1c6db04a5SJed Brown #include <../src/mat/impls/aij/mpi/mpiaij.h> 296c04dfeSSatish Balay 396c04dfeSSatish Balay #undef __FUNCT__ 496c04dfeSSatish Balay #define __FUNCT__ "MatGetMultiProcBlock_MPIAIJ" 5fc08c53fSHong Zhang PetscErrorCode MatGetMultiProcBlock_MPIAIJ(Mat mat, MPI_Comm subComm, MatReuse scall,Mat* subMat) 696c04dfeSSatish Balay { 796c04dfeSSatish Balay PetscErrorCode ierr; 896c04dfeSSatish Balay Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 996c04dfeSSatish Balay Mat_SeqAIJ* aijB = (Mat_SeqAIJ*)aij->B->data; 1096c04dfeSSatish Balay PetscMPIInt commRank,subCommSize,subCommRank; 115d22bc9fSHong Zhang PetscMPIInt *commRankMap,subRank,rank,commsize; 1296c04dfeSSatish Balay PetscInt *garrayCMap,col,i,j,*nnz,newRow,newCol; 1396c04dfeSSatish Balay 1496c04dfeSSatish Balay PetscFunctionBegin; 155d22bc9fSHong Zhang ierr = MPI_Comm_size(((PetscObject)mat)->comm,&commsize);CHKERRQ(ierr); 165d22bc9fSHong Zhang ierr = MPI_Comm_size(subComm,&subCommSize);CHKERRQ(ierr); 1796c04dfeSSatish Balay 1896c04dfeSSatish Balay /* create subMat object with the relavent layout */ 19*ce1d8e51SHong Zhang if (scall == MAT_INITIAL_MATRIX){ 2096c04dfeSSatish Balay ierr = MatCreate(subComm,subMat);CHKERRQ(ierr); 2196c04dfeSSatish Balay ierr = MatSetType(*subMat,MATMPIAIJ);CHKERRQ(ierr); 2296c04dfeSSatish Balay ierr = MatSetSizes(*subMat,mat->rmap->n,mat->cmap->n,PETSC_DECIDE,PETSC_DECIDE);CHKERRQ(ierr); 23fc08c53fSHong Zhang 2496c04dfeSSatish Balay /* need to setup rmap and cmap before Preallocation */ 2596c04dfeSSatish Balay ierr = PetscLayoutSetBlockSize((*subMat)->rmap,mat->rmap->bs);CHKERRQ(ierr); 2696c04dfeSSatish Balay ierr = PetscLayoutSetBlockSize((*subMat)->cmap,mat->cmap->bs);CHKERRQ(ierr); 2796c04dfeSSatish Balay ierr = PetscLayoutSetUp((*subMat)->rmap);CHKERRQ(ierr); 2896c04dfeSSatish Balay ierr = PetscLayoutSetUp((*subMat)->cmap);CHKERRQ(ierr); 29*ce1d8e51SHong Zhang } 3096c04dfeSSatish Balay 31*ce1d8e51SHong Zhang /* create a map of comm_rank from subComm to comm - should commRankMap and garrayCMap be kept for reused? */ 3296c04dfeSSatish Balay ierr = MPI_Comm_rank(((PetscObject)mat)->comm,&commRank);CHKERRQ(ierr); 3396c04dfeSSatish Balay ierr = MPI_Comm_rank(subComm,&subCommRank);CHKERRQ(ierr); 3496c04dfeSSatish Balay ierr = PetscMalloc(subCommSize*sizeof(PetscMPIInt),&commRankMap);CHKERRQ(ierr); 3596c04dfeSSatish Balay ierr = MPI_Allgather(&commRank,1,MPI_INT,commRankMap,1,MPI_INT,subComm);CHKERRQ(ierr); 3696c04dfeSSatish Balay 3796c04dfeSSatish Balay /* Traverse garray and identify column indices [of offdiag mat] that 3896c04dfeSSatish Balay should be discarded. For the ones not discarded, store the newCol+1 3996c04dfeSSatish Balay value in garrayCMap */ 4096c04dfeSSatish Balay ierr = PetscMalloc(aij->B->cmap->n*sizeof(PetscInt),&garrayCMap);CHKERRQ(ierr); 4196c04dfeSSatish Balay ierr = PetscMemzero(garrayCMap,aij->B->cmap->n*sizeof(PetscInt));CHKERRQ(ierr); 4296c04dfeSSatish Balay for (i=0; i<aij->B->cmap->n; i++) { 4396c04dfeSSatish Balay col = aij->garray[i]; 4496c04dfeSSatish Balay for (subRank=0; subRank<subCommSize; subRank++) { 4596c04dfeSSatish Balay rank = commRankMap[subRank]; 