1e8271787SHong Zhang #include <../src/mat/impls/baij/mpi/mpibaij.h> 2e8271787SHong Zhang 3*d71ae5a4SJacob Faibussowitsch PetscErrorCode MatGetMultiProcBlock_MPIBAIJ(Mat mat, MPI_Comm subComm, MatReuse scall, Mat *subMat) 4*d71ae5a4SJacob Faibussowitsch { 5e8271787SHong Zhang Mat_MPIBAIJ *aij = (Mat_MPIBAIJ *)mat->data; 6e8271787SHong Zhang Mat_SeqBAIJ *aijB = (Mat_SeqBAIJ *)aij->B->data; 7e8271787SHong Zhang PetscMPIInt commRank, subCommSize, subCommRank; 8e8271787SHong Zhang PetscMPIInt *commRankMap, subRank, rank, commsize; 902e9429cSHong Zhang PetscInt *garrayCMap, col, i, j, *nnz, newRow, newCol, *newbRow, *newbCol, k, k1; 10e8271787SHong Zhang PetscInt bs = mat->rmap->bs; 1102e9429cSHong Zhang PetscScalar *vals, *aijBvals; 12e8271787SHong Zhang 13e8271787SHong Zhang PetscFunctionBegin; 149566063dSJacob Faibussowitsch PetscCallMPI(MPI_Comm_size(PetscObjectComm((PetscObject)mat), &commsize)); 159566063dSJacob Faibussowitsch PetscCallMPI(MPI_Comm_size(subComm, &subCommSize)); 16e8271787SHong Zhang 17a5b23f4aSJose E. Roman /* create subMat object with the relevant layout */ 18e8271787SHong Zhang if (scall == MAT_INITIAL_MATRIX) { 199566063dSJacob Faibussowitsch PetscCall(MatCreate(subComm, subMat)); 209566063dSJacob Faibussowitsch PetscCall(MatSetType(*subMat, MATMPIBAIJ)); 219566063dSJacob Faibussowitsch PetscCall(MatSetSizes(*subMat, mat->rmap->n, mat->cmap->n, PETSC_DECIDE, PETSC_DECIDE)); 229566063dSJacob Faibussowitsch PetscCall(MatSetBlockSizes(*subMat, mat->rmap->bs, mat->cmap->bs)); 23e8271787SHong Zhang 24e8271787SHong Zhang /* need to setup rmap and cmap before Preallocation */ 259566063dSJacob Faibussowitsch PetscCall(PetscLayoutSetBlockSize((*subMat)->rmap, mat->rmap->bs)); 269566063dSJacob Faibussowitsch PetscCall(PetscLayoutSetBlockSize((*subMat)->cmap, mat->cmap->bs)); 279566063dSJacob Faibussowitsch PetscCall(PetscLayoutSetUp((*subMat)->rmap)); 289566063dSJacob Faibussowitsch PetscCall(PetscLayoutSetUp((*subMat)->cmap)); 29e8271787SHong Zhang } 30e8271787SHong Zhang 31e8271787SHong Zhang /* create a map of comm_rank from subComm to comm - should commRankMap and garrayCMap be kept for reused? */ 329566063dSJacob Faibussowitsch PetscCallMPI(MPI_Comm_rank(PetscObjectComm((PetscObject)mat), &commRank)); 339566063dSJacob Faibussowitsch PetscCallMPI(MPI_Comm_rank(subComm, &subCommRank)); 349566063dSJacob Faibussowitsch PetscCall(PetscMalloc1(subCommSize, &commRankMap)); 359566063dSJacob Faibussowitsch PetscCallMPI(MPI_Allgather(&commRank, 1, MPI_INT, commRankMap, 1, MPI_INT, subComm)); 36e8271787SHong Zhang 37e8271787SHong Zhang /* Traverse garray and identify blocked column indices [of offdiag mat] that 38e8271787SHong Zhang should be discarded. For the ones not discarded, store the newCol+1 39e8271787SHong Zhang value in garrayCMap */ 409566063dSJacob Faibussowitsch PetscCall(PetscCalloc1(aij->B->cmap->n / bs, &garrayCMap)); 41e8271787SHong Zhang for (i = 0; i < aij->B->cmap->n / bs; i++) { 42e8271787SHong Zhang col = aij->garray[i]; /* blocked column index */ 43e8271787SHong Zhang for (subRank = 0; subRank < subCommSize; subRank++) { 44e8271787SHong Zhang rank = commRankMap[subRank]; 45e8271787SHong Zhang if ((col >= mat->cmap->range[rank] / bs) && (col < mat->cmap->range[rank + 1] / bs)) { 4602e9429cSHong Zhang garrayCMap[i] = (((*subMat)->cmap->range[subRank] - mat->cmap->range[rank]) / bs + col + 1); 47e8271787SHong Zhang break; 48e8271787SHong