PetscErrorCode MatSetLocalToGlobalMapping_IS(Mat A,ISLocalToGlobalMapping rmapping,ISLocalToGlobalMapping cmapping)
{
PetscErrorCode ierr;
PetscInt n,bs;
Mat_IS *is = (Mat_IS*)A->data;
IS from,to;
Vec global;
PetscFunctionBegin;
PetscCheckSameComm(A,1,rmapping,2);
if (rmapping != cmapping) SETERRQ(PetscObjectComm((PetscObject)A),PETSC_ERR_ARG_INCOMP,"MATIS requires the row and column mappings to be identical");
if (is->mapping) { /* Currenly destroys the objects that will be created by this routine. Is there anything else that should be checked? */
ierr = ISLocalToGlobalMappingDestroy(&is->mapping);CHKERRQ(ierr);
ierr = VecDestroy(&is->x);CHKERRQ(ierr);
ierr = VecDestroy(&is->y);CHKERRQ(ierr);
ierr = VecScatterDestroy(&is->ctx);CHKERRQ(ierr);
ierr = MatDestroy(&is->A);CHKERRQ(ierr);
}
ierr = PetscObjectReference((PetscObject)rmapping);CHKERRQ(ierr);
ierr = ISLocalToGlobalMappingDestroy(&is->mapping);CHKERRQ(ierr);
is->mapping = rmapping;
/*
ierr = PetscLayoutSetISLocalToGlobalMapping(A->rmap,rmapping);CHKERRQ(ierr);
ierr = PetscLayoutSetISLocalToGlobalMapping(A->cmap,cmapping);CHKERRQ(ierr);
*/
/* Create the local matrix A */
ierr = ISLocalToGlobalMappingGetSize(rmapping,&n);CHKERRQ(ierr);
ierr = ISLocalToGlobalMappingGetBlockSize(rmapping,&bs);CHKERRQ(ierr);
ierr = MatCreate(PETSC_COMM_SELF,&is->A);CHKERRQ(ierr);
if (bs > 1) {
ierr = MatSetType(is->A,MATSEQBAIJ);CHKERRQ(ierr);
} else {
ierr = MatSetType(is->A,MATSEQAIJ);CHKERRQ(ierr);
}
ierr = MatSetSizes(is->A,n,n,n,n);CHKERRQ(ierr);
ierr = MatSetBlockSize(is->A,bs);CHKERRQ(ierr);
ierr = MatSetOptionsPrefix(is->A,((PetscObject)A)->prefix);CHKERRQ(ierr);
ierr = MatAppendOptionsPrefix(is->A,"is_");CHKERRQ(ierr);
ierr = MatSetFromOptions(is->A);CHKERRQ(ierr);
/* Create the local work vectors */
ierr = VecCreate(PETSC_COMM_SELF,&is->x);CHKERRQ(ierr);
ierr = VecSetBlockSize(is->x,bs);CHKERRQ(ierr);
ierr = VecSetSizes(is->x,n,n);CHKERRQ(ierr);
ierr = VecSetOptionsPrefix(is->x,((PetscObject)A)->prefix);CHKERRQ(ierr);
ierr = VecAppendOptionsPrefix(is->x,"is_");CHKERRQ(ierr);
ierr = VecSetFromOptions(is->x);CHKERRQ(ierr);
ierr = VecDuplicate(is->x,&is->y);CHKERRQ(ierr);
/* setup the global to local scatter */
ierr = ISCreateStride(PETSC_COMM_SELF,n,0,1,&to);CHKERRQ(ierr);
ierr = ISLocalToGlobalMappingApplyIS(rmapping,to,&from);CHKERRQ(ierr);
ierr = MatCreateVecs(A,&global,NULL);CHKERRQ(ierr);
ierr = VecScatterCreate(global,from,is->x,to,&is->ctx);CHKERRQ(ierr);
ierr = VecDestroy(&global);CHKERRQ(ierr);
ierr = ISDestroy(&to);CHKERRQ(ierr);
ierr = ISDestroy(&from);CHKERRQ(ierr);
PetscFunctionReturn(0);
}
PetscErrorCode MatSetLocalToGlobalMapping_IS(Mat A,ISLocalToGlobalMapping rmapping,ISLocalToGlobalMapping cmapping)
{
PetscErrorCode ierr;
PetscInt nr,rbs,nc,cbs;
Mat_IS *is = (Mat_IS*)A->data;
IS from,to;
Vec cglobal,rglobal;
PetscFunctionBegin;
PetscCheckSameComm(A,1,rmapping,2);
PetscCheckSameComm(A,1,cmapping,3);
/* Destroy any previous data */
ierr = VecDestroy(&is->x);CHKERRQ(ierr);
ierr = VecDestroy(&is->y);CHKERRQ(ierr);
ierr = VecScatterDestroy(&is->rctx);CHKERRQ(ierr);
ierr = VecScatterDestroy(&is->cctx);CHKERRQ(ierr);
ierr = MatDestroy(&is->A);CHKERRQ(ierr);
ierr = PetscSFDestroy(&is->sf);CHKERRQ(ierr);
ierr = PetscFree2(is->sf_rootdata,is->sf_leafdata);CHKERRQ(ierr);
/* Setup Layout and set local to global maps */
ierr = PetscLayoutSetUp(A->rmap);CHKERRQ(ierr);
ierr = PetscLayoutSetUp(A->cmap);CHKERRQ(ierr);
ierr = PetscLayoutSetISLocalToGlobalMapping(A->rmap,rmapping);CHKERRQ(ierr);
ierr = PetscLayoutSetISLocalToGlobalMapping(A->cmap,cmapping);CHKERRQ(ierr);
/* Create the local matrix A */
ierr = ISLocalToGlobalMappingGetSize(rmapping,&nr);CHKERRQ(ierr);
ierr = ISLocalToGlobalMappingGetBlockSize(rmapping,&rbs);CHKERRQ(ierr);
ierr = ISLocalToGlobalMappingGetSize(cmapping,&nc);CHKERRQ(ierr);
ierr = ISLocalToGlobalMappingGetBlockSize(cmapping,&cbs);CHKERRQ(ierr);
ierr = MatCreate(PETSC_COMM_SELF,&is->A);CHKERRQ(ierr);
ierr = MatSetType(is->A,MATAIJ);CHKERRQ(ierr);
ierr = MatSetSizes(is->A,nr,nc,nr,nc);CHKERRQ(ierr);
ierr = MatSetBlockSizes(is->A,rbs,cbs);CHKERRQ(ierr);
ierr = MatSetOptionsPrefix(is->A,((PetscObject)A)->prefix);CHKERRQ(ierr);
ierr = MatAppendOptionsPrefix(is->A,"is_");CHKERRQ(ierr);
ierr = MatSetFromOptions(is->A);CHKERRQ(ierr);
/* Create the local work vectors */
ierr = MatCreateVecs(is->A,&is->x,&is->y);CHKERRQ(ierr);
/* setup the global to local scatters */
ierr = MatCreateVecs(A,&cglobal,&rglobal);CHKERRQ(ierr);
ierr = ISCreateStride(PETSC_COMM_SELF,nr,0,1,&to);CHKERRQ(ierr);
ierr = ISLocalToGlobalMappingApplyIS(rmapping,to,&from);CHKERRQ(ierr);
ierr = VecScatterCreate(rglobal,from,is->y,to,&is->rctx);CHKERRQ(ierr);
if (rmapping != cmapping) {
ierr = ISDestroy(&to);CHKERRQ(ierr);
ierr = ISDestroy(&from);CHKERRQ(ierr);
ierr = ISCreateStride(PETSC_COMM_SELF,nc,0,1,&to);CHKERRQ(ierr);
ierr = ISLocalToGlobalMappingApplyIS(cmapping,to,&from);CHKERRQ(ierr);
ierr = VecScatterCreate(cglobal,from,is->x,to,&is->cctx);CHKERRQ(ierr);
} else {
ierr = PetscObjectReference((PetscObject)is->rctx);CHKERRQ(ierr);
is->cctx = is->rctx;
}
ierr = VecDestroy(&rglobal);CHKERRQ(ierr);
ierr = VecDestroy(&cglobal);CHKERRQ(ierr);
ierr = ISDestroy(&to);CHKERRQ(ierr);
ierr = ISDestroy(&from);CHKERRQ(ierr);
PetscFunctionReturn(0);
}
PetscErrorCode PCBDDCSetupFETIDPMatContext(FETIDPMat_ctx fetidpmat_ctx )
{
PetscErrorCode ierr;
