SStefano Zampini PetscBool ismatis,isdense,newisdense,destroy_mat; 4311e7931f94SStefano Zampini ISLocalToGlobalMapping l2gmap; 4312e7931f94SStefano Zampini PetscInt* l2gmap_indices; 4313e7931f94SStefano Zampini const PetscInt* is_indices; 4314e7931f94SStefano Zampini MatType new_local_type; 4315e7931f94SStefano Zampini /* buffers */ 4316e7931f94SStefano Zampini PetscInt *ptr_idxs,*send_buffer_idxs,*recv_buffer_idxs; 431728143c3dSStefano Zampini PetscInt *ptr_idxs_is,*send_buffer_idxs_is,*recv_buffer_idxs_is; 43189d30be91SStefano Zampini PetscInt *recv_buffer_idxs_local; 4319e7931f94SStefano Zampini PetscScalar *ptr_vals,*send_buffer_vals,*recv_buffer_vals; 4320e7931f94SStefano Zampini /* MPI */ 432128143c3dSStefano Zampini MPI_Comm comm,comm_n; 432228143c3dSStefano Zampini PetscSubcomm subcomm; 4323e7931f94SStefano Zampini PetscMPIInt n_sends,n_recvs,commsize; 432428143c3dSStefano Zampini PetscMPIInt *iflags,*ilengths_idxs,*ilengths_vals,*ilengths_idxs_is; 432528143c3dSStefano Zampini PetscMPIInt *onodes,*onodes_is,*olengths_idxs,*olengths_idxs_is,*olengths_vals; 432628143c3dSStefano Zampini PetscMPIInt len,tag_idxs,tag_idxs_is,tag_vals,source_dest; 432728143c3dSStefano Zampini MPI_Request *send_req_idxs,*send_req_idxs_is,*send_req_vals; 432828143c3dSStefano Zampini MPI_Request *recv_req_idxs,*recv_req_idxs_is,*recv_req_vals; 4329e7931f94SStefano Zampini PetscErrorCode ierr; 4330e7931f94SStefano Zampini 4331e7931f94SStefano Zampini PetscFunctionBegin; 433228143c3dSStefano Zampini /* TODO: add missing checks */ 433328143c3dSStefano Zampini PetscValidLogicalCollectiveInt(mat,n_subdomains,3); 433428143c3dSStefano Zampini PetscValidLogicalCollectiveBool(mat,restrict_comm,4); 433528143c3dSStefano Zampini PetscValidLogicalCollectiveEnum(mat,reuse,5); 433628143c3dSStefano Zampini PetscValidLogicalCollectiveInt(mat,nis,7); 4337e7931f94SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)mat,MATIS,&ismatis);CHKERRQ(ierr); 433828143c3dSStefano Zampini if (!ismatis) SETERRQ1(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot use %s on a matrix object which is not of type MATIS",__FUNCT__); 4339e7931f94SStefano Zampini ierr = MatISGetLocalMat(mat,&local_mat);CHKERRQ(ierr); 4340e7931f94SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)local_mat,MATSEQDENSE,&isdense);CHKERRQ(ierr); 4341e7931f94SStefano Zampini if (!isdense) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Currently cannot subassemble MATIS when local matrix type is not of type SEQDENSE"); 4342e7931f94SStefano Zampini ierr = MatGetSize(local_mat,&rows,&cols);CHKERRQ(ierr); 4343e7931f94SStefano Zampini if (rows != cols) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Local MATIS matrices should be square"); 434428143c3dSStefano Zampini if (reuse == MAT_REUSE_MATRIX && *mat_n) { 434570cf5478SStefano Zampini PetscInt mrows,mcols,mnrows,mncols; 434670cf5478SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)*mat_n,MATIS,&ismatis);CHKERRQ(ierr); 434728143c3dSStefano Zampini if (!ismatis) SETERRQ(PetscObjectComm((PetscObject)*mat_n),PETSC_ERR_SUP,"Cannot reuse a matrix which is not of type MATIS"); 434870cf5478SStefano Zampini ierr = MatGetSize(mat,&mrows,&mcols);CHKERRQ(ierr); 434970cf5478SStefano Zampini ierr = MatGetSize(*mat_n,&mnrows,&mncols);CHKERRQ(ierr); 435070cf5478SStefano Zampini if (mrows != mnrows) SETERRQ2(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot reuse matrix! Wrong number of rows %D != %D",mrows,mnrows); 435170cf5478SStefano Zampini if (mcols != mncols) SETERRQ2(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot reuse matrix! Wrong number of cols %D != %D",mcols,mncols); 435270cf5478SStefano Zampini } 4353e7931f94SStefano Zampini ierr = MatGetBlockSize(local_mat,&bs);CHKERRQ(ierr); 4354e7931f94SStefano Zampini PetscValidLogicalCollectiveInt(mat,bs,0); 4355e7931f94SStefano Zampini /* prepare IS for sending if not provided */ 4356e7931f94SStefano Zampini if (!is_sends) { 435728143c3dSStefano Zampini if (!n_subdomains) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"You should specify either an IS or a target number of subdomains"); 4358b0c7d250SStefano Zampini ierr = MatISGetSubassemblingPattern(mat,n_subdomains,0,&is_sends_internal);CHKERRQ(ierr); 4359c8587f34SStefano Zampini } else { 4360e7931f94SStefano Zampini ierr = PetscObjectReference((PetscObject)is_sends);CHKERRQ(ierr); 4361e7931f94SStefano Zampini is_sends_internal = is_sends; 4362c8587f34SStefano Zampini } 4363e7931f94SStefano Zampini 4364e7931f94SStefano Zampini /* get comm */ 4365a316fed8SStefano Zampini ierr = PetscObjectGetComm((PetscObject)mat,&comm);CHKERRQ(ierr); 4366e7931f94SStefano Zampini 4367e7931f94SStefano Zampini /* compute number of sends */ 4368e7931f94SStefano Zampini ierr = ISGetLocalSize(is_sends_internal,&i);CHKERRQ(ierr); 4369e7931f94SStefano Zampini ierr = PetscMPIIntCast(i,&n_sends);CHKERRQ(ierr); 4370e7931f94SStefano Zampini 4371e7931f94SStefano Zampini /* compute number of receives */ 4372e7931f94SStefano Zampini ierr = MPI_Comm_size(comm,&commsize);CHKERRQ(ierr); 4373785e854fSJed Brown ierr = PetscMalloc1(commsize,&iflags);CHKERRQ(ierr); 4374e7931f94SStefano Zampini ierr = PetscMemzero(iflags,commsize*sizeof(*iflags));CHKERRQ(ierr); 4375e7931f94SStefano Zampini ierr = ISGetIndices(is_sends_internal,&is_indices);CHKERRQ(ierr); 4376e7931f94SStefano Zampini for (i=0;i<n_sends;i++) iflags[is_indices[i]] = 1; 4377e7931f94SStefano Zampini ierr = PetscGatherNumberOfMessages(comm,iflags,NULL,&n_recvs);CHKERRQ(ierr); 4378e7931f94SStefano Zampini ierr = PetscFree(iflags);CHKERRQ(ierr); 4379e7931f94SStefano Zampini 438028143c3dSStefano Zampini /* restrict comm if requested */ 438128143c3dSStefano Zampini subcomm = 0; 438228143c3dSStefano Zampini destroy_mat = PETSC_FALSE; 438328143c3dSStefano Zampini if (restrict_comm) { 4384779c1cceSStefano Zampini PetscMPIInt color,subcommsize; 4385779c1cceSStefano Zampini 438628143c3dSStefano Zampini color = 0; 438753a05cb3SStefano Zampini if (restrict_full) { 438853a05cb3SStefano Zampini if (!n_recvs) color = 1; /* processes not receiving anything will not partecipate in new comm (full restriction) */ 438953a05cb3SStefano Zampini } else { 439053a05cb3SStefano Zampini if (!n_recvs && n_sends) color = 1; /* just those processes that are sending but not receiving anything will not partecipate in new comm */ 439153a05cb3SStefano Zampini } 439228143c3dSStefano Zampini ierr = MPI_Allreduce(&color,&subcommsize,1,MPI_INT,MPI_SUM,comm);CHKERRQ(ierr); 439328143c3dSStefano Zampini subcommsize = commsize - subcommsize; 439428143c3dSStefano Zampini /* check if reuse has been requested */ 439528143c3dSStefano Zampini if (reuse == MAT_REUSE_MATRIX) { 439628143c3dSStefano Zampini if (*mat_n) { 439728143c3dSStefano Zampini PetscMPIInt subcommsize2; 439828143c3dSStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)*mat_n),&subcommsize2);CHKERRQ(ierr); 439928143c3dSStefano Zampini if (subcommsize != subcommsize2) SETERRQ2(PetscObjectComm((PetscObject)*mat_n),PETSC_ERR_PLIB,"Cannot reuse matrix! wrong subcomm size %d != %d",subcommsize,subcommsize2); 440028143c3dSStefano Zampini comm_n = PetscObjectComm((PetscObject)*mat_n); 440128143c3dSStefano Zampini } else { 440228143c3dSStefano Zampini comm_n = PETSC_COMM_SELF; 440328143c3dSStefano Zampini } 440428143c3dSStefano Zampini } else { /* MAT_INITIAL_MATRIX */ 4405779c1cceSStefano Zampini PetscMPIInt rank; 4406779c1cceSStefano Zampini 4407779c1cceSStefano Zampini ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 440828143c3dSStefano Zampini ierr = PetscSubcommCreate(comm,&subcomm);CHKERRQ(ierr); 440928143c3dSStefano Zampini ierr = PetscSubcommSetNumber(subcomm,2);CHKERRQ(ierr); 441028143c3dSStefano Zampini ierr = PetscSubcommSetTypeGeneral(subcomm,color,rank);CHKERRQ(ierr); 4411306c2d5bSBarry Smith comm_n = PetscSubcommChild(subcomm); 441228143c3dSStefano Zampini } 441328143c3dSStefano Zampini /* flag to destroy *mat_n if not significative */ 441428143c3dSStefano Zampini if (color) destroy_mat = PETSC_TRUE; 441528143c3dSStefano Zampini } else { 441628143c3dSStefano Zampini comm_n = comm; 441728143c3dSStefano Zampini } 441828143c3dSStefano Zampini 4419e7931f94SStefano Zampini /* prepare send/receive buffers */ 4420785e854fSJed Brown ierr = PetscMalloc1(commsize,&ilengths_idxs);CHKERRQ(ierr); 4421e7931f94SStefano Zampini ierr = PetscMemzero(ilengths_idxs,commsize*sizeof(*ilengths_idxs));CHKERRQ(ierr); 4422785e854fSJed Brown ierr = PetscMalloc1(commsize,&ilengths_vals);CHKERRQ(ierr); 4423e7931f94SStefano Zampini ierr = PetscMemzero(ilengths_vals,commsize*sizeof(*ilengths_vals));CHKERRQ(ierr); 442428143c3dSStefano Zampini if (nis) { 4425854ce69bSBarry Smith ierr = PetscCalloc1(commsize,&ilengths_idxs_is);CHKERRQ(ierr); 442628143c3dSStefano Zampini } 4427e7931f94SStefano Zampini 442828143c3dSStefano Zampini /* Get data from local matrices */ 4429e7931f94SStefano Zampini if (!isdense) { 4430a26c9d0eSStefano Zampini SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Subassembling of AIJ local matrices not yet implemented"); 4431e7931f94SStefano Zampini /* TODO: See below some guidelines on how to prepare the local buffers */ 4432e7931f94SStefano Zampini /* 4433e7931f94SStefano Zampini send_buffer_vals should contain the raw values of the local matrix 4434e7931f94SStefano Zampini send_buffer_idxs should contain: 4435e7931f94SStefano Zampini - MatType_PRIVATE type 4436e7931f94SStefano Zampini - PetscInt size_of_l2gmap 4437e7931f94SStefano Zampini - PetscInt global_row_indices[size_of_l2gmap] 4438e7931f94SStefano Zampini - PetscInt all_other_info_which_is_needed_to_compute_preallocation_and_set_values 4439e7931f94SStefano Zampini */ 4440e7931f94SStefano Zampini } else { 4441e7931f94SStefano Zampini ierr = MatDenseGetArray(local_mat,&send_buffer_vals);CHKERRQ(ierr); 44423bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetSize(mat->rmap->mapping,&i);CHKERRQ(ierr); 4443854ce69bSBarry Smith ierr = PetscMalloc1(i+2,&send_buffer_idxs);CHKERRQ(ierr); 4444e7931f94SStefano Zampini send_buffer_idxs[0] = (PetscInt)MATDENSE_PRIVATE; 4445e7931f94SStefano Zampini send_buffer_idxs[1] = i; 44463bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetIndices(mat->rmap->mapping,(const PetscInt**)&ptr_idxs);CHKERRQ(ierr); 4447e7931f94SStefano Zampini ierr = PetscMemcpy(&send_buffer_idxs[2],ptr_idxs,i*sizeof(PetscInt));CHKERRQ(ierr); 44483bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingRestoreIndices(mat->rmap->mapping,(const PetscInt**)&ptr_idxs);CHKERRQ(ierr); 4449e7931f94SStefano Zampini ierr = PetscMPIIntCast(i,&len);CHKERRQ(ierr); 4450e7931f94SStefano Zampini for (i=0;i<n_sends;i++) { 4451e7931f94SStefano Zampini ilengths_vals[is_indices[i]] = len*len; 4452e7931f94SStefano Zampini ilengths_idxs[is_indices[i]] = len+2; 4453c8587f34SStefano Zampini } 4454c8587f34SStefano Zampini } 4455e7931f94SStefano Zampini ierr = PetscGatherMessageLengths2(comm,n_sends,n_recvs,ilengths_idxs,ilengths_vals,&onodes,&olengths_idxs,&olengths_vals);CHKERRQ(ierr); 445628143c3dSStefano Zampini /* additional is (if any) */ 445728143c3dSStefano Zampini if (nis) { 445828143c3dSStefano Zampini PetscMPIInt psum; 445928143c3dSStefano Zampini PetscInt j; 446028143c3dSStefano Zampini for (j=0,psum=0;j<nis;j++) { 446128143c3dSStefano Zampini PetscInt plen; 446228143c3dSStefano Zampini ierr = ISGetLocalSize(isarray[j],&plen);CHKERRQ(ierr); 446328143c3dSStefano Zampini ierr = PetscMPIIntCast(plen,&len);CHKERRQ(ierr); 446428143c3dSStefano Zampini psum += len+1; /* indices + lenght */ 446528143c3dSStefano Zampini } 4466854ce69bSBarry Smith ierr = PetscMalloc1(psum,&send_buffer_idxs_is);CHKERRQ(ierr); 446728143c3dSStefano Zampini for (j=0,psum=0;j<nis;j++) { 446828143c3dSStefano Zampini PetscInt plen; 446928143c3dSStefano Zampini const PetscInt *is_array_idxs; 447028143c3dSStefano Zampini ierr = ISGetLocalSize(isarray[j],&plen);CHKERRQ(ierr); 447128143c3dSStefano Zampini send_buffer_idxs_is[psum] = plen; 447228143c3dSStefano Zampini ierr = ISGetIndices(isarray[j],&is_array_idxs);CHKERRQ(ierr); 447328143c3dSStefano Zampini ierr = PetscMemcpy(&send_buffer_idxs_is[psum+1],is_array_idxs,plen*sizeof(PetscInt));CHKERRQ(ierr); 447428143c3dSStefano Zampini ierr = ISRestoreIndices(isarray[j],&is_array_idxs);CHKERRQ(ierr); 447528143c3dSStefano Zampini psum += plen+1; /* indices + lenght */ 447628143c3dSStefano Zampini } 447728143c3dSStefano Zampini for (i=0;i<n_sends;i++) { 447828143c3dSStefano Zampini ilengths_idxs_is[is_indices[i]] = psum; 447928143c3dSStefano Zampini } 448028143c3dSStefano Zampini ierr = PetscGatherMessageLengths(comm,n_sends,n_recvs,ilengths_idxs_is,&onodes_is,&olengths_idxs_is);CHKERRQ(ierr); 448128143c3dSStefano Zampini } 448228143c3dSStefano Zampini 4483e7931f94SStefano Zampini buf_size_idxs = 0; 4484e7931f94SStefano Zampini buf_size_vals = 0; 448528143c3dSStefano Zampini buf_size_idxs_is = 0; 4486e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 4487e7931f94SStefano Zampini buf_size_idxs += (PetscInt)olengths_idxs[i]; 4488e7931f94SStefano Zampini buf_size_vals += (PetscInt)olengths_vals[i]; 448928143c3dSStefano Zampini if (nis) buf_size_idxs_is += (PetscInt)olengths_idxs_is[i]; 4490e7931f94SStefano Zampini } 4491785e854fSJed Brown ierr = PetscMalloc1(buf_size_idxs,&recv_buffer_idxs);CHKERRQ(ierr); 4492785e854fSJed Brown ierr = PetscMalloc1(buf_size_vals,&recv_buffer_vals);CHKERRQ(ierr); 449395ecbf38SStefano Zampini ierr = PetscMalloc1(buf_size_idxs_is,&recv_buffer_idxs_is);CHKERRQ(ierr); 4494e7931f94SStefano Zampini 4495e7931f94SStefano Zampini /* get new tags for clean communications */ 4496e7931f94SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_idxs);CHKERRQ(ierr); 4497e7931f94SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_vals);CHKERRQ(ierr); 449828143c3dSStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_idxs_is);CHKERRQ(ierr); 4499e7931f94SStefano Zampini 