From: Martin Kronbichler Date: Thu, 28 Mar 2024 14:57:15 +0000 (+0100) Subject: MappingQ: Do not assume contiguous storage of Tensor<2, dim> X-Git-Tag: v9.6.0-rc1~439^2 X-Git-Url: https://gitweb.dealii.org/cgi-bin/gitweb.cgi?a=commitdiff_plain;h=c00cd7802b52445eed3a83c12ac3b9a5b5f37876;p=dealii.git MappingQ: Do not assume contiguous storage of Tensor<2, dim> --- diff --git a/include/deal.II/fe/mapping_q_internal.h b/include/deal.II/fe/mapping_q_internal.h index 5bc48229fa..90ba787dbc 100644 --- a/include/deal.II/fe/mapping_q_internal.h +++ b/include/deal.II/fe/mapping_q_internal.h @@ -1233,40 +1233,6 @@ namespace internal - template - inline DEAL_II_ALWAYS_INLINE void - store_vectorized_tensor( - const unsigned int n_points, - const unsigned int cur_index, - const DerivativeForm<1, dim, spacedim, VectorizedArray> - &derivative, - std::vector> &result_array) - { - AssertDimension(result_array.size(), n_points); - constexpr unsigned int n_lanes = VectorizedArray::size(); - if (cur_index + n_lanes <= n_points) - { - std::array indices; - for (unsigned int j = 0; j < n_lanes; ++j) - indices[j] = j * dim * spacedim; - const unsigned int even = (dim * spacedim) / 4 * 4; - double *result_ptr = &result_array[cur_index][0][0]; - const VectorizedArray *derivative_ptr = &derivative[0][0]; - vectorized_transpose_and_store( - false, even, derivative_ptr, indices.data(), result_ptr); - for (unsigned int d = even; d < dim * spacedim; ++d) - for (unsigned int j = 0; j < n_lanes; ++j) - result_ptr[j * dim * spacedim + d] = derivative_ptr[d][j]; - } - else - for (unsigned int j = 0; j < n_lanes && cur_index + j < n_points; ++j) - for (unsigned int d = 0; d < spacedim; ++d) - for (unsigned int e = 0; e < dim; ++e) - result_array[cur_index + j][d][e] = derivative[d][e][j]; - } - - - template inline void maybe_update_q_points_Jacobians_generic( @@ -1354,7 +1320,10 @@ namespace internal continue; if (update_flags & update_contravariant_transformation) - store_vectorized_tensor(n_points, i, derivative, jacobians); + for (unsigned int j = 0; j < n_lanes && i + j < n_points; ++j) + for (unsigned int d = 0; d < spacedim; ++d) + for (unsigned int e = 0; e < dim; ++e) + jacobians[i + j][d][e] = derivative[d][e][j]; if (update_flags & update_volume_elements) { @@ -1366,10 +1335,10 @@ namespace internal if (update_flags & update_covariant_transformation) { const auto covariant = derivative.covariant_form(); - store_vectorized_tensor(n_points, - i, - covariant.transpose(), - inverse_jacobians); + for (unsigned int j = 0; j < n_lanes && i + j < n_points; ++j) + for (unsigned int d = 0; d < dim; ++d) + for (unsigned int e = 0; e < spacedim; ++e) + inverse_jacobians[i + j][d][e] = covariant[e][d][j]; } } else