fix permutation of gathered matrix

src/hyteg/ginkgo/GinkgoBlockSolver.hpp

@@ -147,6 +147,8 @@ class GinkgoBlockSolver : public Solver< OperatorType >
       std::vector< int32_t > pIndices;
 
       gatherIndices( vIndices, pIndices, *storage_, level, num_ );
+      std::sort( vIndices.begin(), vIndices.end() );
+      std::sort( pIndices.begin(), pIndices.end() );
 
       std::vector< int32_t > perm_vec;
 
@@ -179,11 +181,11 @@ class GinkgoBlockSolver : public Solver< OperatorType >
 
       monolithic_matrix_ = gko::share( csr::create( solver_exec_ ) );
       gko::as< mtx >( host_monolithic_matrix_ )->convert_to( monolithic_matrix_.get() );
-
       gko::Array< gko::int32 > perm{ solver_exec_, perm_vec.begin(), perm_vec.end() };
-      if ( monolithic_matrix_->get_size() )
+      if ( comm_->size() > 1 && monolithic_matrix_->get_size() )
       {
-         monolithic_matrix_->permute( &perm );
+         monolithic_matrix_ = gko::as< csr >( monolithic_matrix_->permute( &perm ) );
+         monolithic_matrix_->sort_by_column_index();
       }
 
       {
@@ -249,10 +251,17 @@ class GinkgoBlockSolver : public Solver< OperatorType >
 
       if ( monolithic_matrix_->get_size() )
       {
-         global_rhs->row_permute( &perm );
-         global_x0->row_permute( &perm );
-         solver_->apply( gko::lend( global_rhs ), gko::lend( global_x0 ) );
-         global_x0->inverse_row_permute( &perm );
+         if ( comm_->size() > 1 )
+         {
+            auto permuted_global_rhs = gko::as< dense >( global_rhs->row_permute( &perm ) );
+            auto permuted_global_x0  = gko::as< dense >( global_x0->row_permute( &perm ) );
+            solver_->apply( gko::lend( permuted_global_rhs ), gko::lend( permuted_global_x0 ) );
+            permuted_global_x0->inverse_row_permute( &perm, global_x0.get() );
+         }
+         else
+         {
+            solver_->apply( gko::lend( global_rhs ), gko::lend( global_x0 ) );
+         }
       }
       gather_idxs_ = compute_gather_idxs( part_ );
       scatter_global_vector( global_x0.get(), x0, gather_idxs_, comm_ );