4500e7931f94SStefano Zampini /* allocate for requests */ 4501785e854fSJed Brown ierr = PetscMalloc1(n_sends,&send_req_idxs);CHKERRQ(ierr); 4502785e854fSJed Brown ierr = PetscMalloc1(n_sends,&send_req_vals);CHKERRQ(ierr); 450395ecbf38SStefano Zampini ierr = PetscMalloc1(n_sends,&send_req_idxs_is);CHKERRQ(ierr); 4504785e854fSJed Brown ierr = PetscMalloc1(n_recvs,&recv_req_idxs);CHKERRQ(ierr); 4505785e854fSJed Brown ierr = PetscMalloc1(n_recvs,&recv_req_vals);CHKERRQ(ierr); 450695ecbf38SStefano Zampini ierr = PetscMalloc1(n_recvs,&recv_req_idxs_is);CHKERRQ(ierr); 4507e7931f94SStefano Zampini 4508e7931f94SStefano Zampini /* communications */ 4509e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 4510e7931f94SStefano Zampini ptr_vals = recv_buffer_vals; 451128143c3dSStefano Zampini ptr_idxs_is = recv_buffer_idxs_is; 4512e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 4513e7931f94SStefano Zampini source_dest = onodes[i]; 4514e7931f94SStefano Zampini ierr = MPI_Irecv(ptr_idxs,olengths_idxs[i],MPIU_INT,source_dest,tag_idxs,comm,&recv_req_idxs[i]);CHKERRQ(ierr); 4515e7931f94SStefano Zampini ierr = MPI_Irecv(ptr_vals,olengths_vals[i],MPIU_SCALAR,source_dest,tag_vals,comm,&recv_req_vals[i]);CHKERRQ(ierr); 4516e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 4517e7931f94SStefano Zampini ptr_vals += olengths_vals[i]; 451828143c3dSStefano Zampini if (nis) { 451928143c3dSStefano 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); 452028143c3dSStefano Zampini ptr_idxs_is += olengths_idxs_is[i]; 452128143c3dSStefano Zampini } 4522e7931f94SStefano Zampini } 4523e7931f94SStefano Zampini for (i=0;i<n_sends;i++) { 4524e7931f94SStefano Zampini ierr = PetscMPIIntCast(is_indices[i],&source_dest);CHKERRQ(ierr); 4525e7931f94SStefano Zampini ierr = MPI_Isend(send_buffer_idxs,ilengths_idxs[source_dest],MPIU_INT,source_dest,tag_idxs,comm,&send_req_idxs[i]);CHKERRQ(ierr); 4526e7931f94SStefano Zampini ierr = MPI_Isend(send_buffer_vals,ilengths_vals[source_dest],MPIU_SCALAR,source_dest,tag_vals,comm,&send_req_vals[i]);CHKERRQ(ierr); 452728143c3dSStefano Zampini if (nis) { 452828143c3dSStefano 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); 452928143c3dSStefano Zampini } 4530e7931f94SStefano Zampini } 4531e7931f94SStefano Zampini ierr = ISRestoreIndices(is_sends_internal,&is_indices);CHKERRQ(ierr); 4532e7931f94SStefano Zampini ierr = ISDestroy(&is_sends_internal);CHKERRQ(ierr); 4533e7931f94SStefano Zampini 4534e7931f94SStefano Zampini /* assemble new l2g map */ 4535e7931f94SStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_idxs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 4536e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 45379d30be91SStefano Zampini new_local_rows = 0; 4538e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 45399d30be91SStefano Zampini new_local_rows += *(ptr_idxs+1); /* second element is the local size of the l2gmap */ 4540e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 4541e7931f94SStefano Zampini } 45429d30be91SStefano Zampini ierr = PetscMalloc1(new_local_rows,&l2gmap_indices);CHKERRQ(ierr); 4543e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 45449d30be91SStefano Zampini new_local_rows = 0; 4545e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 45469d30be91SStefano Zampini ierr = PetscMemcpy(&l2gmap_indices[new_local_rows],ptr_idxs+2,(*(ptr_idxs+1))*sizeof(PetscInt));CHKERRQ(ierr); 45479d30be91SStefano Zampini new_local_rows += *(ptr_idxs+1); /* second element is the local size of the l2gmap */ 4548e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 4549e7931f94SStefano Zampini } 45509d30be91SStefano Zampini ierr = PetscSortRemoveDupsInt(&new_local_rows,l2gmap_indices);CHKERRQ(ierr); 45519d30be91SStefano Zampini ierr = ISLocalToGlobalMappingCreate(comm_n,1,new_local_rows,l2gmap_indices,PETSC_COPY_VALUES,&l2gmap);CHKERRQ(ierr); 4552e7931f94SStefano Zampini ierr = PetscFree(l2gmap_indices);CHKERRQ(ierr); 4553e7931f94SStefano Zampini 4554e7931f94SStefano Zampini /* infer new local matrix type from received local matrices type */ 4555e7931f94SStefano Zampini /* currently if all local matrices are of type X, then the resulting matrix will be of type X, except for the dense case */ 4556e7931f94SStefano 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) */ 4557e7931f94SStefano Zampini if (n_recvs) { 455828143c3dSStefano Zampini MatTypePrivate new_local_type_private = (MatTypePrivate)send_buffer_idxs[0]; 4559e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 4560e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 4561e7931f94SStefano Zampini if ((PetscInt)new_local_type_private != *ptr_idxs) { 4562e7931f94SStefano Zampini new_local_type_private = MATAIJ_PRIVATE; 4563e7931f94SStefano Zampini break; 4564e7931f94SStefano Zampini } 4565e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 4566e7931f94SStefano Zampini } 4567e7931f94SStefano Zampini switch (new_local_type_private) { 456828143c3dSStefano Zampini case MATDENSE_PRIVATE: 456928143c3dSStefano Zampini if (n_recvs>1) { /* subassembling of dense matrices does not give a dense matrix! */ 4570e7931f94SStefano Zampini new_local_type = MATSEQAIJ; 4571e7931f94SStefano Zampini bs = 1; 457228143c3dSStefano Zampini } else { /* if I receive only 1 dense matrix */ 457328143c3dSStefano Zampini new_local_type = MATSEQDENSE; 457428143c3dSStefano Zampini bs = 1; 457528143c3dSStefano Zampini } 4576e7931f94SStefano Zampini break; 4577e7931f94SStefano Zampini case MATAIJ_PRIVATE: 4578e7931f94SStefano Zampini new_local_type = MATSEQAIJ; 4579e7931f94SStefano Zampini bs = 1; 4580e7931f94SStefano Zampini break; 4581e7931f94SStefano Zampini case MATBAIJ_PRIVATE: 4582e7931f94SStefano Zampini new_local_type = MATSEQBAIJ; 4583e7931f94SStefano Zampini break; 4584e7931f94SStefano Zampini case MATSBAIJ_PRIVATE: 4585e7931f94SStefano Zampini new_local_type = MATSEQSBAIJ; 4586e7931f94SStefano Zampini break; 4587e7931f94SStefano Zampini default: 45889d30be91SStefano Zampini SETERRQ2(comm,PETSC_ERR_SUP,"Unsupported private type %d in %s",new_local_type_private,__FUNCT__); 4589e7931f94SStefano Zampini break; 4590e7931f94SStefano Zampini } 459128143c3dSStefano Zampini } else { /* by default, new_local_type is seqdense */ 459228143c3dSStefano Zampini new_local_type = MATSEQDENSE; 459328143c3dSStefano Zampini bs = 1; 4594e7931f94SStefano Zampini } 4595e7931f94SStefano Zampini 459670cf5478SStefano Zampini /* create MATIS object if needed */ 459770cf5478SStefano Zampini if (reuse == MAT_INITIAL_MATRIX) { 4598e7931f94SStefano Zampini ierr = MatGetSize(mat,&rows,&cols);CHKERRQ(ierr); 4599e176bc59SStefano Zampini ierr = MatCreateIS(comm_n,bs,PETSC_DECIDE,PETSC_DECIDE,rows,cols,l2gmap,NULL,mat_n);CHKERRQ(ierr); 460070cf5478SStefano Zampini } else { 460170cf5478SStefano Zampini /* it also destroys the local matrices */ 460270cf5478SStefano Zampini ierr = MatSetLocalToGlobalMapping(*mat_n,l2gmap,l2gmap);CHKERRQ(ierr); 460370cf5478SStefano Zampini } 460470cf5478SStefano Zampini ierr = MatISGetLocalMat(*mat_n,&local_mat);CHKERRQ(ierr); 4605e7931f94SStefano Zampini ierr = MatSetType(local_mat,new_local_type);CHKERRQ(ierr); 46069d30be91SStefano Zampini 46079d30be91SStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_vals,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 46089d30be91SStefano Zampini 46099d30be91SStefano Zampini /* Global to local map of received indices */ 46109d30be91SStefano Zampini ierr = PetscMalloc1(buf_size_idxs,&recv_buffer_idxs_local);CHKERRQ(ierr); /* needed for values insertion */ 46119d30be91SStefano Zampini ierr = ISGlobalToLocalMappingApply(l2gmap,IS_GTOLM_MASK,buf_size_idxs,recv_buffer_idxs,&i,recv_buffer_idxs_local);CHKERRQ(ierr); 46129d30be91SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&l2gmap);CHKERRQ(ierr); 46139d30be91SStefano Zampini 46149d30be91SStefano Zampini /* restore attributes -> type of incoming data and its size */ 46159d30be91SStefano Zampini buf_size_idxs = 0; 46169d30be91SStefano Zampini for (i=0;i<n_recvs;i++) { 46179d30be91SStefano Zampini recv_buffer_idxs_local[buf_size_idxs] = recv_buffer_idxs[buf_size_idxs]; 46189d30be91SStefano Zampini recv_buffer_idxs_local[buf_size_idxs+1] = recv_buffer_idxs[buf_size_idxs+1]; 46199d30be91SStefano Zampini buf_size_idxs += (PetscInt)olengths_idxs[i]; 46209d30be91SStefano Zampini } 46219d30be91SStefano Zampini ierr = PetscFree(recv_buffer_idxs);CHKERRQ(ierr); 46229d30be91SStefano Zampini 46239d30be91SStefano Zampini /* set preallocation */ 46249d30be91SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)local_mat,MATSEQDENSE,&newisdense);CHKERRQ(ierr); 46259d30be91SStefano Zampini if (!newisdense) { 46269d30be91SStefano Zampini PetscInt *new_local_nnz=0; 46279d30be91SStefano Zampini 46289d30be91SStefano Zampini ptr_vals = recv_buffer_vals; 46299d30be91SStefano Zampini ptr_idxs = recv_buffer_idxs_local; 46309d30be91SStefano Zampini if (n_recvs) { 46319d30be91SStefano Zampini ierr = PetscCalloc1(new_local_rows,&new_local_nnz);CHKERRQ(ierr); 46329d30be91SStefano Zampini } 46339d30be91SStefano Zampini for (i=0;i<n_recvs;i++) { 46349d30be91SStefano Zampini PetscInt j; 46359d30be91SStefano Zampini if (*ptr_idxs == (PetscInt)MATDENSE_PRIVATE) { /* preallocation provided for dense case only */ 46369d30be91SStefano Zampini for (j=0;j<*(ptr_idxs+1);j++) { 46379d30be91SStefano Zampini new_local_nnz[*(ptr_idxs+2+j)] += *(ptr_idxs+1); 46389d30be91SStefano Zampini } 46399d30be91SStefano Zampini } else { 46409d30be91SStefano Zampini /* TODO */ 46419d30be91SStefano Zampini } 46429d30be91SStefano Zampini ptr_idxs += olengths_idxs[i]; 46439d30be91SStefano Zampini } 46449d30be91SStefano Zampini if (new_local_nnz) { 46459d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] = PetscMin(new_local_nnz[i],new_local_rows); 46469d30be91SStefano Zampini ierr = MatSeqAIJSetPreallocation(local_mat,0,new_local_nnz);CHKERRQ(ierr); 46479d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] /= bs; 46489d30be91SStefano Zampini ierr = MatSeqBAIJSetPreallocation(local_mat,bs,0,new_local_nnz);CHKERRQ(ierr); 46499d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] = PetscMax(new_local_nnz[i]-i,0); 46509d30be91SStefano Zampini ierr = MatSeqSBAIJSetPreallocation(local_mat,bs,0,new_local_nnz);CHKERRQ(ierr); 46519d30be91SStefano Zampini } else { 46529d30be91SStefano Zampini ierr = MatSetUp(local_mat);CHKERRQ(ierr); 46539d30be91SStefano Zampini } 46549d30be91SStefano Zampini ierr = PetscFree(new_local_nnz);CHKERRQ(ierr); 46559d30be91SStefano Zampini } else { 46569d30be91SStefano Zampini ierr = MatSetUp(local_mat);CHKERRQ(ierr); 46579d30be91SStefano Zampini } 4658e7931f94SStefano Zampini 4659e7931f94SStefano Zampini /* set values */ 4660e7931f94SStefano Zampini ptr_vals = recv_buffer_vals; 46619d30be91SStefano Zampini ptr_idxs = recv_buffer_idxs_local; 4662e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 4663e7931f94SStefano Zampini if (*ptr_idxs == (PetscInt)MATDENSE_PRIVATE) { /* values insertion provided for dense case only */ 4664e7931f94SStefano Zampini ierr = MatSetOption(local_mat,MAT_ROW_ORIENTED,PETSC_FALSE);CHKERRQ(ierr); 46659d30be91SStefano Zampini ierr = MatSetValues(local_mat,*(ptr_idxs+1),ptr_idxs+2,*(ptr_idxs+1),ptr_idxs+2,ptr_vals,ADD_VALUES);CHKERRQ(ierr); 4666e7931f94SStefano Zampini ierr = MatAssemblyBegin(local_mat,MAT_FLUSH_ASSEMBLY);CHKERRQ(ierr); 4667e7931f94SStefano Zampini ierr = MatAssemblyEnd(local_mat,MAT_FLUSH_ASSEMBLY);CHKERRQ(ierr); 4668e7931f94SStefano Zampini ierr = MatSetOption(local_mat,MAT_ROW_ORIENTED,PETSC_TRUE);CHKERRQ(ierr); 466928143c3dSStefano Zampini } else { 467028143c3dSStefano Zampini /* TODO */ 4671e7931f94SStefano Zampini } 4672e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 4673e7931f94SStefano Zampini ptr_vals += olengths_vals[i]; 4674e7931f94SStefano Zampini } 4675e7931f94SStefano Zampini ierr = MatAssemblyBegin(local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4676e7931f94SStefano Zampini ierr = MatAssemblyEnd(local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 467770cf5478SStefano Zampini ierr = MatAssemblyBegin(*mat_n,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 467870cf5478SStefano Zampini ierr = MatAssemblyEnd(*mat_n,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 46799d30be91SStefano Zampini ierr = PetscFree(recv_buffer_vals);CHKERRQ(ierr); 46809d30be91SStefano Zampini ierr = PetscFree(recv_buffer_idxs_local);CHKERRQ(ierr); 4681e7931f94SStefano Zampini 4682dfd14d43SStefano Zampini #if 0 468328143c3dSStefano Zampini if (!restrict_comm) { /* check */ 4684e7931f94SStefano Zampini Vec lvec,rvec; 4685e7931f94SStefano Zampini PetscReal infty_error; 4686e7931f94SStefano Zampini 46872a7a6963SBarry Smith ierr = MatCreateVecs(mat,&rvec,&lvec);CHKERRQ(ierr); 4688e7931f94SStefano Zampini ierr = VecSetRandom(rvec,NULL);CHKERRQ(ierr); 4689e7931f94SStefano Zampini ierr = MatMult(mat,rvec,lvec);CHKERRQ(ierr); 4690e7931f94SStefano Zampini ierr = VecScale(lvec,-1.0);CHKERRQ(ierr); 469170cf5478SStefano Zampini ierr = MatMultAdd(*mat_n,rvec,lvec,lvec);CHKERRQ(ierr); 4692e7931f94SStefano Zampini ierr = VecNorm(lvec,NORM_INFINITY,&infty_error);CHKERRQ(ierr); 4693e7931f94SStefano Zampini ierr = PetscPrintf(PetscObjectComm((PetscObject)mat),"Infinity error subassembling %1.6e\n",infty_error); 4694e7931f94SStefano Zampini ierr = VecDestroy(&rvec);CHKERRQ(ierr); 4695e7931f94SStefano Zampini ierr = VecDestroy(&lvec);CHKERRQ(ierr); 4696e7931f94SStefano Zampini } 469728143c3dSStefano Zampini #endif 4698e7931f94SStefano Zampini 469928143c3dSStefano Zampini /* assemble new additional is (if any) */ 470028143c3dSStefano Zampini if (nis) { 470128143c3dSStefano Zampini PetscInt **temp_idxs,*count_is,j,psum; 470228143c3dSStefano Zampini 470328143c3dSStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_idxs_is,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 4704854ce69bSBarry Smith ierr = PetscCalloc1(nis,&count_is);CHKERRQ(ierr); 470528143c3dSStefano Zampini ptr_idxs = recv_buffer_idxs_is; 470628143c3dSStefano Zampini psum = 0; 470728143c3dSStefano Zampini for (i=0;i<n_recvs;i++) { 470828143c3dSStefano Zampini for (j=0;j<nis;j++) { 470928143c3dSStefano Zampini PetscInt plen = *(ptr_idxs); /* first element is the local size of IS's indices */ 471028143c3dSStefano Zampini count_is[j] += plen; /* increment counting of buffer for j-th IS */ 471128143c3dSStefano Zampini psum += plen; 471228143c3dSStefano Zampini ptr_idxs += plen+1; /* shift pointer to received data */ 471328143c3dSStefano Zampini } 471428143c3dSStefano Zampini } 4715854ce69bSBarry Smith ierr = PetscMalloc1(nis,&temp_idxs);CHKERRQ(ierr); 4716854ce69bSBarry Smith ierr = PetscMalloc1(psum,&temp_idxs[0]);CHKERRQ(ierr); 471728143c3dSStefano Zampini for (i=1;i<nis;i++) { 471828143c3dSStefano Zampini temp_idxs[i] = temp_idxs[i-1]+count_is[i-1]; 471928143c3dSStefano Zampini } 472028143c3dSStefano Zampini ierr = PetscMemzero(count_is,nis*sizeof(PetscInt));CHKERRQ(ierr); 472128143c3dSStefano Zampini ptr_idxs = recv_buffer_idxs_is; 472228143c3dSStefano Zampini for (i=0;i<n_recvs;i++) { 472328143c3dSStefano Zampini for (j=0;j<nis;j++) { 472428143c3dSStefano Zampini PetscInt plen = *(ptr_idxs); /* first element is the local size of IS's indices */ 472528143c3dSStefano Zampini ierr = PetscMemcpy(&temp_idxs[j][count_is[j]],ptr_idxs+1,plen*sizeof(PetscInt));CHKERRQ(ierr); 472628143c3dSStefano Zampini count_is[j] += plen; /* increment starting point of buffer for j-th IS */ 472728143c3dSStefano Zampini ptr_idxs += plen+1; /* shift pointer to received data */ 472828143c3dSStefano Zampini } 472928143c3dSStefano Zampini } 473028143c3dSStefano Zampini for (i=0;i<nis;i++) { 473128143c3dSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 473228143c3dSStefano Zampini ierr = PetscSortRemoveDupsInt(&count_is[i],temp_idxs[i]);CHKERRQ(ierr);CHKERRQ(ierr); 473328143c3dSStefano Zampini ierr = ISCreateGeneral(comm_n,count_is[i],temp_idxs[i],PETSC_COPY_VALUES,&isarray[i]);CHKERRQ(ierr); 473428143c3dSStefano Zampini } 473528143c3dSStefano Zampini ierr = PetscFree(count_is);CHKERRQ(ierr); 473628143c3dSStefano Zampini ierr = PetscFree(temp_idxs[0]);CHKERRQ(ierr); 473728143c3dSStefano Zampini ierr = PetscFree(temp_idxs);CHKERRQ(ierr); 473828143c3dSStefano Zampini } 4739e7931f94SStefano Zampini /* free workspace */ 474028143c3dSStefano Zampini ierr = PetscFree(recv_buffer_idxs_is);CHKERRQ(ierr); 4741e7931f94SStefano Zampini ierr = MPI_Waitall(n_sends,send_req_idxs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 4742e7931f94SStefano Zampini ierr = PetscFree(send_buffer_idxs);CHKERRQ(ierr); 4743e7931f94SStefano Zampini ierr = MPI_Waitall(n_sends,send_req_vals,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 4744e7931f94SStefano Zampini if (isdense) { 4745e7931f94SStefano Zampini ierr = MatISGetLocalMat(mat,&local_mat);CHKERRQ(ierr); 4746e7931f94SStefano Zampini ierr = MatDenseRestoreArray(local_mat,&send_buffer_vals);CHKERRQ(ierr); 4747e7931f94SStefano Zampini } else { 4748e7931f94SStefano Zampini /* ierr = PetscFree(send_buffer_vals);CHKERRQ(ierr); */ 4749e7931f94SStefano Zampini } 475028143c3dSStefano Zampini if (nis) { 475128143c3dSStefano Zampini ierr = MPI_Waitall(n_sends,send_req_idxs_is,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 475228143c3dSStefano Zampini ierr = PetscFree(send_buffer_idxs_is);CHKERRQ(ierr); 475328143c3dSStefano Zampini } 4754e7931f94SStefano Zampini ierr = PetscFree(recv_req_idxs);CHKERRQ(ierr); 4755e7931f94SStefano Zampini ierr = PetscFree(recv_req_vals);CHKERRQ(ierr); 475628143c3dSStefano Zampini ierr = PetscFree(recv_req_idxs_is);CHKERRQ(ierr); 4757e7931f94SStefano Zampini ierr = PetscFree(send_req_idxs);CHKERRQ(ierr); 4758e7931f94SStefano Zampini ierr = PetscFree(send_req_vals);CHKERRQ(ierr); 475928143c3dSStefano Zampini ierr = PetscFree(send_req_idxs_is);CHKERRQ(ierr); 4760e7931f94SStefano Zampini ierr = PetscFree(ilengths_vals);CHKERRQ(ierr); 4761e7931f94SStefano Zampini ierr = PetscFree(ilengths_idxs);CHKERRQ(ierr); 4762e7931f94SStefano Zampini ierr = PetscFree(olengths_vals);CHKERRQ(ierr); 4763e7931f94SStefano Zampini ierr = PetscFree(olengths_idxs);CHKERRQ(ierr); 4764e7931f94SStefano Zampini ierr = PetscFree(onodes);CHKERRQ(ierr); 476528143c3dSStefano Zampini if (nis) { 476628143c3dSStefano Zampini ierr = PetscFree(ilengths_idxs_is);CHKERRQ(ierr); 476728143c3dSStefano Zampini ierr = PetscFree(olengths_idxs_is);CHKERRQ(ierr); 476828143c3dSStefano Zampini ierr = PetscFree(onodes_is);CHKERRQ(ierr); 476928143c3dSStefano Zampini } 477028143c3dSStefano Zampini ierr = PetscSubcommDestroy(&subcomm);CHKERRQ(ierr); 477128143c3dSStefano Zampini if (destroy_mat) { /* destroy mat is true only if restrict comm is true and process will not partecipate */ 477228143c3dSStefano Zampini ierr = MatDestroy(mat_n);CHKERRQ(ierr); 477328143c3dSStefano Zampini for (i=0;i<nis;i++) { 477428143c3dSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 477528143c3dSStefano Zampini } 477653a05cb3SStefano Zampini *mat_n = NULL; 477728143c3dSStefano Zampini } 4778e7931f94SStefano Zampini PetscFunctionReturn(0); 4779e7931f94SStefano Zampini } 4780a57a6d2fSStefano Zampini 478112edc857SStefano Zampini /* temporary hack into ksp private data structure */ 4782af0996ceSBarry Smith #include <petsc/private/kspimpl.h> 478312edc857SStefano Zampini 4784c8587f34SStefano Zampini #undef __FUNCT__ 4785c8587f34SStefano Zampini #define __FUNCT__ "PCBDDCSetUpCoarseSolver" 4786c8587f34SStefano Zampini PetscErrorCode PCBDDCSetUpCoarseSolver(PC pc,PetscScalar* coarse_submat_vals) 4787c8587f34SStefano Zampini { 4788c8587f34SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 4789c8587f34SStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 479020a2ab83SStefano Zampini Mat coarse_mat,coarse_mat_is,coarse_submat_dense; 47919881197aSStefano Zampini MatNullSpace CoarseNullSpace=NULL; 479220a2ab83SStefano Zampini ISLocalToGlobalMapping coarse_islg; 47936e683305SStefano Zampini IS coarse_is,*isarray; 47946e683305SStefano Zampini PetscInt i,im_active=-1,active_procs=-1; 479530368db7SStefano Zampini PetscInt nis,nisdofs,nisneu,nisvert; 4796f9eb5b7dSStefano Zampini PC pc_temp; 4797c8587f34SStefano Zampini PCType coarse_pc_type; 4798c8587f34SStefano Zampini KSPType coarse_ksp_type; 4799f9eb5b7dSStefano Zampini PetscBool multilevel_requested,multilevel_allowed; 48004f3a063dSStefano Zampini PetscBool isredundant,isbddc,isnn,coarse_reuse; 48016e683305SStefano Zampini Mat t_coarse_mat_is; 48026e683305SStefano Zampini PetscInt void_procs,ncoarse_ml,ncoarse_ds,ncoarse; 48036e683305SStefano Zampini PetscMPIInt all_procs; 480474e2c79eSStefano Zampini PetscBool csin_ml,csin_ds,csin,csin_type_simple,redist; 480568457ee5SStefano Zampini PetscBool compute_vecs = PETSC_FALSE; 480622bc73bbSStefano Zampini PetscScalar *array; 48079881197aSStefano Zampini PetscErrorCode ierr; 4808fdc09c96SStefano Zampini 4809c8587f34SStefano Zampini PetscFunctionBegin; 4810c8587f34SStefano Zampini /* Assign global numbering to coarse dofs */ 481168457ee5SStefano 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 */ 4812fa7f1dd8SStefano Zampini PetscInt ocoarse_size; 48135a75c04eSSatish Balay compute_vecs = PETSC_TRUE; 4814fa7f1dd8SStefano Zampini ocoarse_size = pcbddc->coarse_size; 4815f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->global_primal_indices);CHKERRQ(ierr); 4816f4ddd8eeSStefano Zampini ierr = PCBDDCComputePrimalNumbering(pc,&pcbddc->coarse_size,&pcbddc->global_primal_indices);CHKERRQ(ierr); 4817f4ddd8eeSStefano Zampini /* see if we can avoid some work */ 4818fa7f1dd8SStefano Zampini if (pcbddc->coarse_ksp) { /* coarse ksp has already been created */ 481951bea450SStefano Zampini /* if the coarse size is different or we are using adaptive selection, better to not reuse the coarse matrix */ 482051bea450SStefano Zampini if (ocoarse_size != pcbddc->coarse_size || pcbddc->adaptive_selection) { 4821dc4bcba2SStefano Zampini PC pc; 4822dc4bcba2SStefano Zampini PetscBool isbddc; 4823dc4bcba2SStefano Zampini 4824dc4bcba2SStefano Zampini /* temporary workaround since PCBDDC does not have a reset method so far */ 4825dc4bcba2SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&pc);CHKERRQ(ierr); 4826dc4bcba2SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc,PCBDDC,&isbddc);CHKERRQ(ierr); 4827dc4bcba2SStefano Zampini if (isbddc) { 4828dc4bcba2SStefano Zampini ierr = PCDestroy(&pc);CHKERRQ(ierr); 4829dc4bcba2SStefano Zampini } 4830727cdba6SStefano Zampini ierr = KSPReset(pcbddc->coarse_ksp);CHKERRQ(ierr); 4831fa7f1dd8SStefano Zampini coarse_reuse = PETSC_FALSE; 4832fa7f1dd8SStefano Zampini } else { /* we can safely reuse already computed coarse matrix */ 4833fa7f1dd8SStefano Zampini coarse_reuse = PETSC_TRUE; 4834f4ddd8eeSStefano Zampini } 4835fa7f1dd8SStefano Zampini } else { /* there's no coarse ksp, so we need to create the coarse matrix too */ 4836fa7f1dd8SStefano Zampini coarse_reuse = PETSC_FALSE; 4837f4ddd8eeSStefano Zampini } 483870cf5478SStefano Zampini /* reset any subassembling information */ 483970cf5478SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling);CHKERRQ(ierr); 48406e683305SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 48416e683305SStefano Zampini } else { /* primal space is unchanged, so we can reuse coarse matrix */ 4842fa7f1dd8SStefano Zampini coarse_reuse = PETSC_TRUE; 4843f4ddd8eeSStefano Zampini } 4844c8587f34SStefano Zampini 48456e683305SStefano Zampini /* count "active" (i.e. with positive local size) and "void" processes */ 48462b510759SStefano Zampini im_active = !!(pcis->n); 48472b510759SStefano Zampini ierr = MPI_Allreduce(&im_active,&active_procs,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 48486e683305SStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)pc),&all_procs);CHKERRQ(ierr); 48496e683305SStefano Zampini void_procs = all_procs-active_procs; 48506e683305SStefano Zampini csin_type_simple = PETSC_TRUE; 485174e2c79eSStefano Zampini redist = PETSC_FALSE; 485222bc73bbSStefano Zampini if (pcbddc->current_level && void_procs) { 48536e683305SStefano Zampini csin_ml = PETSC_TRUE; 48546e683305SStefano Zampini ncoarse_ml = void_procs; 4855779c1cceSStefano Zampini /* it has no sense to redistribute on a set of processors larger than the number of active processes */ 4856779c1cceSStefano Zampini if (pcbddc->redistribute_coarse > 0 && pcbddc->redistribute_coarse < active_procs) { 48576e683305SStefano Zampini csin_ds = PETSC_TRUE; 485818a45a71SStefano Zampini ncoarse_ds = pcbddc->redistribute_coarse; 485918a45a71SStefano Zampini redist = PETSC_TRUE; 486018a45a71SStefano Zampini } else { 48616e683305SStefano Zampini csin_ds = PETSC_TRUE; 4862779c1cceSStefano Zampini ncoarse_ds = active_procs; 4863779c1cceSStefano Zampini redist = PETSC_TRUE; 486418a45a71SStefano Zampini } 48656e683305SStefano Zampini } else { 48666e683305SStefano Zampini csin_ml = PETSC_FALSE; 48676e683305SStefano Zampini ncoarse_ml = all_procs; 48686e683305SStefano Zampini if (void_procs) { 48696e683305SStefano Zampini csin_ds = PETSC_TRUE; 48706e683305SStefano Zampini ncoarse_ds = void_procs; 48716e683305SStefano Zampini csin_type_simple = PETSC_FALSE; 48726e683305SStefano Zampini } else { 4873779c1cceSStefano Zampini if (pcbddc->redistribute_coarse > 0 && pcbddc->redistribute_coarse < all_procs) { 487474e2c79eSStefano Zampini csin_ds = PETSC_TRUE; 487574e2c79eSStefano Zampini ncoarse_ds = pcbddc->redistribute_coarse; 487674e2c79eSStefano Zampini redist = PETSC_TRUE; 487774e2c79eSStefano Zampini } else { 48786e683305SStefano Zampini csin_ds = PETSC_FALSE; 48796e683305SStefano Zampini ncoarse_ds = all_procs; 48806e683305SStefano Zampini } 48816e683305SStefano Zampini } 488274e2c79eSStefano Zampini } 48836e683305SStefano Zampini 48846e683305SStefano Zampini /* 48856e683305SStefano Zampini test if we can go multilevel: three conditions must be satisfied: 48866e683305SStefano Zampini - we have not exceeded the number of levels requested 48876e683305SStefano Zampini - we can actually subassemble the active processes 48886e683305SStefano Zampini - we can find a suitable number of MPI processes where we can place the subassembled problem 48896e683305SStefano Zampini */ 48906e683305SStefano Zampini multilevel_allowed = PETSC_FALSE; 48916e683305SStefano Zampini multilevel_requested = PETSC_FALSE; 48926e683305SStefano Zampini if (pcbddc->current_level < pcbddc->max_levels) { 48936e683305SStefano Zampini multilevel_requested = PETSC_TRUE; 48946e683305SStefano Zampini if (active_procs/pcbddc->coarsening_ratio < 2 || ncoarse_ml/pcbddc->coarsening_ratio < 2) { 4895f9eb5b7dSStefano Zampini multilevel_allowed = PETSC_FALSE; 48962b510759SStefano Zampini } else { 4897f9eb5b7dSStefano Zampini multilevel_allowed = PETSC_TRUE; 4898c8587f34SStefano Zampini } 4899c8587f34SStefano Zampini } 49006e683305SStefano Zampini /* determine number of process partecipating to coarse solver */ 49016e683305SStefano Zampini if (multilevel_allowed) { 49026e683305SStefano Zampini ncoarse = ncoarse_ml; 49036e683305SStefano Zampini csin = csin_ml; 490458da7f69SStefano Zampini redist = PETSC_FALSE; 49056e683305SStefano Zampini } else { 49066e683305SStefano Zampini ncoarse = ncoarse_ds; 49076e683305SStefano Zampini csin = csin_ds; 49086e683305SStefano Zampini } 4909e7931f94SStefano Zampini 4910abbbba34SStefano Zampini /* creates temporary l2gmap and IS for coarse indexes */ 4911abbbba34SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),pcbddc->local_primal_size,pcbddc->global_primal_indices,PETSC_COPY_VALUES,&coarse_is);CHKERRQ(ierr); 4912abbbba34SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(coarse_is,&coarse_islg);CHKERRQ(ierr); 4913abbbba34SStefano Zampini 4914abbbba34SStefano Zampini /* creates temporary MATIS object for coarse matrix */ 