Example 6: Sparse DataΒΆ

The FleCSI sparse storage class allows the representation of sparse data fields, i.e., fields on which zero, one or more data may be defined at each index of the associated index space. The sparse storage class is suitable for representing sparse matrices, sparse materials, or any other logical data structure that utilizes a compressed storage appraoch. Note that although the current implementation does use a compressed storage scheme, there is no guaruntee that this will be the case. The sparse storage class defines the interface to the registered field data, not the storage mechanism.

Sparse storage differs from simple contiguous data in that, not only are the data at each index mutable, but the sparsity pattern of the data is also mutable. Modifying the sparsity pattern of the data is a more expensive operation than simply modifying the values of existing sparse data. To allow for more efficient access patterns, FleCSI splits operations that mutate the sparsity pattern from those that only modify the data (without changing the sparsity pattern). Access to these is defined using separate types:

  1. A mutator allows both modification of the sparsity structure and of the data values.

  2. A handle only allows modification to the data values.

The data handle interface is the same as for dense data. For the mutator, the user must specify an additional argument that declares how many slots should be pre-allocated for adding new non-sparse entries:

//                                                                v
auto f = flecsi_get_mutator(m, example, field, double, sparse, 0, 5);

This argument is only a hint to the runtime: Slots added within this allocation will be more efficient than those that exceed the allocation. However, both operations will be correct, i.e., if the user has specified 5 additional slots, but adds 6 non-sparse entries, the program will still operate correctly. (FleCSI uses an overflow buffer to manage additional entries that exceed the slot allocation.)


  • Tasks that do not need to mutate the sparsity structure should always use an handle rather than a mutator for efficiency.

  • We are investigating design changes that will allow specialization developers to add new storage classes and storage class implementations.

#include <cstdlib>
#include <iostream>


using namespace flecsi;
using namespace flecsi::tutorial;

// Field registration is as usual (but specifying the 'sparse'
// storage class).

flecsi_register_field(mesh_t, example, field, double, sparse, 1, cells);

namespace example {

// This task takes a mesh and a sparse mutator and randomly populates
// field entries into the sparse field structure.

void initialize_sparse_field(mesh<ro> mesh, sparse_field_mutator f) {

  for(auto c: mesh.cells()) {
    const size_t random = (rand()/double{RAND_MAX}) * 5;

    for(size_t i{0}; i<random; ++i) {
      const size_t entry = (rand()/double{RAND_MAX}) * 5;

      // Note that the field operator () takes a cell index and
      // an entry. This provides logically dense access to sparsley
      // stored data.

      f(c, entry) = entry;

    } // for
  } // for
} // initialize_pressure

flecsi_register_task(initialize_sparse_field, example, loc, single);

// This task prints the non-zero entries of the sparse field.

void print_sparse_field(mesh<ro> mesh, sparse_field<ro> f) {
  for(auto c: mesh.cells()) {
    for(auto m: f.entries(c)) {
      std::cout << f(c,m) << " ";
    } // for
    std::cout << std::endl;
  } // for
} // print_pressure

flecsi_register_task(print_sparse_field, example, loc, single);

} // namespace example

namespace flecsi {
namespace execution {

void driver(int argc, char ** argv) {

  // Get a handle to the mesh

  auto m = flecsi_get_client_handle(mesh_t, clients, mesh);

  // Get a mutator to modify the sparsity structure of the data.

  auto f = flecsi_get_mutator(m, example, field, double, sparse, 0, 5);

  flecsi_execute_task(initialize_sparse_field, example, single, m, f);
  } // scope

  // Get a handle to modify only the values of the data.

  auto f = flecsi_get_handle(m, example, field, double, sparse, 0);

  flecsi_execute_task(print_sparse_field, example, single, m, f);
  } // scope

} // driver

} // namespace execution
} // namespace flecsi