1*96c04dfeSSatish Balay #include "../src/mat/impls/aij/mpi/mpiaij.h" 2*96c04dfeSSatish Balay 3*96c04dfeSSatish Balay /* 4*96c04dfeSSatish Balay 5*96c04dfeSSatish Balay This routine creates multiple [bjacobi] 'parallel submatrices' from 6*96c04dfeSSatish Balay a given 'mat' object. Each submatrix can span multiple procs. 7*96c04dfeSSatish Balay 8*96c04dfeSSatish Balay The submatrix partition across processors is dicated by 'subComm' a 9*96c04dfeSSatish Balay communicator obtained by com_split(comm). Note: the comm_split 10*96c04dfeSSatish Balay is not restriced to be grouped with consequitive original ranks. 11*96c04dfeSSatish Balay 12*96c04dfeSSatish Balay Due the comm_split() usage, the parallel layout of the submatrices 13*96c04dfeSSatish Balay map directly to the layout of the original matrix [wrt the local 14*96c04dfeSSatish Balay row,col partitioning]. So the original 'DiagonalMat' naturally maps 15*96c04dfeSSatish Balay into the 'DiagonalMat' of the subMat, hence it is used directly from 16*96c04dfeSSatish Balay the subMat. However the offDiagMat looses some columns - and this is 17*96c04dfeSSatish Balay reconstructed with MatSetValues() 18*96c04dfeSSatish Balay 19*96c04dfeSSatish Balay */ 20*96c04dfeSSatish Balay 21*96c04dfeSSatish Balay #undef __FUNCT__ 22*96c04dfeSSatish Balay #define __FUNCT__ "MatGetMultiProcBlock_MPIAIJ" 23*96c04dfeSSatish Balay PetscErrorCode MatGetMultiProcBlock_MPIAIJ(Mat mat, MPI_Comm subComm, Mat* subMat) 24*96c04dfeSSatish Balay { 25*96c04dfeSSatish Balay PetscErrorCode ierr; 26*96c04dfeSSatish Balay Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 27*96c04dfeSSatish Balay Mat_SeqAIJ* aijB = (Mat_SeqAIJ*)aij->B->data; 28*96c04dfeSSatish Balay PetscMPIInt commRank,subCommSize,subCommRank; 29*96c04dfeSSatish Balay PetscMPIInt *commRankMap,subRank,rank; 30*96c04dfeSSatish Balay PetscInt *garrayCMap,col,i,j,*nnz,newRow,newCol; 31*96c04dfeSSatish Balay 32*96c04dfeSSatish Balay PetscFunctionBegin; 33*96c04dfeSSatish Balay 34*96c04dfeSSatish Balay /* create subMat object with the relavent layout */ 35*96c04dfeSSatish Balay ierr = MatCreate(subComm,subMat);CHKERRQ(ierr); 36*96c04dfeSSatish Balay ierr = MatSetType(*subMat,MATMPIAIJ);CHKERRQ(ierr); 37*96c04dfeSSatish Balay ierr = MatSetSizes(*subMat,mat->rmap->n,mat->cmap->n,PETSC_DECIDE,PETSC_DECIDE);CHKERRQ(ierr); 38*96c04dfeSSatish Balay /* need to setup rmap and cmap before Preallocation */ 39*96c04dfeSSatish Balay ierr = PetscLayoutSetBlockSize((*subMat)->rmap,mat->rmap->bs);CHKERRQ(ierr); 40*96c04dfeSSatish Balay ierr = PetscLayoutSetBlockSize((*subMat)->cmap,mat->cmap->bs);CHKERRQ(ierr); 41*96c04dfeSSatish Balay ierr = PetscLayoutSetUp((*subMat)->rmap);CHKERRQ(ierr); 42*96c04dfeSSatish Balay ierr = PetscLayoutSetUp((*subMat)->cmap);CHKERRQ(ierr); 43*96c04dfeSSatish Balay 44*96c04dfeSSatish Balay /* create a map of comm_rank from subComm to comm */ 45*96c04dfeSSatish Balay ierr = MPI_Comm_rank(((PetscObject)mat)->comm,&commRank);CHKERRQ(ierr); 46*96c04dfeSSatish Balay ierr = MPI_Comm_size(subComm,&subCommSize);CHKERRQ(ierr); 47*96c04dfeSSatish Balay ierr = MPI_Comm_rank(subComm,&subCommRank);CHKERRQ(ierr); 48*96c04dfeSSatish Balay ierr = PetscMalloc(subCommSize*sizeof(PetscMPIInt),&commRankMap);CHKERRQ(ierr); 49*96c04dfeSSatish Balay ierr = MPI_Allgather(&commRank,1,MPI_INT,commRankMap,1,MPI_INT,subComm);CHKERRQ(ierr); 50*96c04dfeSSatish