PetscErrorCode MatPtAPSymbolic_SeqAIJ_SeqAIJ_DenseAxpy(Mat A,Mat P,PetscReal fill,Mat *C)
{
PetscErrorCode ierr;
Mat_SeqAIJ *ap,*c;
PetscInt *api,*apj,*ci,pn=P->cmap->N;
MatScalar *ca;
Mat_PtAP *ptap;
Mat Pt,AP;
PetscFunctionBegin;
/* Get symbolic Pt = P^T */
ierr = MatTransposeSymbolic_SeqAIJ(P,&Pt);CHKERRQ(ierr);
/* Get symbolic AP = A*P */
ierr = MatMatMultSymbolic_SeqAIJ_SeqAIJ(A,P,fill,&AP);CHKERRQ(ierr);
ap = (Mat_SeqAIJ*)AP->data;
api = ap->i;
apj = ap->j;
ap->free_ij = PETSC_FALSE; /* api and apj are kept in struct ptap, cannot be destroyed with AP */
/* Get C = Pt*AP */
ierr = MatMatMultSymbolic_SeqAIJ_SeqAIJ(Pt,AP,fill,C);CHKERRQ(ierr);
c = (Mat_SeqAIJ*)(*C)->data;
ci = c->i;
ierr = PetscCalloc1(ci[pn]+1,&ca);CHKERRQ(ierr);
c->a = ca;
c->free_a = PETSC_TRUE;
/* Create a supporting struct for reuse by MatPtAPNumeric() */
ierr = PetscNew(&ptap);CHKERRQ(ierr);
c->ptap = ptap;
ptap->destroy = (*C)->ops->destroy;
(*C)->ops->destroy = MatDestroy_SeqAIJ_PtAP;
/* Allocate temporary array for storage of one row of A*P */
ierr = PetscCalloc1(pn+1,&ptap->apa);CHKERRQ(ierr);
(*C)->ops->ptapnumeric = MatPtAPNumeric_SeqAIJ_SeqAIJ;
ptap->api = api;
ptap->apj = apj;
/* Clean up. */
ierr = MatDestroy(&Pt);CHKERRQ(ierr);
ierr = MatDestroy(&AP);CHKERRQ(ierr);
#if defined(PETSC_USE_INFO)
ierr = PetscInfo1((*C),"given fill %g\n",(double)fill);CHKERRQ(ierr);
#endif
PetscFunctionReturn(0);
}
PetscErrorCode MatPtAPSymbolic_SeqAIJ_SeqAIJ(Mat A,Mat P,PetscReal fill,Mat *C)
{
PetscErrorCode ierr;
Mat_SeqAIJ *ap,*c;
PetscInt *api,*apj,*ci,pn=P->cmap->N;
MatScalar *ca;
Mat_PtAP *ptap;
Mat Pt,AP;
PetscBool sparse_axpy=PETSC_TRUE;
PetscFunctionBegin;
ierr = PetscObjectOptionsBegin((PetscObject)A);CHKERRQ(ierr);
/* flag 'sparse_axpy' determines which implementations to be used:
0: do dense axpy in MatPtAPNumeric() - fastest, but requires storage of struct A*P;
1: do two sparse axpy in MatPtAPNumeric() - slowest, does not store structure of A*P. */
ierr = PetscOptionsBool("-matptap_scalable","Use sparse axpy but slower MatPtAPNumeric()","",sparse_axpy,&sparse_axpy,NULL);CHKERRQ(ierr);
ierr = PetscOptionsEnd();CHKERRQ(ierr);
if (sparse_axpy) {
ierr = MatPtAPSymbolic_SeqAIJ_SeqAIJ_SparseAxpy(A,P,fill,C);CHKERRQ(ierr);
PetscFunctionReturn(0);
}
/* Get symbolic Pt = P^T */
ierr = MatTransposeSymbolic_SeqAIJ(P,&Pt);CHKERRQ(ierr);
/* Get symbolic AP = A*P */
ierr = MatMatMultSymbolic_SeqAIJ_SeqAIJ(A,P,fill,&AP);CHKERRQ(ierr);
ap = (Mat_SeqAIJ*)AP->data;
api = ap->i;
apj = ap->j;
ap->free_ij = PETSC_FALSE; /* api and apj are kept in struct ptap, cannot be destroyed with AP */
/* Get C = Pt*AP */
ierr = MatMatMultSymbolic_SeqAIJ_SeqAIJ(Pt,AP,fill,C);CHKERRQ(ierr);
c = (Mat_SeqAIJ*)(*C)->data;
ci = c->i;
ierr = PetscMalloc((ci[pn]+1)*sizeof(MatScalar),&ca);CHKERRQ(ierr);
ierr = PetscMemzero(ca,(ci[pn]+1)*sizeof(MatScalar));CHKERRQ(ierr);
c->a = ca;
c->free_a = PETSC_TRUE;
/* Create a supporting struct for reuse by MatPtAPNumeric() */