PC_IS *pcis=(PC_IS*)fetidpmat_ctx->pc->data;
PC_BDDC *pcbddc=(PC_BDDC*)fetidpmat_ctx->pc->data;
PCBDDCGraph mat_graph=pcbddc->mat_graph;
Mat_IS *matis = (Mat_IS*)fetidpmat_ctx->pc->pmat->data;
MPI_Comm comm;
Mat ScalingMat;
Vec lambda_global;
IS IS_l2g_lambda;
IS subset,subset_mult,subset_n;
PetscBool skip_node,fully_redundant;
PetscInt i,j,k,s,n_boundary_dofs,n_global_lambda,n_vertices,partial_sum;
PetscInt cum,n_local_lambda,n_lambda_for_dof,dual_size,n_neg_values,n_pos_values;
PetscMPIInt rank,size,buf_size,neigh;
PetscScalar scalar_value;
PetscInt *vertex_indices;
PetscInt *dual_dofs_boundary_indices,*aux_local_numbering_1;
const PetscInt *aux_global_numbering;
PetscInt *aux_sums,*cols_B_delta,*l2g_indices;
PetscScalar *array,*scaling_factors,*vals_B_delta;
PetscInt *aux_local_numbering_2;
/* For communication of scaling factors */
PetscInt *ptrs_buffer,neigh_position;
PetscScalar **all_factors,*send_buffer,*recv_buffer;
MPI_Request *send_reqs,*recv_reqs;
/* tests */
Vec test_vec;
PetscBool test_fetidp;
PetscViewer viewer;
PetscFunctionBegin;
ierr = PetscObjectGetComm((PetscObject)(fetidpmat_ctx->pc),&comm);CHKERRQ(ierr);
ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr);
ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr);
/* Default type of lagrange multipliers is non-redundant */
fully_redundant = fetidpmat_ctx->fully_redundant;
/* Evaluate local and global number of lagrange multipliers */
ierr = VecSet(pcis->vec1_N,0.0);CHKERRQ(ierr);
n_local_lambda = 0;
partial_sum = 0;
n_boundary_dofs = 0;
s = 0;
/* Get Vertices used to define the BDDC */
n_vertices = pcbddc->n_vertices;
vertex_indices = pcbddc->local_primal_ref_node;
dual_size = pcis->n_B-n_vertices;
ierr = PetscMalloc1(dual_size,&dual_dofs_boundary_indices);CHKERRQ(ierr);
ierr = PetscMalloc1(dual_size,&aux_local_numbering_1);CHKERRQ(ierr);
ierr = PetscMalloc1(dual_size,&aux_local_numbering_2);CHKERRQ(ierr);
ierr = VecGetArray(pcis->vec1_N,&array);CHKERRQ(ierr);
for (i=0;i<pcis->n;i++){
j = mat_graph->count[i]; /* RECALL: mat_graph->count[i] does not count myself */
if ( j > 0 ) {
n_boundary_dofs++;
}
skip_node = PETSC_FALSE;
if ( s < n_vertices && vertex_indices[s]==i) { /* it works for a sorted set of vertices */
skip_node = PETSC_TRUE;
s++;
}
if (j < 1) {
skip_node = PETSC_TRUE;
}
if ( !skip_node ) {
if (fully_redundant) {
/* fully redundant set of lagrange multipliers */
n_lambda_for_dof = (j*(j+1))/2;
} else {
n_lambda_for_dof = j;
}
n_local_lambda += j;
/* needed to evaluate global number of lagrange multipliers */
array[i]=(1.0*n_lambda_for_dof)/(j+1.0); /* already scaled for the next global sum */
/* store some data needed */
dual_dofs_boundary_indices[partial_sum] = n_boundary_dofs-1;
aux_local_numbering_1[partial_sum] = i;
aux_local_numbering_2[partial_sum] = n_lambda_for_dof;
partial_sum++;
}
}
ierr = VecRestoreArray(pcis->vec1_N,&array);CHKERRQ(ierr);
ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr);
ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr);
ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr);
ierr = VecSum(pcis->vec1_global,&scalar_value);CHKERRQ(ierr);
fetidpmat_ctx->n_lambda = (PetscInt)PetscRealPart(scalar_value);
/* compute global ordering of lagrange multipliers and associate l2g map */
ierr = ISCreateGeneral(comm,partial_sum,aux_local_numbering_1,PETSC_COPY_VALUES,&subset_n);CHKERRQ(ierr);
ierr = ISLocalToGlobalMappingApplyIS(pcis->mapping,subset_n,&subset);CHKERRQ(ierr);
ierr = ISDestroy(&subset_n);CHKERRQ(ierr);
ierr = ISCreateGeneral(comm,partial_sum,aux_local_numbering_2,PETSC_OWN_POINTER,&subset_mult);CHKERRQ(ierr);
ierr = ISRenumber(subset,subset_mult,&i,&subset_n);CHKERRQ(ierr);
ierr = ISDestroy(&subset);CHKERRQ(ierr);
if (i != fetidpmat_ctx->n_lambda) SETERRQ2(PETSC_COMM_WORLD,PETSC_ERR_PLIB,"Global number of multipliers mismatch! (%d!=%d)\n",fetidpmat_ctx->n_lambda,i);
/* init data for scaling factors exchange */
partial_sum = 0;
ierr = PetscMalloc1(pcis->n_neigh,&ptrs_buffer);CHKERRQ(ierr);
ierr = PetscMalloc1(pcis->n_neigh-1,&send_reqs);CHKERRQ(ierr);
ierr = PetscMalloc1(pcis->n_neigh-1,&recv_reqs);CHKERRQ(ierr);
ierr = PetscMalloc1(pcis->n,&all_factors);CHKERRQ(ierr);
ptrs_buffer[0]=0;
for (i=1;i<pcis->n_neigh;i++) {
partial_sum += pcis->n_shared[i];
ptrs_buffer[i] = ptrs_buffer[i-1]+pcis->n_shared[i];
}
ierr = PetscMalloc1(partial_sum,&send_buffer);CHKERRQ(ierr);
ierr = PetscMalloc1(partial_sum,&recv_buffer);CHKERRQ(ierr);
ierr = PetscMalloc1(partial_sum,&all_factors[0]);CHKERRQ(ierr);
for (i=0;i<pcis->n-1;i++) {
j = mat_graph->count[i];
all_factors[i+1]=all_factors[i]+j;
}
/* scatter B scaling to N vec */
ierr = VecScatterBegin(pcis->N_to_B,pcis->D,pcis->vec1_N,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr);
ierr = VecScatterEnd(pcis->N_to_B,pcis->D,pcis->vec1_N,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr);
/* communications */
ierr = VecGetArrayRead(pcis->vec1_N,(const PetscScalar**)&array);CHKERRQ(ierr);
for (i=1;i<pcis->n_neigh;i++) {
for (j=0;j<pcis->n_shared[i];j++) {
send_buffer[ptrs_buffer[i-1]+j]=array[pcis->shared[i][j]];
}
ierr = PetscMPIIntCast(ptrs_buffer[i]-ptrs_buffer[i-1],&buf_size);CHKERRQ(ierr);
ierr = PetscMPIIntCast(pcis->neigh[i],&neigh);CHKERRQ(ierr);
ierr = MPI_Isend(&send_buffer[ptrs_buffer[i-1]],buf_size,MPIU_SCALAR,neigh,0,comm,&send_reqs[i-1]);CHKERRQ(ierr);