Balay 51*96c04dfeSSatish Balay /* Traverse garray and identify column indices [of offdiag mat] that 52*96c04dfeSSatish Balay should be discarded. For the ones not discarded, store the newCol+1 53*96c04dfeSSatish Balay value in garrayCMap */ 54*96c04dfeSSatish Balay ierr = PetscMalloc(aij->B->cmap->n*sizeof(PetscInt),&garrayCMap);CHKERRQ(ierr); 55*96c04dfeSSatish Balay ierr = PetscMemzero(garrayCMap,aij->B->cmap->n*sizeof(PetscInt));CHKERRQ(ierr); 56*96c04dfeSSatish Balay for (i=0; i<aij->B->cmap->n; i++) { 57*96c04dfeSSatish Balay col = aij->garray[i]; 58*96c04dfeSSatish Balay for (subRank=0; subRank<subCommSize; subRank++) { 59*96c04dfeSSatish Balay rank = commRankMap[subRank]; 60*96c04dfeSSatish Balay if ((col >= mat->cmap->range[rank]) && (col < mat->cmap->range[rank+1])) { 61*96c04dfeSSatish Balay garrayCMap[i] = (*subMat)->cmap->range[subRank] + col - mat->cmap->range[rank]+1; 62*96c04dfeSSatish Balay break; 63*96c04dfeSSatish Balay } 64*96c04dfeSSatish Balay } 65*96c04dfeSSatish Balay } 66*96c04dfeSSatish Balay 67*96c04dfeSSatish Balay /* Now compute preallocation for the offdiag mat */ 68*96c04dfeSSatish Balay ierr = PetscMalloc(aij->B->rmap->n*sizeof(PetscInt),&nnz);CHKERRQ(ierr); 69*96c04dfeSSatish Balay ierr = PetscMemzero(nnz,aij->B->rmap->n*sizeof(PetscInt));CHKERRQ(ierr); 70*96c04dfeSSatish Balay for (i=0; i<aij->B->rmap->n; i++) { 71*96c04dfeSSatish Balay for (j=aijB->i[i]; j<aijB->i[i+1]; j++) { 72*96c04dfeSSatish Balay if (garrayCMap[aijB->j[j]]) nnz[i]++; 73*96c04dfeSSatish Balay } 74*96c04dfeSSatish Balay } 75*96c04dfeSSatish Balay ierr = MatMPIAIJSetPreallocation(*(subMat),PETSC_NULL,PETSC_NULL,PETSC_NULL,nnz);CHKERRQ(ierr); 76*96c04dfeSSatish Balay 77*96c04dfeSSatish Balay /* reuse diag block with the new submat */ 78*96c04dfeSSatish Balay ierr = MatDestroy(((Mat_MPIAIJ*)((*subMat)->data))->A);CHKERRQ(ierr); 79*96c04dfeSSatish Balay ((Mat_MPIAIJ*)((*subMat)->data))->A = aij->A; 80*96c04dfeSSatish Balay ierr = PetscObjectReference((PetscObject)aij->A);CHKERRQ(ierr); 81*96c04dfeSSatish Balay 82*96c04dfeSSatish Balay /* Now traverse aij->B and insert values into subMat */ 83*96c04dfeSSatish Balay for (i=0; i<aij->B->rmap->n; i++) { 84*96c04dfeSSatish Balay newRow = (*subMat)->rmap->range[subCommRank] + i; 85*96c04dfeSSatish Balay for (j=aijB->i[i]; j<aijB->i[i+1]; j++) { 86*96c04dfeSSatish Balay newCol = garrayCMap[aijB->j[j]]; 87*96c04dfeSSatish Balay if (newCol) { 88*96c04dfeSSatish Balay newCol--; /* remove the increment */ 89*96c04dfeSSatish Balay ierr = MatSetValues(*subMat,1,&newRow,1,&newCol,(aijB->a+j),INSERT_VALUES);CHKERRQ(ierr); 90*96c04dfeSSatish Balay } 91*96c04dfeSSatish Balay } 92*96c04dfeSSatish Balay } 93*96c04dfeSSatish Balay 94*96c04dfeSSatish Balay /* assemble the submat */ 95*96c04dfeSSatish Balay ierr = MatAssemblyBegin(*subMat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 96*96c04dfeSSatish Balay ierr = MatAssemblyEnd(*subMat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 97*96c04dfeSSatish Balay 98*96c04dfeSSatish Balay /* deallocate temporary data */ 99*96c04dfeSSatish Balay ierr = PetscFree(commRankMap);CHKERRQ(ierr); 100*96c04dfeSSatish Balay ierr = PetscFree(garrayCMap);CHKERRQ(ierr); 101*96c04dfeSSatish Balay ierr = PetscFree(nnz);CHKERRQ(ierr); 102*96c04dfeSSatish Balay 103*96c04dfeSSatish Balay PetscFunctionReturn(0); 104*96c04dfeSSatish Balay } 105