4696c04dfeSSatish Balay if ((col >= mat->cmap->range[rank]) && (col < mat->cmap->range[rank+1])) { 4796c04dfeSSatish Balay garrayCMap[i] = (*subMat)->cmap->range[subRank] + col - mat->cmap->range[rank]+1; 4896c04dfeSSatish Balay break; 4996c04dfeSSatish Balay } 5096c04dfeSSatish Balay } 5196c04dfeSSatish Balay } 5296c04dfeSSatish Balay 53*ce1d8e51SHong Zhang if (scall == MAT_INITIAL_MATRIX){ 5496c04dfeSSatish Balay /* Now compute preallocation for the offdiag mat */ 5596c04dfeSSatish Balay ierr = PetscMalloc(aij->B->rmap->n*sizeof(PetscInt),&nnz);CHKERRQ(ierr); 5696c04dfeSSatish Balay ierr = PetscMemzero(nnz,aij->B->rmap->n*sizeof(PetscInt));CHKERRQ(ierr); 5796c04dfeSSatish Balay for (i=0; i<aij->B->rmap->n; i++) { 5896c04dfeSSatish Balay for (j=aijB->i[i]; j<aijB->i[i+1]; j++) { 5996c04dfeSSatish Balay if (garrayCMap[aijB->j[j]]) nnz[i]++; 6096c04dfeSSatish Balay } 6196c04dfeSSatish Balay } 62ecd8bba6SJed Brown ierr = MatMPIAIJSetPreallocation(*(subMat),0,PETSC_NULL,0,nnz);CHKERRQ(ierr); 6396c04dfeSSatish Balay 6496c04dfeSSatish Balay /* reuse diag block with the new submat */ 656bf464f9SBarry Smith ierr = MatDestroy(&((Mat_MPIAIJ*)((*subMat)->data))->A);CHKERRQ(ierr); 6696c04dfeSSatish Balay ((Mat_MPIAIJ*)((*subMat)->data))->A = aij->A; 6796c04dfeSSatish Balay ierr = PetscObjectReference((PetscObject)aij->A);CHKERRQ(ierr); 68*ce1d8e51SHong Zhang } else if ( ((Mat_MPIAIJ*)(*subMat)->data)->A != aij->A ){ 69*ce1d8e51SHong Zhang PetscObject obj = (PetscObject)((Mat_MPIAIJ*)((*subMat)->data))->A; 70*ce1d8e51SHong Zhang ierr = PetscObjectReference((PetscObject)obj);CHKERRQ(ierr); 71*ce1d8e51SHong Zhang ((Mat_MPIAIJ*)((*subMat)->data))->A = aij->A; 72*ce1d8e51SHong Zhang ierr = PetscObjectReference((PetscObject)aij->A);CHKERRQ(ierr); 73*ce1d8e51SHong Zhang } 7496c04dfeSSatish Balay 7596c04dfeSSatish Balay /* Now traverse aij->B and insert values into subMat */ 7696c04dfeSSatish Balay for (i=0; i<aij->B->rmap->n; i++) { 7796c04dfeSSatish Balay newRow = (*subMat)->rmap->range[subCommRank] + i; 7896c04dfeSSatish Balay for (j=aijB->i[i]; j<aijB->i[i+1]; j++) { 7996c04dfeSSatish Balay newCol = garrayCMap[aijB->j[j]]; 8096c04dfeSSatish Balay if (newCol) { 8196c04dfeSSatish Balay newCol--; /* remove the increment */ 8296c04dfeSSatish Balay ierr = MatSetValues(*subMat,1,&newRow,1,&newCol,(aijB->a+j),INSERT_VALUES);CHKERRQ(ierr); 8396c04dfeSSatish Balay } 8496c04dfeSSatish Balay } 8596c04dfeSSatish Balay } 8696c04dfeSSatish Balay 8796c04dfeSSatish Balay /* assemble the submat */ 8896c04dfeSSatish Balay ierr = MatAssemblyBegin(*subMat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 8996c04dfeSSatish Balay ierr = MatAssemblyEnd(*subMat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 9096c04dfeSSatish Balay 9196c04dfeSSatish Balay /* deallocate temporary data */ 9296c04dfeSSatish Balay ierr = PetscFree(commRankMap);CHKERRQ(ierr); 9396c04dfeSSatish Balay ierr = PetscFree(garrayCMap);CHKERRQ(ierr); 94*ce1d8e51SHong Zhang if (scall == MAT_INITIAL_MATRIX){ 9596c04dfeSSatish Balay ierr = PetscFree(nnz);CHKERRQ(ierr); 96*ce1d8e51SHong Zhang } 9796c04dfeSSatish Balay PetscFunctionReturn(0); 9896c04dfeSSatish Balay } 99