ierr = MPI_Irecv(&recv_buffer[ptrs_buffer[i-1]],buf_size,MPIU_SCALAR,neigh,0,comm,&recv_reqs[i-1]);CHKERRQ(ierr);
}
ierr = VecRestoreArrayRead(pcis->vec1_N,(const PetscScalar**)&array);CHKERRQ(ierr);
ierr = MPI_Waitall((pcis->n_neigh-1),recv_reqs,MPI_STATUSES_IGNORE);CHKERRQ(ierr);
/* put values in correct places */
for (i=1;i<pcis->n_neigh;i++) {
for (j=0;j<pcis->n_shared[i];j++) {
k = pcis->shared[i][j];
neigh_position = 0;
while(mat_graph->neighbours_set[k][neigh_position] != pcis->neigh[i]) {neigh_position++;}
all_factors[k][neigh_position]=recv_buffer[ptrs_buffer[i-1]+j];
}
}
ierr = MPI_Waitall((pcis->n_neigh-1),send_reqs,MPI_STATUSES_IGNORE);CHKERRQ(ierr);
ierr = PetscFree(send_reqs);CHKERRQ(ierr);
ierr = PetscFree(recv_reqs);CHKERRQ(ierr);
ierr = PetscFree(send_buffer);CHKERRQ(ierr);
ierr = PetscFree(recv_buffer);CHKERRQ(ierr);
ierr = PetscFree(ptrs_buffer);CHKERRQ(ierr);
/* Compute B and B_delta (local actions) */
ierr = PetscMalloc1(pcis->n_neigh,&aux_sums);CHKERRQ(ierr);
ierr = PetscMalloc1(n_local_lambda,&l2g_indices);CHKERRQ(ierr);
ierr = PetscMalloc1(n_local_lambda,&vals_B_delta);CHKERRQ(ierr);
ierr = PetscMalloc1(n_local_lambda,&cols_B_delta);CHKERRQ(ierr);
ierr = PetscMalloc1(n_local_lambda,&scaling_factors);CHKERRQ(ierr);
ierr = ISGetIndices(subset_n,&aux_global_numbering);CHKERRQ(ierr);
partial_sum=0;
cum = 0;
for (i=0;i<dual_size;i++) {
n_global_lambda = aux_global_numbering[cum];
j = mat_graph->count[aux_local_numbering_1[i]];
aux_sums[0]=0;
for (s=1;s<j;s++) {
aux_sums[s]=aux_sums[s-1]+j-s+1;
}
array = all_factors[aux_local_numbering_1[i]];
n_neg_values = 0;
while(n_neg_values < j && mat_graph->neighbours_set[aux_local_numbering_1[i]][n_neg_values] < rank) {n_neg_values++;}
n_pos_values = j - n_neg_values;
if (fully_redundant) {
for (s=0;s<n_neg_values;s++) {
l2g_indices [partial_sum+s]=aux_sums[s]+n_neg_values-s-1+n_global_lambda;
cols_B_delta [partial_sum+s]=dual_dofs_boundary_indices[i];
vals_B_delta [partial_sum+s]=-1.0;
scaling_factors[partial_sum+s]=array[s];
}
for (s=0;s<n_pos_values;s++) {
l2g_indices [partial_sum+s+n_neg_values]=aux_sums[n_neg_values]+s+n_global_lambda;
cols_B_delta [partial_sum+s+n_neg_values]=dual_dofs_boundary_indices[i];
vals_B_delta [partial_sum+s+n_neg_values]=1.0;
scaling_factors[partial_sum+s+n_neg_values]=array[s+n_neg_values];
}
partial_sum += j;
} else {
/* l2g_indices and default cols and vals of B_delta */
for (s=0;s<j;s++) {
l2g_indices [partial_sum+s]=n_global_lambda+s;
cols_B_delta [partial_sum+s]=dual_dofs_boundary_indices[i];
vals_B_delta [partial_sum+s]=0.0;
}
/* B_delta */
if ( n_neg_values > 0 ) { /* there's a rank next to me to the left */
vals_B_delta [partial_sum+n_neg_values-1]=-1.0;
}
if ( n_neg_values < j ) { /* there's a rank next to me to the right */
vals_B_delta [partial_sum+n_neg_values]=1.0;
}
/* scaling as in Klawonn-Widlund 1999*/
for (s=0;s<n_neg_values;s++) {
scalar_value = 0.0;
for (k=0;k<s+1;k++) {
scalar_value += array[k];
}
scaling_factors[partial_sum+s] = -scalar_value;
}
for (s=0;s<n_pos_values;s++) {
scalar_value = 0.0;
for (k=s+n_neg_values;k<j;k++) {
scalar_value += array[k];
}
scaling_factors[partial_sum+s+n_neg_values] = scalar_value;
}
partial_sum += j;
}
cum += aux_local_numbering_2[i];
}
ierr = ISRestoreIndices(subset_n,&aux_global_numbering);CHKERRQ(ierr);
ierr = ISDestroy(&subset_mult);CHKERRQ(ierr);
ierr = ISDestroy(&subset_n);CHKERRQ(ierr);
ierr = PetscFree(aux_sums);CHKERRQ(ierr);
ierr = PetscFree(aux_local_numbering_1);CHKERRQ(ierr);
ierr = PetscFree(dual_dofs_boundary_indices);CHKERRQ(ierr);
ierr = PetscFree(all_factors[0]);CHKERRQ(ierr);
ierr = PetscFree(all_factors);CHKERRQ(ierr);
/* Local to global mapping of fetidpmat */
ierr = VecCreate(PETSC_COMM_SELF,&fetidpmat_ctx->lambda_local);CHKERRQ(ierr);
ierr = VecSetSizes(fetidpmat_ctx->lambda_local,n_local_lambda,n_local_lambda);CHKERRQ(ierr);
ierr = VecSetType(fetidpmat_ctx->lambda_local,VECSEQ);CHKERRQ(ierr);
ierr = VecCreate(comm,&lambda_global);CHKERRQ(ierr);
ierr = VecSetSizes(lambda_global,PETSC_DECIDE,fetidpmat_ctx->n_lambda);CHKERRQ(ierr);
ierr = VecSetType(lambda_global,VECMPI);CHKERRQ(ierr);
ierr = ISCreateGeneral(comm,n_local_lambda,l2g_indices,PETSC_OWN_POINTER,&IS_l2g_lambda);CHKERRQ(ierr);
ierr = VecScatterCreate(fetidpmat_ctx->lambda_local,(IS)0,lambda_global,IS_l2g_lambda,&fetidpmat_ctx->l2g_lambda);CHKERRQ(ierr);
ierr = ISDestroy(&IS_l2g_lambda);CHKERRQ(ierr);
/* Create local part of B_delta */
ierr = MatCreate(PETSC_COMM_SELF,&fetidpmat_ctx->B_delta);CHKERRQ(ierr);
ierr = MatSetSizes(fetidpmat_ctx->B_delta,n_local_lambda,pcis->n_B,n_local_lambda,pcis->n_B);CHKERRQ(ierr);
ierr = MatSetType(fetidpmat_ctx->B_delta,MATSEQAIJ);CHKERRQ(ierr);
ierr = MatSeqAIJSetPreallocation(fetidpmat_ctx->B_delta,1,NULL);CHKERRQ(ierr);
ierr = MatSetOption(fetidpmat_ctx->B_delta,MAT_IGNORE_ZERO_ENTRIES,PETSC_TRUE);CHKERRQ(ierr);
for (i=0;i<n_local_lambda;i++) {
ierr = MatSetValue(fetidpmat_ctx->B_delta,i,cols_B_delta[i],vals_B_delta[i],INSERT_VALUES);CHKERRQ(ierr);
}
ierr = PetscFree(vals_B_delta);CHKERRQ(ierr);
ierr = MatAssemblyBegin(fetidpmat_ctx->B_delta,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
ierr = MatAssemblyEnd (fetidpmat_ctx->B_delta,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
if (fully_redundant) {
ierr = MatCreate(PETSC_COMM_SELF,&ScalingMat);CHKERRQ(ierr);
ierr = MatSetSizes(ScalingMat,n_local_lambda,n_local_lambda,n_local_lambda,n_local_lambda);CHKERRQ(ierr);
