1 /* 2 * overlapsplit.c: increase the overlap of a 'big' subdomain across several processor cores 3 * 4 * Author: Fande Kong <fdkong.jd@gmail.com> 5 */ 6 7 #include <petscsf.h> 8 #include <petsc/private/matimpl.h> 9 10 11 #undef __FUNCT__ 12 #define __FUNCT__ "MatIncreaseOverlapSplit_Single" 13 14 /* 15 * Increase overlap for the sub-matrix across sub communicator 16 * sub-matrix could be a graph or numerical matrix 17 * */ 18 PetscErrorCode MatIncreaseOverlapSplit_Single(Mat mat,IS *is,PetscInt ov) 19 { 20 PetscInt i,nindx,*indices_sc,*indices_ov,localsize,*localsizes_sc,localsize_tmp; 21 PetscInt *indices_ov_rd,nroots,nleaves,*localoffsets,*indices_recv,*sources_sc,*sources_sc_rd; 22 const PetscInt *indices; 23 PetscMPIInt srank,ssize,issamecomm,k,grank; 24 IS is_sc,allis_sc,partitioning; 25 MPI_Comm gcomm,dcomm,scomm; 26 PetscSF sf; 27 PetscSFNode *remote; 28 Mat *smat; 29 MatPartitioning part; 30 PetscErrorCode ierr; 31 32 PetscFunctionBegin; 33 /* get a sub communicator before call individual MatIncreaseOverlap 34 * since the sub communicator may be changed. 35 * */ 36 ierr = PetscObjectGetComm((PetscObject)(*is),&dcomm);CHKERRQ(ierr); 37 /* make a copy before the original one is deleted */ 38 ierr = PetscCommDuplicate(dcomm,&scomm,NULL);CHKERRQ(ierr); 39 /* get a global communicator, where mat should be a global matrix */ 40 ierr = PetscObjectGetComm((PetscObject)mat,&gcomm);CHKERRQ(ierr); 41 ierr = (*mat->ops->increaseoverlap)(mat,1,is,ov);CHKERRQ(ierr); 42 ierr = MPI_Comm_compare(gcomm,scomm,&issamecomm);CHKERRQ(ierr); 43 /* if the sub-communicator is the same as the global communicator, 44 * user does not want to use a sub-communicator 45 * */ 46 if(issamecomm == MPI_IDENT || issamecomm == MPI_CONGRUENT) PetscFunctionReturn(0); 47 /* if the sub-communicator is petsc_comm_self, 48 * user also does not care the sub-communicator 49 * */ 50 ierr = MPI_Comm_compare(scomm,PETSC_COMM_SELF,&issamecomm);CHKERRQ(ierr); 51 if(issamecomm == MPI_IDENT || issamecomm == MPI_CONGRUENT){PetscFunctionReturn(0);} 52 ierr = MPI_Comm_rank(scomm,&srank);CHKERRQ(ierr); 53 ierr = MPI_Comm_size(scomm,&ssize);CHKERRQ(ierr); 54 ierr = MPI_Comm_rank(gcomm,&grank);CHKERRQ(ierr); 55 /* create a new IS based on sub-communicator 56 * since the old IS is often based on petsc_comm_self 57 * */ 58 ierr = ISGetLocalSize(*is,&nindx);CHKERRQ(ierr); 59 ierr = PetscCalloc1(nindx,&indices_sc);CHKERRQ(ierr); 60 ierr = ISGetIndices(*is,&indices);CHKERRQ(ierr); 61 ierr = PetscMemcpy(indices_sc,indices,sizeof(PetscInt)*nindx);CHKERRQ(ierr); 62 ierr = ISRestoreIndices(*is,&indices);CHKERRQ(ierr); 63 /* we do not need any more */ 64 ierr = ISDestroy(is);CHKERRQ(ierr); 