dealii - drop extra copy on op inputs/outputs

Author: jeremylt

examples/deal.II/bps.h

@@ -249,15 +249,14 @@ class OperatorCeed : public OperatorBase<Number>
 
           for (const auto i : dof_mapping)
             indices.emplace_back(
-              partitioner->global_to_local(local_indices[fe.component_to_system_index(0, i)]) /
-              n_components);
+              partitioner->global_to_local(local_indices[fe.component_to_system_index(0, i)]));
         }
 
     CeedElemRestrictionCreate(ceed,
                               n_local_active_cells,
                               fe.n_dofs_per_cell() / n_components,
                               n_components,
-                              std::max<unsigned int>(this->extended_local_size() / n_components, 1),
+                              1,
                               this->extended_local_size(),
                               CEED_MEM_HOST,
                               CEED_COPY_VALUES,
@@ -344,46 +343,18 @@ class OperatorCeed : public OperatorBase<Number>
     // communicate: update ghost values
     src.update_ghost_values();
 
-    if (dof_handler.get_fe().n_components() == 1)
-      {
-        // pass memory buffers to libCEED
-        VectorTypeCeed x(src_ceed);
-        VectorTypeCeed y(dst_ceed);
-        x.import_array(src, CEED_MEM_HOST);
-        y.import_array(dst, CEED_MEM_HOST);
-
-        // apply operator
-        CeedOperatorApply(op_apply, x(), y(), CEED_REQUEST_IMMEDIATE);
-
-        // pull arrays back to deal.II
-        x.sync_array();
-        y.sync_array();
-      }
-    else // TODO: needed for multiple components
-      {
-        // allocate space for block vectors
-        src_tmp.reinit(this->extended_local_size(), true);
-        dst_tmp.reinit(this->extended_local_size(), true);
-
-        // copy to block vector
-        copy_to_block_vector(src_tmp, src);
-
-        // pass memory buffers to libCEED
-        VectorTypeCeed x(src_ceed);
-        VectorTypeCeed y(dst_ceed);
-        x.import_array(src_tmp, CEED_MEM_HOST);
-        y.import_array(dst_tmp, CEED_MEM_HOST);
-
-        // apply operator
-        CeedOperatorApply(op_apply, x(), y(), CEED_REQUEST_IMMEDIATE);
+    // pass memory buffers to libCEED
+    VectorTypeCeed x(src_ceed);
+    VectorTypeCeed y(dst_ceed);
+    x.import_array(src, CEED_MEM_HOST);
+    y.import_array(dst, CEED_MEM_HOST);
 
-        // pull arrays back to deal.II
-        x.sync_array();
-        y.sync_array();
+    // apply operator
+    CeedOperatorApply(op_apply, x(), y(), CEED_REQUEST_IMMEDIATE);
 
-        // copy from block vector
-        copy_from_block_vector(dst, dst_tmp);
-      }
+    // pull arrays back to deal.II
+    x.sync_array();
+    y.sync_array();
 
     // communicate: compress
     src.zero_out_ghost_values();
@@ -419,17 +390,6 @@ class OperatorCeed : public OperatorBase<Number>
     // pull array back to deal.II
     y.sync_array();
 
-    const unsigned int n_components = dof_handler.get_fe().n_components();
-
-    if (n_components > 1) // TODO: needed for multiple components
-      {
-        VectorType tmp(diagonal);
-
-        copy_from_block_vector(tmp, diagonal);
-
-        std::swap(tmp, diagonal);
-      }
-
     diagonal.compress(VectorOperation::add);
 
     for (auto &i : diagonal)
@@ -503,36 +463,6 @@ class OperatorCeed : public OperatorBase<Number>
     CeedVector  vec_ceed;
   };
 
-  /**
-   * Copy from block vector.
-   *
-   * @note Only needed for multiple components.
-   */
-  void
-  copy_from_block_vector(VectorType &dst, const VectorType &src) const
-  {
-    const unsigned int scalar_size = this->extended_local_size() / dim;
-
-    for (unsigned int i = 0; i < scalar_size; ++i)
-      for (unsigned int j = 0; j < dim; ++j)
-        dst.get_values()[j + i * dim] = src.get_values()[j * scalar_size + i];
-  }
-
-  /**
-   * Copy to block vector.
-   *
-   * @note Only needed for multiple components.
-   */
-  void
-  copy_to_block_vector(VectorType &dst, const VectorType &src) const
-  {
-    const unsigned int scalar_size = this->extended_local_size() / dim;
-
-    for (unsigned int i = 0; i < scalar_size; ++i)
-      for (unsigned int j = 0; j < dim; ++j)
-        dst.get_values()[j * scalar_size + i] = src.get_values()[j + i * dim];
-  }
-
   /**
    * Number of locally active DoFs.
    */