ierr = MatSetType(ScalingMat,MATSEQAIJ);CHKERRQ(ierr);
ierr = MatSeqAIJSetPreallocation(ScalingMat,1,NULL);CHKERRQ(ierr);
for (i=0;i<n_local_lambda;i++) {
ierr = MatSetValue(ScalingMat,i,i,scaling_factors[i],INSERT_VALUES);CHKERRQ(ierr);
}
ierr = MatAssemblyBegin(ScalingMat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
ierr = MatAssemblyEnd (ScalingMat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
ierr = MatMatMult(ScalingMat,fetidpmat_ctx->B_delta,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&fetidpmat_ctx->B_Ddelta);CHKERRQ(ierr);
ierr = MatDestroy(&ScalingMat);CHKERRQ(ierr);
} else {
ierr = MatCreate(PETSC_COMM_SELF,&fetidpmat_ctx->B_Ddelta);CHKERRQ(ierr);
ierr = MatSetSizes(fetidpmat_ctx->B_Ddelta,n_local_lambda,pcis->n_B,n_local_lambda,pcis->n_B);CHKERRQ(ierr);
ierr = MatSetType(fetidpmat_ctx->B_Ddelta,MATSEQAIJ);CHKERRQ(ierr);
ierr = MatSeqAIJSetPreallocation(fetidpmat_ctx->B_Ddelta,1,NULL);CHKERRQ(ierr);
for (i=0;i<n_local_lambda;i++) {
ierr = MatSetValue(fetidpmat_ctx->B_Ddelta,i,cols_B_delta[i],scaling_factors[i],INSERT_VALUES);CHKERRQ(ierr);
}
ierr = MatAssemblyBegin(fetidpmat_ctx->B_Ddelta,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
ierr = MatAssemblyEnd (fetidpmat_ctx->B_Ddelta,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
}
ierr = PetscFree(scaling_factors);CHKERRQ(ierr);
ierr = PetscFree(cols_B_delta);CHKERRQ(ierr);
/* Create some vectors needed by fetidp */
ierr = VecDuplicate(pcis->vec1_B,&fetidpmat_ctx->temp_solution_B);CHKERRQ(ierr);
ierr = VecDuplicate(pcis->vec1_D,&fetidpmat_ctx->temp_solution_D);CHKERRQ(ierr);
test_fetidp = PETSC_FALSE;
ierr = PetscOptionsGetBool(NULL,NULL,"-fetidp_check",&test_fetidp,NULL);CHKERRQ(ierr);
if (test_fetidp && !pcbddc->use_deluxe_scaling) {
PetscReal real_value;
ierr = PetscViewerASCIIGetStdout(comm,&viewer);CHKERRQ(ierr);
ierr = PetscViewerASCIIPushSynchronized(viewer);CHKERRQ(ierr);
ierr = PetscViewerASCIIPrintf(viewer,"----------FETI_DP TESTS--------------\n");CHKERRQ(ierr);
ierr = PetscViewerASCIIPrintf(viewer,"All tests should return zero!\n");CHKERRQ(ierr);
ierr = PetscViewerASCIIPrintf(viewer,"FETIDP MAT context in the ");CHKERRQ(ierr);
if (fully_redundant) {
ierr = PetscViewerASCIIPrintf(viewer,"fully redundant case for lagrange multipliers.\n");CHKERRQ(ierr);
} else {
ierr = PetscViewerASCIIPrintf(viewer,"Non-fully redundant case for lagrange multiplier.\n");CHKERRQ(ierr);
}
ierr = PetscViewerFlush(viewer);CHKERRQ(ierr);
/******************************************************************/
/* TEST A/B: Test numbering of global lambda dofs */
/******************************************************************/
ierr = VecDuplicate(fetidpmat_ctx->lambda_local,&test_vec);CHKERRQ(ierr);
ierr = VecSet(lambda_global,1.0);CHKERRQ(ierr);
ierr = VecSet(test_vec,1.0);CHKERRQ(ierr);
ierr = VecScatterBegin(fetidpmat_ctx->l2g_lambda,lambda_global,fetidpmat_ctx->lambda_local,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr);
ierr = VecScatterEnd (fetidpmat_ctx->l2g_lambda,lambda_global,fetidpmat_ctx->lambda_local,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr);
scalar_value = -1.0;
ierr = VecAXPY(test_vec,scalar_value,fetidpmat_ctx->lambda_local);CHKERRQ(ierr);
ierr = VecNorm(test_vec,NORM_INFINITY,&real_value);CHKERRQ(ierr);
ierr = VecDestroy(&test_vec);CHKERRQ(ierr);
ierr = PetscViewerASCIISynchronizedPrintf(viewer,"A[%04d]: CHECK glob to loc: % 1.14e\n",rank,real_value);CHKERRQ(ierr);
ierr = PetscViewerFlush(viewer);CHKERRQ(ierr);
if (fully_redundant) {
ierr = VecSet(lambda_global,0.0);CHKERRQ(ierr);
ierr = VecSet(fetidpmat_ctx->lambda_local,0.5);CHKERRQ(ierr);
ierr = VecScatterBegin(fetidpmat_ctx->l2g_lambda,fetidpmat_ctx->lambda_local,lambda_global,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
ierr = VecScatterEnd (fetidpmat_ctx->l2g_lambda,fetidpmat_ctx->lambda_local,lambda_global,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
ierr = VecSum(lambda_global,&scalar_value);CHKERRQ(ierr);
ierr = PetscViewerASCIISynchronizedPrintf(viewer,"B[%04d]: CHECK loc to glob: % 1.14e\n",rank,PetscRealPart(scalar_value)-fetidpmat_ctx->n_lambda);CHKERRQ(ierr);
ierr = PetscViewerFlush(viewer);CHKERRQ(ierr);
}
/******************************************************************/
/* TEST C: It should holds B_delta*w=0, w\in\widehat{W} */
/* This is the meaning of the B matrix */
/******************************************************************/
ierr = VecSetRandom(pcis->vec1_N,NULL);CHKERRQ(ierr);
ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr);
ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr);
ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr);
ierr = VecScatterBegin(matis->rctx,pcis->vec1_global,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
ierr = VecScatterEnd(matis->rctx,pcis->vec1_global,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
ierr = VecScatterBegin(pcis->N_to_B,pcis->vec1_N,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
ierr = VecScatterEnd (pcis->N_to_B,pcis->vec1_N,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
/* Action of B_delta */
ierr = MatMult(fetidpmat_ctx->B_delta,pcis->vec1_B,fetidpmat_ctx->lambda_local);CHKERRQ(ierr);
ierr = VecSet(lambda_global,0.0);CHKERRQ(ierr);
ierr = VecScatterBegin(fetidpmat_ctx->l2g_lambda,fetidpmat_ctx->lambda_local,lambda_global,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
