From: Martin Kronbichler Date: Tue, 31 May 2022 07:15:40 +0000 (+0200) Subject: Speed up the function MatrixFree::n_active_entries_per_cell_batch X-Git-Tag: v9.4.0-rc1~100^2 X-Git-Url: https://gitweb.dealii.org/cgi-bin/gitweb.cgi?a=commitdiff_plain;h=47e3e3cdfc2647b18f328a4b82ff07e614a00509;p=dealii.git Speed up the function MatrixFree::n_active_entries_per_cell_batch --- diff --git a/include/deal.II/matrix_free/matrix_free.h b/include/deal.II/matrix_free/matrix_free.h index 498d0e828e..903d551b04 100644 --- a/include/deal.II/matrix_free/matrix_free.h +++ b/include/deal.II/matrix_free/matrix_free.h @@ -2626,16 +2626,14 @@ inline unsigned int MatrixFree::n_active_entries_per_cell_batch( const unsigned int cell_batch_index) const { + Assert(!dof_info.empty(), ExcNotInitialized()); AssertIndexRange(cell_batch_index, task_info.cell_partition_data.back()); - unsigned int n_lanes = VectorizedArrayType::size(); - while (n_lanes > 1 && - cell_level_index[cell_batch_index * VectorizedArrayType::size() + - n_lanes - 1] == - cell_level_index[cell_batch_index * VectorizedArrayType::size() + - n_lanes - 2]) - --n_lanes; - AssertIndexRange(n_lanes - 1, VectorizedArrayType::size()); - return n_lanes; + const std::vector &n_lanes_filled = + dof_info[0].n_vectorization_lanes_filled + [internal::MatrixFreeFunctions::DoFInfo::dof_access_cell]; + AssertIndexRange(cell_batch_index, n_lanes_filled.size()); + + return n_lanes_filled[cell_batch_index]; } @@ -2646,13 +2644,12 @@ MatrixFree::n_active_entries_per_face_batch( const unsigned int face_batch_index) const { AssertIndexRange(face_batch_index, face_info.faces.size()); - unsigned int n_lanes = VectorizedArrayType::size(); - while (n_lanes > 1 && - face_info.faces[face_batch_index].cells_interior[n_lanes - 1] == - numbers::invalid_unsigned_int) - --n_lanes; - AssertIndexRange(n_lanes - 1, VectorizedArrayType::size()); - return n_lanes; + Assert(!dof_info.empty(), ExcNotInitialized()); + const std::vector &n_lanes_filled = + dof_info[0].n_vectorization_lanes_filled + [internal::MatrixFreeFunctions::DoFInfo::dof_access_face_interior]; + AssertIndexRange(face_batch_index, n_lanes_filled.size()); + return n_lanes_filled[face_batch_index]; } diff --git a/include/deal.II/matrix_free/matrix_free.templates.h b/include/deal.II/matrix_free/matrix_free.templates.h index e6adbab2b0..5cf7a73d11 100644 --- a/include/deal.II/matrix_free/matrix_free.templates.h +++ b/include/deal.II/matrix_free/matrix_free.templates.h @@ -514,11 +514,24 @@ MatrixFree::internal_reinit( dof_info[i].dofs_per_cell.push_back( dof_handler[i]->get_fe(0).n_dofs_per_cell()); + const unsigned int n_regular_cells = cell_level_index.size(); // if indices are not initialized, the cell_level_index might not be // divisible by the vectorization length. But it must be for // mapping_info... while (cell_level_index.size() % VectorizedArrayType::size() != 0) cell_level_index.push_back(cell_level_index.back()); + + // adjust lengths for vectorization + dof_info[i] + .n_vectorization_lanes_filled + [internal::MatrixFreeFunctions::DoFInfo::dof_access_cell] + .resize(cell_level_index.size() / VectorizedArrayType::size(), + VectorizedArrayType::size()); + if (n_regular_cells < cell_level_index.size()) + dof_info[i].n_vectorization_lanes_filled + [internal::MatrixFreeFunctions::DoFInfo::dof_access_cell] + [n_regular_cells / VectorizedArrayType::size()] = + n_regular_cells % VectorizedArrayType::size(); } }