ierr = PetscNew(Mat_PtAP,&ptap);CHKERRQ(ierr);
c->ptap = ptap;
ptap->destroy = (*C)->ops->destroy;
(*C)->ops->destroy = MatDestroy_SeqAIJ_PtAP;
/* Allocate temporary array for storage of one row of A*P */
ierr = PetscMalloc((pn+1)*sizeof(PetscScalar),&ptap->apa);CHKERRQ(ierr);
ierr = PetscMemzero(ptap->apa,(pn+1)*sizeof(PetscScalar));CHKERRQ(ierr);
(*C)->ops->ptapnumeric = MatPtAPNumeric_SeqAIJ_SeqAIJ;
ptap->api = api;
ptap->apj = apj;
/* Clean up. */
ierr = MatDestroy(&Pt);CHKERRQ(ierr);
ierr = MatDestroy(&AP);CHKERRQ(ierr);
#if defined(PETSC_USE_INFO)
ierr = PetscInfo2((*C),"given fill %G, use scalable %d\n",fill,sparse_axpy);CHKERRQ(ierr);
#endif
PetscFunctionReturn(0);
}
PetscErrorCode MatPtAPSymbolic_SeqAIJ_SeqAIJ(Mat A,Mat P,PetscReal fill,Mat *C)
{
PetscErrorCode ierr;
Mat_SeqAIJ *ap,*c;
PetscInt *api,*apj,*ci,pn=P->cmap->N,sparse_axpy=0;
MatScalar *ca;
Mat_PtAP *ptap;
Mat Pt,AP;
PetscFunctionBegin;
/* flag 'sparse_axpy' determines which implementations to be used:
0: do dense axpy in MatPtAPNumeric() - fastest, but requires storage of struct A*P; (default)
1: do one sparse axpy - uses same memory as sparse_axpy=0 and might execute less flops
(apnz vs. cnz in the outerproduct), slower than case '0' when cnz is not too large than apnz;
2: do two sparse axpy in MatPtAPNumeric() - slowest, does not store structure of A*P. */
ierr = PetscOptionsGetInt(PETSC_NULL,"-matptap_sparseaxpy",&sparse_axpy,PETSC_NULL);CHKERRQ(ierr);
if (sparse_axpy == 2){
ierr = MatPtAPSymbolic_SeqAIJ_SeqAIJ_SparseAxpy2(A,P,fill,C);CHKERRQ(ierr);
PetscFunctionReturn(0);
}
/* Get symbolic Pt = P^T */
ierr = MatTransposeSymbolic_SeqAIJ(P,&Pt);CHKERRQ(ierr);
/* Get symbolic AP = A*P */
ierr = MatMatMultSymbolic_SeqAIJ_SeqAIJ(A,P,fill,&AP);CHKERRQ(ierr);
ap = (Mat_SeqAIJ*)AP->data;
api = ap->i;
apj = ap->j;
ap->free_ij = PETSC_FALSE; /* api and apj are kept in struct ptap, cannot be destroyed with AP */
/* Get C = Pt*AP */
ierr = MatMatMultSymbolic_SeqAIJ_SeqAIJ(Pt,AP,fill,C);CHKERRQ(ierr);
c = (Mat_SeqAIJ*)(*C)->data;
ci = c->i;
ierr = PetscMalloc((ci[pn]+1)*sizeof(MatScalar),&ca);CHKERRQ(ierr);
ierr = PetscMemzero(ca,(ci[pn]+1)*sizeof(MatScalar));CHKERRQ(ierr);
c->a = ca;
c->free_a = PETSC_TRUE;
/* Create a supporting struct for reuse by MatPtAPNumeric() */
ierr = PetscNew(Mat_PtAP,&ptap);CHKERRQ(ierr);
c->ptap = ptap;
ptap->destroy = (*C)->ops->destroy;
(*C)->ops->destroy = MatDestroy_SeqAIJ_PtAP;
/* Allocate temporary array for storage of one row of A*P */
ierr = PetscMalloc((pn+1)*sizeof(PetscScalar),&ptap->apa);CHKERRQ(ierr);
ierr = PetscMemzero(ptap->apa,(pn+1)*sizeof(PetscScalar));CHKERRQ(ierr);
if (sparse_axpy == 1){
A->ops->ptapnumeric = MatPtAPNumeric_SeqAIJ_SeqAIJ_SparseAxpy;
} else {
A->ops->ptapnumeric = MatPtAPNumeric_SeqAIJ_SeqAIJ;
}
ptap->api = api;
ptap->apj = apj;
/* Clean up. */
ierr = MatDestroy(&Pt);CHKERRQ(ierr);
ierr = MatDestroy(&AP);CHKERRQ(ierr);
PetscFunctionReturn(0);
}