ierr = VecScatterEnd (fetidpmat_ctx->l2g_lambda,fetidpmat_ctx->lambda_local,lambda_global,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
ierr = VecNorm(lambda_global,NORM_INFINITY,&real_value);CHKERRQ(ierr);
ierr = PetscViewerASCIIPrintf(viewer,"C[coll]: CHECK infty norm of B_delta*w (w continuous): % 1.14e\n",real_value);CHKERRQ(ierr);
ierr = PetscViewerFlush(viewer);CHKERRQ(ierr);
/******************************************************************/
/* TEST D: It should holds E_Dw = w - P_Dw w\in\widetilde{W} */
/* E_D = R_D^TR */
/* P_D = B_{D,delta}^T B_{delta} */
/* eq.44 Mandel Tezaur and Dohrmann 2005 */
/******************************************************************/
/* compute a random vector in \widetilde{W} */
ierr = VecSetRandom(pcis->vec1_N,NULL);CHKERRQ(ierr);
scalar_value = 0.0; /* set zero at vertices */
ierr = VecGetArray(pcis->vec1_N,&array);CHKERRQ(ierr);
for (i=0;i<n_vertices;i++) { array[vertex_indices[i]]=scalar_value; }
ierr = VecRestoreArray(pcis->vec1_N,&array);CHKERRQ(ierr);
/* store w for final comparison */
ierr = VecDuplicate(pcis->vec1_B,&test_vec);CHKERRQ(ierr);
ierr = VecScatterBegin(pcis->N_to_B,pcis->vec1_N,test_vec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
ierr = VecScatterEnd (pcis->N_to_B,pcis->vec1_N,test_vec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
/* Jump operator P_D : results stored in pcis->vec1_B */
ierr = VecScatterBegin(pcis->N_to_B,pcis->vec1_N,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
ierr = VecScatterEnd (pcis->N_to_B,pcis->vec1_N,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
/* Action of B_delta */
ierr = MatMult(fetidpmat_ctx->B_delta,pcis->vec1_B,fetidpmat_ctx->lambda_local);CHKERRQ(ierr);
ierr = VecSet(lambda_global,0.0);CHKERRQ(ierr);
ierr = VecScatterBegin(fetidpmat_ctx->l2g_lambda,fetidpmat_ctx->lambda_local,lambda_global,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
ierr = VecScatterEnd (fetidpmat_ctx->l2g_lambda,fetidpmat_ctx->lambda_local,lambda_global,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
/* Action of B_Ddelta^T */
ierr = VecScatterBegin(fetidpmat_ctx->l2g_lambda,lambda_global,fetidpmat_ctx->lambda_local,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr);
ierr = VecScatterEnd (fetidpmat_ctx->l2g_lambda,lambda_global,fetidpmat_ctx->lambda_local,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr);
ierr = MatMultTranspose(fetidpmat_ctx->B_Ddelta,fetidpmat_ctx->lambda_local,pcis->vec1_B);CHKERRQ(ierr);
/* Average operator E_D : results stored in pcis->vec2_B */
ierr = VecScatterBegin(pcis->N_to_B,pcis->vec1_N,pcis->vec2_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
ierr = VecScatterEnd (pcis->N_to_B,pcis->vec1_N,pcis->vec2_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
ierr = PCBDDCScalingExtension(fetidpmat_ctx->pc,pcis->vec2_B,pcis->vec1_global);CHKERRQ(ierr);
ierr = VecScatterBegin(pcis->global_to_B,pcis->vec1_global,pcis->vec2_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
ierr = VecScatterEnd (pcis->global_to_B,pcis->vec1_global,pcis->vec2_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
/* test E_D=I-P_D */
scalar_value = 1.0;
ierr = VecAXPY(pcis->vec1_B,scalar_value,pcis->vec2_B);CHKERRQ(ierr);
scalar_value = -1.0;
ierr = VecAXPY(pcis->vec1_B,scalar_value,test_vec);CHKERRQ(ierr);
ierr = VecNorm(pcis->vec1_B,NORM_INFINITY,&real_value);CHKERRQ(ierr);
ierr = VecDestroy(&test_vec);CHKERRQ(ierr);
ierr = PetscViewerASCIISynchronizedPrintf(viewer,"D[%04d] CHECK infty norm of E_D + P_D - I: % 1.14e\n",rank,real_value);CHKERRQ(ierr);
ierr = PetscViewerFlush(viewer);CHKERRQ(ierr);
/******************************************************************/
/* TEST E: It should holds R_D^TP_Dw=0 w\in\widetilde{W} */
/* eq.48 Mandel Tezaur and Dohrmann 2005 */
/******************************************************************/
ierr = VecSetRandom(pcis->vec1_N,NULL);CHKERRQ(ierr);
ierr = VecGetArray(pcis->vec1_N,&array);CHKERRQ(ierr);
scalar_value = 0.0; /* set zero at vertices */
for (i=0;i<n_vertices;i++) { array[vertex_indices[i]]=scalar_value; }
ierr = VecRestoreArray(pcis->vec1_N,&array);CHKERRQ(ierr);
/* Jump operator P_D : results stored in pcis->vec1_B */
ierr = VecScatterBegin(pcis->N_to_B,pcis->vec1_N,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
ierr = VecScatterEnd (pcis->N_to_B,pcis->vec1_N,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
/* Action of B_delta */
ierr = MatMult(fetidpmat_ctx->B_delta,pcis->vec1_B,fetidpmat_ctx->lambda_local);CHKERRQ(ierr);
ierr = VecSet(lambda_global,0.0);CHKERRQ(ierr);
ierr = VecScatterBegin(fetidpmat_ctx->l2g_lambda,fetidpmat_ctx->lambda_local,lambda_global,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
ierr = VecScatterEnd (fetidpmat_ctx->l2g_lambda,fetidpmat_ctx->lambda_local,lambda_global,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
/* Action of B_Ddelta^T */
ierr = VecScatterBegin(fetidpmat_ctx->l2g_lambda,lambda_global,fetidpmat_ctx->lambda_local,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr);
ierr = VecScatterEnd (fetidpmat_ctx->l2g_lambda,lambda_global,fetidpmat_ctx->lambda_local,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr);
ierr = MatMultTranspose(fetidpmat_ctx->B_Ddelta,fetidpmat_ctx->lambda_local,pcis->vec1_B);CHKERRQ(ierr);
/* scaling */