491522bc73bbSStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,pcbddc->local_primal_size,pcbddc->local_primal_size,NULL,&coarse_submat_dense);CHKERRQ(ierr); 491622bc73bbSStefano Zampini ierr = MatDenseGetArray(coarse_submat_dense,&array);CHKERRQ(ierr); 491722bc73bbSStefano Zampini ierr = PetscMemcpy(array,coarse_submat_vals,sizeof(*coarse_submat_vals)*pcbddc->local_primal_size*pcbddc->local_primal_size);CHKERRQ(ierr); 491822bc73bbSStefano Zampini ierr = MatDenseRestoreArray(coarse_submat_dense,&array);CHKERRQ(ierr); 4919e176bc59SStefano 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); 49206e683305SStefano Zampini ierr = MatISSetLocalMat(t_coarse_mat_is,coarse_submat_dense);CHKERRQ(ierr); 49216e683305SStefano Zampini ierr = MatAssemblyBegin(t_coarse_mat_is,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 49226e683305SStefano Zampini ierr = MatAssemblyEnd(t_coarse_mat_is,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4923abbbba34SStefano Zampini ierr = MatDestroy(&coarse_submat_dense);CHKERRQ(ierr); 4924abbbba34SStefano Zampini 49256e683305SStefano Zampini /* compute dofs splitting and neumann boundaries for coarse dofs */ 492630368db7SStefano Zampini if (multilevel_allowed && (pcbddc->n_ISForDofsLocal || pcbddc->NeumannBoundariesLocal || (pcbddc->benign_saddle_point && pcbddc->user_primal_vertices_local))) { /* protects from unneded computations */ 49276e683305SStefano Zampini PetscInt *tidxs,*tidxs2,nout,tsize,i; 49286e683305SStefano Zampini const PetscInt *idxs; 49296e683305SStefano Zampini ISLocalToGlobalMapping tmap; 49306e683305SStefano Zampini 49316e683305SStefano Zampini /* create map between primal indices (in local representative ordering) and local primal numbering */ 49320be93d54SStefano Zampini ierr = ISLocalToGlobalMappingCreate(PETSC_COMM_SELF,1,pcbddc->local_primal_size,pcbddc->primal_indices_local_idxs,PETSC_COPY_VALUES,&tmap);CHKERRQ(ierr); 49336e683305SStefano Zampini /* allocate space for temporary storage */ 4934854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->local_primal_size,&tidxs);CHKERRQ(ierr); 4935854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->local_primal_size,&tidxs2);CHKERRQ(ierr); 49366e683305SStefano Zampini /* allocate for IS array */ 49376e683305SStefano Zampini nisdofs = pcbddc->n_ISForDofsLocal; 49386e683305SStefano Zampini nisneu = !!pcbddc->NeumannBoundariesLocal; 493930368db7SStefano Zampini nisvert = 0; 494030368db7SStefano Zampini if (pcbddc->benign_saddle_point && pcbddc->user_primal_vertices_local) { 494130368db7SStefano Zampini nisvert = 1; 494230368db7SStefano Zampini } 494330368db7SStefano Zampini nis = nisdofs + nisneu + nisvert; 4944854ce69bSBarry Smith ierr = PetscMalloc1(nis,&isarray);CHKERRQ(ierr); 49456e683305SStefano Zampini /* dofs splitting */ 49466e683305SStefano Zampini for (i=0;i<nisdofs;i++) { 49476e683305SStefano Zampini /* ierr = ISView(pcbddc->ISForDofsLocal[i],0);CHKERRQ(ierr); */ 49486e683305SStefano Zampini ierr = ISGetLocalSize(pcbddc->ISForDofsLocal[i],&tsize);CHKERRQ(ierr); 49496e683305SStefano Zampini ierr = ISGetIndices(pcbddc->ISForDofsLocal[i],&idxs);CHKERRQ(ierr); 49506e683305SStefano Zampini ierr = ISGlobalToLocalMappingApply(tmap,IS_GTOLM_DROP,tsize,idxs,&nout,tidxs);CHKERRQ(ierr); 49516e683305SStefano Zampini ierr = ISRestoreIndices(pcbddc->ISForDofsLocal[i],&idxs);CHKERRQ(ierr); 49526e683305SStefano Zampini ierr = ISLocalToGlobalMappingApply(coarse_islg,nout,tidxs,tidxs2);CHKERRQ(ierr); 495330368db7SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),nout,tidxs2,PETSC_COPY_VALUES,&isarray[i]);CHKERRQ(ierr); 49546e683305SStefano Zampini /* ierr = ISView(isarray[i],0);CHKERRQ(ierr); */ 49556e683305SStefano Zampini } 49566e683305SStefano Zampini /* neumann boundaries */ 49576e683305SStefano Zampini if (pcbddc->NeumannBoundariesLocal) { 49586e683305SStefano Zampini /* ierr = ISView(pcbddc->NeumannBoundariesLocal,0);CHKERRQ(ierr); */ 49596e683305SStefano Zampini ierr = ISGetLocalSize(pcbddc->NeumannBoundariesLocal,&tsize);CHKERRQ(ierr); 49606e683305SStefano Zampini ierr = ISGetIndices(pcbddc->NeumannBoundariesLocal,&idxs);CHKERRQ(ierr); 49616e683305SStefano Zampini ierr = ISGlobalToLocalMappingApply(tmap,IS_GTOLM_DROP,tsize,idxs,&nout,tidxs);CHKERRQ(ierr); 49626e683305SStefano Zampini ierr = ISRestoreIndices(pcbddc->NeumannBoundariesLocal,&idxs);CHKERRQ(ierr); 49636e683305SStefano Zampini ierr = ISLocalToGlobalMappingApply(coarse_islg,nout,tidxs,tidxs2);CHKERRQ(ierr); 496430368db7SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),nout,tidxs2,PETSC_COPY_VALUES,&isarray[nisdofs]);CHKERRQ(ierr); 49656e683305SStefano Zampini /* ierr = ISView(isarray[nisdofs],0);CHKERRQ(ierr); */ 49666e683305SStefano Zampini } 496730368db7SStefano Zampini /* primal vertices (benign) */ 496830368db7SStefano Zampini if (pcbddc->benign_saddle_point && pcbddc->user_primal_vertices_local) { 496930368db7SStefano Zampini ierr = ISGetLocalSize(pcbddc->user_primal_vertices_local,&tsize);CHKERRQ(ierr); 497030368db7SStefano Zampini ierr = ISGetIndices(pcbddc->user_primal_vertices_local,&idxs);CHKERRQ(ierr); 497130368db7SStefano Zampini ierr = ISGlobalToLocalMappingApply(tmap,IS_GTOLM_DROP,tsize,idxs,&nout,tidxs);CHKERRQ(ierr); 497230368db7SStefano Zampini ierr = ISRestoreIndices(pcbddc->user_primal_vertices_local,&idxs);CHKERRQ(ierr); 497330368db7SStefano Zampini ierr = ISLocalToGlobalMappingApply(coarse_islg,nout,tidxs,tidxs2);CHKERRQ(ierr); 497430368db7SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),nout,tidxs2,PETSC_COPY_VALUES,&isarray[nis-1]);CHKERRQ(ierr); 497530368db7SStefano Zampini } 49766e683305SStefano Zampini /* free memory */ 49776e683305SStefano Zampini ierr = PetscFree(tidxs);CHKERRQ(ierr); 49786e683305SStefano Zampini ierr = PetscFree(tidxs2);CHKERRQ(ierr); 49796e683305SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&tmap);CHKERRQ(ierr); 49806e683305SStefano Zampini } else { 49816e683305SStefano Zampini nis = 0; 49826e683305SStefano Zampini nisdofs = 0; 49836e683305SStefano Zampini nisneu = 0; 498430368db7SStefano Zampini nisvert = 0; 49856e683305SStefano Zampini isarray = NULL; 49866e683305SStefano Zampini } 49876e683305SStefano Zampini /* destroy no longer needed map */ 49886e683305SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&coarse_islg);CHKERRQ(ierr); 49896e683305SStefano Zampini 49906e683305SStefano Zampini /* restrict on coarse candidates (if needed) */ 49916e683305SStefano Zampini coarse_mat_is = NULL; 49926e683305SStefano Zampini if (csin) { 49936e683305SStefano Zampini if (!pcbddc->coarse_subassembling_init ) { /* creates subassembling init pattern if not present */ 499474e2c79eSStefano Zampini if (redist) { 499574e2c79eSStefano Zampini PetscMPIInt rank; 4996779c1cceSStefano Zampini PetscInt spc,n_spc_p1,dest[1],destsize; 499774e2c79eSStefano Zampini 499874e2c79eSStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)pc),&rank);CHKERRQ(ierr); 499958da7f69SStefano Zampini spc = active_procs/ncoarse; 500058da7f69SStefano Zampini n_spc_p1 = active_procs%ncoarse; 5001779c1cceSStefano Zampini if (im_active) { 5002779c1cceSStefano Zampini destsize = 1; 500374e2c79eSStefano Zampini if (rank > n_spc_p1*(spc+1)-1) { 500474e2c79eSStefano Zampini dest[0] = n_spc_p1+(rank-(n_spc_p1*(spc+1)))/spc; 500574e2c79eSStefano Zampini } else { 500674e2c79eSStefano Zampini dest[0] = rank/(spc+1); 500774e2c79eSStefano Zampini } 500874e2c79eSStefano Zampini } else { 5009779c1cceSStefano Zampini destsize = 0; 50106e683305SStefano Zampini } 5011779c1cceSStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),destsize,dest,PETSC_COPY_VALUES,&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 5012779c1cceSStefano Zampini } else if (csin_type_simple) { 50136e683305SStefano Zampini PetscMPIInt rank; 50146e683305SStefano Zampini PetscInt issize,isidx; 5015779c1cceSStefano Zampini 50166e683305SStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)pc),&rank);CHKERRQ(ierr); 50176e683305SStefano Zampini if (im_active) { 50186e683305SStefano Zampini issize = 1; 50196e683305SStefano Zampini isidx = (PetscInt)rank; 50206e683305SStefano Zampini } else { 50216e683305SStefano Zampini issize = 0; 50226e683305SStefano Zampini isidx = -1; 50236e683305SStefano Zampini } 50246e683305SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),issize,&isidx,PETSC_COPY_VALUES,&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 5025779c1cceSStefano Zampini } else { /* get a suitable subassembling pattern from MATIS code */ 5026b0c7d250SStefano Zampini ierr = MatISGetSubassemblingPattern(t_coarse_mat_is,ncoarse,pcbddc->coarse_adj_red,&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 50276e683305SStefano Zampini } 5028779c1cceSStefano Zampini 5029779c1cceSStefano Zampini /* we need to shift on coarse candidates either if we are not redistributing or we are redistributing and we have enough void processes */ 5030779c1cceSStefano Zampini if (!redist || ncoarse <= void_procs) { 5031779c1cceSStefano Zampini PetscInt ncoarse_cand,tissize,*nisindices; 5032779c1cceSStefano Zampini PetscInt *coarse_candidates; 5033779c1cceSStefano Zampini const PetscInt* tisindices; 5034779c1cceSStefano Zampini 5035779c1cceSStefano Zampini /* get coarse candidates' ranks in pc communicator */ 5036779c1cceSStefano Zampini ierr = PetscMalloc1(all_procs,&coarse_candidates);CHKERRQ(ierr); 5037779c1cceSStefano Zampini ierr = MPI_Allgather(&im_active,1,MPIU_INT,coarse_candidates,1,MPIU_INT,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 5038779c1cceSStefano Zampini for (i=0,ncoarse_cand=0;i<all_procs;i++) { 5039779c1cceSStefano Zampini if (!coarse_candidates[i]) { 5040779c1cceSStefano Zampini coarse_candidates[ncoarse_cand++]=i; 5041779c1cceSStefano Zampini } 5042779c1cceSStefano Zampini } 5043779c1cceSStefano Zampini if (ncoarse_cand < ncoarse) SETERRQ2(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"This should not happen! %d < %d",ncoarse_cand,ncoarse); 5044779c1cceSStefano Zampini 5045779c1cceSStefano Zampini 50466e683305SStefano Zampini if (pcbddc->dbg_flag) { 50476e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 50486e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Subassembling pattern init (before shift)\n");CHKERRQ(ierr); 50496e683305SStefano Zampini ierr = ISView(pcbddc->coarse_subassembling_init,pcbddc->dbg_viewer);CHKERRQ(ierr); 50506e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Coarse candidates\n");CHKERRQ(ierr); 5051779c1cceSStefano Zampini for (i=0;i<ncoarse_cand;i++) { 50526e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"%d ",coarse_candidates[i]);CHKERRQ(ierr); 50536e683305SStefano Zampini } 50546e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"\n");CHKERRQ(ierr); 50556e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 50566e683305SStefano Zampini } 50576e683305SStefano Zampini /* shift the pattern on coarse candidates */ 50586e683305SStefano Zampini ierr = ISGetLocalSize(pcbddc->coarse_subassembling_init,&tissize);CHKERRQ(ierr); 50596e683305SStefano Zampini ierr = ISGetIndices(pcbddc->coarse_subassembling_init,&tisindices);CHKERRQ(ierr); 5060854ce69bSBarry Smith ierr = PetscMalloc1(tissize,&nisindices);CHKERRQ(ierr); 50616e683305SStefano Zampini for (i=0;i<tissize;i++) nisindices[i] = coarse_candidates[tisindices[i]]; 50626e683305SStefano Zampini ierr = ISRestoreIndices(pcbddc->coarse_subassembling_init,&tisindices);CHKERRQ(ierr); 50636e683305SStefano Zampini ierr = ISGeneralSetIndices(pcbddc->coarse_subassembling_init,tissize,nisindices,PETSC_OWN_POINTER);CHKERRQ(ierr); 50646e683305SStefano Zampini ierr = PetscFree(coarse_candidates);CHKERRQ(ierr); 50656e683305SStefano Zampini } 50666e683305SStefano Zampini if (pcbddc->dbg_flag) { 50676e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 50686e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Subassembling pattern init\n");CHKERRQ(ierr); 50696e683305SStefano Zampini ierr = ISView(pcbddc->coarse_subassembling_init,pcbddc->dbg_viewer);CHKERRQ(ierr); 50706e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 50716e683305SStefano Zampini } 5072779c1cceSStefano Zampini } 50736e683305SStefano Zampini /* get temporary coarse mat in IS format restricted on coarse procs (plus additional index sets of isarray) */ 507453a05cb3SStefano Zampini if (multilevel_allowed) { /* we need to keep tracking of void processes for future placements */ 507553a05cb3SStefano 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); 507653a05cb3SStefano Zampini } else { /* this is the last level, so use just receiving processes in subcomm */ 507753a05cb3SStefano 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); 507853a05cb3SStefano Zampini } 50796e683305SStefano Zampini } else { 50806e683305SStefano Zampini if (pcbddc->dbg_flag) { 50816e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 50826e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Subassembling pattern init not needed\n");CHKERRQ(ierr); 50836e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 50846e683305SStefano Zampini } 50856e683305SStefano Zampini ierr = PetscObjectReference((PetscObject)t_coarse_mat_is);CHKERRQ(ierr); 50866e683305SStefano Zampini coarse_mat_is = t_coarse_mat_is; 50876e683305SStefano Zampini } 50886e683305SStefano Zampini 50896e683305SStefano Zampini /* create local to global scatters for coarse problem */ 509068457ee5SStefano Zampini if (compute_vecs) { 50916e683305SStefano Zampini PetscInt lrows; 50926e683305SStefano Zampini ierr = VecDestroy(&pcbddc->coarse_vec);CHKERRQ(ierr); 50936e683305SStefano Zampini if (coarse_mat_is) { 50946e683305SStefano Zampini ierr = MatGetLocalSize(coarse_mat_is,&lrows,NULL);CHKERRQ(ierr); 50956e683305SStefano Zampini } else { 50966e683305SStefano Zampini lrows = 0; 50976e683305SStefano Zampini } 50986e683305SStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pc),&pcbddc->coarse_vec);CHKERRQ(ierr); 50996e683305SStefano Zampini ierr = VecSetSizes(pcbddc->coarse_vec,lrows,PETSC_DECIDE);CHKERRQ(ierr); 