Zhang } 49e8271787SHong Zhang } 50e8271787SHong Zhang } 51e8271787SHong Zhang 52e8271787SHong Zhang if (scall == MAT_INITIAL_MATRIX) { 53e8271787SHong Zhang /* Now compute preallocation for the offdiag mat */ 549566063dSJacob Faibussowitsch PetscCall(PetscCalloc1(aij->B->rmap->n / bs, &nnz)); 55e8271787SHong Zhang for (i = 0; i < aij->B->rmap->n / bs; i++) { 56e8271787SHong Zhang for (j = aijB->i[i]; j < aijB->i[i + 1]; j++) { 57e8271787SHong Zhang if (garrayCMap[aijB->j[j]]) nnz[i]++; 58e8271787SHong Zhang } 59e8271787SHong Zhang } 609566063dSJacob Faibussowitsch PetscCall(MatMPIBAIJSetPreallocation(*(subMat), bs, 0, NULL, 0, nnz)); 61e8271787SHong Zhang 62e8271787SHong Zhang /* reuse diag block with the new submat */ 639566063dSJacob Faibussowitsch PetscCall(MatDestroy(&((Mat_MPIBAIJ *)((*subMat)->data))->A)); 64e8271787SHong Zhang 65e8271787SHong Zhang ((Mat_MPIBAIJ *)((*subMat)->data))->A = aij->A; 66e8271787SHong Zhang 679566063dSJacob Faibussowitsch PetscCall(PetscObjectReference((PetscObject)aij->A)); 68e8271787SHong Zhang } else if (((Mat_MPIBAIJ *)(*subMat)->data)->A != aij->A) { 69e8271787SHong Zhang PetscObject obj = (PetscObject)((Mat_MPIBAIJ *)((*subMat)->data))->A; 70e8271787SHong Zhang 719566063dSJacob Faibussowitsch PetscCall(PetscObjectReference((PetscObject)obj)); 72e8271787SHong Zhang 73e8271787SHong Zhang ((Mat_MPIBAIJ *)((*subMat)->data))->A = aij->A; 74e8271787SHong Zhang 759566063dSJacob Faibussowitsch PetscCall(PetscObjectReference((PetscObject)aij->A)); 76e8271787SHong Zhang } 77e8271787SHong Zhang 78e8271787SHong Zhang /* Now traverse aij->B and insert values into subMat */ 799566063dSJacob Faibussowitsch PetscCall(PetscMalloc3(bs, &newbRow, bs, &newbCol, bs * bs, &vals)); 80e8271787SHong Zhang for (i = 0; i < aij->B->rmap->n / bs; i++) { 81e8271787SHong Zhang newRow = (*subMat)->rmap->range[subCommRank] + i * bs; 82e8271787SHong Zhang for (j = aijB->i[i]; j < aijB->i[i + 1]; j++) { 83e8271787SHong Zhang newCol = garrayCMap[aijB->j[j]]; 84e8271787SHong Zhang if (newCol) { 85e8271787SHong Zhang newCol--; /* remove the increment */ 86e8271787SHong Zhang newCol *= bs; 87e8271787SHong Zhang for (k = 0; k < bs; k++) { 88e8271787SHong Zhang newbRow[k] = newRow + k; 89e8271787SHong Zhang newbCol[k] = newCol + k; 90e8271787SHong Zhang } 91e8271787SHong Zhang /* copy column-oriented aijB->a into row-oriented vals */ 9202e9429cSHong Zhang aijBvals = aijB->a + j * bs * bs; 93e8271787SHong Zhang for (k1 = 0; k1 < bs; k1++) { 94ad540459SPierre Jolivet for (k = 0; k < bs; k++) vals[k1 + k * bs] = *aijBvals++; 95e8271787SHong Zhang } 969566063dSJacob Faibussowitsch PetscCall(MatSetValues(*subMat, bs, newbRow, bs, newbCol, vals, INSERT_VALUES)); 97e8271787SHong Zhang } 98e8271787SHong Zhang } 99e8271787SHong Zhang } 1009566063dSJacob Faibussowitsch PetscCall(MatAssemblyBegin(*subMat, MAT_FINAL_ASSEMBLY)); 1019566063dSJacob Faibussowitsch PetscCall(MatAssemblyEnd(*subMat, MAT_FINAL_ASSEMBLY)); 102e8271787SHong Zhang 103e8271787SHong Zhang /* deallocate temporary data */ 1049566063dSJacob Faibussowitsch PetscCall(PetscFree3(newbRow, newbCol, vals)); 1059566063dSJacob Faibussowitsch PetscCall(PetscFree(commRankMap)); 1069566063dSJacob Faibussowitsch PetscCall(PetscFree(garrayCMap)); 10748a46eb9SPierre Jolivet if (scall == MAT_INITIAL_MATRIX) PetscCall(PetscFree(nnz)); 108e8271787SHong Zhang PetscFunctionReturn(0); 109e8271787SHong Zhang } 110