ierr = PCBDDCScalingExtension(fetidpmat_ctx->pc,pcis->vec1_B,pcis->vec1_global);CHKERRQ(ierr);
ierr = VecNorm(pcis->vec1_global,NORM_INFINITY,&real_value);CHKERRQ(ierr);
ierr = PetscViewerASCIIPrintf(viewer,"E[coll]: CHECK infty norm of R^T_D P_D: % 1.14e\n",real_value);CHKERRQ(ierr);
ierr = PetscViewerFlush(viewer);CHKERRQ(ierr);
if (!fully_redundant) {
/******************************************************************/
/* TEST F: It should holds B_{delta}B^T_{D,delta}=I */
/* Corollary thm 14 Mandel Tezaur and Dohrmann 2005 */
/******************************************************************/
ierr = VecDuplicate(lambda_global,&test_vec);CHKERRQ(ierr);
ierr = VecSetRandom(lambda_global,NULL);CHKERRQ(ierr);
/* Action of B_Ddelta^T */
ierr = VecScatterBegin(fetidpmat_ctx->l2g_lambda,lambda_global,fetidpmat_ctx->lambda_local,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr);
ierr = VecScatterEnd (fetidpmat_ctx->l2g_lambda,lambda_global,fetidpmat_ctx->lambda_local,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr);
ierr = MatMultTranspose(fetidpmat_ctx->B_Ddelta,fetidpmat_ctx->lambda_local,pcis->vec1_B);CHKERRQ(ierr);
/* Action of B_delta */
ierr = MatMult(fetidpmat_ctx->B_delta,pcis->vec1_B,fetidpmat_ctx->lambda_local);CHKERRQ(ierr);
ierr = VecSet(test_vec,0.0);CHKERRQ(ierr);
ierr = VecScatterBegin(fetidpmat_ctx->l2g_lambda,fetidpmat_ctx->lambda_local,test_vec,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
ierr = VecScatterEnd (fetidpmat_ctx->l2g_lambda,fetidpmat_ctx->lambda_local,test_vec,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
scalar_value = -1.0;
ierr = VecAXPY(lambda_global,scalar_value,test_vec);CHKERRQ(ierr);
ierr = VecNorm(lambda_global,NORM_INFINITY,&real_value);CHKERRQ(ierr);
ierr = PetscViewerASCIIPrintf(viewer,"E[coll]: CHECK infty norm of P^T_D - I: % 1.14e\n",real_value);CHKERRQ(ierr);
ierr = PetscViewerFlush(viewer);CHKERRQ(ierr);
ierr = PetscViewerFlush(viewer);CHKERRQ(ierr);
ierr = VecDestroy(&test_vec);CHKERRQ(ierr);
}
}
/* final cleanup */
ierr = VecDestroy(&lambda_global);CHKERRQ(ierr);
PetscFunctionReturn(0);
}
PetscErrorCode PCBDDCGraphSetUp(PCBDDCGraph graph, PetscInt custom_minimal_size, IS neumann_is, IS dirichlet_is, PetscInt n_ISForDofs, IS ISForDofs[], IS custom_primal_vertices)
{
IS subset,subset_n;
MPI_Comm comm;
const PetscInt *is_indices;
PetscInt n_neigh,*neigh,*n_shared,**shared,*queue_global;
PetscInt i,j,k,s,total_counts,nodes_touched,is_size;
PetscMPIInt commsize;
PetscBool same_set,mirrors_found;
PetscErrorCode ierr;
PetscFunctionBegin;
PetscValidLogicalCollectiveInt(graph->l2gmap,custom_minimal_size,2);
if (neumann_is) {
PetscValidHeaderSpecific(neumann_is,IS_CLASSID,3);
PetscCheckSameComm(graph->l2gmap,1,neumann_is,3);
}
graph->has_dirichlet = PETSC_FALSE;
if (dirichlet_is) {
PetscValidHeaderSpecific(dirichlet_is,IS_CLASSID,4);
PetscCheckSameComm(graph->l2gmap,1,dirichlet_is,4);
graph->has_dirichlet = PETSC_TRUE;
}
PetscValidLogicalCollectiveInt(graph->l2gmap,n_ISForDofs,5);
for (i=0;i<n_ISForDofs;i++) {
PetscValidHeaderSpecific(ISForDofs[i],IS_CLASSID,6);
PetscCheckSameComm(graph->l2gmap,1,ISForDofs[i],6);
}
if (custom_primal_vertices) {
PetscValidHeaderSpecific(custom_primal_vertices,IS_CLASSID,6);
PetscCheckSameComm(graph->l2gmap,1,custom_primal_vertices,7);
}
ierr = PetscObjectGetComm((PetscObject)(graph->l2gmap),&comm);CHKERRQ(ierr);
ierr = MPI_Comm_size(comm,&commsize);CHKERRQ(ierr);
/* custom_minimal_size */
graph->custom_minimal_size = custom_minimal_size;
/* get info l2gmap and allocate work vectors */
ierr = ISLocalToGlobalMappingGetInfo(graph->l2gmap,&n_neigh,&neigh,&n_shared,&shared);CHKERRQ(ierr);
/* check if we have any local periodic nodes (periodic BCs) */
mirrors_found = PETSC_FALSE;
if (graph->nvtxs && n_neigh) {
for (i=0; i<n_shared[0]; i++) graph->count[shared[0][i]] += 1;
for (i=0; i<n_shared[0]; i++) {
if (graph->count[shared[0][i]] > 1) {
mirrors_found = PETSC_TRUE;
break;
}
}
}
/* compute local mirrors (if any) */
if (mirrors_found) {
IS to,from;
PetscInt *local_indices,*global_indices;
ierr = ISCreateStride(PETSC_COMM_SELF,graph->nvtxs,0,1,&to);CHKERRQ(ierr);
ierr = ISLocalToGlobalMappingApplyIS(graph->l2gmap,to,&from);CHKERRQ(ierr);
/* get arrays of local and global indices */
ierr = PetscMalloc1(graph->nvtxs,&local_indices);CHKERRQ(ierr);
ierr = ISGetIndices(to,(const PetscInt**)&is_indices);CHKERRQ(ierr);
ierr = PetscMemcpy(local_indices,is_indices,graph->nvtxs*sizeof(PetscInt));CHKERRQ(ierr);
ierr = ISRestoreIndices(to,(const PetscInt**)&is_indices);CHKERRQ(ierr);
ierr = PetscMalloc1(graph->nvtxs,&global_indices);CHKERRQ(ierr);
ierr = ISGetIndices(from,(const PetscInt**)&is_indices);CHKERRQ(ierr);
ierr = PetscMemcpy(global_indices,is_indices,graph->nvtxs*sizeof(PetscInt));CHKERRQ(ierr);
ierr = ISRestoreIndices(from,(const PetscInt**)&is_indices);CHKERRQ(ierr);
/* allocate space for mirrors */
ierr = PetscMalloc2(graph->nvtxs,&graph->mirrors,graph->nvtxs,&graph->mirrors_set);CHKERRQ(ierr);
ierr = PetscMemzero(graph->mirrors,graph->nvtxs*sizeof(PetscInt));CHKERRQ(ierr);
graph->mirrors_set[0] = 0;
k=0;
for (i=0;i<n_shared[0];i++) {
j=shared[0][i];
if (graph->count[j] > 1) {
graph->mirrors[j]++;
k++;
}
}
/* allocate space for set of mirrors */
ierr = PetscMalloc1(k,&graph->mirrors_set[0]);CHKERRQ(ierr);
for (i=1;i<graph->nvtxs;i++)
graph->mirrors_set[i]=graph->mirrors_set[i-1]+graph->mirrors[i-1];