51006e683305SStefano Zampini ierr = VecSetType(pcbddc->coarse_vec,VECSTANDARD);CHKERRQ(ierr); 51016e683305SStefano Zampini ierr = VecScatterDestroy(&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 51026e683305SStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_P,NULL,pcbddc->coarse_vec,coarse_is,&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 51036e683305SStefano Zampini } 51046e683305SStefano Zampini ierr = ISDestroy(&coarse_is);CHKERRQ(ierr); 51056e683305SStefano Zampini ierr = MatDestroy(&t_coarse_mat_is);CHKERRQ(ierr); 5106c8587f34SStefano Zampini 5107f9eb5b7dSStefano Zampini /* set defaults for coarse KSP and PC */ 5108f9eb5b7dSStefano Zampini if (multilevel_allowed) { 5109f9eb5b7dSStefano Zampini coarse_ksp_type = KSPRICHARDSON; 5110f9eb5b7dSStefano Zampini coarse_pc_type = PCBDDC; 5111f9eb5b7dSStefano Zampini } else { 5112f9eb5b7dSStefano Zampini coarse_ksp_type = KSPPREONLY; 5113f9eb5b7dSStefano Zampini coarse_pc_type = PCREDUNDANT; 5114c8587f34SStefano Zampini } 5115c8587f34SStefano Zampini 51166e683305SStefano Zampini /* print some info if requested */ 51176e683305SStefano Zampini if (pcbddc->dbg_flag) { 51186e683305SStefano Zampini if (!multilevel_allowed) { 51196e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 51206e683305SStefano Zampini if (multilevel_requested) { 51216e683305SStefano 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); 51226e683305SStefano Zampini } else if (pcbddc->max_levels) { 51236e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Maximum number of requested levels reached (%d)\n",pcbddc->max_levels);CHKERRQ(ierr); 51246e683305SStefano Zampini } 51256e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 51266e683305SStefano Zampini } 51276e683305SStefano Zampini } 51286e683305SStefano Zampini 5129f9eb5b7dSStefano Zampini /* create the coarse KSP object only once with defaults */ 51306e683305SStefano Zampini if (coarse_mat_is) { 51316e683305SStefano Zampini MatReuse coarse_mat_reuse; 51326a1308c2SStefano Zampini PetscViewer dbg_viewer = NULL; 51336e683305SStefano Zampini if (pcbddc->dbg_flag) { 51346e683305SStefano Zampini dbg_viewer = PETSC_VIEWER_STDOUT_(PetscObjectComm((PetscObject)coarse_mat_is)); 51356e683305SStefano Zampini ierr = PetscViewerASCIIAddTab(dbg_viewer,2*pcbddc->current_level);CHKERRQ(ierr); 51366e683305SStefano Zampini } 5137f9eb5b7dSStefano Zampini if (!pcbddc->coarse_ksp) { 5138312be037SStefano Zampini char prefix[256],str_level[16]; 5139e604994aSStefano Zampini size_t len; 51406e683305SStefano Zampini ierr = KSPCreate(PetscObjectComm((PetscObject)coarse_mat_is),&pcbddc->coarse_ksp);CHKERRQ(ierr); 5141422a814eSBarry Smith ierr = KSPSetErrorIfNotConverged(pcbddc->coarse_ksp,pc->erroriffailure);CHKERRQ(ierr); 5142c8587f34SStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->coarse_ksp,(PetscObject)pc,1);CHKERRQ(ierr); 5143f9eb5b7dSStefano Zampini ierr = KSPSetTolerances(pcbddc->coarse_ksp,PETSC_DEFAULT,PETSC_DEFAULT,PETSC_DEFAULT,1);CHKERRQ(ierr); 51445f76c7aeSStefano Zampini ierr = KSPSetOperators(pcbddc->coarse_ksp,coarse_mat_is,coarse_mat_is);CHKERRQ(ierr); 5145c8587f34SStefano Zampini ierr = KSPSetType(pcbddc->coarse_ksp,coarse_ksp_type);CHKERRQ(ierr); 51466e683305SStefano Zampini ierr = KSPSetNormType(pcbddc->coarse_ksp,KSP_NORM_NONE);CHKERRQ(ierr); 5147c8587f34SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&pc_temp);CHKERRQ(ierr); 5148c8587f34SStefano Zampini ierr = PCSetType(pc_temp,coarse_pc_type);CHKERRQ(ierr); 5149e604994aSStefano Zampini /* prefix */ 5150e604994aSStefano Zampini ierr = PetscStrcpy(prefix,"");CHKERRQ(ierr); 5151e604994aSStefano Zampini ierr = PetscStrcpy(str_level,"");CHKERRQ(ierr); 5152e604994aSStefano Zampini if (!pcbddc->current_level) { 5153e604994aSStefano Zampini ierr = PetscStrcpy(prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 5154e604994aSStefano Zampini ierr = PetscStrcat(prefix,"pc_bddc_coarse_");CHKERRQ(ierr); 5155c8587f34SStefano Zampini } else { 5156e604994aSStefano Zampini ierr = PetscStrlen(((PetscObject)pc)->prefix,&len);CHKERRQ(ierr); 5157312be037SStefano Zampini if (pcbddc->current_level>1) len -= 3; /* remove "lX_" with X level number */ 5158312be037SStefano Zampini if (pcbddc->current_level>10) len -= 1; /* remove another char from level number */ 515934d6797cSStefano Zampini ierr = PetscStrncpy(prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 5160312be037SStefano Zampini sprintf(str_level,"l%d_",(int)(pcbddc->current_level)); 5161e604994aSStefano Zampini ierr = PetscStrcat(prefix,str_level);CHKERRQ(ierr); 5162e604994aSStefano Zampini } 5163e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->coarse_ksp,prefix);CHKERRQ(ierr); 51643e3c6dadSStefano Zampini /* propagate BDDC info to the next level (these are dummy calls if pc_temp is not of type PCBDDC) */ 51653e3c6dadSStefano Zampini ierr = PCBDDCSetLevel(pc_temp,pcbddc->current_level+1);CHKERRQ(ierr); 51663e3c6dadSStefano Zampini ierr = PCBDDCSetCoarseningRatio(pc_temp,pcbddc->coarsening_ratio);CHKERRQ(ierr); 51673e3c6dadSStefano Zampini ierr = PCBDDCSetLevels(pc_temp,pcbddc->max_levels);CHKERRQ(ierr); 5168f9eb5b7dSStefano Zampini /* allow user customization */ 5169f9eb5b7dSStefano Zampini ierr = KSPSetFromOptions(pcbddc->coarse_ksp);CHKERRQ(ierr); 51703e3c6dadSStefano Zampini } 51713e3c6dadSStefano Zampini /* propagate BDDC info to the next level (these are dummy calls if pc_temp is not of type PCBDDC) */ 517251bea450SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&pc_temp);CHKERRQ(ierr); 51733e3c6dadSStefano Zampini if (nisdofs) { 51743e3c6dadSStefano Zampini ierr = PCBDDCSetDofsSplitting(pc_temp,nisdofs,isarray);CHKERRQ(ierr); 51753e3c6dadSStefano Zampini for (i=0;i<nisdofs;i++) { 51763e3c6dadSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 51773e3c6dadSStefano Zampini } 51783e3c6dadSStefano Zampini } 51793e3c6dadSStefano Zampini if (nisneu) { 51803e3c6dadSStefano Zampini ierr = PCBDDCSetNeumannBoundaries(pc_temp,isarray[nisdofs]);CHKERRQ(ierr); 51813e3c6dadSStefano Zampini ierr = ISDestroy(&isarray[nisdofs]);CHKERRQ(ierr); 5182312be037SStefano Zampini } 518330368db7SStefano Zampini if (nisvert) { 518430368db7SStefano Zampini ierr = PCBDDCSetPrimalVerticesIS(pc_temp,isarray[nis-1]);CHKERRQ(ierr); 518530368db7SStefano Zampini ierr = ISDestroy(&isarray[nis-1]);CHKERRQ(ierr); 518630368db7SStefano Zampini } 5187f9eb5b7dSStefano Zampini 5188f9eb5b7dSStefano Zampini /* get some info after set from options */ 5189f9eb5b7dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCNN,&isnn);CHKERRQ(ierr); 5190f9eb5b7dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCBDDC,&isbddc);CHKERRQ(ierr); 51914f3a063dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCREDUNDANT,&isredundant);CHKERRQ(ierr); 51926e683305SStefano Zampini if (isbddc && !multilevel_allowed) { /* multilevel can only be requested via pc_bddc_set_levels */ 5193f9eb5b7dSStefano Zampini ierr = PCSetType(pc_temp,coarse_pc_type);CHKERRQ(ierr); 5194f9eb5b7dSStefano Zampini isbddc = PETSC_FALSE; 5195f9eb5b7dSStefano Zampini } 519639f0f02cSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 51974f3a063dSStefano Zampini if (isredundant) { 51984f3a063dSStefano Zampini KSP inner_ksp; 51994f3a063dSStefano Zampini PC inner_pc; 52004f3a063dSStefano Zampini ierr = PCRedundantGetKSP(pc_temp,&inner_ksp);CHKERRQ(ierr); 52014f3a063dSStefano Zampini ierr = KSPGetPC(inner_ksp,&inner_pc);CHKERRQ(ierr); 52024f3a063dSStefano Zampini ierr = PCFactorSetReuseFill(inner_pc,PETSC_TRUE);CHKERRQ(ierr); 52034f3a063dSStefano Zampini } 5204f9eb5b7dSStefano Zampini 5205f9eb5b7dSStefano Zampini /* assemble coarse matrix */ 5206fa7f1dd8SStefano Zampini if (coarse_reuse) { 520781d9aea3SBarry Smith ierr = KSPGetOperators(pcbddc->coarse_ksp,&coarse_mat,NULL);CHKERRQ(ierr); 5208fa7f1dd8SStefano Zampini ierr = PetscObjectReference((PetscObject)coarse_mat);CHKERRQ(ierr); 52096e683305SStefano Zampini coarse_mat_reuse = MAT_REUSE_MATRIX; 5210fa7f1dd8SStefano Zampini } else { 52116e683305SStefano Zampini coarse_mat_reuse = MAT_INITIAL_MATRIX; 5212fa7f1dd8SStefano Zampini } 5213c8587f34SStefano Zampini if (isbddc || isnn) { 521422bc73bbSStefano Zampini if (pcbddc->coarsening_ratio > 1) { 521570cf5478SStefano Zampini if (!pcbddc->coarse_subassembling) { /* subassembling info is not present */ 5216b0c7d250SStefano Zampini ierr = MatISGetSubassemblingPattern(coarse_mat_is,active_procs/pcbddc->coarsening_ratio,pcbddc->coarse_adj_red,&pcbddc->coarse_subassembling);CHKERRQ(ierr); 521722b6e8a2SStefano Zampini if (pcbddc->dbg_flag) { 52186e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 52196e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Subassembling pattern\n");CHKERRQ(ierr); 52206e683305SStefano Zampini ierr = ISView(pcbddc->coarse_subassembling,dbg_viewer);CHKERRQ(ierr); 52216e683305SStefano Zampini ierr = PetscViewerFlush(dbg_viewer);CHKERRQ(ierr); 522222b6e8a2SStefano Zampini } 522370cf5478SStefano Zampini } 522453a05cb3SStefano Zampini ierr = MatISSubassemble(coarse_mat_is,pcbddc->coarse_subassembling,0,PETSC_FALSE,PETSC_FALSE,coarse_mat_reuse,&coarse_mat,0,NULL);CHKERRQ(ierr); 522570cf5478SStefano Zampini } else { 522622bc73bbSStefano Zampini ierr = PetscObjectReference((PetscObject)coarse_mat_is);CHKERRQ(ierr); 522722bc73bbSStefano Zampini coarse_mat = coarse_mat_is; 522822bc73bbSStefano Zampini } 522922bc73bbSStefano Zampini } else { 52302e1e5fa4SStefano Zampini ierr = MatISGetMPIXAIJ(coarse_mat_is,coarse_mat_reuse,&coarse_mat);CHKERRQ(ierr); 5231c8587f34SStefano Zampini } 5232c8587f34SStefano Zampini ierr = MatDestroy(&coarse_mat_is);CHKERRQ(ierr); 5233c8587f34SStefano Zampini 52343301b35fSStefano Zampini /* propagate symmetry info of coarse matrix */ 52355a16e3a0SStefano Zampini ierr = MatSetOption(coarse_mat,MAT_STRUCTURALLY_SYMMETRIC,PETSC_TRUE);CHKERRQ(ierr); 52363301b35fSStefano Zampini if (pc->pmat->symmetric_set) { 52373301b35fSStefano Zampini ierr = MatSetOption(coarse_mat,MAT_SYMMETRIC,pc->pmat->symmetric);CHKERRQ(ierr); 52383301b35fSStefano Zampini } 52393301b35fSStefano Zampini if (pc->pmat->hermitian_set) { 52403301b35fSStefano Zampini ierr = MatSetOption(coarse_mat,MAT_HERMITIAN,pc->pmat->hermitian);CHKERRQ(ierr); 52413301b35fSStefano Zampini } 52423301b35fSStefano Zampini if (pc->pmat->spd_set) { 52433301b35fSStefano Zampini ierr = MatSetOption(coarse_mat,MAT_SPD,pc->pmat->spd);CHKERRQ(ierr); 52443301b35fSStefano Zampini } 52456e683305SStefano Zampini /* set operators */ 52465f76c7aeSStefano Zampini ierr = KSPSetOperators(pcbddc->coarse_ksp,coarse_mat,coarse_mat);CHKERRQ(ierr); 52476e683305SStefano Zampini if (pcbddc->dbg_flag) { 52486e683305SStefano Zampini ierr = PetscViewerASCIISubtractTab(dbg_viewer,2*pcbddc->current_level);CHKERRQ(ierr); 52496e683305SStefano Zampini } 52506e683305SStefano Zampini } else { /* processes non partecipating to coarse solver (if any) */ 52516e683305SStefano Zampini coarse_mat = 0; 52526e683305SStefano Zampini } 52536e683305SStefano Zampini ierr = PetscFree(isarray);CHKERRQ(ierr); 5254b1ecc7b1SStefano Zampini #if 0 5255b9b85e73SStefano Zampini { 5256b9b85e73SStefano Zampini PetscViewer viewer; 5257b9b85e73SStefano Zampini char filename[256]; 5258b1ecc7b1SStefano Zampini sprintf(filename,"coarse_mat_level%d.m",pcbddc->current_level); 5259b1ecc7b1SStefano Zampini ierr = PetscViewerASCIIOpen(PetscObjectComm((PetscObject)coarse_mat),filename,&viewer);CHKERRQ(ierr); 5260b9b85e73SStefano Zampini ierr = PetscViewerSetFormat(viewer,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 5261b9b85e73SStefano Zampini ierr = MatView(coarse_mat,viewer);CHKERRQ(ierr); 5262b9b85e73SStefano Zampini ierr = PetscViewerDestroy(&viewer);CHKERRQ(ierr); 5263b9b85e73SStefano Zampini } 5264b9b85e73SStefano Zampini #endif 5265c8587f34SStefano Zampini 5266c8587f34SStefano Zampini /* Compute coarse null space (special handling by BDDC only) */ 5267298c0119SStefano Zampini #if 0 5268c8587f34SStefano Zampini if (pcbddc->NullSpace) { 5269c8587f34SStefano Zampini ierr = PCBDDCNullSpaceAssembleCoarse(pc,coarse_mat,&CoarseNullSpace);CHKERRQ(ierr); 527098a51de6SStefano Zampini } 5271298c0119SStefano Zampini #endif 5272b0f5fe93SStefano Zampini /* hack */ 527398a51de6SStefano Zampini if (pcbddc->coarse_ksp) { 527498a51de6SStefano Zampini Vec crhs,csol; 527504708bb6SStefano Zampini 5276f347579bSStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&csol);CHKERRQ(ierr); 5277f347579bSStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&crhs);CHKERRQ(ierr); 5278f347579bSStefano Zampini if (!csol) { 52792a7a6963SBarry Smith ierr = MatCreateVecs(coarse_mat,&((pcbddc->coarse_ksp)->vec_sol),NULL);CHKERRQ(ierr); 5280f9eb5b7dSStefano Zampini } 5281f347579bSStefano Zampini if (!crhs) { 52822a7a6963SBarry Smith ierr = MatCreateVecs(coarse_mat,NULL,&((pcbddc->coarse_ksp)->vec_rhs));CHKERRQ(ierr); 5283f347579bSStefano Zampini } 5284b0f5fe93SStefano Zampini } 5285b0f5fe93SStefano Zampini 5286b0f5fe93SStefano Zampini /* compute null space for coarse solver if the benign trick has been requested */ 5287b0f5fe93SStefano Zampini if (pcbddc->benign_null) { 5288b0f5fe93SStefano Zampini 5289b0f5fe93SStefano Zampini ierr = VecSet(pcbddc->vec1_P,0.);CHKERRQ(ierr); 52904f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 52914f1b2e48SStefano Zampini ierr = VecSetValue(pcbddc->vec1_P,pcbddc->local_primal_size-pcbddc->benign_n+i,1.0,INSERT_VALUES);CHKERRQ(ierr); 52924f1b2e48SStefano Zampini } 5293b0f5fe93SStefano Zampini ierr = VecAssemblyBegin(pcbddc->vec1_P);CHKERRQ(ierr); 5294b0f5fe93SStefano Zampini ierr = VecAssemblyEnd(pcbddc->vec1_P);CHKERRQ(ierr); 5295b0f5fe93SStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 5296b0f5fe93SStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 5297b0f5fe93SStefano Zampini if (coarse_mat) { 5298b0f5fe93SStefano Zampini Vec nullv; 5299b0f5fe93SStefano Zampini PetscScalar *array,*array2; 5300b0f5fe93SStefano