65 /* create a index set based on the sub communicator */ 66 ierr = ISCreateGeneral(scomm,nindx,indices_sc,PETSC_OWN_POINTER,&is_sc);CHKERRQ(ierr); 67 /* gather all indices within the sub communicator */ 68 ierr = ISAllGather(is_sc,&allis_sc);CHKERRQ(ierr); 69 ierr = ISDestroy(&is_sc);CHKERRQ(ierr); 70 /* gather local sizes */ 71 ierr = PetscMalloc1(ssize,&localsizes_sc);CHKERRQ(ierr); 72 /* get individual local sizes for all index sets */ 73 ierr = MPI_Gather(&nindx,1,MPIU_INT,localsizes_sc,1,MPIU_INT,0,scomm);CHKERRQ(ierr); 74 /* only root does these computations */ 75 if(!srank){ 76 /* get local size for the big index set */ 77 ierr = ISGetLocalSize(allis_sc,&localsize);CHKERRQ(ierr); 78 ierr = PetscCalloc2(localsize,&indices_ov,localsize,&sources_sc);CHKERRQ(ierr); 79 ierr = PetscCalloc2(localsize,&indices_ov_rd,localsize,&sources_sc_rd);CHKERRQ(ierr); 80 ierr = ISGetIndices(allis_sc,&indices);CHKERRQ(ierr); 81 ierr = PetscMemcpy(indices_ov,indices,sizeof(PetscInt)*localsize);CHKERRQ(ierr); 82 ierr = ISRestoreIndices(allis_sc,&indices);CHKERRQ(ierr); 83 ierr = ISDestroy(&allis_sc);CHKERRQ(ierr); 84 /* assign corresponding sources */ 85 localsize_tmp = 0; 86 for(k=0; k<ssize; k++){ 87 for(i=0; i<localsizes_sc[k]; i++){ 88 sources_sc[localsize_tmp++] = k; 89 } 90 } 91 /* record where indices come from */ 92 ierr = PetscSortIntWithArray(localsize,indices_ov,sources_sc);CHKERRQ(ierr); 93 /* count local sizes for reduced indices */ 94 ierr = PetscMemzero(localsizes_sc,sizeof(PetscInt)*ssize);CHKERRQ(ierr); 95 /* initialize the first entity */ 96 if(localsize){ 97 indices_ov_rd[0] = indices_ov[0]; 98 sources_sc_rd[0] = sources_sc[0]; 99 localsizes_sc[sources_sc[0]]++; 100 } 101 localsize_tmp = 1; 102 /* remove duplicate integers */ 103 for(i=1; i<localsize; i++){ 104 if(indices_ov[i] != indices_ov[i-1]){ 105 indices_ov_rd[localsize_tmp] = indices_ov[i]; 106 sources_sc_rd[localsize_tmp++] = sources_sc[i]; 107 localsizes_sc[sources_sc[i]]++; 108 } 109 } 110 ierr = PetscFree2(indices_ov,sources_sc);CHKERRQ(ierr); 111 ierr = PetscCalloc1(ssize+1,&localoffsets);CHKERRQ(ierr); 112 for(k=0; k<ssize; k++){ 113 localoffsets[k+1] = localoffsets[k] + localsizes_sc[k]; 114 } 115 nleaves = localoffsets[ssize]; 116 ierr = PetscMemzero(localoffsets,(ssize+1)*sizeof(PetscInt));CHKERRQ(ierr); 117 nroots = localsizes_sc[srank]; 118 ierr = PetscCalloc1(nleaves,&remote);CHKERRQ(ierr); 119 for(i=0; i<nleaves; i++){ 120 remote[i].rank = sources_sc_rd[i]; 121 remote[i].index = localoffsets[sources_sc_rd[i]]++; 122 } 123 ierr = PetscFree(localoffsets);CHKERRQ(ierr); 124 }else{ 125 ierr = ISDestroy(&allis_sc);CHKERRQ(ierr); 126 /* Allocate