/* fill arrays */
ierr = PetscMemzero(graph->mirrors,graph->nvtxs*sizeof(PetscInt));CHKERRQ(ierr);
for (j=0;j<n_shared[0];j++) {
i=shared[0][j];
if (graph->count[i] > 1)
graph->mirrors_set[i][graph->mirrors[i]++]=global_indices[i];
}
ierr = PetscSortIntWithArray(graph->nvtxs,global_indices,local_indices);CHKERRQ(ierr);
for (i=0;i<graph->nvtxs;i++) {
if (graph->mirrors[i] > 0) {
ierr = PetscFindInt(graph->mirrors_set[i][0],graph->nvtxs,global_indices,&k);CHKERRQ(ierr);
j = global_indices[k];
while ( k > 0 && global_indices[k-1] == j) k--;
for (j=0;j<graph->mirrors[i];j++) {
graph->mirrors_set[i][j]=local_indices[k+j];
}
ierr = PetscSortInt(graph->mirrors[i],graph->mirrors_set[i]);CHKERRQ(ierr);
}
}
ierr = PetscFree(local_indices);CHKERRQ(ierr);
ierr = PetscFree(global_indices);CHKERRQ(ierr);
ierr = ISDestroy(&to);CHKERRQ(ierr);
ierr = ISDestroy(&from);CHKERRQ(ierr);
}
ierr = PetscMemzero(graph->count,graph->nvtxs*sizeof(*graph->count));CHKERRQ(ierr);
/* Count total number of neigh per node */
k = 0;
for (i=1;i<n_neigh;i++) {
k += n_shared[i];
for (j=0;j<n_shared[i];j++) {
graph->count[shared[i][j]] += 1;
}
}
/* Allocate space for storing the set of neighbours for each node */
if (graph->nvtxs) {
ierr = PetscMalloc1(k,&graph->neighbours_set[0]);CHKERRQ(ierr);
}
for (i=1;i<graph->nvtxs;i++) { /* dont count myself */
graph->neighbours_set[i]=graph->neighbours_set[i-1]+graph->count[i-1];
}
/* Get information for sharing subdomains */
ierr = PetscMemzero(graph->count,graph->nvtxs*sizeof(*graph->count));CHKERRQ(ierr);
for (i=1;i<n_neigh;i++) { /* dont count myself */
s = n_shared[i];
for (j=0;j<s;j++) {
k = shared[i][j];
graph->neighbours_set[k][graph->count[k]] = neigh[i];
graph->count[k] += 1;
}
}
/* sort set of sharing subdomains */
for (i=0;i<graph->nvtxs;i++) {
ierr = PetscSortRemoveDupsInt(&graph->count[i],graph->neighbours_set[i]);CHKERRQ(ierr);
}
/* free memory allocated by ISLocalToGlobalMappingGetInfo */
ierr = ISLocalToGlobalMappingRestoreInfo(graph->l2gmap,&n_neigh,&neigh,&n_shared,&shared);CHKERRQ(ierr);
/*
Get info for dofs splitting
User can specify just a subset; an additional field is considered as a complementary field
*/
for (i=0;i<graph->nvtxs;i++) graph->which_dof[i] = n_ISForDofs; /* by default a dof belongs to the complement set */
for (i=0;i<n_ISForDofs;i++) {
ierr = ISGetLocalSize(ISForDofs[i],&is_size);CHKERRQ(ierr);
ierr = ISGetIndices(ISForDofs[i],(const PetscInt**)&is_indices);CHKERRQ(ierr);
for (j=0;j<is_size;j++) {
if (is_indices[j] > -1 && is_indices[j] < graph->nvtxs) { /* out of bounds indices (if any) are skipped */
graph->which_dof[is_indices[j]] = i;
}
}
ierr = ISRestoreIndices(ISForDofs[i],(const PetscInt**)&is_indices);CHKERRQ(ierr);
}
/* Take into account Neumann nodes */
if (neumann_is) {
ierr = ISGetLocalSize(neumann_is,&is_size);CHKERRQ(ierr);
ierr = ISGetIndices(neumann_is,(const PetscInt**)&is_indices);CHKERRQ(ierr);
for (i=0;i<is_size;i++) {
if (is_indices[i] > -1 && is_indices[i] < graph->nvtxs) { /* out of bounds indices (if any) are skipped */
graph->special_dof[is_indices[i]] = PCBDDCGRAPH_NEUMANN_MARK;
}
}
ierr = ISRestoreIndices(neumann_is,(const PetscInt**)&is_indices);CHKERRQ(ierr);
}
/* Take into account Dirichlet nodes (they overwrite any neumann boundary mark previously set) */
if (dirichlet_is) {
ierr = ISGetLocalSize(dirichlet_is,&is_size);CHKERRQ(ierr);
ierr = ISGetIndices(dirichlet_is,(const PetscInt**)&is_indices);CHKERRQ(ierr);
for (i=0;i<is_size;i++){
if (is_indices[i] > -1 && is_indices[i] < graph->nvtxs) { /* out of bounds indices (if any) are skipped */
if (commsize > graph->commsizelimit) { /* dirichlet nodes treated as internal */
ierr = PetscBTSet(graph->touched,is_indices[i]);CHKERRQ(ierr);
graph->subset[is_indices[i]] = 0;
}
graph->special_dof[is_indices[i]] = PCBDDCGRAPH_DIRICHLET_MARK;
}
}
ierr = ISRestoreIndices(dirichlet_is,(const PetscInt**)&is_indices);CHKERRQ(ierr);
}
/* mark local periodic nodes (if any) and adapt CSR graph (if any) */
if (graph->mirrors) {
for (i=0;i<graph->nvtxs;i++)
if (graph->mirrors[i])
graph->special_dof[i] = PCBDDCGRAPH_LOCAL_PERIODIC_MARK;
if (graph->xadj) {
PetscInt *new_xadj,*new_adjncy;
/* sort CSR graph */
for (i=0;i<graph->nvtxs;i++)
ierr = PetscSortInt(graph->xadj[i+1]-graph->xadj[i],&graph->adjncy[graph->xadj[i]]);CHKERRQ(ierr);
/* adapt local CSR graph in case of local periodicity */
k = 0;
for (i=0;i<graph->nvtxs;i++)
for (j=graph->xadj[i];j<graph->xadj[i+1];j++)
k += graph->mirrors[graph->adjncy[j]];
ierr = PetscMalloc1(graph->nvtxs+1,&new_xadj);CHKERRQ(ierr);
ierr = PetscMalloc1(k+graph->xadj[graph->nvtxs],&new_adjncy);CHKERRQ(ierr);
new_xadj[0] = 0;
for (i=0;i<graph->nvtxs;i++) {
k = graph->xadj[i+1]-graph->xadj[i];
ierr = PetscMemcpy(&new_adjncy[new_xadj[i]],&graph->adjncy[graph->xadj[i]],k*sizeof(PetscInt));CHKERRQ(ierr);
new_xadj[i+1] = new_xadj[i]+k;
for (j=graph->xadj[i];j<graph->xadj[i+1];j++) {
k = graph->mirrors[graph->adjncy[j]];
ierr = PetscMemcpy(&new_adjncy[new_xadj[i+1]],graph->mirrors_set[graph->adjncy[j]],k*sizeof(PetscInt));CHKERRQ(ierr);
new_xadj[i+1] += k;
}
k = new_xadj[i+1]-new_xadj[i];
ierr = PetscSortRemoveDupsInt(&k,&new_adjncy[new_xadj[i]]);CHKERRQ(ierr);
new_xadj[i+1] = new_xadj[i]+k;
}
/* set new CSR into graph */
ierr = PetscFree(graph->xadj);CHKERRQ(ierr);
ierr = PetscFree(graph->adjncy);CHKERRQ(ierr);
graph->xadj = new_xadj;