Zampini PetscInt nl; 5301b0f5fe93SStefano Zampini 5302b0f5fe93SStefano Zampini ierr = MatCreateVecs(coarse_mat,&nullv,NULL);CHKERRQ(ierr); 5303b0f5fe93SStefano Zampini ierr = VecGetLocalSize(nullv,&nl);CHKERRQ(ierr); 5304b0f5fe93SStefano Zampini ierr = VecGetArrayRead(pcbddc->coarse_vec,(const PetscScalar**)&array);CHKERRQ(ierr); 5305b0f5fe93SStefano Zampini ierr = VecGetArray(nullv,&array2);CHKERRQ(ierr); 5306b0f5fe93SStefano Zampini ierr = PetscMemcpy(array2,array,nl*sizeof(*array));CHKERRQ(ierr); 5307b0f5fe93SStefano Zampini ierr = VecRestoreArray(nullv,&array2);CHKERRQ(ierr); 5308b0f5fe93SStefano Zampini ierr = VecRestoreArrayRead(pcbddc->coarse_vec,(const PetscScalar**)&array);CHKERRQ(ierr); 5309b0f5fe93SStefano Zampini ierr = VecNormalize(nullv,NULL);CHKERRQ(ierr); 5310b0f5fe93SStefano Zampini ierr = MatNullSpaceCreate(PetscObjectComm((PetscObject)coarse_mat),PETSC_FALSE,1,&nullv,&CoarseNullSpace);CHKERRQ(ierr); 5311b0f5fe93SStefano Zampini ierr = VecDestroy(&nullv);CHKERRQ(ierr); 5312b0f5fe93SStefano Zampini } 5313b0f5fe93SStefano Zampini } 5314b0f5fe93SStefano Zampini 5315b0f5fe93SStefano Zampini if (pcbddc->coarse_ksp) { 5316b0f5fe93SStefano Zampini PetscBool ispreonly; 5317b0f5fe93SStefano Zampini 5318b0f5fe93SStefano Zampini if (CoarseNullSpace) { 5319b0f5fe93SStefano Zampini PetscBool isnull; 5320b0f5fe93SStefano Zampini ierr = MatNullSpaceTest(CoarseNullSpace,coarse_mat,&isnull);CHKERRQ(ierr); 5321b0f5fe93SStefano Zampini if (isnull) { 532230368db7SStefano Zampini if (isbddc && !pcbddc->benign_saddle_point) { 5323b0f5fe93SStefano Zampini ierr = PCBDDCSetNullSpace(pc_temp,CoarseNullSpace);CHKERRQ(ierr); 5324b0f5fe93SStefano Zampini } else { 5325b0f5fe93SStefano Zampini ierr = MatSetNullSpace(coarse_mat,CoarseNullSpace);CHKERRQ(ierr); 5326b0f5fe93SStefano Zampini } 5327b0f5fe93SStefano Zampini } else { 5328b0f5fe93SStefano Zampini ierr = MatNullSpaceDestroy(&CoarseNullSpace);CHKERRQ(ierr); 5329b0f5fe93SStefano Zampini } 5330b0f5fe93SStefano Zampini } 5331b0f5fe93SStefano Zampini /* setup coarse ksp */ 5332b0f5fe93SStefano Zampini ierr = KSPSetUp(pcbddc->coarse_ksp);CHKERRQ(ierr); 5333cbcc2c2aSStefano Zampini /* Check coarse problem if in debug mode or if solving with an iterative method */ 5334cbcc2c2aSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcbddc->coarse_ksp,KSPPREONLY,&ispreonly);CHKERRQ(ierr); 53356e683305SStefano Zampini if (pcbddc->dbg_flag || (!ispreonly && pcbddc->use_coarse_estimates) ) { 5336c8587f34SStefano Zampini KSP check_ksp; 53372b510759SStefano Zampini KSPType check_ksp_type; 5338c8587f34SStefano Zampini PC check_pc; 53396e683305SStefano Zampini Vec check_vec,coarse_vec; 53406a1308c2SStefano Zampini PetscReal abs_infty_error,infty_error,lambda_min=1.0,lambda_max=1.0; 53412b510759SStefano Zampini PetscInt its; 53426e683305SStefano Zampini PetscBool compute_eigs; 53436e683305SStefano Zampini PetscReal *eigs_r,*eigs_c; 53446e683305SStefano Zampini PetscInt neigs; 53458e185a42SStefano Zampini const char *prefix; 5346c8587f34SStefano Zampini 53472b510759SStefano Zampini /* Create ksp object suitable for estimation of extreme eigenvalues */ 53486e683305SStefano Zampini ierr = KSPCreate(PetscObjectComm((PetscObject)pcbddc->coarse_ksp),&check_ksp);CHKERRQ(ierr); 5349422a814eSBarry Smith ierr = KSPSetErrorIfNotConverged(pcbddc->coarse_ksp,pc->erroriffailure);CHKERRQ(ierr); 535023ee1639SBarry Smith ierr = KSPSetOperators(check_ksp,coarse_mat,coarse_mat);CHKERRQ(ierr); 5351f4ddd8eeSStefano Zampini ierr = KSPSetTolerances(check_ksp,1.e-12,1.e-12,PETSC_DEFAULT,pcbddc->coarse_size);CHKERRQ(ierr); 53522b510759SStefano Zampini if (ispreonly) { 53532b510759SStefano Zampini check_ksp_type = KSPPREONLY; 53546e683305SStefano Zampini compute_eigs = PETSC_FALSE; 53552b510759SStefano Zampini } else { 5356cbcc2c2aSStefano Zampini check_ksp_type = KSPGMRES; 53576e683305SStefano Zampini compute_eigs = PETSC_TRUE; 5358c8587f34SStefano Zampini } 5359c8587f34SStefano Zampini ierr = KSPSetType(check_ksp,check_ksp_type);CHKERRQ(ierr); 53606e683305SStefano Zampini ierr = KSPSetComputeSingularValues(check_ksp,compute_eigs);CHKERRQ(ierr); 53616e683305SStefano Zampini ierr = KSPSetComputeEigenvalues(check_ksp,compute_eigs);CHKERRQ(ierr); 53626e683305SStefano Zampini ierr = KSPGMRESSetRestart(check_ksp,pcbddc->coarse_size+1);CHKERRQ(ierr); 5363a7dc3881SStefano Zampini ierr = KSPGetOptionsPrefix(pcbddc->coarse_ksp,&prefix);CHKERRQ(ierr); 5364a7dc3881SStefano Zampini ierr = KSPSetOptionsPrefix(check_ksp,prefix);CHKERRQ(ierr); 5365a7dc3881SStefano Zampini ierr = KSPAppendOptionsPrefix(check_ksp,"check_");CHKERRQ(ierr); 5366a7dc3881SStefano Zampini ierr = KSPSetFromOptions(check_ksp);CHKERRQ(ierr); 5367c8587f34SStefano Zampini ierr = KSPSetUp(check_ksp);CHKERRQ(ierr); 5368c8587f34SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&check_pc);CHKERRQ(ierr); 5369c8587f34SStefano Zampini ierr = KSPSetPC(check_ksp,check_pc);CHKERRQ(ierr); 5370c8587f34SStefano Zampini /* create random vec */ 53716e683305SStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&coarse_vec);CHKERRQ(ierr); 53726e683305SStefano Zampini ierr = VecDuplicate(coarse_vec,&check_vec);CHKERRQ(ierr); 5373c8587f34SStefano Zampini ierr = VecSetRandom(check_vec,NULL);CHKERRQ(ierr); 5374c8587f34SStefano Zampini if (CoarseNullSpace) { 5375c8587f34SStefano Zampini ierr = MatNullSpaceRemove(CoarseNullSpace,check_vec);CHKERRQ(ierr); 5376c8587f34SStefano Zampini } 53776e683305SStefano Zampini ierr = MatMult(coarse_mat,check_vec,coarse_vec);CHKERRQ(ierr); 5378c8587f34SStefano Zampini /* solve coarse problem */ 53796e683305SStefano Zampini ierr = KSPSolve(check_ksp,coarse_vec,coarse_vec);CHKERRQ(ierr); 5380c8587f34SStefano Zampini if (CoarseNullSpace) { 53816e683305SStefano Zampini ierr = MatNullSpaceRemove(CoarseNullSpace,coarse_vec);CHKERRQ(ierr); 5382c8587f34SStefano Zampini } 5383cbcc2c2aSStefano Zampini /* set eigenvalue estimation if preonly has not been requested */ 53846e683305SStefano Zampini if (compute_eigs) { 5385854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->coarse_size+1,&eigs_r);CHKERRQ(ierr); 5386854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->coarse_size+1,&eigs_c);CHKERRQ(ierr); 53876e683305SStefano Zampini ierr = KSPComputeEigenvalues(check_ksp,pcbddc->coarse_size+1,eigs_r,eigs_c,&neigs);CHKERRQ(ierr); 53886e683305SStefano Zampini lambda_max = eigs_r[neigs-1]; 53896e683305SStefano Zampini lambda_min = eigs_r[0]; 53906e683305SStefano Zampini if (pcbddc->use_coarse_estimates) { 53916e683305SStefano Zampini if (lambda_max>lambda_min) { 5392cbcc2c2aSStefano Zampini ierr = KSPChebyshevSetEigenvalues(pcbddc->coarse_ksp,lambda_max,lambda_min);CHKERRQ(ierr); 5393cbcc2c2aSStefano Zampini ierr = KSPRichardsonSetScale(pcbddc->coarse_ksp,2.0/(lambda_max+lambda_min));CHKERRQ(ierr); 5394cbcc2c2aSStefano Zampini } 5395c8587f34SStefano Zampini } 5396c8587f34SStefano Zampini } 5397cbcc2c2aSStefano Zampini 5398c8587f34SStefano Zampini /* check coarse problem residual error */ 53996e683305SStefano Zampini if (pcbddc->dbg_flag) { 54006e683305SStefano Zampini PetscViewer dbg_viewer = PETSC_VIEWER_STDOUT_(PetscObjectComm((PetscObject)pcbddc->coarse_ksp)); 54016e683305SStefano Zampini ierr = PetscViewerASCIIAddTab(dbg_viewer,2*(pcbddc->current_level+1));CHKERRQ(ierr); 54026e683305SStefano Zampini ierr = VecAXPY(check_vec,-1.0,coarse_vec);CHKERRQ(ierr); 5403c8587f34SStefano Zampini ierr = VecNorm(check_vec,NORM_INFINITY,&infty_error);CHKERRQ(ierr); 54046e683305SStefano Zampini ierr = MatMult(coarse_mat,check_vec,coarse_vec);CHKERRQ(ierr); 54056e683305SStefano Zampini ierr = VecNorm(coarse_vec,NORM_INFINITY,&abs_infty_error);CHKERRQ(ierr); 5406c8587f34SStefano Zampini ierr = VecDestroy(&check_vec);CHKERRQ(ierr); 5407779c1cceSStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem details (use estimates %d)\n",pcbddc->use_coarse_estimates);CHKERRQ(ierr); 54086e683305SStefano Zampini ierr = PetscObjectPrintClassNamePrefixType((PetscObject)(pcbddc->coarse_ksp),dbg_viewer);CHKERRQ(ierr); 54096e683305SStefano Zampini ierr = PetscObjectPrintClassNamePrefixType((PetscObject)(check_pc),dbg_viewer);CHKERRQ(ierr); 54106e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem exact infty_error : %1.6e\n",infty_error);CHKERRQ(ierr); 54116e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem residual infty_error: %1.6e\n",abs_infty_error);CHKERRQ(ierr); 5412b0f5fe93SStefano Zampini if (CoarseNullSpace) { 5413b0f5fe93SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem is singular\n");CHKERRQ(ierr); 5414b0f5fe93SStefano Zampini } 54156e683305SStefano Zampini if (compute_eigs) { 54166e683305SStefano Zampini PetscReal lambda_max_s,lambda_min_s; 5417deec49d1SStefano Zampini ierr = KSPGetType(check_ksp,&check_ksp_type);CHKERRQ(ierr); 5418c8587f34SStefano Zampini ierr = KSPGetIterationNumber(check_ksp,&its);CHKERRQ(ierr); 54196e683305SStefano Zampini ierr = KSPComputeExtremeSingularValues(check_ksp,&lambda_max_s,&lambda_min_s);CHKERRQ(ierr); 54206e683305SStefano 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); 54216e683305SStefano Zampini for (i=0;i<neigs;i++) { 54226e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"%1.6e %1.6ei\n",eigs_r[i],eigs_c[i]);CHKERRQ(ierr); 5423c8587f34SStefano Zampini } 54246e683305SStefano Zampini } 54256e683305SStefano Zampini ierr = PetscViewerFlush(dbg_viewer);CHKERRQ(ierr); 54266e683305SStefano Zampini ierr = PetscViewerASCIISubtractTab(dbg_viewer,2*(pcbddc->current_level+1));CHKERRQ(ierr); 54276e683305SStefano Zampini } 5428c8587f34SStefano Zampini ierr = KSPDestroy(&check_ksp);CHKERRQ(ierr); 54296e683305SStefano Zampini if (compute_eigs) { 54306e683305SStefano Zampini ierr = PetscFree(eigs_r);CHKERRQ(ierr); 54316e683305SStefano Zampini ierr = PetscFree(eigs_c);CHKERRQ(ierr); 5432c8587f34SStefano Zampini } 54336e683305SStefano Zampini } 54346e683305SStefano Zampini } 5435cbcc2c2aSStefano Zampini /* print additional info */ 5436cbcc2c2aSStefano Zampini if (pcbddc->dbg_flag) { 54376e683305SStefano Zampini /* waits until all processes reaches this point */ 54386e683305SStefano Zampini ierr = PetscBarrier((PetscObject)pc);CHKERRQ(ierr); 5439cbcc2c2aSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Coarse solver setup completed at level %d\n",pcbddc->current_level);CHKERRQ(ierr); 5440cbcc2c2aSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 5441cbcc2c2aSStefano Zampini } 5442cbcc2c2aSStefano Zampini 54432b510759SStefano Zampini /* free memory */ 5444c8587f34SStefano Zampini ierr = MatNullSpaceDestroy(&CoarseNullSpace);CHKERRQ(ierr); 5445fdc635d7SStefano Zampini ierr = MatDestroy(&coarse_mat);CHKERRQ(ierr); 5446c8587f34SStefano Zampini PetscFunctionReturn(0); 5447c8587f34SStefano Zampini } 5448674ae819SStefano Zampini 5449f34684f1SStefano Zampini #undef __FUNCT__ 5450f34684f1SStefano Zampini #define __FUNCT__ "PCBDDCComputePrimalNumbering" 5451f34684f1SStefano Zampini PetscErrorCode PCBDDCComputePrimalNumbering(PC pc,PetscInt* coarse_size_n,PetscInt** local_primal_indices_n) 5452f34684f1SStefano Zampini { 5453f34684f1SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 5454f34684f1SStefano Zampini PC_IS* pcis = (PC_IS*)pc->data; 5455f34684f1SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 5456dc456d91SStefano Zampini IS subset,subset_mult,subset_n; 5457dc456d91SStefano Zampini PetscInt local_size,coarse_size=0; 545873be2a3aSStefano Zampini PetscInt *local_primal_indices=NULL; 5459dc456d91SStefano Zampini const PetscInt *t_local_primal_indices; 5460f34684f1SStefano Zampini PetscErrorCode ierr; 5461f34684f1SStefano Zampini 5462f34684f1SStefano Zampini PetscFunctionBegin; 5463f34684f1SStefano Zampini /* Compute global number of coarse dofs */ 54640e6343abSStefano Zampini if (pcbddc->local_primal_size && !pcbddc->local_primal_ref_node) { 54650e6343abSStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"BDDC ConstraintsSetUp should be called first"); 5466727cdba6SStefano Zampini } 5467dc456d91SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)(pc->pmat)),pcbddc->local_primal_size_cc,pcbddc->local_primal_ref_node,PETSC_COPY_VALUES,&subset_n);CHKERRQ(ierr); 54683bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(pcis->mapping,subset_n,&subset);CHKERRQ(ierr); 5469dc456d91SStefano Zampini ierr = ISDestroy(&subset_n);CHKERRQ(ierr); 5470dc456d91SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)(pc->pmat)),pcbddc->local_primal_size_cc,pcbddc->local_primal_ref_mult,PETSC_COPY_VALUES,&subset_mult);CHKERRQ(ierr); 5471dc456d91SStefano Zampini ierr = PCBDDCSubsetNumbering(subset,subset_mult,&coarse_size,&subset_n);CHKERRQ(ierr); 5472dc456d91SStefano Zampini ierr = ISDestroy(&subset);CHKERRQ(ierr); 5473dc456d91SStefano Zampini ierr = ISDestroy(&subset_mult);CHKERRQ(ierr); 5474dc456d91SStefano Zampini ierr = ISGetLocalSize(subset_n,&local_size);CHKERRQ(ierr); 54750e6343abSStefano Zampini if (local_size != pcbddc->local_primal_size) { 54760e6343abSStefano Zampini SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Invalid number of local primal indices computed %d != %d",local_size,pcbddc->local_primal_size); 54770e6343abSStefano Zampini } 5478dc456d91SStefano Zampini ierr = PetscMalloc1(local_size,&local_primal_indices);CHKERRQ(ierr); 5479dc456d91SStefano Zampini ierr = ISGetIndices(subset_n,&t_local_primal_indices);CHKERRQ(ierr); 5480dc456d91SStefano Zampini ierr = PetscMemcpy(local_primal_indices,t_local_primal_indices,local_size*sizeof(PetscInt));CHKERRQ(ierr); 5481dc456d91SStefano Zampini ierr = ISRestoreIndices(subset_n,&t_local_primal_indices);CHKERRQ(ierr); 5482dc456d91SStefano Zampini ierr = ISDestroy(&subset_n);CHKERRQ(ierr); 5483f34684f1SStefano Zampini 5484f34684f1SStefano Zampini /* check numbering */ 5485f34684f1SStefano Zampini if (pcbddc->dbg_flag) { 5486019a44ceSStefano Zampini PetscScalar coarsesum,*array,*array2; 5487dc456d91SStefano Zampini PetscInt i; 5488b9b85e73SStefano