a 'zero' pointer to avoid using uninitialized variable */ 127 ierr = PetscCalloc1(0,&remote);CHKERRQ(ierr); 128 nleaves = 0; 129 indices_ov_rd = 0; 130 sources_sc_rd = 0; 131 } 132 /* scatter sizes to everybody */ 133 ierr = MPI_Scatter(localsizes_sc,1, MPIU_INT,&nroots,1, MPIU_INT,0,scomm);CHKERRQ(ierr); 134 ierr = PetscFree(localsizes_sc);CHKERRQ(ierr); 135 ierr = PetscCalloc1(nroots,&indices_recv);CHKERRQ(ierr); 136 /* set data back to every body */ 137 ierr = PetscSFCreate(scomm,&sf);CHKERRQ(ierr); 138 ierr = PetscSFSetType(sf,PETSCSFBASIC);CHKERRQ(ierr); 139 ierr = PetscSFSetFromOptions(sf);CHKERRQ(ierr); 140 ierr = PetscSFSetGraph(sf,nroots,nleaves,PETSC_NULL,PETSC_OWN_POINTER,remote,PETSC_OWN_POINTER);CHKERRQ(ierr); 141 ierr = PetscSFReduceBegin(sf,MPIU_INT,indices_ov_rd,indices_recv,MPIU_REPLACE);CHKERRQ(ierr); 142 ierr = PetscSFReduceEnd(sf,MPIU_INT,indices_ov_rd,indices_recv,MPIU_REPLACE);CHKERRQ(ierr); 143 ierr = PetscSFDestroy(&sf);CHKERRQ(ierr); 144 ierr = PetscFree2(indices_ov_rd,sources_sc_rd);CHKERRQ(ierr); 145 ierr = ISCreateGeneral(scomm,nroots,indices_recv,PETSC_OWN_POINTER,&is_sc);CHKERRQ(ierr); 146 ierr = MatGetSubMatricesMPI(mat,1,&is_sc,&is_sc,MAT_INITIAL_MATRIX,&smat);CHKERRQ(ierr); 147 ierr = ISDestroy(&allis_sc);CHKERRQ(ierr); 148 /* create a partitioner to repartition the sub-matrix */ 149 ierr = MatPartitioningCreate(scomm,&part);CHKERRQ(ierr); 150 ierr = MatPartitioningSetAdjacency(part,smat[0]);CHKERRQ(ierr); 151 #if PETSC_HAVE_PARMETIS 152 /* if there exists a ParMETIS installation, we try to use ParMETIS 153 * because a repartition routine possibly work better 154 * */ 155 ierr = MatPartitioningSetType(part,MATPARTITIONINGPARMETIS);CHKERRQ(ierr); 156 /* try to use reparition function, instead of partition function */ 157 ierr = MatPartitioningParmetisSetRepartition(part);CHKERRQ(ierr); 158 #else 159 /* we at least provide a default partitioner to rebalance the computation */ 160 ierr = MatPartitioningSetType(part,MATPARTITIONINGAVERAGE);CHKERRQ(ierr); 161 #endif 162 /* user can pick up any partitioner by using an option */ 163 ierr = MatPartitioningSetFromOptions(part);CHKERRQ(ierr); 164 ierr = MatPartitioningApply(part,&partitioning);CHKERRQ(ierr); 165 ierr = MatPartitioningDestroy(&part);CHKERRQ(ierr); 166 ierr = MatDestroy(&(smat[0]));CHKERRQ(ierr); 167 ierr = PetscFree(smat);CHKERRQ(ierr); 168 /* get local rows including overlap */ 169 ierr = ISBuildTwoSided(partitioning,is_sc,is);CHKERRQ(ierr); 170 ierr = ISDestroy(&is_sc);CHKERRQ(ierr); 171 ierr = ISDestroy(&partitioning);CHKERRQ(ierr); 172 ierr = PetscCommDestroy(&scomm);CHKERRQ(ierr); 173 PetscFunctionReturn(0); 174 } 175 176 177