graph->adjncy = new_adjncy;
}
}
/* mark special nodes (if any) -> each will become a single node equivalence class */
if (custom_primal_vertices) {
ierr = ISGetLocalSize(custom_primal_vertices,&is_size);CHKERRQ(ierr);
ierr = ISGetIndices(custom_primal_vertices,(const PetscInt**)&is_indices);CHKERRQ(ierr);
for (i=0,j=0;i<is_size;i++){
if (is_indices[i] > -1 && is_indices[i] < graph->nvtxs && graph->special_dof[is_indices[i]] != PCBDDCGRAPH_DIRICHLET_MARK) { /* out of bounds indices (if any) are skipped */
graph->special_dof[is_indices[i]] = PCBDDCGRAPH_SPECIAL_MARK-j;
j++;
}
}
ierr = ISRestoreIndices(custom_primal_vertices,(const PetscInt**)&is_indices);CHKERRQ(ierr);
}
/* mark interior nodes (if commsize > graph->commsizelimit) as touched and belonging to partition number 0 */
if (commsize > graph->commsizelimit) {
for (i=0;i<graph->nvtxs;i++) {
if (!graph->count[i]) {
ierr = PetscBTSet(graph->touched,i);CHKERRQ(ierr);
graph->subset[i] = 0;
}
}
}
/* init graph structure and compute default subsets */
nodes_touched = 0;
for (i=0;i<graph->nvtxs;i++) {
if (PetscBTLookup(graph->touched,i)) {
nodes_touched++;
}
}
i = 0;
graph->ncc = 0;
total_counts = 0;
/* allocated space for queues */
if (commsize == graph->commsizelimit) {
ierr = PetscMalloc2(graph->nvtxs+1,&graph->cptr,graph->nvtxs,&graph->queue);CHKERRQ(ierr);
} else {
PetscInt nused = graph->nvtxs - nodes_touched;
ierr = PetscMalloc2(nused+1,&graph->cptr,nused,&graph->queue);CHKERRQ(ierr);
}
while (nodes_touched<graph->nvtxs) {
/* find first untouched node in local ordering */
while (PetscBTLookup(graph->touched,i)) i++;
ierr = PetscBTSet(graph->touched,i);CHKERRQ(ierr);
graph->subset[i] = graph->ncc+1;
graph->cptr[graph->ncc] = total_counts;
graph->queue[total_counts] = i;
total_counts++;
nodes_touched++;
/* now find all other nodes having the same set of sharing subdomains */
for (j=i+1;j<graph->nvtxs;j++) {
/* check for same number of sharing subdomains, dof number and same special mark */
if (!PetscBTLookup(graph->touched,j) && graph->count[i] == graph->count[j] && graph->which_dof[i] == graph->which_dof[j] && graph->special_dof[i] == graph->special_dof[j]) {
/* check for same set of sharing subdomains */
same_set = PETSC_TRUE;
for (k=0;k<graph->count[j];k++){
if (graph->neighbours_set[i][k] != graph->neighbours_set[j][k]) {
same_set = PETSC_FALSE;
}
}
/* I found a friend of mine */
if (same_set) {
ierr = PetscBTSet(graph->touched,j);CHKERRQ(ierr);
graph->subset[j] = graph->ncc+1;
nodes_touched++;
graph->queue[total_counts] = j;
total_counts++;
}
}
}
graph->ncc++;
}
/* set default number of subsets (at this point no info on csr and/or local_subs has been taken into account, so n_subsets = ncc */
graph->n_subsets = graph->ncc;
ierr = PetscMalloc1(graph->n_subsets,&graph->subset_ncc);CHKERRQ(ierr);
for (i=0;i<graph->n_subsets;i++) {
graph->subset_ncc[i] = 1;
}
/* final pointer */
graph->cptr[graph->ncc] = total_counts;
/* For consistency reasons (among neighbours), I need to sort (by global ordering) each connected component */
/* Get a reference node (min index in global ordering) for each subset for tagging messages */
ierr = PetscMalloc1(graph->ncc,&graph->subset_ref_node);CHKERRQ(ierr);
ierr = PetscMalloc1(graph->cptr[graph->ncc],&queue_global);CHKERRQ(ierr);
ierr = ISLocalToGlobalMappingApply(graph->l2gmap,graph->cptr[graph->ncc],graph->queue,queue_global);CHKERRQ(ierr);
for (j=0;j<graph->ncc;j++) {
ierr = PetscSortIntWithArray(graph->cptr[j+1]-graph->cptr[j],&queue_global[graph->cptr[j]],&graph->queue[graph->cptr[j]]);CHKERRQ(ierr);
graph->subset_ref_node[j] = graph->queue[graph->cptr[j]];
}
ierr = PetscFree(queue_global);CHKERRQ(ierr);
graph->queue_sorted = PETSC_TRUE;
/* save information on subsets (needed when analyzing the connected components) */
if (graph->ncc) {
ierr = PetscMalloc2(graph->ncc,&graph->subset_size,graph->ncc,&graph->subset_idxs);CHKERRQ(ierr);
ierr = PetscMalloc1(graph->cptr[graph->ncc],&graph->subset_idxs[0]);CHKERRQ(ierr);
ierr = PetscMemzero(graph->subset_idxs[0],graph->cptr[graph->ncc]*sizeof(PetscInt));CHKERRQ(ierr);
for (j=1;j<graph->ncc;j++) {
graph->subset_size[j-1] = graph->cptr[j] - graph->cptr[j-1];
graph->subset_idxs[j] = graph->subset_idxs[j-1] + graph->subset_size[j-1];
}
graph->subset_size[graph->ncc-1] = graph->cptr[graph->ncc] - graph->cptr[graph->ncc-1];
ierr = PetscMemcpy(graph->subset_idxs[0],graph->queue,graph->cptr[graph->ncc]*sizeof(PetscInt));CHKERRQ(ierr);
}
/* renumber reference nodes */
ierr = ISCreateGeneral(PetscObjectComm((PetscObject)(graph->l2gmap)),graph->ncc,graph->subset_ref_node,PETSC_COPY_VALUES,&subset_n);CHKERRQ(ierr);
ierr = ISLocalToGlobalMappingApplyIS(graph->l2gmap,subset_n,&subset);CHKERRQ(ierr);
ierr = ISDestroy(&subset_n);CHKERRQ(ierr);
ierr = ISRenumber(subset,NULL,NULL,&subset_n);CHKERRQ(ierr);
ierr = ISDestroy(&subset);CHKERRQ(ierr);
ierr = ISGetLocalSize(subset_n,&k);CHKERRQ(ierr);
if (k != graph->ncc) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Invalid size of new subset! %D != %D",k,graph->ncc);
ierr = ISGetIndices(subset_n,&is_indices);CHKERRQ(ierr);
ierr = PetscMemcpy(graph->subset_ref_node,is_indices,graph->ncc*sizeof(PetscInt));CHKERRQ(ierr);
ierr = ISRestoreIndices(subset_n,&is_indices);CHKERRQ(ierr);
ierr = ISDestroy(&subset_n);CHKERRQ(ierr);
/* free workspace */
graph->setupcalled = PETSC_TRUE;
PetscFunctionReturn(0);
}