Zampini PetscBool set_error = PETSC_FALSE,set_error_reduced = PETSC_FALSE; 5489f34684f1SStefano Zampini 5490f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 5491f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 5492f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Check coarse indices\n");CHKERRQ(ierr); 54930fccc4e9SStefano Zampini ierr = PetscViewerASCIISynchronizedAllow(pcbddc->dbg_viewer,PETSC_TRUE);CHKERRQ(ierr); 5494019a44ceSStefano Zampini /* counter */ 5495019a44ceSStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 5496019a44ceSStefano Zampini ierr = VecSet(pcis->vec1_N,1.0);CHKERRQ(ierr); 5497019a44ceSStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 5498019a44ceSStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 5499019a44ceSStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_global,pcis->vec2_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 5500019a44ceSStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_global,pcis->vec2_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 5501019a44ceSStefano Zampini 5502f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_N,0.0);CHKERRQ(ierr); 5503f34684f1SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 5504727cdba6SStefano Zampini ierr = VecSetValue(pcis->vec1_N,pcbddc->primal_indices_local_idxs[i],1.0,INSERT_VALUES);CHKERRQ(ierr); 5505f34684f1SStefano Zampini } 5506f34684f1SStefano Zampini ierr = VecAssemblyBegin(pcis->vec1_N);CHKERRQ(ierr); 5507f34684f1SStefano Zampini ierr = VecAssemblyEnd(pcis->vec1_N);CHKERRQ(ierr); 5508f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 5509e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 5510e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 5511e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_global,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 5512e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_global,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 5513f34684f1SStefano Zampini ierr = VecGetArray(pcis->vec1_N,&array);CHKERRQ(ierr); 5514019a44ceSStefano Zampini ierr = VecGetArray(pcis->vec2_N,&array2);CHKERRQ(ierr); 5515f34684f1SStefano Zampini for (i=0;i<pcis->n;i++) { 5516019a44ceSStefano Zampini if (array[i] != 0.0 && array[i] != array2[i]) { 551775c01103SStefano Zampini PetscInt owned = (PetscInt)PetscRealPart(array[i]); 551875c01103SStefano Zampini PetscInt neigh = (PetscInt)PetscRealPart(array2[i]); 5519b9b85e73SStefano Zampini set_error = PETSC_TRUE; 5520019a44ceSStefano 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); 5521f34684f1SStefano Zampini } 5522f34684f1SStefano Zampini } 5523019a44ceSStefano Zampini ierr = VecRestoreArray(pcis->vec2_N,&array2);CHKERRQ(ierr); 5524b9b85e73SStefano Zampini ierr = MPI_Allreduce(&set_error,&set_error_reduced,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 5525f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 5526f34684f1SStefano Zampini for (i=0;i<pcis->n;i++) { 5527f34684f1SStefano Zampini if (PetscRealPart(array[i]) > 0.0) array[i] = 1.0/PetscRealPart(array[i]); 5528f34684f1SStefano Zampini } 5529f34684f1SStefano Zampini ierr = VecRestoreArray(pcis->vec1_N,&array);CHKERRQ(ierr); 5530f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 5531e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 5532e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 5533f34684f1SStefano Zampini ierr = VecSum(pcis->vec1_global,&coarsesum);CHKERRQ(ierr); 5534f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Size of coarse problem is %d (%lf)\n",coarse_size,PetscRealPart(coarsesum));CHKERRQ(ierr); 5535b9b85e73SStefano Zampini if (pcbddc->dbg_flag > 1 || set_error_reduced) { 5536ca8b9ea9SStefano Zampini PetscInt *gidxs; 5537ca8b9ea9SStefano Zampini 5538ca8b9ea9SStefano Zampini ierr = PetscMalloc1(pcbddc->local_primal_size,&gidxs);CHKERRQ(ierr); 55393bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingApply(pcis->mapping,pcbddc->local_primal_size,pcbddc->primal_indices_local_idxs,gidxs);CHKERRQ(ierr); 5540f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Distribution of local primal indices\n");CHKERRQ(ierr); 5541f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 5542f34684f1SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d\n",PetscGlobalRank);CHKERRQ(ierr); 5543f34684f1SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 55444bc2dc4bSStefano 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); 5545f34684f1SStefano Zampini } 5546f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 5547ca8b9ea9SStefano Zampini ierr = PetscFree(gidxs);CHKERRQ(ierr); 5548f34684f1SStefano Zampini } 5549f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 5550302440fdSBarry Smith if (set_error_reduced) SETERRQ(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"BDDC Numbering of coarse dofs failed"); 5551f34684f1SStefano Zampini } 55528bec7fa6SStefano Zampini /* ierr = PetscPrintf(PetscObjectComm((PetscObject)pc),"Size of coarse problem is %d\n",coarse_size);CHKERRQ(ierr); */ 5553f34684f1SStefano Zampini /* get back data */ 5554f34684f1SStefano Zampini *coarse_size_n = coarse_size; 5555f34684f1SStefano Zampini *local_primal_indices_n = local_primal_indices; 5556674ae819SStefano Zampini PetscFunctionReturn(0); 5557674ae819SStefano Zampini } 5558674ae819SStefano Zampini 5559e456f2a8SStefano Zampini #undef __FUNCT__ 5560e456f2a8SStefano Zampini #define __FUNCT__ "PCBDDCGlobalToLocal" 5561a7dc3881SStefano Zampini PetscErrorCode PCBDDCGlobalToLocal(VecScatter g2l_ctx,Vec gwork, Vec lwork, IS globalis, IS* localis) 5562e456f2a8SStefano Zampini { 5563e456f2a8SStefano Zampini IS localis_t; 5564a7dc3881SStefano Zampini PetscInt i,lsize,*idxs,n; 5565e456f2a8SStefano Zampini PetscScalar *vals; 5566e456f2a8SStefano Zampini PetscErrorCode ierr; 5567e456f2a8SStefano Zampini 5568e456f2a8SStefano Zampini PetscFunctionBegin; 5569a7dc3881SStefano Zampini /* get indices in local ordering exploiting local to global map */ 5570e456f2a8SStefano Zampini ierr = ISGetLocalSize(globalis,&lsize);CHKERRQ(ierr); 5571854ce69bSBarry Smith ierr = PetscMalloc1(lsize,&vals);CHKERRQ(ierr); 5572e456f2a8SStefano Zampini for (i=0;i<lsize;i++) vals[i] = 1.0; 5573e456f2a8SStefano Zampini ierr = ISGetIndices(globalis,(const PetscInt**)&idxs);CHKERRQ(ierr); 5574a7dc3881SStefano Zampini ierr = VecSet(gwork,0.0);CHKERRQ(ierr); 5575a7dc3881SStefano Zampini ierr = VecSet(lwork,0.0);CHKERRQ(ierr); 55761035eff8SStefano Zampini if (idxs) { /* multilevel guard */ 5577a7dc3881SStefano Zampini ierr = VecSetValues(gwork,lsize,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 55781035eff8SStefano Zampini } 5579a7dc3881SStefano Zampini ierr = VecAssemblyBegin(gwork);CHKERRQ(ierr); 5580e456f2a8SStefano Zampini ierr = ISRestoreIndices(globalis,(const PetscInt**)&idxs);CHKERRQ(ierr); 5581e456f2a8SStefano Zampini ierr = PetscFree(vals);CHKERRQ(ierr); 5582a7dc3881SStefano Zampini ierr = VecAssemblyEnd(gwork);CHKERRQ(ierr); 5583a7dc3881SStefano Zampini /* now compute set in local ordering */ 5584a7dc3881SStefano Zampini ierr = VecScatterBegin(g2l_ctx,gwork,lwork,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 5585a7dc3881SStefano Zampini ierr = VecScatterEnd(g2l_ctx,gwork,lwork,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 5586a7dc3881SStefano Zampini ierr = VecGetArrayRead(lwork,(const PetscScalar**)&vals);CHKERRQ(ierr); 5587a7dc3881SStefano Zampini ierr = VecGetSize(lwork,&n);CHKERRQ(ierr); 5588a7dc3881SStefano Zampini for (i=0,lsize=0;i<n;i++) { 5589ff92baa0SMatthew G. Knepley if (PetscRealPart(vals[i]) > 0.5) { 5590e456f2a8SStefano Zampini lsize++; 5591e456f2a8SStefano Zampini } 5592e456f2a8SStefano Zampini } 5593854ce69bSBarry Smith ierr = PetscMalloc1(lsize,&idxs);CHKERRQ(ierr); 5594a7dc3881SStefano Zampini for (i=0,lsize=0;i<n;i++) { 5595ff92baa0SMatthew G. Knepley if (PetscRealPart(vals[i]) > 0.5) { 5596e456f2a8SStefano Zampini idxs[lsize++] = i; 5597e456f2a8SStefano Zampini } 5598e456f2a8SStefano Zampini } 5599a7dc3881SStefano Zampini ierr = VecRestoreArrayRead(lwork,(const PetscScalar**)&vals);CHKERRQ(ierr); 5600a7dc3881SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)gwork),lsize,idxs,PETSC_OWN_POINTER,&localis_t);CHKERRQ(ierr); 5601e456f2a8SStefano Zampini *localis = localis_t; 5602e456f2a8SStefano Zampini PetscFunctionReturn(0); 5603e456f2a8SStefano Zampini } 5604906d46d4SStefano Zampini 5605906d46d4SStefano Zampini /* the next two functions will be called in KSPMatMult if a change of basis has been requested */ 5606906d46d4SStefano Zampini #undef __FUNCT__ 5607906d46d4SStefano Zampini #define __FUNCT__ "PCBDDCMatMult_Private" 5608906d46d4SStefano Zampini static PetscErrorCode PCBDDCMatMult_Private(Mat A, Vec x, Vec y) 5609906d46d4SStefano Zampini { 5610906d46d4SStefano Zampini PCBDDCChange_ctx change_ctx; 5611906d46d4SStefano Zampini PetscErrorCode ierr; 5612906d46d4SStefano Zampini 5613906d46d4SStefano Zampini PetscFunctionBegin; 5614906d46d4SStefano Zampini ierr = MatShellGetContext(A,&change_ctx);CHKERRQ(ierr); 5615906d46d4SStefano Zampini ierr = MatMult(change_ctx->global_change,x,change_ctx->work[0]);CHKERRQ(ierr); 5616906d46d4SStefano Zampini ierr = MatMult(change_ctx->original_mat,change_ctx->work[0],change_ctx->work[1]);CHKERRQ(ierr); 5617906d46d4SStefano Zampini ierr = MatMultTranspose(change_ctx->global_change,change_ctx->work[1],y);CHKERRQ(ierr); 5618906d46d4SStefano Zampini PetscFunctionReturn(0); 5619906d46d4SStefano Zampini } 5620906d46d4SStefano Zampini 5621906d46d4SStefano Zampini #undef __FUNCT__ 5622906d46d4SStefano Zampini #define __FUNCT__ "PCBDDCMatMultTranspose_Private" 5623906d46d4SStefano Zampini static PetscErrorCode PCBDDCMatMultTranspose_Private(Mat A, Vec x, Vec y) 5624906d46d4SStefano Zampini { 5625906d46d4SStefano Zampini PCBDDCChange_ctx change_ctx; 5626906d46d4SStefano Zampini PetscErrorCode ierr; 5627906d46d4SStefano Zampini 5628906d46d4SStefano Zampini PetscFunctionBegin; 5629906d46d4SStefano Zampini ierr = MatShellGetContext(A,&change_ctx);CHKERRQ(ierr); 5630906d46d4SStefano Zampini ierr = MatMult(change_ctx->global_change,x,change_ctx->work[0]);CHKERRQ(ierr); 5631906d46d4SStefano Zampini ierr = MatMultTranspose(change_ctx->original_mat,change_ctx->work[0],change_ctx->work[1]);CHKERRQ(ierr); 5632906d46d4SStefano Zampini ierr = MatMultTranspose(change_ctx->global_change,change_ctx->work[1],y);CHKERRQ(ierr); 5633906d46d4SStefano Zampini PetscFunctionReturn(0); 5634906d46d4SStefano Zampini } 5635b96c3477SStefano Zampini 5636b96c3477SStefano Zampini #undef __FUNCT__ 5637b96c3477SStefano Zampini #define __FUNCT__ "PCBDDCSetUpSubSchurs" 563808122e43SStefano Zampini PetscErrorCode PCBDDCSetUpSubSchurs(PC pc) 5639b96c3477SStefano Zampini { 5640a64f4aa4SStefano Zampini PC_IS *pcis=(PC_IS*)pc->data; 5641b96c3477SStefano Zampini PC_BDDC *pcbddc=(PC_BDDC*)pc->data; 5642b96c3477SStefano Zampini PCBDDCSubSchurs sub_schurs=pcbddc->sub_schurs; 5643a64f4aa4SStefano Zampini Mat S_j; 5644b96c3477SStefano Zampini PetscInt *used_xadj,*used_adjncy; 5645b96c3477SStefano Zampini PetscBool free_used_adj; 5646b96c3477SStefano Zampini PetscErrorCode ierr; 5647b96c3477SStefano Zampini 5648b96c3477SStefano Zampini PetscFunctionBegin; 5649b96c3477SStefano Zampini /* decide the adjacency to be used for determining internal problems for local schur on subsets */ 5650b96c3477SStefano Zampini free_used_adj = PETSC_FALSE; 565108122e43SStefano Zampini if (pcbddc->sub_schurs_layers == -1) { 5652b96c3477SStefano Zampini used_xadj = NULL; 5653b96c3477SStefano Zampini used_adjncy = NULL; 5654b96c3477SStefano Zampini } else { 565508122e43SStefano Zampini if (pcbddc->sub_schurs_use_useradj && pcbddc->mat_graph->xadj) { 565608122e43SStefano Zampini used_xadj = pcbddc->mat_graph->xadj; 565708122e43SStefano Zampini used_adjncy = pcbddc->mat_graph->adjncy; 565808122e43SStefano Zampini } else if (pcbddc->computed_rowadj) { 5659b96c3477SStefano Zampini used_xadj = pcbddc->mat_graph->xadj; 5660b96c3477SStefano Zampini used_adjncy = pcbddc->mat_graph->adjncy; 5661b96c3477SStefano Zampini } else { 56622fffb893SStefano Zampini PetscBool flg_row=PETSC_FALSE; 5663b96c3477SStefano Zampini const PetscInt *xadj,*adjncy; 5664b96c3477SStefano Zampini PetscInt nvtxs; 5665b96c3477SStefano Zampini 56662fffb893SStefano Zampini ierr = MatGetRowIJ(pcbddc->local_mat,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 56672fffb893SStefano Zampini if (flg_row) { 5668b96c3477SStefano Zampini ierr = PetscMalloc2(nvtxs+1,&used_xadj,xadj[nvtxs],&used_adjncy);CHKERRQ(ierr); 5669b96c3477SStefano Zampini ierr = PetscMemcpy(used_xadj,xadj,(nvtxs+1)*sizeof(*xadj));CHKERRQ(ierr); 5670b96c3477SStefano Zampini ierr = PetscMemcpy(used_adjncy,adjncy,(xadj[nvtxs])*sizeof(*adjncy));CHKERRQ(ierr); 5671b96c3477SStefano Zampini free_used_adj = PETSC_TRUE; 56722fffb893SStefano Zampini } else { 56732fffb893SStefano Zampini pcbddc->sub_schurs_layers = -1; 56742fffb893SStefano Zampini used_xadj = NULL; 56752fffb893SStefano Zampini used_adjncy = NULL; 56762fffb893SStefano Zampini } 56772fffb893SStefano Zampini ierr = MatRestoreRowIJ(pcbddc->local_mat,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 5678b96c3477SStefano Zampini } 5679b96c3477SStefano Zampini } 5680d5574798SStefano Zampini 5681d5574798SStefano Zampini /* setup sub_schurs data */ 5682a64f4aa4SStefano Zampini ierr = MatCreateSchurComplement(pcis->A_II,pcis->A_II,pcis->A_IB,pcis->A_BI,pcis->A_BB,&S_j);CHKERRQ(ierr); 5683a64f4aa4SStefano Zampini if (!sub_schurs->use_mumps) { 5684a64f4aa4SStefano Zampini /* pcbddc->ksp_D up to date only if not using MUMPS */ 5685a64f4aa4SStefano Zampini ierr = MatSchurComplementSetKSP(S_j,pcbddc->ksp_D);CHKERRQ(ierr); 5686*683d3df6SStefano 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); 5687a64f4aa4SStefano Zampini } else { 56886816873aSStefano Zampini PetscBool reuse_solvers = (PetscBool)!pcbddc->use_change_of_basis; 568904708bb6SStefano Zampini PetscBool isseqaij; 56905feab87aSStefano Zampini if (!pcbddc->use_vertices && reuse_solvers) { 56915feab87aSStefano Zampini PetscInt n_vertices; 56925feab87aSStefano Zampini 56935feab87aSStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_vertices,&n_vertices);CHKERRQ(ierr); 56942034aafcSStefano Zampini reuse_solvers = (PetscBool)!n_vertices; 56955feab87aSStefano Zampini } 569604708bb6SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcbddc->local_mat,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 569704708bb6SStefano Zampini if (!isseqaij) { 569804708bb6SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 569904708bb6SStefano Zampini if (matis->A == pcbddc->local_mat) { 570004708bb6SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 570104708bb6SStefano Zampini ierr = MatConvert(matis->A,MATSEQAIJ,MAT_INITIAL_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 570204708bb6SStefano Zampini } else { 570304708bb6SStefano Zampini ierr = MatConvert(pcbddc->local_mat,MATSEQAIJ,MAT_REUSE_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 570404708bb6SStefano Zampini } 570504708bb6SStefano Zampini } 5706*683d3df6SStefano 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); 5707a64f4aa4SStefano Zampini } 5708a64f4aa4SStefano Zampini ierr = MatDestroy(&S_j);CHKERRQ(ierr); 5709b96c3477SStefano Zampini 5710b96c3477SStefano Zampini /* free adjacency */ 5711b96c3477SStefano Zampini if (free_used_adj) { 5712b96c3477SStefano Zampini ierr = PetscFree2(used_xadj,used_adjncy);CHKERRQ(ierr); 5713b96c3477SStefano Zampini } 5714b96c3477SStefano Zampini PetscFunctionReturn(0); 5715b96c3477SStefano Zampini } 5716b96c3477SStefano Zampini 5717b96c3477SStefano Zampini #undef __FUNCT__ 5718b96c3477SStefano Zampini #define __FUNCT__ "PCBDDCInitSubSchurs" 571908122e43SStefano Zampini PetscErrorCode PCBDDCInitSubSchurs(PC pc) 5720b96c3477SStefano Zampini { 5721b96c3477SStefano Zampini PC_IS *pcis=(PC_IS*)pc->data; 5722b96c3477SStefano Zampini PC_BDDC *pcbddc=(PC_BDDC*)pc->data; 5723b96c3477SStefano Zampini PCBDDCSubSchurs sub_schurs=pcbddc->sub_schurs; 5724b96c3477SStefano Zampini PCBDDCGraph graph; 5725b96c3477SStefano Zampini PetscErrorCode ierr; 5726b96c3477SStefano Zampini 5727b96c3477SStefano Zampini PetscFunctionBegin; 5728b96c3477SStefano Zampini /* attach interface graph for determining subsets */ 572908122e43SStefano Zampini if (pcbddc->sub_schurs_rebuild) { /* in case rebuild has been requested, it uses a graph generated only by the neighbouring information */ 57303301b35fSStefano Zampini IS verticesIS,verticescomm; 57313301b35fSStefano Zampini PetscInt vsize,*idxs; 5732b96c3477SStefano Zampini 5733b96c3477SStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,NULL,NULL,NULL,NULL,&verticesIS);CHKERRQ(ierr); 57343301b35fSStefano Zampini ierr = ISGetSize(verticesIS,&vsize);CHKERRQ(ierr); 57353301b35fSStefano Zampini ierr = ISGetIndices(verticesIS,(const PetscInt**)&idxs);CHKERRQ(ierr); 57363301b35fSStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),vsize,idxs,PETSC_COPY_VALUES,&verticescomm);CHKERRQ(ierr); 57373301b35fSStefano Zampini ierr = ISRestoreIndices(verticesIS,(const PetscInt**)&idxs);CHKERRQ(ierr); 57383301b35fSStefano Zampini ierr = ISDestroy(&verticesIS);CHKERRQ(ierr); 5739b96c3477SStefano Zampini ierr = PCBDDCGraphCreate(&graph);CHKERRQ(ierr); 57407fb0e2dbSStefano Zampini ierr = PCBDDCGraphInit(graph,pcbddc->mat_graph->l2gmap,pcbddc->mat_graph->nvtxs_global);CHKERRQ(ierr); 57413301b35fSStefano Zampini ierr = PCBDDCGraphSetUp(graph,0,NULL,pcbddc->DirichletBoundariesLocal,0,NULL,verticescomm);CHKERRQ(ierr); 57423301b35fSStefano Zampini ierr = ISDestroy(&verticescomm);CHKERRQ(ierr); 5743b96c3477SStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(graph);CHKERRQ(ierr); 5744b96c3477SStefano Zampini /* 5745b96c3477SStefano Zampini if (pcbddc->dbg_flag) { 5746b96c3477SStefano Zampini ierr = PCBDDCGraphASCIIView(graph,pcbddc->dbg_flag,pcbddc->dbg_viewer);CHKERRQ(ierr); 5747b96c3477SStefano Zampini } 5748b96c3477SStefano Zampini */ 5749b96c3477SStefano Zampini } else { 5750b96c3477SStefano Zampini graph = pcbddc->mat_graph; 5751b96c3477SStefano Zampini } 5752b96c3477SStefano Zampini 5753b96c3477SStefano Zampini /* sub_schurs init */ 5754a64f4aa4SStefano Zampini ierr = PCBDDCSubSchursInit(sub_schurs,pcis->is_I_local,pcis->is_B_local,graph,pcis->BtoNmap);CHKERRQ(ierr); 5755a64f4aa4SStefano Zampini 5756b96c3477SStefano Zampini /* free graph struct */ 575708122e43SStefano Zampini if (pcbddc->sub_schurs_rebuild) { 5758b96c3477SStefano Zampini ierr = PCBDDCGraphDestroy(&graph);CHKERRQ(ierr); 5759b96c3477SStefano Zampini } 5760b96c3477SStefano Zampini PetscFunctionReturn(0); 5761b96c3477SStefano Zampini } 5762fa34dd3eSStefano Zampini 5763fa34dd3eSStefano Zampini #undef __FUNCT__ 5764fa34dd3eSStefano Zampini #define __FUNCT__ "PCBDDCCheckOperator" 5765fa34dd3eSStefano Zampini PetscErrorCode PCBDDCCheckOperator(PC pc) 5766fa34dd3eSStefano Zampini { 5767fa34dd3eSStefano Zampini PC_IS *pcis=(PC_IS*)pc->data; 5768fa34dd3eSStefano Zampini PC_BDDC *pcbddc=(PC_BDDC*)pc->data; 5769fa34dd3eSStefano Zampini PetscErrorCode ierr; 5770fa34dd3eSStefano Zampini 5771fa34dd3eSStefano Zampini PetscFunctionBegin; 5772fa34dd3eSStefano Zampini if (pcbddc->n_vertices == pcbddc->local_primal_size) { 5773fa34dd3eSStefano Zampini IS zerodiag = NULL; 57744f1b2e48SStefano Zampini Mat S_j,B0_B=NULL; 5775fa34dd3eSStefano Zampini Vec dummy_vec=NULL,vec_check_B,vec_scale_P; 57764f1b2e48SStefano Zampini PetscScalar *p0_check,*array,*array2; 577775c01103SStefano Zampini PetscReal norm; 5778fa34dd3eSStefano Zampini PetscInt i; 5779fa34dd3eSStefano Zampini 5780fa34dd3eSStefano Zampini /* B0 and B0_B */ 5781fa34dd3eSStefano Zampini if (zerodiag) { 5782fa34dd3eSStefano Zampini IS dummy; 5783fa34dd3eSStefano Zampini 57844f1b2e48SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcbddc->benign_n,0,1,&dummy);CHKERRQ(ierr); 57854f1b2e48SStefano Zampini ierr = MatGetSubMatrix(pcbddc->benign_B0,dummy,pcis->is_B_local,MAT_INITIAL_MATRIX,&B0_B);CHKERRQ(ierr); 5786fa34dd3eSStefano Zampini ierr = MatCreateVecs(B0_B,NULL,&dummy_vec);CHKERRQ(ierr); 5787fa34dd3eSStefano Zampini ierr = ISDestroy(&dummy);CHKERRQ(ierr); 5788fa34dd3eSStefano Zampini } 5789fa34dd3eSStefano Zampini /* I need a primal vector to scale primal nodes since BDDC sums contibutions */ 5790fa34dd3eSStefano Zampini ierr = VecDuplicate(pcbddc->vec1_P,&vec_scale_P);CHKERRQ(ierr); 5791fa34dd3eSStefano Zampini ierr = VecSet(pcbddc->vec1_P,1.0);CHKERRQ(ierr); 5792fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 5793fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 5794fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,vec_scale_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 5795fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,vec_scale_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 5796fa34dd3eSStefano Zampini ierr = VecReciprocal(vec_scale_P);CHKERRQ(ierr); 5797fa34dd3eSStefano Zampini /* S_j */ 5798fa34dd3eSStefano Zampini ierr = MatCreateSchurComplement(pcis->A_II,pcis->A_II,pcis->A_IB,pcis->A_BI,pcis->A_BB,&S_j);CHKERRQ(ierr); 5799fa34dd3eSStefano Zampini ierr = MatSchurComplementSetKSP(S_j,pcbddc->ksp_D);CHKERRQ(ierr); 5800fa34dd3eSStefano Zampini 5801fa34dd3eSStefano Zampini /* mimic vector in \widetilde{W}_\Gamma */ 5802fa34dd3eSStefano Zampini ierr = VecSetRandom(pcis->vec1_N,NULL);CHKERRQ(ierr); 5803fa34dd3eSStefano Zampini /* continuous in primal space */ 5804fa34dd3eSStefano Zampini ierr = VecSetRandom(pcbddc->coarse_vec,NULL);CHKERRQ(ierr); 5805fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,pcbddc->vec1_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 5806fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,pcbddc->vec1_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 5807fa34dd3eSStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 58084f1b2e48SStefano Zampini ierr = PetscCalloc1(pcbddc->benign_n,&p0_check);CHKERRQ(ierr); 58094f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) p0_check[i] = array[pcbddc->local_primal_size-pcbddc->benign_n+i]; 5810fa34dd3eSStefano Zampini ierr = VecSetValues(pcis->vec1_N,pcbddc->local_primal_size,pcbddc->local_primal_ref_node,array,INSERT_VALUES);CHKERRQ(ierr); 5811fa34dd3eSStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 5812fa34dd3eSStefano Zampini ierr = VecAssemblyBegin(pcis->vec1_N);CHKERRQ(ierr); 5813fa34dd3eSStefano Zampini ierr = VecAssemblyEnd(pcis->vec1_N);CHKERRQ(ierr); 5814fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcis->N_to_B,pcis->vec1_N,pcis->vec2_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 5815fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcis->N_to_B,pcis->vec1_N,pcis->vec2_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 5816fa34dd3eSStefano Zampini ierr = VecDuplicate(pcis->vec2_B,&vec_check_B);CHKERRQ(ierr); 5817fa34dd3eSStefano Zampini ierr = VecCopy(pcis->vec2_B,vec_check_B);CHKERRQ(ierr); 5818fa34dd3eSStefano Zampini 5819fa34dd3eSStefano Zampini /* assemble rhs for coarse problem */ 5820fa34dd3eSStefano Zampini /* widetilde{S}_\Gamma w_\Gamma + \widetilde{B0}^T_B p0 */ 5821fa34dd3eSStefano Zampini /* local with Schur */ 5822fa34dd3eSStefano Zampini ierr = MatMult(S_j,pcis->vec2_B,pcis->vec1_B);CHKERRQ(ierr); 5823fa34dd3eSStefano Zampini if (zerodiag) { 5824fa34dd3eSStefano Zampini ierr = VecGetArray(dummy_vec,&array);CHKERRQ(ierr); 58254f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) array[i] = p0_check[i]; 5826fa34dd3eSStefano Zampini ierr = VecRestoreArray(dummy_vec,&array);CHKERRQ(ierr); 5827fa34dd3eSStefano Zampini ierr = MatMultTransposeAdd(B0_B,dummy_vec,pcis->vec1_B,pcis->vec1_B);CHKERRQ(ierr); 5828fa34dd3eSStefano Zampini } 5829fa34dd3eSStefano Zampini /* sum on primal nodes the local contributions */ 5830fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcis->N_to_B,pcis->vec1_B,pcis->vec1_N,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 5831fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcis->N_to_B,pcis->vec1_B,pcis->vec1_N,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 5832fa34dd3eSStefano Zampini ierr = VecGetArray(pcis->vec1_N,&array);CHKERRQ(ierr); 5833fa34dd3eSStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array2);CHKERRQ(ierr); 5834fa34dd3eSStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) array2[i] = array[pcbddc->local_primal_ref_node[i]]; 5835fa34dd3eSStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array2);CHKERRQ(ierr); 5836fa34dd3eSStefano Zampini ierr = VecRestoreArray(pcis->vec1_N,&array);CHKERRQ(ierr); 5837fa34dd3eSStefano Zampini ierr = VecSet(pcbddc->coarse_vec,0.);CHKERRQ(ierr); 5838fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 5839fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 5840fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,pcbddc->vec1_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 5841fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,pcbddc->vec1_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 5842fa34dd3eSStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 5843fa34dd3eSStefano Zampini /* scale primal nodes (BDDC sums contibutions) */ 5844fa34dd3eSStefano Zampini ierr = VecPointwiseMult(pcbddc->vec1_P,vec_scale_P,pcbddc->vec1_P);CHKERRQ(ierr); 5845fa34dd3eSStefano Zampini ierr = VecSetValues(pcis->vec1_N,pcbddc->local_primal_size,pcbddc->local_primal_ref_node,array,INSERT_VALUES);CHKERRQ(ierr); 5846fa34dd3eSStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 5847fa34dd3eSStefano Zampini ierr = VecAssemblyBegin(pcis->vec1_N);CHKERRQ(ierr); 5848fa34dd3eSStefano Zampini ierr = VecAssemblyEnd(pcis->vec1_N);CHKERRQ(ierr); 5849fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcis->N_to_B,pcis->vec1_N,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 5850fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcis->N_to_B,pcis->vec1_N,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 5851fa34dd3eSStefano Zampini /* global: \widetilde{B0}_B w_\Gamma */ 5852fa34dd3eSStefano Zampini if (zerodiag) { 5853fa34dd3eSStefano Zampini ierr = MatMult(B0_B,pcis->vec2_B,dummy_vec);CHKERRQ(ierr); 5854fa34dd3eSStefano Zampini ierr = VecGetArray(dummy_vec,&array);CHKERRQ(ierr); 58554f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) pcbddc->benign_p0[i] = array[i]; 5856fa34dd3eSStefano Zampini ierr = VecRestoreArray(dummy_vec,&array);CHKERRQ(ierr); 5857fa34dd3eSStefano Zampini } 5858fa34dd3eSStefano Zampini /* BDDC */ 5859fa34dd3eSStefano Zampini ierr = VecSet(pcis->vec1_D,0.);CHKERRQ(ierr); 5860fa34dd3eSStefano Zampini ierr = PCBDDCApplyInterfacePreconditioner(pc,PETSC_FALSE);CHKERRQ(ierr); 5861fa34dd3eSStefano Zampini 5862fa34dd3eSStefano Zampini ierr = VecCopy(pcis->vec1_B,pcis->vec2_B);CHKERRQ(ierr); 5863fa34dd3eSStefano Zampini ierr = VecAXPY(pcis->vec1_B,-1.0,vec_check_B);CHKERRQ(ierr); 5864fa34dd3eSStefano Zampini ierr = VecNorm(pcis->vec1_B,NORM_INFINITY,&norm);CHKERRQ(ierr); 5865fa34dd3eSStefano Zampini PetscPrintf(PETSC_COMM_SELF,"[%d] BDDC local error is %1.4e\n",PetscGlobalRank,norm); 58664f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 58674f1b2e48SStefano Zampini PetscPrintf(PETSC_COMM_SELF,"[%d] BDDC p0[%d] error is %1.4e\n",PetscGlobalRank,i,PetscAbsScalar(pcbddc->benign_p0[i]-p0_check[i])); 5868fa34dd3eSStefano Zampini } 58694f1b2e48SStefano Zampini ierr = PetscFree(p0_check);CHKERRQ(ierr); 5870fa34dd3eSStefano Zampini ierr = VecDestroy(&vec_scale_P);CHKERRQ(ierr); 5871fa34dd3eSStefano Zampini ierr = VecDestroy(&vec_check_B);CHKERRQ(ierr); 5872fa34dd3eSStefano Zampini ierr = VecDestroy(&dummy_vec);CHKERRQ(ierr); 5873fa34dd3eSStefano Zampini ierr = MatDestroy(&S_j);CHKERRQ(ierr); 5874fa34dd3eSStefano Zampini ierr = MatDestroy(&B0_B);CHKERRQ(ierr); 5875fa34dd3eSStefano Zampini } 5876fa34dd3eSStefano Zampini PetscFunctionReturn(0); 5877fa34dd3eSStefano Zampini } 5878